Yoshiharu Ishikawa Jianzhong Li

Wei Wang Rui Zhang Wenjie Zhang (Eds.)

Web Technologies
LNCS 7808

and Applications
15th Asia-Pacific Web Conference, APWeb 2013
Sydney, Australia, April 2013
Proceedings

123
Lecture Notes in Computer Science 7808
Commenced Publication in 1973
Founding and Former Series Editors:
Gerhard Goos, Juris Hartmanis, and Jan van Leeuwen

Editorial Board
David Hutchison
Lancaster University, UK
Takeo Kanade
Carnegie Mellon University, Pittsburgh, PA, USA
Josef Kittler
University of Surrey, Guildford, UK
Jon M. Kleinberg
Cornell University, Ithaca, NY, USA
Alfred Kobsa
University of California, Irvine, CA, USA
Friedemann Mattern
ETH Zurich, Switzerland
John C. Mitchell
Stanford University, CA, USA
Moni Naor
Weizmann Institute of Science, Rehovot, Israel
Oscar Nierstrasz
University of Bern, Switzerland
C. Pandu Rangan
Indian Institute of Technology, Madras, India
Bernhard Steffen
TU Dortmund University, Germany
Madhu Sudan
Microsoft Research, Cambridge, MA, USA
Demetri Terzopoulos
University of California, Los Angeles, CA, USA
Doug Tygar
University of California, Berkeley, CA, USA
Gerhard Weikum
Max Planck Institute for Informatics, Saarbruecken, Germany
Yoshiharu Ishikawa Jianzhong Li
Wei Wang Rui Zhang Wenjie Zhang (Eds.)

Web Technologies
and Applications
15th Asia-Pacific Web Conference, APWeb 2013
Sydney, Australia, April 4-6, 2013
Proceedings

13
Volume Editors
Yoshiharu Ishikawa
Nagoya University
Graduate School of Information Science
Nagoya 464-8601, Japan
E-mail: [email protected]
Jianzhong Li
Harbin Institute of Technology
Department of Computer Science and Technology
Harbin 150006, China
E-mail: [email protected]
Wei Wang
Wenjie Zhang
University of New South Wales
School of Computer Science and Engineering
Sydney, NSW 2052, Australia
E-mail: {weiw, zhangw}@cse.unsw.edu.au
Rui Zhang
University of Melbourne
Department of Computing and Information Systems
Melbourne, VIC 3052, Australia
E-mail: [email protected]

ISSN 0302-9743 e-ISSN 1611-3349

ISBN 978-3-642-37400-5 e-ISBN 978-3-642-37401-2
DOI 10.1007/978-3-642-37401-2
Springer Heidelberg Dordrecht London New York

Library of Congress Control Number: 2013934117

CR Subject Classification (1998): H.2.8, H.2, H.3, H.5, H.4, J.1, K.4, I.2

LNCS Sublibrary: SL 3 Information Systems and Application, incl. Internet/Web

and HCI
Springer-Verlag Berlin Heidelberg 2013

This work is subject to copyright. All rights are reserved, whether the whole or part of the material is
concerned, specifically the rights of translation, reprinting, re-use of illustrations, recitation, broadcasting,
reproduction on microfilms or in any other way, and storage in data banks. Duplication of this publication
or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965,
in ist current version, and permission for use must always be obtained from Springer. Violations are liable
to prosecution under the German Copyright Law.
The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply,
even in the absence of a specific statement, that such names are exempt from the relevant protective laws
and regulations and therefore free for general use.
Typesetting: Camera-ready by author, data conversion by Scientific Publishing Services, Chennai, India
Printed on acid-free paper
Springer is part of Springer Science+Business Media (www.springer.com)
 Message from the General Chairs

Welcome to APWeb 2013, the 15th Edition of the Asia Pacific Web Confer-
ence. APWeb is a leading international conference on research, development, and
applications of Web technologies, database systems, information management,
and software engineering, with a focus on the Asia-Pacific region. Previous AP-
Web conferences were held in Kunming (2012), Beijing (2011), Busan (2010),
Suzhou (2009), Shenyang (2008), Huangshan (2007), Harbin (2006), Shanghai
(2005), Hangzhou (2004), Xian (2003), Changsha (2001), Xian (2000), Hong
Kong (1999), and Beijing (1998).
The APWeb 2013 conference was, for the first time, held in Sydney, Australia
a city blessed with a temperate climate, a beautiful harbor, and natural at-
tractions surrounding it. These proceedings collect the technical papers selected
for presentation at the conference, during April 46, 2013.
The APWeb 2013 program featured a main conference, a special track, and
four satellite workshops. The main conference had three keynotes by eminent re-
searchers H.V. Jagadish from the University of Michigan, USA, Mark Sanderson
from RMIT University, Australia, and Dan Suciu from the University of Wash-
ington, USA. Three tutorials were oered by Haixun Wang, Microsoft Research
Asia, China, Yuqing Wu, Indiana University, USA, George Fletcher, Eindhoven
University of Technology, The Netherlands, and Lei Chen, Hong Kong University
of Science and Technology, Hong Kong, China. The conference received 165 paper
submissions from North America, South America, Europe, Asia, and Oceania.
Each submitted paper underwent a rigorous review by at least three indepen-
dent referees, with detailed review reports. Finally, 39 full research papers and
22 short research papers were accepted, from Australia, Bangladesh, Canada,
China, India, Ireland, Italy, Japan, New Zealand, Saudia Arabia, Sweden, Nor-
way, UK, and USA. The special track of Distributed Processing of Graph, XML
and RDF Data: Theory and Practice was organized by Alfredo Cuzzocrea. The
conference had four workshops

The Second International Workshop on Data Management for Emerging Net-

work Infrastructure (DaMEN 2013)
International Workshop on Location-Based Data Management (LBDM 2013)
International Workshop on Management of Spatial Temporal Data
(MSTD 2013)
International Workshop on Social Media Analytics and Recommendation
Technologies (SMART 2013)

We were extremely excited with our strong Program Committee, comprising out-
standing researchers in the APWeb research areas. We would like to extend our
sincere gratitude to the Program Committee members and external reviewers.
Last but not least, we would like to thank the sponsors, for their strong support
VI Message from the General Chairs

of this conference, making it a big success. Special thanks go to the Chinese Uni-
versity of Hong Kong, the University of New South Wales, Macquarie University,
and the University of Sydney.
Finally, we wish to thank the APWeb Steering Committee, led by Xuemin
Lin, for oering us the opportunity to organize APWeb 2013 in Sydney. We also
wish to thank the host organization, the University of New South Wales, and
Local Arrangements Committee and volunteers for their assistance in organizing
this conference.

February 2013 Vijay Varadharajan

Jerey Xu Yu
 Conference Organization

Conference Co-chairs
Vijay Varadharajan Macquarie University, Australia
Jerey Xu Yu Chinese University of Hong Kong, China

Program Committee Co-chairs

Yoshiharu Ishikawa Nagoya University, Japan
Jianzhong Li Harbin Institute of Technology, China
Wei Wang University of New South Wales, Australia

Local Organization Co-chairs

Muhammad Aamir Cheema University of New South Wales, Australia
Ying Zhang University of New South Wales, Australia

Workshop Co-chairs
James Bailey University of Melbourne, Australia
Xiaochun Yang Northeastern University, China

Tutorial/Panel Co-chairs
Sanjay Chawla University of Sydney, Australia
Xiaofeng Meng Renmin University of China, China

Industrial Co-chairs
Marek Kowalkiewicz SAP Research in Brisbane, Australia
Mukesh Mohania IBM Research, India

Publication Co-chairs
Rui Zhang University of Melbourne, Australia
Wenjie Zhang University of New South Wales, Australia
VIII Conference Organization

Publicity Co-chairs
Alfredo Cuzzocrea University of Calabria, Italy
Jiaheng Lu Renmin University of China, China

Demo Co-chairs
Wook-Shin Han Kyungpook National University, Korea
Helen Huang University of Queensland, Australia

APWeb Steering Committee Liaison

Xuemin Lin University of New South Wales, Australia

WISE Society Liaison

Yanchun Zhang Victoria University, Australia

WAIM Steering Committee Liaison

Qing Li City University of Hong Kong, China

Webmasters
Yu Zheng East China Normal University, China
Chen Chen University of New South Wales, Australia

Program Committee
Toshiyuki Amagasa University of Tsukuba
Djamal Benslimane University of Lyon
Jae-Woo Chang Chonbuk National University
Haiming Chen Chinese Academy of Sciences
Jinchuan Chen Renmin University of China
David Cheung The University of Hong Kong
Bin Cui Beijing University
Alfredo Cuzzocrea ICAR-CNR & University of Calabria
Ting Deng Beihang University
Jianlin Feng Sun Yat-Sen University
Yaokai Feng Kyushu University
 Conference Organization IX

Sergio Flesca University of Calabria

Hong Gao Harbin Institute of Technology
Yunjun Gao Zhejiang University
Stephane Grumbach INRIA
Giovanna Guerrini University of Genoa
Mohand-Said Hacid University of Lyon 1
Qi He IBM
Jun Hong Queens University Belfast
Michael Houle National Institute of Informatics
Bin Hu Lanzhou University
Zi Huang University of Queensland
Jeong-Hyon Hwang State University of New York at Albany
Seung-won Hwang POSTECH
Mizuho Iwaihara Waseda University
Adam Jatowt Kyoto University
Cheqing Jin East China Normal University
Anastasios Kementsietsidis IBM T.J. Watson Research Center
Jin-Ho Kim Kangwon National University
Markus Kirchberg Institute for Infocomm Research
Manolis Koubarakis University of Athens
Byung Lee Vermont University
Chiang Lee National Cheng Kung University
Jae-Gil Lee KAIST
SangKeun Lee Korea University
Carson Leung University of Manitoba
Jianxin Li Swinburne University
Xue Li Queensland University
Yingshu Li Georgia State University
Yinsheng Li Fudan University
Zhanhuai Li Northwestern Polytechnical University
Xiang Lian University of Texas - Pan American
Guimei Liu National University of Singapore
Mengchi Liu Carleton University
Chengfei Liu Swinburne University of Technology
Bo Luo University of Kansas
Jiangang Ma University of Adelaide
Qiang Ma Kyoto University
Shuai Ma Beihang University
Zakaria Maamar Zayed University
Sanjay Madria University of Missouri-Rolla
Weiyi Meng State University of New York at Binghamton
Yang-Sae Moon Kangwon National University
Michael Mrissa University of Lyon
Akiyo Nadamoto Konan University
Shinsuke Nakajima Kyoto Sangyo University
X Conference Organization

Miyuki Nakano University of Tokyo

Werner Nutt Free University of Bozen-Bolzano
Satoshi Oyama Hokkaido University
Helen Paik University of New South Wales
Chaoyi Pang CSIRO
Apostolos Papadopoulos Aristotle University
Eric Pardede Latrobe University
Sanghyun Park Yonsei University
Zhiyong Peng Wuhan University
Tieyun Qian Wuhan University
Weining Qian East China Normal University
Joao Rocha-Junior Univ. Estadual de Feira de Santana
KeunHo Ryu Chungbuk National University
Markus Schneider University of Florida
Marc Scholl Universitat Konstanz
Aviv Segev KAIST
Bin Shao Microsoft Research Asia
Derong Shen Northeastern University
Heng Shen Queensland University
Jialie Shen Singapore Management University
Timothy Shih National Taipei University of Education
Lidan Shou Zhejiang University
Shaoxu Song Tsinghua University
Konstantinos Stefanidis Norwegian University of Science and
Technology
Kazutoshi Sumiya University of Hyogo
Aixin Sun Nanyang Technological University
Claudia Szabo University of Adelaide
Changjie Tang Sichuan University
Nan Tang University of Edinburgh
David Taniar Monash University
Alex Thomo University of Victoria
Chaokun Wang Tsinghua University
Daling Wang Northeastern University
Fan Wang Microsoft
Guoren Wang Northeastern University
Hongzhi Wang Harbin Institute of Technology
Hua Wang University of Southern Queensland
Jianyong Wang Tsinghua University
X. Wang Fudan University
Xiaoling Wang East China Normal University
Jef Wijsen University of Mons-Hainaut
Jianliang Xu Hong Kong Baptist University
Xiaochun Yang Northeastern University
Jian Yin Sun Yat-Sen University
Haruo Yokota Tokyo Institute of Technology
 Conference Organization XI

Jian Yu Swinburne University of Technology

Ming Zhang Beijing University
Xiao Zhang Renmin University of China
Baihua Zheng Singapore Management University
Rui Zhou Swinburne University
Shuigeng Zhou Fudan University
Xiaofang Zhou University of Queensland
Xuan Zhou Renmin University
Zhaonian Zou Harbin Institute of Technology

External Reviewers
Xuefei Li Mahmoud Barhamgi
Hongyun Cai Xian Li
Jingkuan Song Yu Jiang
Yang Yang Saurav Acharya
Xiaofeng Zhu Syed K. Tanbeer
Scott Bourne Hongda Ren
Yasser Salem Wei Shen
Shi Feng Zhenhua Song
Jianwei Zhang Jianhua Yin
Kenta Oku Liu Chen
Sukhwan Jung Wei Song
 An Overview of Probabilistic Databases

Dan Suciu
University of Washington
[email protected]
http://homes.cs.washington.edu/ suciu/

A major challenge in modern data management is how to cope with uncertainty

in the data. Uncertainty may exists because the data was extracted automatically
from text, or was derived from the physical world such as RFID data, or was
obtained by integrating several data sets using fuzzy matches, or may be the
result of complex stochastic models. In a probabilistic database uncertainty is
modeled using probabilities, and data management techniques are extended to
cope with probabilistic data.
The main challenge is query evaluation. For each answer to the query, its de-
gree of uncertainty is the probability that its lineage formula is true. Thus, query
evaluation reduces to the problem of computing the probability of a Boolean
formula. This problem generalizes model counting, which has been extensively
studied in the AI and model checking literature. Todays state of the art methods
for computing the exact probability are extensions of Davis Putnams (DP) pro-
cedure [3, 2, 1, 4]. In probabilistic databases we can take a new approach, because
here we can fix the query, and consider only the database as variable input (called
data complexity [7]). An interesting dichotomy theorem holds: for every query,
either its complexity is in PTIME or is #P-hard. A new probabilistic inference
algorithm was needed in order to compute all PTIME queries, which uses the
inclusion/exclusion principle [6]. This technique is missing from todays exten-
sions of DP, yet necessary: without it one can show that probabilistic inference
for certain simple PTIME queries requires exponential time [5].

References
1. Bacchus, F., Dalmao, S., Pitassi, T.: Algorithms and complexity results for #sat
and bayesian inference. In: FOCS, pp. 340351 (2003)
2. Birnbaum, E., Lozinskii, E.L.: The good old davis-putnam procedure helps counting
models. J. Artif. Int. Res. 10(1), 457477 (1999)
3. Davis, M., Logemann, G., Loveland, D.: A machine program for theorem-proving.
Commun. ACM 5(7), 394397 (1962)
4. Gomes, C.P., Sabharwal, A., Selman, B.: Model counting. In: Handbook of Satisfi-
ability, pp. 633654 (2009)
5. Jha, A.K., Suciu, D.: Knowledge compilation meets database theory: compiling
queries to decision diagrams. In: ICDT, pp. 162173 (2011)
6. Suciu, D., Olteanu, D., Re, C., Koch, C.: Probabilistic Databases. In: Synthesis
Lectures on Data Management. Morgan & Claypool Publishers (2011)
7. Vardi, M.Y.: The complexity of relational query languages (extended abstract).
In: STOC, pp. 137146 (1982)
 Challenges with Big Data on the Web

H.V. Jagadish
University of Michigan
[email protected]

The promise of data-driven decision-making is now being recognized broadly,

and there is growing enthusiasm for the notion of Big Data.O O This is true
of Big Data in the enterprise, but this is even more true of Big Data on the
web. While the promise of Big Data is real for example, it is estimated that
Google alone contributed 54 billion dollars to the US economy in 2009 there
is currently a wide gap between its potential and its realization.
Heterogeneity, scale, timeliness, complexity, and privacy problems with Big
Data impede progress at all phases of the pipeline that can create value from data.
The problems start right away during data acquisition, when the data tsunami
requires us to make decisions, currently in an ad hoc manner, about what data
to keep and what to discard, and how to store what we keep reliably with the
right metadata. Much data today is not natively in structured format; for exam-
ple, tweets and blogs are weakly structured pieces of text, while images and video
are structured for storage and display, but not for semantic content and search:
transforming such content into a structured format for later analysis is a major
challenge. The value of data explodes when it can be linked with other data, thus
data integration is a major creator of value. Since most data is directly generated
in digital format today, we have the opportunity and the challenge both to influ-
ence the creation to facilitate later linkage and to automatically link previously
created data. Data analysis, organization, retrieval, and modeling are other foun-
dational challenges. Finally, presentation of the results and its interpretation by
non-technical domain experts is crucial to extracting actionable knowledge.
A recent white paper[CCC12] mapped out the many challenges in this space.
In this talk, drawing upon this white paper, I will present these challenges,
particularly as they relate to the web. I will draw upon examples from database
usability to show how size and complexity of Big Data can create diculties
for a user, and mention some directions of work in this regard. In particular,
I will highlight how Big Data issues arise in surprising contexts, such as in
browsing[SIGMOD12].

References
[CCC12] Jagadish, H.V., et al: Challenges and Opportunities with Big Data,
http://cra.org/ccc/docs/init/bigdatawhitepaper.pdf
[SIGMOD12] Singh, M., Nandi, A., Jagadish, H.V.: Skimmer: rapid scrolling of rela-
tional query results. In: SIGMOD Conference, pp. 181192 (2012)

Supported in part by NSF under grant IIS-1017296.

 Twenty Years of Web Search Where to Next?

Mark Sanderson
School of Computer Science and Information Technology
RMIT University
GPO Box 2476, Melbourne 3001
Victoria, Australia

Abstract. This year, (2013) marks the 20th anniversary of the first public web
search engine JumpStation launched in late 1993. For those who were around
in those early days, it was becoming clear that an information provision and an
information access revolution was on its way; though very few, if any would have
predicted the state of the information society we have today. It is perhaps worth
reflecting on what has been achieved in the field of information retrieval since
these systems were first created, and consider what remains to be accomplished.
It is perhaps easy to see the success of systems like Google and ask what else is
there to achieve? However, in some ways, Google has it easy. In this talk, I will
explain why Web search can be viewed as a relatively easy task and why other
forms of search are much harder to perform accurately.
Search engines require a great deal of tuning, currently achieved empirically.
The tuning carried out depends greatly on the types of queries submitted to a
search engine and the types of document collections the queries will search over.
It should be possible to study the population of queries and documents and
predictively configure a search engine. However, there is little understanding
in either the research or practitioner communities on how query and collection
properties map to search engine configurations. I will present the some of the
early work we have conducted at RMIT to start charting the problems in this
particular space.
Another crucial challenge for search engine companies is how to ensure that
users are delivered the best quality content. There is a growth in systems that
recommend content based not only on queries, but also on user context. The
problem is that the quality of these systems is highly variable; one way of tackling
this problem is gathering context from a wider range of places. I will present some
of the possible new approaches to providing that context to search engines. Here
diverse social media, and advances in location technologies will be emphasized.
Finally, I will describe what I see as one of the more important challenges
that face the whole of the information community, namely the penetration of
computer systems to virtually every person on the planet and the challenges
that such an expansion presents.
 Table of Contents

Tutorials
Understanding Short Texts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Haixun Wang

Managing the Wisdom of Crowds on Social Media Services . . . . . . . . . . . . 2

Lei Chen

Search on Graphs: Theory Meets Engineering . . . . . . . . . . . . . . . . . . . . . . . . 3

Yuqing Wu and George H.L. Fletcher

Distributed Processing:
A Simple XSLT Processor for Distributed XML . . . . . . . . . . . . . . . . . . . . . . 7
Hiroki Mizumoto and Nobutaka Suzuki

Ontology Usage Network Analysis Framework . . . . . . . . . . . . . . . . . . . . . . . 19

Jamshaid Ashraf and Omar Khadeer Hussain

Energy Eciency in W-Grid Data-Centric Sensor Networks via

Workload Balancing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
Alfredo Cuzzocrea, Gianluca Moro, and Claudio Sartori

Update Semantics for Interoperability among XML, RDF and RDB:

A Case Study of Semantic Presence in CISCOs Unified Presence
Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
Muhammad Intizar Ali, Nuno Lopes, Owen Friel, and
Alessandra Mileo

Graphs
GPU-Accelerated Bidirected De Bruijn Graph Construction
for Genome Assembly . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
Mian Lu, Qiong Luo, Bingqiang Wang, Junkai Wu, and Jiuxin Zhao

K Hops Frequent Subgraphs Mining for Large Attribute Graph . . . . . . . . 63

Haiwei Zhang, Simeng Jin, Xiangyu Hu, Ying Zhang,
Yanlong Wen, and Xiaojie Yuan

Privacy Preserving Graph Publication in a Distributed Environment . . . . 75

Mingxuan Yuan, Lei Chen, Philip S. Yu, and Hong Mei
XX Table of Contents

Correlation Mining in Graph Databases with a New Measure . . . . . . . . . . 88

Md. Samiullah, Chowdhury Farhan Ahmed, Manziba Akanda Nishi,
Anna Fariha, S M Abdullah, and Md. Rafiqul Islam

Improved Parallel Processing of Massive De Bruijn Graph for Genome

Assembly . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
Li Zeng, Jiefeng Cheng, Jintao Meng, Bingqiang Wang, and
Shengzhong Feng

B3Clustering: Identifying Protein Complexes from Protein-Protein

Interaction Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
Eunjung Chin and Jia Zhu

Detecting Event Rumors on Sina Weibo Automatically . . . . . . . . . . . . . . . 120

Shengyun Sun, Hongyan Liu, Jun He, and Xiaoyong Du

Uncertain Subgraph Query Processing over Uncertain Graphs . . . . . . . . . 132

Wenjing Ruan, Chaokun Wang, Lu Han, Zhuo Peng, and Yiyuan Bai

Web Search and Web Mining

Improving Keyphrase Extraction from Web News by Exploiting
Comments Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
Zhunchen Luo, Jintao Tang, and Ting Wang

A Two-Layer Multi-dimensional Trustworthiness Metric for Web

Service Composition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
Han Jiao, Jixue Liu, Jiuyong Li, and Chengfei Liu

An Integrated Approach for Large-Scale Relation Extraction from the

Web . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
Naimdjon Takhirov, Fabien Duchateau, Trond Aalberg, and
Ingeborg Slvberg

Multi-QoS Eective Prediction in Web Service Selection . . . . . . . . . . . . . . 176

Zhongjun Liang, Hua Zou, Jing Guo, Fangchun Yang, and
Rongheng Lin

Accelerating Topic Model Training on a Single Machine . . . . . . . . . . . . . . . 184

Mian Lu, Ge Bai, Qiong Luo, Jie Tang, and Jiuxin Zhao

Collusion Detection in Online Rating Systems . . . . . . . . . . . . . . . . . . . . . . . 196

Mohammad Allahbakhsh, Aleksandar Ignjatovic,
Boualem Benatallah, Seyed-Mehdi-Reza Beheshti, Elisa Bertino, and
Norman Foo

A Recommender System Model Combining Trust with Topic Maps . . . . . 208

Zukun Yu, William Wei Song, Xiaolin Zheng, and Deren Chen
 Table of Contents XXI

A Novel Approach to Large-Scale Services Composition . . . . . . . . . . . . . . . 220

Hongbing Wang and Xiaojun Wang

XML, RDF Data and Query Processing

The Consistency and Absolute Consistency Problems of XML Schema
Mappings between Restricted DTDs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
Hayato Kuwada, Kenji Hashimoto, Yasunori Ishihara, and
Toru Fujiwara
Linking Entities in Unstructured Texts with RDF Knowledge Bases . . . . 240
Fang Du, Yueguo Chen, and Xiaoyong Du
An Approach to Retrieving Similar Source Codes by Control
Structure and Method Identifiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
Yoshihisa Udagawa
Complementary Information for Wikipedia by Comparing Multilingual
Articles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 260
Yuya Fujiwara, Yu Suzuki, Yukio Konishi, and Akiyo Nadamoto

Social Networks
Identification of Sybil Communities Generating Context-Aware Spam
on Online Social Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268
Faraz Ahmed and Muhammad Abulaish
Location-Based Emerging Event Detection in Social Networks . . . . . . . . . 280
Sayan Unankard, Xue Li, and Mohamed A. Sharaf
Measuring Strength of Ties in Social Network . . . . . . . . . . . . . . . . . . . . . . . 292
Dakui Sheng, Tao Sun, Sheng Wang, Ziqi Wang, and Ming Zhang
Finding Diverse Friends in Social Networks . . . . . . . . . . . . . . . . . . . . . . . . . . 301
Syed Khairuzzaman Tanbeer and Carson Kai-Sang Leung
Social Network User Influence Dynamics Prediction . . . . . . . . . . . . . . . . . . 310
Jingxuan Li, Wei Peng, Tao Li, and Tong Sun
Credibility-Based Twitter Social Network Analysis . . . . . . . . . . . . . . . . . . . 323
Jebrin Al-Sharawneh, Suku Sinnappan, and Mary-Anne Williams
Design and Evaluation of Access Control Model Based on Classification
of Users Network Behaviors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 332
Peipeng Liu, Jinqiao Shi, Fei Xu, Lihong Wang, and Li Guo
Two Phase Extraction Method for Extracting Real Life Tweets Using
LDA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 340
Shuhei Yamamoto and Tetsuji Satoh
XXII Table of Contents

Probabilistic Queries
A Probabilistic Model for Diversifying Recommendation Lists . . . . . . . . . 348
Yutaka Kabutoya, Tomoharu Iwata, Hiroyuki Toda, and
Hiroyuki Kitagawa

A Probabilistic Data Replacement Strategy for Flash-Based Hybrid

Storage System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 360
Yanfei Lv, Xuexuan Chen, Guangyu Sun, and Bin Cui

An Influence Strength Measurement via Time-Aware Probabilistic

Generative Model for Microblogs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 372
Zhaoyun Ding, Yan Jia, Bin Zhou, Jianfeng Zhang, Yi Han, and
Chunfeng Yu

A New Similarity Measure Based on Preference Sequences for

Collaborative Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 384
Tianfeng Shang, Qing He, Fuzhen Zhuang, and Zhongzhi Shi

Multimedia and Visualization

Visually Extracting Data Records from Query Result Pages . . . . . . . . . . . 392
Neil Anderson and Jun Hong

Leveraging Visual Features and Hierarchical Dependencies for

Conference Information Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 404
Yue You, Guandong Xu, Jian Cao, Yanchun Zhang, and
Guangyan Huang

Aggregation-Based Probing for Large-Scale Duplicate Image

Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 417
Ziming Feng, Jia Chen, Xian Wu, and Yong Yu

User Interest Based Complex Web Information Visualization . . . . . . . . . . 429

Shibli Saleheen and Wei Lai

Spatial-Temporal Databases
FIMO: A Novel WiFi Localization Method . . . . . . . . . . . . . . . . . . . . . . . . . . 437
Yao Zhou, Leilei Jin, Cheqing Jin, and Aoying Zhou

An Algorithm for Outlier Detection on Uncertain Data Stream . . . . . . . . 449

Keyan Cao, Donghong Han, Guoren Wang, Yachao Hu, and Ye Yuan

Improved Spatial Keyword Search Based on IDF Approximation . . . . . . . 461

Xiaoling Zhou, Yifei Lu, Yifang Sun, and Muhammad Aamir Cheema
 Table of Contents XXIII

Ecient Location-Dependent Skyline Retrieval with Peer-to-Peer

Sharing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 473
Yingyuan Xiao, Yan Shen, Hongya Wang, and Xiaoye Wang

Data Mining and Knowledge Discovery

What Can We Get from Learning Resource Comments on Engineering
Pathway . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 482
Yunlu Zhang, Wei Yu, and Shijun Li

Tuned X-HYBRIDJOIN for Near-Real-Time Data Warehousing . . . . . . . . 494

M. Asif Naeem

Exploiting Interaction Features in User Intent Understanding . . . . . . . . . . 506

Vincenzo Deufemia, Massimiliano Giordano, Giuseppe Polese, and
Luigi Marco Simonetti

Identifying Semantic-Related Search Tasks in Query Log . . . . . . . . . . . . . . 518

Shuai Gong, Jinhua Xiong, Cheng Zhang, and Zhiyong Liu

Privacy and Security

Multi-verifier: A Novel Method for Fact Statement Verification . . . . . . . . . 526
Teng Wang, Qing Zhu, and Shan Wang

An Ecient Privacy-Preserving RFID Ownership Transfer Protocol . . . . 538

Wei Xin, Zhi Guan, Tao Yang, Huiping Sun, and Zhong Chen

Fractal Based Anomaly Detection over Data Streams . . . . . . . . . . . . . . . . . 550

Xueqing Gong, Weining Qian, Shouke Qin, and Aoying Zhou

Preservation of Proximity Privacy in Publishing Categorical Sensitive

Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 563
Yujia Li, Xianmang He, Wei Wang, Huahui Chen, and Zhihui Wang

Performance
S2MART: Smart Sql to Map-Reduce Translators . . . . . . . . . . . . . . . . . . . . . 571
Narayan Gowraj, Prasanna Venkatesh Ravi, Mouniga V, and
M.R. Sumalatha

MisDis: An Ecent Misbehavior Discovering Method

Based on Accountability and State Machine in VANET . . . . . . . . . . . . . . . 583
Tao Yang, Wei Xin, Liangwen Yu, Yong Yang, Jianbin Hu, and
Zhong Chen
XXIV Table of Contents

A Scalable Approach for LRT Computation in GPGPU

Environments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 595
Linsey Xiaolin Pang, Sanjay Chawla, Bernhard Scholz, and
Georgina Wilcox

ASAWA: An Automatic Partition Key Selection Strategy . . . . . . . . . . . . . 609

Xiaoyan Wang, Jinchuan Chen, and Xiaoyong Du

Query Processing and Optimization

An Active Service Reselection Triggering Mechanism . . . . . . . . . . . . . . . . . 621
Ying Yin, Tiancheng Zhang, Bin Zhang, Gang Sheng,
Yuhai Zhao, and Ming Li

Linked Data Informativeness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 629

Rouzbeh Meymandpour and Joseph G. Davis

Harnessing the Wisdom of Crowds for Corpus Annotation

through CAPTCHA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 638
Yini Cao and Xuan Zhou

A Framework for OLAP in Column-Store Database: One-Pass Join and

Pushing the Materialization to the End . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 646
Yuean Zhu, Yansong Zhang, Xuan Zhou, and Shan Wang

A Self-healing Framework for QoS-Aware Web Service Composition via

Case-Based Reasoning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 654
Guoqiang Li, Lejian Liao, Dandan Song, Jingang Wang,
Fuzhen Sun, and Guangcheng Liang

The Second International Workshop on Data

Management for Emerging Network Infrastructure
Workload-Aware Cache for Social Media Data . . . . . . . . . . . . . . . . . . . . . . . 662
Jinxian Wei, Fan Xia, Chaofeng Sha, Chen Xu, Xiaofeng He, and
Aoying Zhou

Shortening the Tour-Length of a Mobile Data Collector in the WSN by

the Method of Linear Shortcut . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 674
Md. Shaifur Rahman and Mahmuda Naznin

Towards Fault-Tolerant Chord P2P System: Analysis of Some

Replication Strategies. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 686
Rafa!l Kapelko

A MapReduce-Based Method for Learning Bayesian Network from

Massive Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 697
Qiyu Fang, Kun Yue, Xiaodong Fu, Hong Wu, and Weiyi Liu
 Table of Contents XXV

Practical Duplicate Bug Reports Detection in a Large Web-Based

Development Community . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 709
Liang Feng, Leyi Song, Chaofeng Sha, and Xueqing Gong

Selecting a Diversified Set of Reviews . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 721

Wenzhe Yu, Rong Zhang, Xiaofeng He, and Chaofeng Sha

International Workshop on Social Media Analytics

and Recommendation Technologies
Detecting Community Structures in Microblogs from Behavioral
Interactions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 734
Ping Zhang, Kun Yue, Jin Li, Xiaodong Fu, and Weiyi Liu

Towards a Novel and Timely Search and Discovery System Using the
Real-Time Social Web. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 746
Owen Phelan, Kevin McCarthy, and Barry Smyth

GWMF: Gradient Weighted Matrix Factorisation for Recommender

Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 758
Nipa Chowdhury and Xiongcai Cai

Collaborative Ranking with Ranking-Based Neighborhood . . . . . . . . . . . . 770

Chaosheng Fan and Zuoquan Lin

International Workshop on Management

of Spatial Temporal Data
Probabilistic Top-k Dominating Query over Sliding Windows . . . . . . . . . . 782
Xing Feng, Xiang Zhao, Yunjun Gao, and Ying Zhang

Distributed Range Querying Moving Objects in Network-Centric

Warfare . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 794
Bin Ge, Chong Zhang, Da-quan Tang, and Wei-dong Xiao

An Ecient Approach on Answering Top-k Queries with Grid

Dominant Graph Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 804
Aiping Li, Jinghu Xu, Liang Gan, Bin Zhou, and Yan Jia

A Survey on Clustering Techniques for Situation Awareness . . . . . . . . . . . 815

Stefan Mitsch, Andreas M uller, Werner Retschitzegger,
Andrea Salfinger, and Wieland Schwinger
XXVI Table of Contents

Parallel k -Skyband Computation on Multicore Architecture . . . . . . . . . . . 827

Xing Feng, Yunjun Gao, Tao Jiang, Lu Chen, Xiaoye Miao, and
Qing Liu

Moving Distance Simulation for Electric Vehicle Sharing Systems

for Jeju City Area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 838
Junghoon Lee and Gyung-Leen Park

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 843

 Understanding Short Texts

Haixun Wang

Microsoft Research Asia, China

[email protected]

Abstract. Many applications handle short texts, and enableing ma-

chines to understand short texts is a big challenge. For example, in Ads
selection, it is is dicult to evaluate the semantic similarity between a
search query and an ad. Clearly, edit distance based string similarity
does not work. Moreover, statistical methods that find latent topic mod-
els from text also fall short because ads and search queries are insucient
to provide enough statistical signals.
In this tutorial, I will talk about a knowledge empowered approach
for text understanding. When the input is sparse, noisy, and ambiguous,
knowledge is needed to fill the gap in understanding. I will introduce
the Probase project at Microsoft Research Asia, whose goal is to enable
machines to understand human communications. Probase is a universal,
probabilistic taxonomy more comprehensive than any current taxonomy.
It contains more than 2 million concepts, harnessed automatically from
a corpus of 1.68 billion web pages and two years worth of search-log
data. It enables probabilistic interpretations of search queries, document
titles, ad keywords, etc. The probabilistic nature also enables it to in-
corporate heterogeneous information naturally. I will explain how the
core taxonomy, which contains hypernym-hyponym relationships, is con-
structed and how it models knowledges inherent uncertainty, ambiguity,
and inconsistency.

1 Speakers
Haixun Wang is a senior researcher at Microsoft Research Asia in Beijing, China,
where he manages the Data Management, Analytics, and Services group. Before
joining Microsoft, he had been a research sta member at IBM T. J. Watson
Research Center for 9 years. He was Technical Assistant to Stuart Feldman (Vice
President of Computer Science of IBM Research) from 2006 to 2007, and Tech-
nical Assistant to Mark Wegman (Head of Computer Science of IBM Research)
from 2007 to 2009. Haixun Wang has published more than 120 research papers
in referred international journals and conference proceedings. He is on the edi-
torial board of Distributed and Parallel Databases (DAPD), IEEE Transactions
of Knowledge and Data Engineering (TKDE), Knowledge and Information Sys-
tem (KAIS), Journal of Computer Science and Technology (JCST). He is PC
co-Chair of WWW 2013 (P&E), ICDE 2013 (Industry), CIKM 2012, ICMLA
2011, WAIM 2011. Haixun Wang got the ER 2008 Conference best paper award
(DKE 25 year award), and ICDM 2009 Best Student Paper run-up award.

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, p. 1, 2013.

c Springer-Verlag Berlin Heidelberg 2013

 Managing the Wisdom

of Crowds on Social Media Services

Lei Chen

Hong Kong University of Science and Technology, Hong Kong

[email protected]

Abstract. Recently, the Wisdom of Crowds has attracted a huge

amount of interests from both research and industrial communities. For
the current stage, most of focus has been put on several specific crowd-
sourcing marketplaces like Amazon MTurk or CrowdFlower, on which
requesters publish tasks and workers select tasks according to their
own benefits. However, users on social media services can also serve as
candidate workers for crowdsourcing tasks, and it is possible for the
requesters to actively manage the quality and cost of such crowds.
In this tutorial, we will first review the basic concept of crowdsourcing
and its applications. Then, we will discuss the current popular crowd-
sourcing platforms and several interesting crowdsourcing related algo-
rithms. Finally, we will discuss the benefits of using social media services
as the crowdsourcing platform and propose several research challenges
on managing the wisdom of crowds on social media.

1 Speakers
Lei Chen received the BS degree in Computer Science and Engineering from
Tianjin University, Tianjin, China, in 1994, the MA degree from Asian Institute
of Technology, Bangkok, Thailand, in 1997, and the PhD degree in computer
science from the University of Waterloo, Waterloo, Ontario, Canada, in 2005.
He is currently an Associate Professor in the Department of Computer Science
and Engineering, Hong Kong University of Science and Technology. His research
interests include crowd sourcing on social media, social media analysis, proba-
bilistic and uncertain databases, and privacy-preserved data publishing. So far,
he has published nearly 200 conference and journal papers. He got the Best
Paper Awards in DASFAA 2009 and 2010. He is PC Track Chairs for VLDB
2014, ICDE 2012, CIKM 2012, SIGMM 2011. He has served as PC members
for SIGMOD, VLDB, ICDE, SIGMM, and WWW. Currently, he serves as an
Associate Editor for IEEE Transaction on Data and Knowledge Engineering and
Distribute and Parallel Databases. He is a member of the ACM and the chairman
of ACM Hong Kong Chapter.

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, p. 2, 2013.

c Springer-Verlag Berlin Heidelberg 2013

 Search on Graphs: Theory Meets Engineering

Yuqing Wu1 and George H.L. Fletcher2

1
Indiana University, Bloomington, USA
[email protected]
2
Eindhoven University of Technology, The Netherlands
[email protected]

Abstract. The last decade has witnessed an explosion of the availability

of and interest in graph structured data. The desire to search and reason
over these increasingly massive data collections pushes the boundaries of
search languages, from pure keyword search to structure-aware searches
in the graph. These phenomena have inspired a rich body of research
on query languages, data management and query evaluation techniques
for graph data, both from the theoretical and engineering angles. In this
tutorial, we present an overview of the progress on graph search queries,
focusing specifically on how the theoretical and engineering perspectives
meet and together advanced the field.

1 Tutorial Overview

Exploratory keyword-style search has been heavily studied in the past decade,
both in the context of structured [18] and semi-structured [9] data. Given the
ubiquity of massive (loosely structured) graph data in domains such as the web,
social networks, biological networks, and linked open data (to name a few), there
recently has been a surge of interest and advances on the problem of search in
graphs (e.g., [1, 3, 12, 15, 17, 19, 20]). As graph exploration leads to deeper
domain understanding, user queries begin to shift from unstructured searching
to richer structure-based exploration of the graph. Consequently, there has been
a flurry of language proposals specifically targeting this style of structure-aware
querying in graphs (e.g., [2, 3, 5, 13, 16]).
In this tutorial, we survey this growing body of work, with an eye towards
both bringing participants up to speed in this field of rapid progress and delim-
iting the boundaries of the state-of-the-art. A particular focus will be on recent
results in the theory of graph languages on the design and structural charac-
terization of simple yet powerful algebraic languages for graph search, which
bridge structure-oblivious and structure-aware graph exploration [46]. At the
heart of these results is the methodology of coupling the expressive power of a
given query language with an appropriate structural notion on data instances.
Here, the idea is to characterize language equivalence of data objects in in-
stances (i.e., the inability of queries in the language to distinguish the objects)
purely in terms of the structure of the instance (i.e., equivalence under no-
tions such as homomorphism or bisimilarity). Recently, first steps towards graph

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 36, 2013.
c Springer-Verlag Berlin Heidelberg 2013

4 Y. Wu and G.H.L. Fletcher

indexing have shown promise in transferring this theoretical framework into

practice [7, 14]. The basic intuition behind this approach is that data should
be organized to optimally reflect the type and style of queries being asked, and
that the optimality of this organization can be formally established, as above.
Recent results have established the practical feasibility of computing and main-
taining these structural organizations on massive graphs [8, 10, 11].

2 Tutorial Outline

The tutorial will be presented as follows:

Part 1: Searching the Graph
We start the tutorial with an overview of variants of graph data and the
evolution of the search queries on graph, leading into the discussion of the
challenges posed by the massive size and complex nature of graph data and
the flexible nature of graph queries and their evaluation.
Part 2: Bridging Theory and Engineering
We next examine the lines of work in the theoretical study of query lan-
guages and the engineering eorts in developing novel techniques for manag-
ing graph data and evaluating various types of search queries on such data.
We then present our methodology of coupling the expressive power of a given
query language with an appropriate structural notion on data instances, as
a tool for reasoning about and guiding engineering eorts.
Part 3: Indexing Graph Data A Case Study
We follow this discussion with a presentation of the design of indices for semi-
structured and graph data, as a case study, to illustrate our methodology.
Part 4: Looking Forward
We close the tutorial with indications of ongoing and future research direc-
tions.

3 Speakers
Yuqing Wu and George Fletcher, together with a group of collaborators in the
USA, the Netherlands, and Belgium, have been conducting research in this area
in recent years and have published several papers in both the theory and engi-
neering branches of database research.

Yuqing Wu, Indiana University, Bloomington, USA

Prof. Wu is an Associate Professor at School of Informatics and Computing,
Indiana University, Bloomington, USA. Prof. Wu received her Ph.D. degree
from University of Michigan, Ann Arbor, in 2004. Her research area is in
data management, especially semi-structure and non-structured data, with
an emphasis on query language, query processing and query optimization.
 Search on Graphs: Theory Meets Engineering 5

George Fletcher, Eindhoven University of Technology, The Netherlands

Dr. Fletcher is an Assistant Professor in the Databases and Hypermedia
group at Eindhoven University of Technology, The Netherlands. Dr. Fletcher
was awarded a doctorate in computer science from Indiana University, Bloom-
ington (2007), with a dissertation on the topic of query learning for data
integration. His current research focuses on the study of database query
languages for data integration and web data.

References
1. Delbru, R., Campinas, S., Tummarello, G.: Searching web data: An entity retrieval
and high-performance indexing model. J. Web Sem. 10, 3358 (2012)
2. Fazzinga, B., Gianforme, G., Gottlob, G., Lukasiewicz, T.: Semantic web search
based on ontological conjunctive queries. J. Web Sem. 9(4), 453473 (2011)
3. Fletcher, G.H.L., Van den Bussche, J., Van Gucht, D., Vansummeren, S.: Towards
a theory of search queries. ACM Trans. Database Syst. 35(4), 28 (2010)
4. Fletcher, G.H.L., Gyssens, M., Leinders, D., Van den Bussche, J., Van Gucht, D.,
Vansummeren, S.: Similarity and bisimilarity notions appropriate for characterizing
indistinguishability in fragments of the calculus of relations. CoRR, abs/1210.2688
(2012)
5. Fletcher, G.H.L., Gyssens, M., Leinders, D., Van den Bussche, J., Van Gucht, D.,
Vansummeren, S., Wu, Y.: Relative expressive power of navigational querying on
graphs. In: Proc. ICDT, Uppsala, Sweden, pp. 197207 (2011)
6. Fletcher, G.H.L., Gyssens, M., Leinders, D., Van den Bussche, J., Van Gucht,
D., Vansummeren, S., Wu, Y.: The impact of transitive closure on the boolean
expressiveness of navigational query languages on graphs. In: Lukasiewicz, T., Sali,
A. (eds.) FoIKS 2012. LNCS, vol. 7153, pp. 124143. Springer, Heidelberg (2012)
7. Fletcher, G.H.L., Hidders, J., Vansummeren, S., Picalausa, F., Luo, Y., De Bra, P.:
On guarded simulations and acyclic first-order languages. In: Proc. DBPL, Seattle,
WA, USA (2011)
8. Hellings, J., Fletcher, G.H.L., Haverkort, H.: Ecient external-memory bisimula-
tion on DAGs. In: Proc. ACM SIGMOD, Scottsdale, AZ, USA, pp. 553564 (2012)
9. Liu, Z., Chen, Y.: Processing keyword search on XML: a survey. World Wide
Web 14(5-6), 671707 (2011)
10. Luo, Y., de Lange, Y., Fletcher, G.H.L., De Bra, P., Hidders, J., Wu, Y.: Bisimu-
lation reduction of big graphs on MapReduce (manuscript in preparation, 2013)
11. Luo, Y., Fletcher, G.H.L., Hidders, J., Wu, Y., De Bra, P.: I/O-ecient algo-
rithms for localized bisimulation partition construction and maintenance on mas-
sive graphs. CoRR, abs/1210.0748 (2012)
12. Mass, Y., Sagiv, Y.: Language models for keyword search over data graphs. In:
Proc. ACM WSDM, Seattle, Washington, USA (2012)
13. Perez, J., Arenas, M., Gutierrez, C.: nSPARQL: A navigational language for RDF.
J. Web Sem. 8(4), 255270 (2010)
14. Picalausa, F., Luo, Y., Fletcher, G.H.L., Hidders, J., Vansummeren, S.: A struc-
tural approach to indexing triples. In: Simperl, E., Cimiano, P., Polleres, A., Cor-
cho, O., Presutti, V. (eds.) ESWC 2012. LNCS, vol. 7295, pp. 406421. Springer,
Heidelberg (2012)
15. Tran, T., Herzig, D.M., Ladwig, G.: SemSearchPro - using semantics throughout
the search process. J. Web Sem. 9(4), 349364 (2011)
6 Y. Wu and G.H.L. Fletcher

16. Tran, T., Wang, H., Rudolph, S., Cimiano, P.: Top-k exploration of query candi-
dates for ecient keyword search on graph-shaped (RDF) data. In: Proc. IEEE
ICDE, Shanghai, pp. 405416 (2009)
17. Wu, Y., Van Gucht, D., Gyssens, M., Paredaens, J.: A study of a positive fragment
of path queries: Expressiveness, normal form and minimization. Comput. J. 54(7),
10911118 (2011)
18. Yu, J.X., Qin, L., Chang, L.: Keyword search in relational databases: A survey.
IEEE Data Eng. Bull. 33(1), 6778 (2010)
19. Zhou, M., Pan, Y., Wu, Y.: Conkar: constraint keyword-based association discov-
ery. In: Proc. ACM CIKM, Glasgow, UK, pp. 25532556 (2011)
20. Zhou, M., Pan, Y., Wu, Y.: Ecient association discovery with keyword-based
constraints on large graph data. In: Proc. ACM CIKM, Glasgow, UK, pp. 2441
2444 (2011)
A Simple XSLT Processor for Distributed XML

Hiroki Mizumoto and Nobutaka Suzuki

University of Tsukuba
1-2, Kasuga, Tsukuba Ibaraki 305-8550, Japan
{s0911654@u,nsuzuki@slis}.tsukuba.ac.jp

Abstract. Recently, the sizes of XML documents have rapidly been

increasing, and due to geographical and administrative reasons XML
documents are tend to be partitioned into fragments and managed sepa-
rately in plural sites. Such a form of XML documents is called distributed
XML. In this paper, we propose a method for performing XSLT transfor-
mation eciently for distributed XML documents. Our method assumes
that the expressive power of XSLT is restricted to unranked top-down
tree transducer, and all the sites storing an XML fragment perform an
XSLT transformation in parallel. We implemented our method in Ruby
and made evaluation experiments. This result suggests that our method
is more ecient than a centralized approach.

1 Introduction
XML has been a de facto standard format in the Web, and the sizes of XML doc-
uments have rapidly been increasing. Recently, due to geographical and admin-
istrative reasons an XML document is partitioned into fragments and managed
separately in plural sites. Such a form of XML documents is called distributed
XML[5,2,1]. For example, Figures 1 and 2 show a distributed XML document
of an auction site. In this example, one XML document is partitioned into four
fragments F0 , F1 , F2 , and F3 , and fragment F1 is stored in site S1 , fragment F2
is stored in site S2 , and so on. We say that S1 and S3 are the child sites of S0
(S0 is the parent site of S1 and S3 ).
In this paper, we consider XSLT transformation for distributed XML docu-
ments. An usual centralized approach for performing an XSLT transformation
of a distributed XML document is to send all the fragments to a specific site,
then merge all the fragments into one XML document, and perform an XSLT
transformation to the merged document. However, this approach is inecient
due to the following reasons. First, in this approach an XSLT transformation
processing is not load-balanced. Second, an XSLT transformation becomes in-
ecient if the size of the target XML document is large[12]. This implies that
the centralized approach is inecient even if the size of each XML fragment is
small, whenever the merged document is large.
In this paper, we propose a method for performing XSLT transformation ef-
ficiently for distributed XML documents. In our method, all the sites having an
XML fragment perform an XSLT transformation in parallel. This avoids cen-
tralized XSLT transformation for distributed XML documents and leads to an

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 718, 2013.
c Springer-Verlag Berlin Heidelberg 2013

8 H. Mizumoto and N. Suzuki

Fig. 1. Auction XML document

ecient XSLT processing. For distributed XML, however, we have to consider

handling modes of XSLT templates carefully, since multiple templates may be
defined to a single element, e.g., in Fig. 3 two templates are applicable to a
and the other two templates are applicable to b. This implies that we cannot
determine the template that should be applied to a fragment F until the trans-
formation of the parent fragment of F is completed. For example, in Fig. 1,
in order to transform F2 we need the mode applied to F2 that is obtained af-
ter the transformation of F1 is completed. Thus, a child site have to wait until
the transformation of its parent site is completed, and thus fragments stored in
distributed sites cannot be transformed in parallel. To handle this problem, our
method takes the following approach.
1. For a site S having fragment F , S applies all the templates applicable to F
simultaneously using native threads.
2. When the parent site S of S completes the transformation of its fragment,
the mode m that should be applied to F is identified. S sends the mode m
to S.
3. S chooses the result transformed in mode m, and returns it as the result.
Since XSLT is Turing-complete[6], it is impossible to plan a complete strategy of
XSLT transformation for distributed XML. Therefore, in this paper we restrict
the expressive power of XSLT to unranked top-down tree transducer[8]. We
implemented our method in Ruby and made evaluation experiments. The result
suggests that our method is more ecient than the centralized approach.

Related Work
A distribution design of XML documents is firstly proposed in [3]. There have
been several studies on evaluations of XPath and other languages for distributed
XML. [4,5] propose an ecient XPath evaluation algorithms for distributed
XML. [7] proposes a method for evaluating XQ, a subset of XPath, for vertically
partitioned XML documents. [10] considers a regular path query evaluation in
an distributed environment. [11] extensively studies the complexities of regular
 A Simple XSLT Processor for Distributed XML 9

Fig. 2. Fragments of an XML document

path query and structural recursion over distributed semistructured data. Be-
sides query languages, [2,1] study on the complexities of schema design problems
for distributed XML. To the best of the authors knowledge, there is no study
on XSLT evaluation for distributed XML.

2 Definitions

Trees, Hedges, and Tree Transducer

Since our method is based on unranked top-down tree transducer, we first show
related definitions. Let be a set of labels. By T we mean the set of unranked
-trees. A tree whose root is labeled with a and has n subtrees t1 , , tn is
denoted by a(t1 tn ). In the following, we always mean -tree whenever we say
tree. A hedge is a finite sequence of trees. The set of hedges is denoted by H .
In the following, we use t, t1 , t2 , to denote trees and h, h1 , h2 , to denote
hedges. For a set Q, by H (Q) (T (Q)) we mean the set of -hedges (resp.,
-trees) such that leaf nodes can be labeled with elements from Q.
A tree transducer is a quadruple (Q, , q0 , R), where Q is a finite set of states,
is the set of labels, q0 Q is the initial state, and R is a finite set of rules of
the form (q, a) h, where a , q Q and h H (Q). If q = q0 , then h is
restricted to T (Q) \ Q. A state corresponds to a mode of XSLT.
The translation defined by a tree transducer T r = (Q, , q0 , R) on a tree t in
state q, denoted by T rq (t), is inductively defined as follows.

R1: If t = , then T rq (t) := .

R2: If t = a(t1 tn ) and there is a rule (q, a) h R, then T rq (t) is obtained
from h by replacing every node u in h labeled with p Q by the hedge
T rp (t1 ) T rp (tn ).
R3: If there is no rule (q, a) h in R, then T rp (t) := .

The transformation of t by T r, denoted by T r(t), is defined as T rq0 (t).

10 H. Mizumoto and N. Suzuki

<xsl:template match="a" mode="p">

<x>
<z/>
</x>
</xsl:template>

<xsl:template match="a" mode="q">

<z>
<xsl:apply-templates mode="q" />
</z>
</xsl:template>

<xsl:template match="b" mode="p">

<x>
<xsl:apply-templates mode="p" />
<xsl:apply-templates mode="q" />
</x>
</xsl:template>

<xsl:template match="b" mode="q">

<y>
<xsl:apply-templates mode="p" />
</y>
</xsl:template>
Fig. 3. An example XSLT script

Example 1. Let T r = (Q, , p, R) be a tree transducer, where

Q = {p, q},
= {a, b, x, y, z},
R = {(p, a) x(z), (q, a) z(q), (p, b) x(p q), (q, b) y(p)}.

Tr corresponds to the XSLT script shown in Fig. 3. For example, consider the
rule (p, a) x(z) in R. This corresponds to the first template in Fig. 3. The
mode p in the left-hand side of the rule corresponds to the mode of the template,
and label a in the left-hand side of the rule corresponds to the match attribute.
The tree t in Fig. 4 is transformed to T r(t) in Fig. 5.

Fig. 4. Input tree t Fig. 5. Output tree T r(t)

 A Simple XSLT Processor for Distributed XML 11

Fig. 6. Fragment F0

Distributed XML
In this paper, we consider settings in which an XML tree t is partitioned into a
set Ft of disjoint subtrees of t, where each subtree is called fragment. We assume
that each fragment Fi Ft is stored in a distinct site. We allow arbitrary
nesting of fragments. Thus, fragments can appear at any level of the tree.
For example, the XML tree t T in Fig. 1 is partitioned into four fragments,
Ft = {F0 , F1 , F2 , F3 }. For a tree t, the fragment containing the root node of t is
called root fragment. In Fig. 2, the root fragment is F0 .
For two fragments Fi and Fj , we say that Fj is a child fragment of Fi if
the root node of Fj corresponds to a leaf node v of Fi . In order to represent a
connection between Fi and Fj , we use an EntityReference node at the position
of v, which refers the root node of Fj . For example, in Fig. 6 EntityReference
nodes &fragment1 and &fragment3 are inserted into F0 at the positions of
regions and people, respectively.
Each fragment is stored in a site. The site having the root fragment is called
root site and the other sites are called slave sites. For example, in Fig. 2 S0 is
the root site and S1 , S2 , S3 are slave sites. We assume that no two fragments are
stored in the same site. If the fragment in site S has a child fragment stored in
S , then S is a child site of S (S is the parent site of S ). For example, in Fig. 2
S0 has two child sites S1 and S3 .

3 Transformation Method
3.1 Outline
In our transformation method, all the sites S transform the fragment stored in
S in parallel, in order to avoid transformation processes being centralized on a
specific site. To realize this strategy, we have a problem to solve. For a fragment
F , the template that should be applied to F cannot be determined until the
transformation of the parent fragment of F is completed. For example, consider
the right child a of the root in Fig.4. Since this node is labeled with a, we have
two rules that can be applied to the node, (p, a) x(z) and (q, a) z(q),
and we cannot determine which of the rules should be applied to this node
12 H. Mizumoto and N. Suzuki

until the transformation of the root node is completed. Thus, for a fragment
F and the rules r1 , , rn that can be applied to F , our method proceeds the
transformation of F as follows.

1. Let S be the site storing F . S applies r1 , , rn to F in parallel, using native

threads. Thus S will obtain n transformation results.
2. When the parent site S of S completes the transformation of the parent
fragment of F , we can identify the mode(s) that should be applied to F . S
sends the mode(s) to S.
3. S receives the mode(s) that should be applied to F from S . S choose the
appropriate transformation result(s) of the n transformation results and send
the result(s) to the root site.

3.2 Our Transformation Method

We now present the details of our method. We use two XSLT possessors Master-
XSLT and Slave-XSLT. First, Master-XSLT is used in the root site and has the
following functions.

Transforming the root fragment.

Sending appropriate modes to its child sites according to the transformation
result of the root fragment.
Merging (a) the transformed root fragment and (b) the transformed frag-
ments received from the slave sites.

Second, Slave-XSLT is used in a slave site and has the following functions.

Transforming a fragment F stored in the slave site.

Applying rules to F in parallel, using native threads of Ruby.
Sending appropriate modes to its child sites according to the transformation
result of F .
Sending the transformed result of F to the root site.

Let us first present Master-XSLT. This procedure first sends a tree transducer
(XSLT styleseet) to each slave site (lines 1 to 3), then transforms the root frag-
ment by procedure Transform shown later (line 5), and receives the transformed
fragments from the slave sites (line 6), and finally merges the fragments (line 7).

Master-XSLT
Input: A tree transducer Tr = (Q, , q0 , R) and the root fragment F .
Output: Tree T T .

1. for each slave site S do

2. Send tree transducer Tr to S.
3. end
4. v the root node of F ;
5. F Transform(Tr , F, v, q0 );
 A Simple XSLT Processor for Distributed XML 13

6. Wait until transformed fragment Fi is received from each slave site Si .

Let F1 , , Fk be the received fragments.
7. Merge F and F1 , , Fk into T .
8. Return T ;
We next present procedure Transform used in line 5 above. This procedure trans-
forms a given fragment recursively. Lines 4 to 21 transform the subfragment
rooted at v of F according to Tr . Line 4 to 13 apply rule R2 and line 21 applies
rule R3 of the definition of tree transducer. In lines 1 to 3, when the procedure
encounters an EntityReference node v , then the procedure identifies the mode(s)
q that should be applied to fragment F (v ) and thus sends the mode(s) to site
S(v ), where F (v ) denotes the fragment that v referrers to and S(v ) denotes
the site storing F (v ).

Procedure Transform
Input: A tree transducer Tr = (Q, , q0 , R), a fragment F , the context node v
of F , and a mode q.
Output: A transformed fragment of F in mode q.
1. if v is an EntityReference node then
2. By referring to the parent node of v , identify the mode q that should
be applied to F (v ).
3. Send the mode q to S(v ).
4. else if (q, v ) h R for some h then
5. Q {q | q is a leaf node of h} Q;
6. if v has a child node v1 , , vk then
7. for each q Q do
8. for each child node vi {v1 , , vk } of v do
9. Fi the subtree rooted at vi of F ;
10. Ti Transform(Tr , Fi , vi , q );
11. end
12. Replace node q in h with hedge T1 Tk .
13. end
14. else
15. for each q Q do
16. q ;
17. end
18. end
19. Replace the subtree rooted at v of F with h.
20. else
21. Replace the subtree rooted at v of F with .
22. end
23. Return F ;
Finally, we present Slave-XSLT. This procedure runs in each slave site S and
transforms the fragment F stored in S. This procedure starts when it receives
a tree transducer from the root site, and then transforms F by using the rules
14 H. Mizumoto and N. Suzuki

that can be applied to F in parallel, using threads (lines 3 to 8). Thus more than
one transformation results may be obtained. Then the procedure waits until the
mode(s) p that should be applied to F is received from the parent site (line 9),
and sends the fragment(s) transformed in mode(s) p to the root site (line 11).
Note that if F contains EntityReference nodes v, we have to tell the appropriate
modes of v to child sites S(v). This is done in lines 12 to 15.

Procedure Slave-XSLT
Input: A fragment F stored in its own site.
Output: none
1. Wait until tree transducer Tr = (Q, , q0 , R) is received from the root site.
2. vr the root node of F ;
3. M odes {q | (q, vr ) h R for some h};
4. for each q M odes do
5. Thread start
6. Tq Transform(Tr , F, vr , q);
7. Thread end
8. end
9. Wait until mode(s) p is received from the parent site.
10. if Tp = null then
11. Send Tp to the root site.
12. for each EntityReference node v in F do
13. Identify the mode q of v in Tp .
14. Send q to S(v).
15. end
16. else
17. Send to the root site.
18. end
For example, let us consider the distributed XML document consisting of two
fragments m and f in Fig. 7. Let Tr be the tree transducer defined in Example 1.
Moreover, let Sr be the root site storing m and Ss be the slave site storing f .
Master-XSLT running in Sr sends Tr to site S(&f ragment) = Ss and starts the
transformation of m with the initial mode p (Figure 8 shows the the transfor-
mation result). Slave-XSLT running in Ss receives Tr and starts to transform
f . In this case, two rules (p, a) x(z) and (q, a) z(q) can be applied to
the root node a of f . Thus, Slave-XSLT applies the two rules to f in parallel,
and the obtained results are shown in Fig. 9. When Master-XSLT encounters
the EntityReference node &f ragment of m, the procedure sends the appro-
priate mode(s) to S(&f ragment). In this case, both modes p and q are send to
S(&f ragment). After Slave-XSLT completes the transformation of f , the pro-
cedure sends the transformed fragments. In this case, the two fragments shown
in Fig. 9 are send to the root site Sr . Finally, Sr merges the three fragments and
we obtain the output tree in Fig.10.
 A Simple XSLT Processor for Distributed XML 15

Fig. 7. A distributed XML document consisting of fragments m and f

Fig. 8. Transformed fragment of m Fig. 9. Transformed fragments of f

4 Evaluation Experiment

In this section, we present experimental results on our method. We implemented

our method in Ruby 1.93. We used 4 Linux machines, distributed over a local
LAN (100base-TX). Each machine has a 2.4GHz Intel Xeon CPU and 4GB of
memory. We used five XML documents generated by XMark[9], each of the
documents were divided into four fragments (Table 1). Each XML document is
divided as follows. As shown in Fig. 1, the root fragment F0 has site as the
root node, fragment F1 has regions as the root node, fragment F2 has asia
as the root node, and fragment F3 has people as the root node. Fragment Fi is
stored in site Si for 0 i 3. We measure the execution time of (a) a centralized
method and (b) our transformation method. In approach (a), fragments F1 , F 2,
and F3 are first sent to S0 , then F0 to F3 are merged into one document T and
an XSLT transformation is performed on T in site S0 .
We used the following three XSLT stylesheets. These are synthetic stylesheets
generated by our Ruby program.

Fig. 10. Transformation result

16 H. Mizumoto and N. Suzuki

Table 1. Sizes of distributed XML documents

Fragment size
Total size F0 F1 F2 F3
50MB 23MB 25MB 5.5MB 2.6MB
100MB 46MB 50MB 11MB 5.1MB
150MB 69MB 75MB 17MB 7.6MB
200MB 92MB 100MB 22MB 11MB
250MB 115MB 125MB 28MB 13MB

Fig. 11. Experimental result with Sheet-1

Fig. 12. Experimental result with Sheet-2

Sheet-1: This stylesheet has, for every element m of an XML document, at

least one template applicable to m. Two modes are used in this stylesheet.
This stylesheet consists of 137 templates.
Sheet-2: This stylesheet has, for every element m of an XML document, a
template applicable to m. A single mode is used in this stylesheet. This
stylesheet consists of 68 templates.
Sheet-3: This stylesheet has, for some element m of an XML document, at
least one template applicable to m. Two modes are used in this stylesheet.
This stylesheet consists of 94 templates.
 A Simple XSLT Processor for Distributed XML 17

Fig. 13. Experimental with Sheet-3

The results are shown in Figs. 11 to 13. As shown in these figures, our method
is about four times faster than the centralized method, regardless the used
stylesheets. This suggests that our method works well for distributed XML doc-
uments.

5 Conclusion
In this paper, we proposed a method for performing XSLT transformation for
distributed XML documents. The experimental results suggest that our method
work well for distributed XML documents.
However, we have a lot of future work to do. First, in this paper the expressive
power of XSLT is restricted to unranked top-down tree transducer. We have
investigated XSLT elements and functions, and we have found that about half
of the elements/functions can easily be incorporated into our method (Type
A of Table 2). The elements/functions f of this type can be calculated within
the fragment in which f is used, e.g., xslt:text. On the other hand, it seems
dicult to incorporate the rest elements/functions into our method (Type B of
Table 2). An example element of this type is xslt:for-each, which accesses
several fragments beyond the fragment in which the xslt:for-each element is
used. Thus, we have to handle XSLT elements/functions of Type B carefully in
order to extend the expressive power of our method. Another future work relates
to experimentation. In our experimentation we use only three synthetic XSLT
stylesheets. Thus we need to make more experiments using real-world XSLT
stylesheets.

Table 2. XSLT 1.0 elements and functions

elements function
Type A 18 18
Type B 17 16
Total 35 34
18 H. Mizumoto and N. Suzuki

Acknowledgement. This research is partially supported by Research Center

for Knowledge Communities, University of Tsukuba.

References
1. Abiteboul, S., Gottlob, G., Manna, M.: Distributed XML design. J. Comput. Syst.
Sci. 77(6), 936964 (2011)
2. Abiteboul, S., Gottlob, G., Manna, M.: Distributed XML design. In: Proc. PODS
2009 Proceedings of the Twenty-Eighth ACM SIGMOD-SIGACT-SIGART Sym-
posium on Principles of Database Systems, pp. 247258 (2009)
3. Bremer, J.M., Gertz, M.: On distributing XML repositories. In: Proc. of WebDB,
pp. 7378 (2003)
4. Buneman, P., Cong, G., Fan, W., Kementsietsidis, A.: Using partial evaluation in
distributed query evaluation. In: Proc. VLDB 2006 (2006)
5. Cong, G., Fan, W., Anastasios: Distributed Query Evaluation with Performance
Guarantees. In: Proc. SIGMOD 2007 Proceedings of the 2007 ACM SIGMOD
International Conference on Management of Data, pp. 509520 (2007)
6. Kepser, S.: A simple proof for the Turing-completeness of XSLT and XQuery. In:
Extreme Markup Languages (2004)

7. Kling, P., Ozsu, M.T., Daudjee, K.: Generating ecient execution plans for verti-
cally partitioned XML databases. Proc. VLDB Endow. 4(1), 111 (2010)
8. Martens, W., Neven, F.: Typechecking Top-Down Uniform Unranked Tree Trans-
ducers. In: Calvanese, D., Lenzerini, M., Motwani, R. (eds.) ICDT 2003. LNCS,
vol. 2572, pp. 6478. Springer, Heidelberg (2002)
9. Schmidt, A., Waas, F., Kersten, M., Carey, M.J., Manolescu, I., Busse, R.: Xmark:
a benchmark for XML data management. In: Proceedings of the 28th International
Conference on Very Large Data Bases, Proc. VLDB 2002, pp. 974985 (2002)
10. Stefanescu, D.C., Thomo, A., Thomo, L.: Distributed evaluation of generalized
path queries. In: Proc. the 2005 ACM Symposium on Applied Computing, SAC
2005, pp. 610616 (2005)
11. Suciu, D.: Distributed query evaluation on semistructured data. ACM Trans.
Database Syst. 27(1), 162 (2002)
12. Zavoral, F., Dvorakovam, J.: Perfomance of XSLT processors on large data sets.
In: Proc. Applications of Digital Information and Web Technologies, pp. 110115
(2009)
 Ontology Usage Network Analysis Framework

Jamshaid Ashraf and Omar Khadeer Hussain

School of Information Systems, CBS, Curtin University, Perth, WA 6845

[email protected], [email protected]

Abstract. Recently, there is tremendous growth in the use of ontologies

to publish semantically rich structured data on the Web. In order to
understand the adoption and uptake of ontologies in real world setting,
it is important to analyse the ontology usage. In this paper, we propose
Ontology Usage Network Analysis Framework (OUN-AF) which models
the ontology usage by dierent data publishers in the form of aliation
network. Metrics are defined to measure the ontology usage, their co-
usability and the cohesive subgroups emerging from the dataset.

1 Introduction
A decade-long eort by the Semantic Web community regarding knowledge
representation and assimilation has resulted in the development of methodolo-
gies, tools and technologies to develop and manage ontologies [1]. As a result,
numerous domain ontologies have been developed to describe the information
pertaining to dierent domains such as Healthcare and Life Science (HCLS),
entertainment, financial services and eCommerce. Consequently, we now have
billions of triples [2] on the Web in dierent domains, annotated by using dif-
ferent domain-specific ontologies [3] to provide structured and semantically rich
data on the Web.
In our previous work [4], Ontology Usage Analysis Framework (OUSAF) is
presented in order to analyse the use of ontologies on the web. The OUSAF
framework is comprised of four phases namely identification, investigation, rep-
resentation and utilization. The identification phase, which is the focus of this
paper, is responsible for identifying dierent ontologies that are being used in a
particular application area or in a given dataset for further analysis. A few of
the common requirements which form the selection criteria for the identification
of ontologies in this scenario are:
1. What are the widely used ontologies in the given application?
2. What ontologies are more interlinked with other ontologies to describe
domain-specific entities?
3. What ontologies are used more frequently and what is their usage percentage
based on the given dataset?
4. Which ontology clusters form cohesive groups?
In order to establish a better understanding of ontology usage and to identify
the ontologies, based on the abovementioned criteria, this paper proposes the

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 1930, 2013.
c Springer-Verlag Berlin Heidelberg 2013

20 J. Ashraf and O.K. Hussain

Ontology Usage Network Analysis Framework (OUN-AF) that models the use of
ontologies by dierent data sources using an Ontology Usage Network (OUN).
OUN represents the relationships between ontologies and data sources based
on the actual usage data available on the Web in the form of a graph-based
relationship structure. This structure is then analysed using metrics to study
the structural, typological and functional characteristics of OUN by applying
Social Network Analysis (SNA) techniques.

2 Background

2.1 Related Work

In [5], the authors have analyzed the social and structural relationships avail-
able on the Semantic Web, focusing mainly on FOAF and DC vocabularies. The
study was performed on approximately 1.5 million FOAF documents to analyze
instance data available on the Web and their usefulness in understanding so-
cial structures and networks. They identified the graphical patterns emerging in
social networks and represented this using the FOAF vocabulary and the de-
gree distribution of the network. As this research provides a detailed analysis of
Semantic Web data by focusing on a specific vocabulary, it does not provide a
framework or methodology to make it applicable to dierent vocabularies.
Other work in the literature has analyzed a dierent number of ontologies
using SNA. For example, [6] covered five ontologies, [7] analyzed 250 ontologies
and [8] used approximately 3,000 vocabularies. In all these work, ontologies were
investigated to measure their structural properties, the distribution of dierent
measures and terminological knowledge encoded in ontologies, but none includes
how they are being used on the Web. The use of SNA and its techniques to anal-
yse the use of ontologies and measure the relationships based on usage has only
been applied marginally. In the identification phase of the OUSAF framework,
the ontologies and their use by dierent data sources are represented in a way
that allows the aliation between ontologies and dierent data publishers to
be measured.

2.2 Aliation Network

Networks, in general sense, record relationships between entities. However, in

aliation network relationship reflect the aliation between entities and the
events in which entities participate. Aliation Networks are essentially two-
mode networks comprised of two disjoint sets of nodes with links which always
are between a node of one set and a node of the other. Aliation networks
are represented through an aliation matrix, A = {aij }. Matrix A is a two-
mode sociomatrix in which rows represent actors and columns represent events.
Generally, aliation network A is defined as: A is a bipartite graph A = (U, V, E)
where U (often known as actors) and V (often known as events) are disjoint
set of nodes and E (U XV ) is the set of edges. With p = |V | and q = |U |,
 Ontology Usage Network Analysis Framework (OUN-AF) 21

A is represented by an incident matrix with p lines and q columns. Formally,

A = {aij } records the aliation of each actor with each event in an aliation
matrix such that:
!
1, if actor i is aliated with event j
aij = (1)
0, otherwise
The value of 1 is put in the (i,j)th cell if ith actor (ith row) is aliated with jth
event (jth column) and an entry of 0 if ith actor is not aliated with jth event
as shown in Fig. 1.

Fig. 1. (a) Aliation matrix of author-paper aliation network, and (b) an example
of an author-paper aliation network

Often, two-mode networks are transformed into a one-mode network by a

procedure called projection which generates a one-mode network by selecting
nodes of one set (e.g. authors in Fig. 1) and linking two nodes from the set if
both are connected to the same node of the other set.

3 Ontology Usage Network Analysis Framework

(OUN-AF)

The objective of the ontology identification is to identify the use of dierent on-
tologies by dierent data publishers to discover hidden usage patterns. Therefore,
in order to mine such analysis, the Ontology Usage Network Analysis Framework
(OUN-AF) is proposed, as shown in Fig. 2. OUN-AF comprises three phases,
namely Input phase, Computation phase and Analysis phase.

3.1 Input Phase

The input phase is responsible for managing the dataset. The two key compo-
nents in this phase are a crawler and an RDF triple store. In order to point the
crawler to relevant and interesting data sources, the bootstrapping process first
builds the seed URIs as multiple starting points. A list of seed URIs is obtained
by accessing semantic search engines which return the URIs (URLs) of the data
sources (web sites) with structured data. The crawler collects the Semantic Web
data (RDF data) and after preprocessing it, loads it to the RDF triple store
(database). From a data management point of view, since RDF data comprises
22 J. Ashraf and O.K. Hussain

Fig. 2. Ontology Usage Network Analysis Framework (OUN-AF) and its phases

triples (statements) which do not provide a default mechanism to group or as-

sociate certain sets of triples to a context, the Named Graph approach is used
which enables the provision of contextualization by introducing an additional
URI (context) to a set of related triples.

3.2 Computation Phase

The computation phase provides the computational architecture to transform

the RDF data to a model so that further analysis can be performed. The OUN-
AF transforms the data into two formats: two-mode aliation network (i.e.
Ontology Usage Network ) and the subsequent one-mode networks (i.e. Ontology
Co-Usability and Data-Source Collaboration networks).
Ontology Usage Network (OUN) : OUN is an aliation network represented
as a bipartite graph. OUN comprises ontology and datasource sets of nodes with
an edge between the ontology node and data-source node if the ontology has
been used by the data source. To formally define the Ontology Usage Network,
first the two sets of nodes, namely ontology set and data-source set are defined
and then the OUN definition is presented.

An ontology set is defined as the set O which represents the nodes of the
first mode of the aliation network. An ontology set O contains the list of
ontologies used on the Web-of-Data such that there is a triple t < s, p, o >
anywhere in the dataset (specifically, otherwise in general, on the Web-of-
Data) where o O is the URI of either p or o.
A data-source set is defined as the set D which has the list of hostnames
on the Web-of-Data such that there exist a triple t < s, p, o > in the dataset
(specifically, otherwise in general, on the Web-of-Data), where s is the host-
name and either p or o is a member of O.
The Ontology Usage Network (OUN) is a bipartite graph, denoted as
OU N (O, D) that represents the aliation network, with a set of ontologies
 Ontology Usage Network Analysis Framework (OUN-AF) 23

O on one side and a set of data- sources D on the other and edge (o, d)
represents the fact that o is used by d.

In order to infer the connectedness present within one set of nodes based on their
co-participation in the other set of nodes, OUN is transformed from a two-mode
network to a one-mode network using the process of projection. Projection is
used to generate two one-mode graphs; one for nodes in the ontology set known
as the Ontology Co-Usability network, and second for the data-source set known
as the Data-Source Collaboration network as shown in Fig. 2.
The Ontology Co-Usability network is an ontology-to-ontology network,
in which two nodes are connected if both of the ontologies are being used by
the same data source. This means that the Ontology Co-Usability network rep-
resents the connectedness of an ontology with other ontologies, based on their
co-membership in the data source.
The Data-Source Collaboration network is a data-source-to-data-source
network in which two nodes are connected if both of them have used the same on-
tology to describe their data. The Data-Source Collaboration network represents
the similarity of data-sources in terms of their need to semantically describe the
information on the Web.

3.3 Analysis Phase

The analysis phase is the third and last phase of OUN-AF. The objective of this
phase is to mine the hidden relationships explicitly or implicitly present in the
two-mode network (i.e. OUN) and one-mode networks (Ontology Co-Usability
and Data-Source Collaboration). This is done by developing metrics for analyse
the two- and one-mode network as described in next section.

4 Metrics for Ontology Identification

The metrics used for the identification phase are presented below.

4.1 Ontology Usage Distribution (OUD)

The metric Ontology Usage Distribution, OU Dk , is used to identify which frac-
tion of the ontologies in the network have a degree k. OUD measures the degree
distribution of ontologies in an aliation network. In aliation network, the
degree of a node is the number of ties it has with the number of nodes of the
other set. The degree centrality CD (oi ) of an ontology oi is measured as:
n1
"
CD (oi ) = d(oi ) = Aij (2)
j=1

Where i = 1, . . . n1 , n1 = |O|, d(oi ) is the degree of ith ontology o, and A is the

aliation matrix representing OUN.
24 J. Ashraf and O.K. Hussain

4.2 Semanticity
Semanticity distribution identifies the fraction of the data sources in the net-
work which have a degree k. Semanticity measures the participation of dierent
ontologies in a given data source. The more ontologies are being used by a data
source, the higher semanticity value it has. Semanticity is measured by calculat-
ing the degree centrality and degree distribution on the second set of nodes in
an aliation network, which is the set representing the data sources present in
the dataset. The degree centrality CD (dsi ) of a data source dsi is measured as:
n2
"
CD (dsi ) = d(dsi ) = Aij (3)
j=1

Where i = 1, . . . n2 , n2 = |D|, d(dsi ) is the degree of ith data source ds, and A
is the aliation matrix representing OUN

4.3 Betweenness and Closeness Centrality

Like Ontology Co-Usability, both these centrality measures are computed on the
projected one-mode network.
Betweenness Centrality is the number of shortest paths between any two nodes
that passes through the given node. The betweenness centrality CB (vi ) of a node
vi is measured as :
" st (vi )
CB (vi ) = (4)
st
vs =vi =vt V

where st is the number of shortest paths between vertices vs and vt .

And st (vi ) is the number of shortest paths between vs and vt that pass
through vi .
Closeness centrality is a measure of the overall position of a node (actor) in
the network, giving an idea about how long it takes to reach other nodes in
the network from a given starting node. Closeness measures reachability, that
is, how fast a given node (actor) can reach everyone in the network [9] and is
defined as:
n1
cv = # (5)
uV d(v, u)

where d(v, u) denotes the length of a shortest-path between v and u. The close-
ness centrality of v is the inverse of the average (shortest path) distance from
v to any other vertex in the graph. The higher the cv , the shorter the average
distance from v to other vertices, and v is more important by this measure.

4.4 Cohesive Subgroups

Generally speaking, cohesive subgroups refer to the areas of the network in
which actors are more closely related to each other than actors outside the
group. Cohesive subgroups are often measures using n-clique, in which it is not
 Ontology Usage Network Analysis Framework (OUN-AF) 25

required that each member of the clique has a direct tie with the others, but
instead that it has to be no more than distance n from each other. Formally,
a clique is the maximum number of actors who have all possible ties present
between them.

5 Analysis of Ontology Usage Network

The detail of dataset which is used to populate OUN and the analysis performed
using metrics defined above, are discussed as follows.

5.1 Dataset and Its Characteristics

In order to build a dataset which has a fair representation of the Semantic Web
data described using domain ontologies, semantic search engines such as Sindice
and Swoogle are used to build the seed URLs. These seed URLs are then used
to crawl the structured data published on the Web using ontologies. The dataset
built for the identification phase comprises 22.3 millions of triples, collected from
211 data sources1 . In this dataset, 44 namespaces are used to describe entities
semantically.
The resulting Ontology Usage Network is comprise of 1390 edges linking 44
ontologies to 211 data sources. In terms of generic OUN properties, the density
of the network is 0.149 and the average degree is 10.90. The average degree shows
that the network is neither too sparse nor too dense which is a common pattern
in information networks. Details on the other properties and metrics are given
in the following subsections.

5.2 Analysing Ontology Usage Distribution (OUD)

Through OUD, we would like to determine how the use of an ontology is dis-
tributed over the data sources in the dataset. Using Eq. 2, Fig. 3(a) shows the
percentage of the ontologies being used by a number of dierent data sources.
The relative frequency of OUD on the dataset shows that there is both extreme
and average ontology usage by data sources. It also shows that 13.6% of on-
tologies are not used by any of the data sources and approximately half of the
ontologies are exclusively used by the data sources. The second row of the Fig.
3(a) shows that 47.7% of ontologies (21 ontologies) are being used by a data
source that has not used any other ontology. This means that there are several
ontologies in the dataset which either conceptualize a very specialized domain,
restricting their reusability, or are of a proprietary nature. From the third row
of Fig. 3(a) onwards, there is an increase in the reusability factor of ontologies.
This is because an increasingly large number of data sources are using them. The
last row shows that 4.5% (2 ontologies) of ontologies are being used equally by
1
https://docs.google.com/spreadsheet/ccc?key=
0AqjAK1TTtaSZdGpIMkVQUTRNenlrTGctR2J1bkl6WEE
26 J. Ashraf and O.K. Hussain

208 data sources. Through this analysis, we can see that there are less ontologies
which are not being used at all and there are a few which have almost optimal
utilization.
Fig. 3(b) shows the degree distribution of ontology usage in a number of data
sources. The value of degree is shown on the x-axis and the number of ontologies
with that degree is shown on the y-axis. It can be seen that there are a large
number of ontologies with a small degree value and only a few ontologies have a
larger degree value.

Fig. 3. (a) Distribution of Ontology Usage in data sources, and (b) Degree distribution
of ontology usage (data sources per ontology)

5.3 Analysing Semanticity

Semanticity measures the richness of a data source in terms of ontology usage.
Using Eq. 3, in the OUN, it is found that on average, 6.6 ontologies per data
source are used in the dataset which, in our view, is an encouragingly high
semanticity value, particularly bearing in mind that there are several ontologies
with very low ontology usage degree values such as 0 and 1, as described in the
previous section on Ontology Usage Distribution section. After determining the
average semanticity of the data sources, we then look at their degree distribution.
Figure 4(a) shows the relative frequency of ontologies being used by a number
of data sources. The degree distribution of ontology usage per data source is
dierent from ontology usage distribution.
At the lowest, in the network, two ontologies www.oettl.it and www.openlinksw.
com are used by one data source, which shows the lowest semanticity value and
14 is the maximum semanticity value which is also used by one data source. When
the data sources degree distribution is plotted, it follows the Gaussian distribu-
tion, as shown in Fig. 4(b). Gaussian distribution [10], which is essentially a bell
 Ontology Usage Network Analysis Framework (OUN-AF) 27

Fig. 4. (a) Distribution of Semanticity (Ontology used per data source), and (b)Degree
distribution of semanticity (Ontologies per Data source)

shaped curve, is normally concentrated in the centre and decreases on either side.
This signifies that degree has less tendency to produce extreme values compared
to power law distribution.

5.4 Analysing Betweenness and Closeness

Betweenness centrality measures the number of shortest paths going through a

certain node. However, on the other hand, the closeness centrality of a node in
the network is the inverse of the average shortest path distance from the node
to any other node in the network [11].
In the context of the Ontology Co-usage network, the interpretation of be-
tweenness and closeness measures are dierent from other collaboration networks
such as co-authorship collaboration networks [12]. In betweenness, which is mea-
sured using Eq. 4, the nodes of larger values are considered to be the hub of the
network, controlling the communication flow (or becoming the major facilitator)
between nodes with a geodesic path passing through these hub. In the Ontol-
ogy Co-usage network, ontologies are linked based on the fact that these are
being co-used by the data sources, therefore we believe the ontologies with max-
imum betweenness centrality act as a semantic gateway and become a major
motivational factor for the usage of other ontologies.
Likewise, in closeness centrality which is measured using Eq. 5, the larger
the value, the shorter the average distance from the node to any other node,
and thus the node (with a larger value) is positioned in the best location to
spread information quickly [13]. This centrality measure in the ontology co-
usage graph enables the establishment of correspondence between ontologies
which have concepts related to each other, supplementing each others concep-
tual model to form an exploded domain. The utilization of ontology indexing
based on closeness centrality is very similar to the features discussed in [14] in
supporting the application specific use of ontologies such as: (i) the ontologies
closer to each other in their usage are better candidates for vocabulary alignment;
28 J. Ashraf and O.K. Hussain

(ii) ontologies closer to each other have more entities which correspond to en-
tities of other ontologies; and (iii) closely related ontologies tend to facilitate
query answering on the Semantic Web.
The betweenness and closeness centrality of ontology co-usage nodes is shown
in Fig. 5(a) and 5(b), respectively and node size reflects the centrality value.

Fig. 5. (a) Betweenness centrality of Ontology Co-Usage network, and (b) Closeness
centrality of Ontology Co-Usage network

As we can see in Fig. 5(a), the Ontology Co-Usage network has very few nodes
with a higher betweenness value which means that the ontologies represented
by the green nodes (which are not many) are the ones falling in between the
geodesic path of many other nodes and acting as the gateway (or hub) in the
communication between other ontologies in the graph. These are the nodes,
namely rdf, rdfs, gr, vCard and foaf which, in our interpretation, are acting as
the semantic gateway by becoming the reason for the adoption of other ontologies
on the Web. However, on the other hand, closeness centrality is approximately
distributed evenly in the network. Thus, it is safe to assume that almost every
node is reachable to other nodes except those which are not connected.

5.5 Analysing Cohesive Subgroups

In a collaboration graph such as the Ontology Co-Usage network, a connected
component is a maximal set of ontologies that are mutually reachable (and con-
nected) through a chain of (co-usage) links. A cohesive sub-group analysis to
identify connected components of the Ontology Co-Usage network shows that
the network is widely connected. The connected component is 86.36% (See Fig-
ure 6), in which only six are not connected in the network (this means 0-core),
while others are connected with varying k-core values) making it a giant net-
work since it encompasses the majority of the nodes. This means that 86.36% of
the ontologies are reachable to each other by following the links (domain names
 Ontology Usage Network Analysis Framework (OUN-AF) 29

Fig. 6. Stacking of k-cores of Ontology Co-Usage network

URIs) of the data sources included in the dataset (or on the Web to generalize
it). Note that the size of the cohesive sub-group, in terms of percentage, closely
matches the findings of [15] for the classical Web which was 91%. Within the
giant connected component, to know the sub-component based on the equal dis-
tribution of the concentration of links around a set of nodes, k-core is computed.
k-core is the maximum sub-graph in which each node has at least degree k within
the sub-graph. Fig. 6 stacks the k-core components, based on ascending k values
from highest to lowest. From Figure 6, it is easy to see which ontologies are
highly linked, based on ontology usage patterns invariance across data sources.

6 Conclusion and Future Work

In this paper, the use ontologies by dierent data sources are analysed by repre-
senting the ontology usage related data in the form of aliation network. Using
dierent metrics the network is analysed to understand the usage patterns in real
world settings. In future work, we would like to enrich the dataset by crawling
more data sources and automating the analyses process.

References
1. Hitzler, P., van Harmelen, F.: A Reasonable Semantic Web. Semantic Web Jour-
nal 1(1-2), 3944 (2010)
2. Bizer, C., Jentzsch, A., Cyganiak, R.: State of the Linked Open Data (LOD) Cloud.
Technical Report April 5, 2011 (March 2011),
http://www4.wiwiss.fu-berlin.de/lodcloud/state/
3. Ashraf, J., Cyganiak, R., ORiain, S., Hadzic, M.: Open eBusiness Ontology Usage:
Investigating Community Implementation of GoodRelations. In: Proceedings of
Linked Data on the Web Workshop (LDOW) at WWW 2011, Hyderabad, India,
March 29. CEUR Workshop Proceedings, vol. 813, pp. 111. CEUR-WS.org (2011)
30 J. Ashraf and O.K. Hussain

4. Ashraf, J.: A Framework for Ontology Usage Analysis. In: Simperl, E., Cimiano,
P., Polleres, A., Corcho, O., Presutti, V. (eds.) ESWC 2012. LNCS, vol. 7295, pp.
813817. Springer, Heidelberg (2012)
5. Ding, L., Zhou, L., Finin, T., Joshi, A.: How the Semantic Web is Being Used:
An Analysis of FOAF Documents. In: Proceedings of the 38th Annual Hawaii
International Conference on System Sciences, vol. 4, pp. 113120. IEEE Computer
Society, Washington, DC (2005)
6. Zhang, H.: The scale-free nature of semantic web ontology. In: Proceeding of the
17th International Conference on World Wide Web, pp. 10471048. ACM (2008)
7. Theoharis, Y., Tzitzikas, Y., Kotzinos, D., Christophides, V.: On Graph Features
of Semantic Web Schemas. IEEE Transactions on Knowledge and Data Engineer-
ing 20(5), 692702 (2008)
8. Cheng, G., Qu, Y.: Term Dependence on the Semantic Web. In: Sheth, A.P., Staab,
S., Dean, M., Paolucci, M., Maynard, D., Finin, T., Thirunarayan, K. (eds.) ISWC
2008. LNCS, vol. 5318, pp. 665680. Springer, Heidelberg (2008)
9. Oliveira, M., Gama, J.: An overview of social network analysis. Wiley Interdisci-
plinary Reviews: Data Mining and Knowledge Discovery 2(2), 99115 (2012)
10. Weisstein, E.: Normal distribution (2005) From MathWorld,A Wolfram Web Re-
source, http://mathworld.wolfram.com/NormalDistribution.html
11. Newman, M.: The mathematics of networks. The New Palgrave Encyclopedia of
Economics 2 (2008)
12. Newman, M.E.J.: Coauthorship networks and patterns of scientific collaboration.
Proceedings of the National Academy of Sciences of the United States of Amer-
ica 101(suppl. 1), 52005205 (2004)
13. Okamoto, K., Chen, W., Li, X.-Y.: Ranking of closeness centrality for large-scale
social networks. In: Preparata, F.P., Wu, X., Yin, J. (eds.) FAW 2008. LNCS,
vol. 5059, pp. 186195. Springer, Heidelberg (2008)
14. David, J., Euzenat, J.: Comparison between Ontology Distances (Preliminary Re-
sults). In: Sheth, A.P., Staab, S., Dean, M., Paolucci, M., Maynard, D., Finin,
T., Thirunarayan, K. (eds.) ISWC 2008. LNCS, vol. 5318, pp. 245260. Springer,
Heidelberg (2008)
15. Broder, A., Kumar, R., Maghoul, F., Raghavan, P., Rajagopalan, S., Stata, R.,
Tomkins, A., Wiener, J.: Graph structure in the web. Computer Networks 33(1),
309320 (2000)
 Energy Eciency in W-Grid Data-Centric
Sensor Networks via Workload Balancing

Alfredo Cuzzocrea1 , Gianluca Moro2 , and Claudio Sartori3

1
ICAR-CNR and University of Calabria, Italy
[email protected]
2
DISI Department Cesena Branch, University of Bologna, Italy
3
DISI Department Bologna Branch, University of Bologna, Italy
{gianluca.moro,claudio.sartori}@unibo.it

Abstract. Wireless sensor networks are usually composed by small

units able to sense and transmit to a sink elementary data which are
then processed by an external machine. However, recent improvements
in the memory and computational power of sensors, together with the re-
duction of energy consumption, are rapidly changing the potential of such
systems, moving the attention towards data-centric sensor networks. In
this kind of networks, nodes are smart enough either to store some data
and to perform basic processing allowing the network itself to supply
higher level information closer to the network user expectations. In other
words, sensors no longer transmit each elementary data sensed, rather
they cooperate in order to assemble them in more complex and syn-
thetic information, which will be locally stored and transmitted accord-
ing to queries and/or events defined by users and external applications.
Recently, we proposed W-Grid, a fully decentralized cross-layer infras-
tructure for self-organizing data-centric sensor networks where wireless
communication occur through multi-hop routing among devices. In this
paper, we show that network trac, and thus the energy consumption,
can be balanced among sensors by assigning multiple virtual coordinates
to nodes trough a fully decentralized workload balancing algorithm, which
extends W-Grid.

1 Introduction
Wireless sensor networks consist of large number of distributed nodes that orga-
nize themselves into a multi-hop (wireless) network. Sensor nodes are generally
powered by batteries with an inherently limited lifetime, hence energy consump-
tion issues play a leading role. In more detail, we focus the attention on so-called
data-centric sensor networks, where nodes are smart enough either to store some
data and to perform basic processing allowing the network itself to supply higher
level information closer to the network user expectations.
Energy saving in wireless sensor networks involves both MAC layer and net-
work layer. In particular, here we state that the routing protocol must be
lightweight, must not require too much computation, such as complex evalu-
ations of possible paths, and must not need a wide knowledge of the network

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 3142, 2013.
c Springer-Verlag Berlin Heidelberg 2013

32 A. Cuzzocrea, G. Moro, and C. Sartori

organization. In previous work, we showed that W-Grid [16,15] routing scheme

satisfies both previously described requirements. In fact, W-Grid routing pro-
tocol needs information about only one-hop away devices and the choice of the
next hop requires bit-a-bit comparison of simple binary strings. We also proved
that W-Grid allows sensors to generate a decentralized wireless network in which
devices, thanks to uni-cast multi-hop transmissions (i.e., no broadcast propaga-
tions), can communicate each other independently of their location in the net-
work. Our solution enhances typical sensor networks, in such a way that queries
can be posed to the network anytime through any sensor by overcoming limita-
tions of a dedicated sink node, which collects raw data in a fixed position and
under time constraints (e.g., sensor synchronization). Thus, the resulting sensor
network is capable of executing, in principle, any distributed algorithm for pro-
cessing raw data they generate, in such a way to supply external users with the
high level information needed or something almost close to their expectations.
Also, the resulting network, which is W-Grid-aware, is able to eciently index
and query multi-dimensional data without reliance either on Global Positioning
System (GPS) or flooding/broadcasting operations [16,15].
As mentioned, energy consumption plays a central role in wireless sensor net-
works. Indeed, in ad-hoc and sensor networks, most of the energy consumption is
due to radio transmissions and hence protocol design for these networks must be
lightweight and directed towards reducing communication cost. Strictly adhering
to this paradigm, in this paper, we show that network trac, and thus the energy
consumption, can be balanced among sensors by assigning multiple virtual coor-
dinates to nodes trough a fully decentralized workload balancing algorithm, which
extends W-Grid. The energy consumption balancing, besides improving the en-
tire network eciency by lowering collisions and resource contentions among
nodes, also prolongs the network life avoiding premature partitions caused by
non-uniform battery discharging.
As a non-secondary contribution, we evaluate network routing performances,
which have direct influence on energy consumption, as the quantity of network
knowledge at nodes (namely routing table size) vary. An extensive number of
experiments have been conducted in order to proof the improvement in energy
consumption allowed by workload balancing in W-Grid.
The remaining part of the paper is organized as follows. Section 2 describes
related work. Section 3 briefly presents W-Grid. Section 4 contains the results
of our comprehensive experimental campaign. Finally, Section 5 concludes the
paper with some considerations and introducing future work.

2 Related Work
Existing routing protocols have been developed according to dierent approaches.
Basically, routing is necessary whenever a data is sensed (we also say gener-
ated) and stored in the system or whenever a query is submitted to the system.
 Energy Eciency in W-Grid Data-Centric Sensor Networks 33

As stated before, we do not consider sensor network systems which store sensor
data externally at a remote base station, but rather we focus on advances wire-
less sensor networks in which data or events are kept at sensors, by means of
representing them in terms of relations in a virtual distributed database and, for
eciency purposes, indexed by suitable attributes. For instance in [12,10,21,20],
data generated at a node is assumed to be stored at the same node, and queries
are either flooded throughout the network. In [19], a Geographic Hash Tables
(GHT) approach is proposed, where data are hashed by name to a location
within the network, enabling highly ecient rendezvous. GHTs are built upon
the GPSR protocol [11] thus leveraging some interesting properties of that proto-
col, such as the ability of routing to sensors nearest to a given location, together
with some of well-known limits, such as the risk of dead ends. Dead end prob-
lems, especially under low density environments or scenarios with obstacles and
holes, are caused by the inherent greedy nature of routing algorithms that can
lead to situation in which a packet gets stuck at a local optimal sensors that
appears closer to the destination than any of its known neighbors. In order to
solve this flaw, correction methods such as perimeter routing, which tries to ex-
ploit the right hand rule, have been implemented. However, packet losses still
remain and, in addition to this, using perimeter routing causes loss of eciency
both in terms of average path length and energy consumption. Besides, another
limitation of GHT-based routing is that it needs sensors to know their physical
position, which causes additional localization costs to the whole system. In [9],
Greenstein et al. have designed a spatially-distributed index, called DIFS, to fa-
cilitate range search over attributes. In [12], Li et al. have built a distributed
index, called DIM, for multidimensional range queries over attributes but they
require nodes to be aware of their physical location and of network perimeter.
Moreover, they exploit GPSR for routing. Our solution extending W-Grid also
behaves like a distributed index, but its indexing feature is cross-layered with
routing, meaning that no physical position nor any external routing protocol is
necessary, routing information is given by index itself.
Another research area that is directly related to our work is represented by
the problem of eectively and eciently managing multidimensional data over
sensor networks, as W-Grid may represent a very ecient indexing layer for tech-
niques and algorithms supporting this critical task. The problem of performing
multidimensional and OLAP analysis of data streams, like the ones originated by
sensor networks, has received great attention recently (e.g., [2]). Due to compu-
tational overheads introduced by these time-consuming tasks, several solutions
have been proposed in literature, such as data compression (e.g., [4]), usage of
high-performance Computational Grids (e.g., [6,5]), extensions to uncertain and
imprecise domains (e.g., [3]) that occur very frequently in sensor environments,
and so forth. This problem also exposes very interesting and challenging corre-
lations with cross-disciplinary scientific areas, such as mobile computing (e.g.,
[7]), thus opening the door to novel research perspectives still poorly explored,
such as data management issues in mobile sensor networks (e.g., [8,14,13]).
34 A. Cuzzocrea, G. Moro, and C. Sartori

3 W-Grid Overview

W-Grid1 can be viewed as a binary tree index cross-layering both routing and
data management features over a wireless sensor network. Two main phases are
performed in W-Grid: (1) implicitly generate coordinates and relations among
nodes that allow ecient message routing; (2) determine a data indexing space
partition by means of so-generated coordinates in order to eciently support
multidimensional data management. Also, each node can have one or more vir-
tual coordinates on which an order relation is defined and through which the
routing occurs. At the same time, each virtual coordinate represents a portion
of the data indexing space for which a device is assigned with the (data) manage-
ment tasks. By assigning multiple coordinates at nodes, we aim at reducing query
path length and latency. This is obtained via bounding the probability that two
nodes physically close have very dierent virtual coordinates, which may happen
whenever a multidimensional space is translated into a one-dimensional space. In
the next Sections, we provide a formal description on the W-Grid main features.

3.1 W-Grid Static Network Properties

The target sensor network is represented as a graph S:

S = (D, L) (1)

such that D is the set of participating devices and L is the set of physical
connectivity between couples of devices, which is defined as follows:

L = {(di , dj ) : two way connection between di and dj } (2)

Each device is assigned one or more (virtual) coordinate(s). We define C as the

set of existing coordinates. Each coordinate ci is represented as a string of bits
starting with . According to the regular expression formalism, coordinates are
defined as follows:
C = {c : c = (0 | 1) } (3)
For instance, 01001 is a valid W-Grid coordinate. Given a coordinate i and a bit
b, their concatenation will be indicated as ci b. For instance, let: ci = 0100, b = 0;
then: ci b = 01000. Given a bit b, we denote its complementary as follows: b.
For instance, 1 = 0.
Some functions are defined on top of C. We provide their description.
Given a coordinate c, length(c) returns the number of bits in c. ( excluded).
For instance, length(01001) = 5. length(c) is formally defined as follows:

length(c) : C (4)
1
From now on, we will refer at devices with the terms sensors or nodes indis-
tinctly.
 Energy Eciency in W-Grid Data-Centric Sensor Networks 35

Given a coordinate c and a positive integer k length(c), bit(c, k) returns the

k-th bit of c. Position 0 is out of the domain since it is assigned to . bit(c, k) is
formally defined as follows:

bit(c, k) : (C, {0}) {0, 1} (5)

Given a coordinate c and a positive integer k length(c), pref (c, k) returns

the first k bits of c. For instance, pref (01001, 3) = 010. pref (c, k) is formally
defined as follows:
pref (c, k) : ( , ) C (6)
Given a coordinate c, we define the complementary of c, denoted by c, as follows:

c = pref (c, length(c) 1)bit(c, length(c)) (7)

For instance, 01001) = 01000.

Given a coordinate c, we define the father of c, denoted by f ather(c), as
follows:
f ather(c) : (C {}) C (8)

f ather(c) = pref (c, length(c) 1) (9)

Given a coordinate c, we define the left child of c, denoted by lChild(c), as
follows:
lChild(c), rChild(c) : (C) C (10)

lChild(c) = c0 (11)
Given a coordinate c, we define the right child of c, denoted by rChild(c), as
follows:
rChild(c) = c1 (12)
To give examples, given a coordinate ci = 011, then: f ather(011) = 01,
lChild(011) = 0110, rChild(011) = 0111.
Given a coordinates in the domain C and devices in the domain D, we in-
troduce the mapping function M that maps each coordinate c to the device d
holding it, as follows:
M :CD (13)
A W-Grid network W is represented in terms of a graph, as follows:

W = (C, P ) (14)

such that P is the set of parent-ships between pair of coordinates, defined as

follows:
P = {(ci , cj ) : cj = ci (0 | 1)} (15)
For instance, pi = (010, 0101).
36 A. Cuzzocrea, G. Moro, and C. Sartori

Given a parent-ship p = (ci , cj ), we define the complementary of parent-ship

p = (ci , cj ), denoted by p, as follows:

p = (ci , cj ) (16)

For instance, if p = (010, 0101), then: p = (010, 0100).

Finally, given a graph W , W is a valid W-Grid network if all the following
properties are satisfied: (i) p = (ci , cj ) P, (M (ci ) = M (cj )) ((M (ci ), M (cj ))
L); (ii) p = (ci , cj ) P : M (ci ) = M (cj ) p = (ci , cj ) P : M (ci ) =
M (cj ).

3.2 W-Grid Dynamic Rules

W-Grid network is generated according to this few simple rules, which we de-
scribe next.
1. The first node that joins the networks (hence it initiates a coordinate space)
gets the coordinate . We say that a node that holds a W-Grid coordinate is
active. Given a device d, the function last returns the last coordinate received
by d, and it is defined as follows:

last(d) : (D) C (17)

If d is not active, last returns {}. Let n1 denote the first node that joined the
network, then last(n1 ) = .
2. l = (di , dj ) L : last(di ) = {}, two parent-ships are generated: (i) p =
(last(di ), c ): M (c ) = dj ; (ii) p, where c = lChild(last(di )) | rChild(last(di )).
Namely, c corresponds to the non-deterministic choice of one of the children of c.
According to a pre-defined coordinate selection strategy, nodes progressively get
new coordinates from each of its physical neighbors, in order to establish parent-
ships with them. Coordinate getting is also called split, and it is actually re-
lated to data management tasks of W-Grid. Actors of the split procedure are the
so-called joining node and giving node, respectively. We say that a coordinate
ci is split by concatenating a bit to it, and, after, one of the new coordinates is
assigned to the joining node, while the other one is kept by the giving node. Ob-
viously, an already split coordinate ci cannot be split anymore since this would
generate duplicates. Besides, in order to guarantee uniqueness of coordinates even
in case of simultaneous requests, each joining node must be acknowledged by the
giving node. Thus, if two nodes ask for the same coordinate to split, only one
request will succeed, while the other one will be temporarily rejected and post-
poned. Coordinate discovering is gradually performed by implicitly overhearing
of neighbor sensor transmissions.

3.3 W-Grid Routing Algorithm

The routing of any message/query is based on the concept of distance among
coordinates. Given two coordinates ci and cj , the distance between ci and cj ,
 Energy Eciency in W-Grid Data-Centric Sensor Networks 37

denoted by d(ci , cj), is measured in terms of logical hops and corresponds to

the sum of the number of bits of ci and cj which are not part of their common
prefix. To give and For instance: d(*0011, *011) = 5.
Given a target binary string ct , each device of the network is able of constantly
getting closer to it by choosing the neighbor that exposes the shortest distance
to ct . It is important to notice that each device needs neither global nor partial
knowledge about network topology to route messages, as its routing table is lim-
ited to information about its direct neighbors coordinates. This means ensuring
scalability with respect to network size, which is a nice amenity supported by
W-Grid.

3.4 W-Grid Data Management

W-Grid distributes data (i.e., tuples of attributes) gathered by sensors among

them in a data-centric manner. Values of tuples are linearized into binary strings
(see [18]) and stored at nodes whose W-Grid coordinates have the longest com-
mon prefix with so-generated strings. Thus, a W-Grid network acts directly as a
distributed database and coordinates c are used as items stored in a data reposi-
tory. This means that each coordinate represents a portion (region) of the global
data space. In order to balance data load at each region, a maximum number
of data items is pre-fixed, namely the bucket size b. Whenever a coordinate c
turns overloaded with respect to b, c splits into two coordinates c and c . Split
operation on coordinates is slightly dierent from the split operation on nodes
described in Section 3.2. The operations supporting the split of c into c and c
are described next: (i) both newly generated coordinates c and c are (temporar-
ily) kept at the node M (c); (ii) data held by c are distributed between c and
c ; (iii) parent-ships p = (c, c ) and p = (c, c ) are generated; (iv ) coordinate c
is labeled as givable, meaning that, in the case of a node asking M (c) for a
coordinate, c is immediately returned without performing any other split.
If the first split has not solved the overload in one between c or c , the split op-
eration is performed again by repeating the previously described steps. In order
to further improve the data distribution balancing, the Storage Load Balancing
algorithm (SLOB) [16] is executed.
In addition to the data management solutions described above, in W-Grid
sensed data are replicated in each of the existing spaces, thus allowing us to
improve data availability as to cope with sensor failures and periodically turn-
o a partition of them without compromising the network functionalities (the
latter for energy eciency purposes).

4 Experimental Assessment and Analysis

An extensive number of experiments have been conducted in order to assess the

eectiveness and the eciency of the W-Grid load balancing solutions presented
in this paper. Again, we stress the concept that the proposed experimental anal-
ysis is focused to intelligent sensor networks in which data collected by sensors
38 A. Cuzzocrea, G. Moro, and C. Sartori

are not simply downloaded at a fixed sink or in which a sink node is in charge of
periodically performing queries but, contrary to this, networks where each active
sensor can be queried at any time.
Eectiveness and eciency of W-Grid load balancing solutions have been
evaluated in terms of the following experimental parameters: query average path
length, data and trac load at sensors, and energy consumption of the resulting
sensor network. The functional parameters chosen are, instead, the following:
device density, number of coordinates per node. In the following, we first describe
the simulation model of our experimental campaign (Section 4.1). Then, we
provide experimental results for the following experimental aspect assessing the
W-Grids performance: query eciency (Section 4.2), network load (Section 4.3)
and energy consumption (Section 4.4).

4.1 Simulation Model

We performed our experiments on top of a Java-based in-laboratory simulation
environment. We adopted the following simulation model. Simulation area size
is 800 800m, where nodes are spread in a uniformly-random manner. Each
sensor has its own identifier (ID), a radio range of 100m, and it adhere to an
ideal transmission mode. As a result, the average number of one-hop neighbors
varies from 11 to 4. Due to space limitation, here we show the results for the
range 11 8. For each scenario, we ran five simulations and, for each simulation,
we submitted 10, 000 queries to the system.

AVERAGE NUMBER OF HOPS AVERAGE NUMBER OF HOPS

Parameters: Avg DENSITY: 11 neighbours per Parameters: Avg DENSITY: 8 neighbours per
node - Balancing OFF node - Balancing OFF
10 10
9 9
8 8
7 7
H 6 H 6
o o
5 5
p p
s 4
s 4
3
3
2
2
1
1
0
1 2 4 6 8 10 OPN 0
Number of Virtual Coordinates 1 2 3 4 5 6 8 OPN
Num ber of Virtual Coordinates

GPSR GPSR Stddev W-Grid W_Grid Stddev GPSR GPSR Stddev W-Grid W_Grid Stddev

Fig. 1. Query eciency due to W-Grid without load balancing solutions, in comparison
to GPSR, for the configurations having 11 (left) and 8 (right) neighbors per node

We let the simulator perform the following tasks: (i) uniformly random place-
ment of n sensors in a user-defined area; (ii) gradual generation of W-Grid
coordinates at sensors that exploit implicit overhearing; (iii) random generation
of data sensing and query in a user-defined ratio q : i.
At each simulator run, we observed the following experimental parameters:
(i) average path length of queries (compared with the one due to GPSR [11]);
(ii) loads at sensors, modeled in terms of the relative dierence between network
trac and managed data space at sensors; (iii) network energy consumption in
each dierent scenario.
 Energy Eciency in W-Grid Data-Centric Sensor Networks 39

4.2 Query Eciency

In the first experiment, we tested the ratio of succeeded queries due to W-Grid
with respect to the ones accomplished by GPSR. Indeed, for the sake of clarity,
here we highlight that the latter comparison is unfair as GPSR assumes to know
sensor physical positions over the network. This means that the probability of
committing a query is always very high, especially in quite dense application
scenarios like the one we tested. However, our main experimental goal was to
test our solution against the most ecient one available in literature, even if
the comparison one was advantaged. Figure 1 and Figure 2 show the average
number of hops needed to succeed queries with respect to the number of virtual
coordinates for W-Grid and GPSR in the two distinctive scenarios of not apply-
ing (Figure 1) and applying (Figure 2) the load balancing solutions. As Figure 1
and Figure 2 show, it clearly follows that W-Grid behavior is absolutely ecient
and not too much distant from the one by GPSR, which reasonably represents
the favorite case.

AVERAGE NUMBER OF HOPS AVERAGE NUMBER OF HOPS

Parameters: Avg DENSITY: 11 neighbours per Parameters: Avg DENSITY: 8 neighbours per
node - Balancing ON node - Balancing ON
10 10
9 9
8 8
7 7
H 6 H 6
o o
5 5
p p
s 4 s 4
3 3
2 2
1 1
0 0
1 2 4 6 8 10 OPN 1 2 3 4 5 6 8 OPN

Number of Virtual Coordinates Number of Virtual Coordinates

GPSR GPSR Stddev W-Grid W_Grid Stddev GPSR GPSR Stddev W-Grid W_Grid Stddev

Fig. 2. Query eciency due to W-Grid with load balancing solutions, in comparison
to GPSR, for the configurations having 11 (left) and 8 (right) neighbors per node

4.3 Network Load

Network load trends of W-Grid has also been observed as well. Network workload
variations are the result of the interaction between the following two factors: (i)
data load balancing among nodes and (ii) increase of transmission costs due to
path lengthening. As the number of coordinates grows, these two factors expose
better trends, as confirmed by Figure 3 and Figure 4, where we depict the average
and standard deviation evolutions of network trac distribution among nodes
with respect to the number of coordinates for W-Grid and GPSR in the two
distinctive scenarios of not applying (Figure 3) and applying (Figure 4) the load
balancing solutions. Similarly, for cross-comparison purposes, we studied the
variation of the standard deviation of data space at sensors (Figure 5) (clearly,
with load balancing solutions). Retrieved results clearly confirm the benefits
deriving from W-Grid.
40 A. Cuzzocrea, G. Moro, and C. Sartori

AVERAGE AND STANDARD DEVIATION OF AVERAGE AND STANDARD DEVIATION OF NET

NET TRAFFIC DISTRIBUTION AMONG NODES TRAFFIC DISTRIBUTION AMONG NODES
Parameters: Avg DENSITY: 11 neighbours per node - Parameters: Avg DENSITY: 8 neighbours per node -
Balancing OFF Balancing OFF
200
200
180
180
160
160
140
Number of data

140

Number of data
120
120
100
100
80
80
60
60
40
40
20
20
0
0
1 2 4 6 8 10 OPN 1 2 3 4 5 6 8 OPN
Number of coordinates Number of coordinates
Average Standard deviation Average Standard Deviation

Fig. 3. Network load due to W-Grid without load balancing solutions for the configu-
rations having 11 (left) and 8 (right) neighbors per node

AVERAGE AND STANDARD DEVIATION OF AVERAGE AND STANDARD DEVIATION OF NET

NET TRAFFIC DISTRIBUTION AMONG NODES TRAFFIC DISTRIBUTION AMONG NODES
Parameters: Avg DENSITY: 11 neighbours per node - Parameters: Avg DENSITY: 8 neighbours per node -
Balancing ON Balancing ON
200 200
180 180
160 160
140 140
Number of data

Number of data
120 120
100 100
80 80
60 60
40 40
20 20
0 0
1 2 4 6 8 10 1 2 3 4 5 6 8 OPN
Number of coordinates Number of coordinates
Average Standard Deviation Average Standard Deviation

Fig. 4. Network load due to W-Grid with load balancing solutions for the configurations
having 11 (left) and 8 (right) neighbors per node

STANDARD DEVIATION OF DATA SPACE STANDARD DEVIATION OF DATA SPACE

DISTRIBUTION AMONG NODES DISTRIBUTION AMONG NODES
DENSITY = 11 neighbours per node DENSITY = 8 neighbours per node
0,03 0,03

0,025 0,025

0,02 0,02

0,015 0,015

0,01 0,01

0,005 0,005

0 0
1 2 4 6 8 10 OPN 1 2 3 4 5 6 8 OPN
Number of Coordinates Num ber of Coordinates
Avg StdDev Load Balancing OFF StdDev Load Balancing ON Avg StdDev Load Balancing OFF StdDev Load Balancing ON

Fig. 5. Variation of data space at sensors due to W-Grid with load balancing solutions
for the configurations having 11 (left) and 8 (right) neighbors per node

4.4 Energy Consumption

Finally, we focused the attention on energy consumption due to W-Grid for
each application scenario considered in the experimental campaign (i.e., 11, 8
and 4 neighbors per node). Energy consumption was measured according to [17].
Figure 6 shows the retrieved observations for the two dierent cases of applying
(left) and not applying (right) load balancing solutions. Even in this study, W-
Grid observations expose a very positive trend.
 Energy Eciency in W-Grid Data-Centric Sensor Networks 41

NETWORK ENERGY CONSUMPTION NETWORK ENERGY CONSUMPTION

Parameters: Load Balancing ON Parameters: Load Balancing OFF
800 800

750 750

700 700
Joules

Joules
650 650
600 600
550 550
500 500
0,10 0,20 0,40 0,60 0,80 1,00 OPN 0,10 0,20 0,40 0,60 0,80 1,00 OPN
Coordinates per neighbor Coordinates per neighbor

11 neighbors 8 neighbors 4 neighbors 11 neighbors 8 neighbors 4 neighbors

Fig. 6. Energy consumption due to W-Grid for dierent configurations with (left) and
without (right) load balancing solutions

5 Conclusions and Future Work

In this paper, we proposed load balancing extensions of W-Grid and assessed

them by means of a comprehensive experimental campaign that has focused the
attention on several interesting experimental parameters, ranging from query
eciency to energy consumption. Experiments have been conducted on top of
a Java-based simulation environment. Retrieved results have clearly confirmed
the benefits due to the usage of W-Grid as cross-layer infrastructure for self-
organizing data-centric sensor networks.
Future work is mainly oriented towards three directions. First, we are con-
sidering the presence of outliers in the data layer of such networks. How their
presence impacts on the performance of W-Grid ? is a challenging question that
demands for solution. Second, we are considering the integration of W-Grids
architecture with modern Cloud computing platforms. Finally we are address-
ing in W-Grid the issue of network security with malicious nodes by extending
decentralized approach based on distributed supervised data mining [1].

References

1. Cerroni, W., Moro, G., Pirini, T., Ramilli, M.: Peer-to-peer data mining classifiers
for decentralized detection of network attacks. In: Wang, H., Zhang, R. (eds.) ADC
2013, Adelaide, South Australia. CRPIT, pp. 18. ACS (2013)
2. Cuzzocrea, A.: CAMS: OLAPing multidimensional data streams eciently. In:
Pedersen, T.B., Mohania, M.K., Tjoa, A.M. (eds.) DaWaK 2009. LNCS, vol. 5691,
pp. 4862. Springer, Heidelberg (2009)
3. Cuzzocrea, A.: Retrieving accurate estimates to OLAP queries over uncertain and
imprecise multidimensional data streams. In: Bayard Cushing, J., French, J., Bow-
ers, S. (eds.) SSDBM 2011. LNCS, vol. 6809, pp. 575576. Springer, Heidelberg
(2011)
4. Cuzzocrea, A., Chakravarthy, S.: Event-based lossy compression for eective and
ecient OLAP over data streams. Data Knowl. Eng. 69(7), 678708 (2010)
5. Cuzzocrea, A., Furfaro, F., Greco, S., Masciari, E., Mazzeo, G.M., Sacc` a, D.: A
distributed system for answering range queries on sensor network data. In: PerCom
Workshops, pp. 369373 (2005)
42 A. Cuzzocrea, G. Moro, and C. Sartori

6. Cuzzocrea, A., Furfaro, F., Mazzeo, G.M., Sacc a, D.: A grid framework for ap-
proximate aggregate query answering on summarized sensor network readings. In:
Meersman, R., Tari, Z., Corsaro, A. (eds.) OTM Workshops 2004. LNCS, vol. 3292,
pp. 144153. Springer, Heidelberg (2004)
7. Cuzzocrea, A., Furfaro, F., Sacc` a, D.: Enabling OLAP in mobile environments
via intelligent data cube compression techniques. J. Intell. Inf. Syst. 33(2), 95143
(2009)
8. El-Moukaddem, F., Torng, E., Xing, G.: Mobile relay configuration in data-
intensive wireless sensor networks. IEEE Trans. Mob. Comput. 12(2), 261273
(2013)
9. Greenstein, B., Estrin, D., Govindan, R., Ratnasamy, S., Shenker, S.: Difs: A
distributed index for features in sensor networks. In: Proceedings of First IEEE
WSNA, pp. 163173. IEEE Computer Society (2003)
10. Intanagonwiwat, C., Govindan, R., Estrin, D., Heidemann, J., Silva, F.: Directed
diusion for wireless sensor networking. IEEE/ACM Trans. Netw. 11(1), 216
(2003)
11. Karp, B., Kung, H.: GPRS: greedy perimeter stateless routing for wireless net-
works. In: MobiCom 2000, pp. 243254. ACM Press (2000)
12. Li, X., Kim, Y., Govindan, R., Hong, W.: Multi-dimensional range queries in sensor
networks. In: SenSys 2003, pp. 6375. ACM Press, New York (2003)
13. Li, Z., Liu, Y., Li, M., Wang, J., Cao, Z.: Exploiting ubiquitous data collection for
mobile users in wireless sensor networks. IEEE Trans. Parallel Distrib. Syst. 24(2),
312326 (2013)
14. Liu, B., Dousse, O., Nain, P., Towsley, D.: Dynamic coverage of mobile sensor
networks. IEEE Trans. Parallel Distrib. Syst. 24(2), 301311 (2013)
15. Monti, G., Moro, G., Sartori, C.: WR -Grid: A scalable cross-layer infrastructure
for routing, multi-dimensional data management and replication in wireless sensor
networks. In: Min, G., Di Martino, B., Yang, L.T., Guo, M., R unger, G. (eds.)
ISPA 2006 Ws. LNCS, vol. 4331, pp. 377386. Springer, Heidelberg (2006)
16. Moro, G., Monti, G.: W-Grid: a self-organizing infrastructure for multi-dimensional
querying and routing in wireless ad-hoc networks. In: IEEE P2P 2006 (2006)
17. Moro, G., Monti, G.: W-grid: A scalable and ecient self-organizing infrastructure
for multi-dimensional data management, querying and routing in wireless data-
centric sensor networks. Journal of Network and Computer Applications 35(4),
12181234 (2012), http://dx.doi.org/10.1016/j.jnca.2011.05.002
18. Ouksel, A.M., Moro, G.: G-Grid: A class of scalable and self-organizing data struc-
tures for multi-dimensional querying and content routing in P2P networks. In:
Moro, G., Sartori, C., Singh, M.P. (eds.) AP2PC 2003. LNCS (LNAI), vol. 2872,
pp. 123137. Springer, Heidelberg (2004)
19. Ratnasamy, S., Karp, B., Shenker, S., Estrin, D., Govindan, R., Yin, L., Yu, F.:
Data-centric storage in sensornets with ght, a geographic hash table. Mob. Netw.
Appl. 8(4), 427442 (2003)
20. Xiao, L., Ouksel, A.: Tolerance of localization imprecision in eciently managing
mobile sensor databases. In: ACM MobiDE 2005, pp. 2532. ACM Press, New York
(2005)
21. Ye, F., Luo, H., Cheng, J., Lu, S., Zhang, L.: A two-tier data dissemination model
for large-scale wireless sensor networks. In: MobiCom 2002, pp. 148159. ACM
Press, New York (2002)
 Update Semantics for Interoperability
among XML, RDF and RDB
A Case Study of Semantic Presence
in CISCOs Unified Presence Systems

Muhammad Intizar Ali1 , Nuno Lopes1 , Owen Friel2 , and Alessandra Mileo1
1
DERI, National University of Ireland, Galway, Ireland
{ali.intizar,nuno.lopes,alessandra.mileo}@deri.org
2
Cisco Systems, Galway, Ireland
[email protected]

Abstract. XSPARQL is a transformation and querying language that

provides an integrated access over heterogeneous data sources on the fly.
It is an extension of XQuery which supports a subset of SPARQL and
SQL to provide unified access over XML, RDF and RDB formats. In
practical applications, data integration does not only require the inte-
grated access over distributed heterogeneous data sources, but also the
update of underlying data.
XSPARQL in its present state is only a querying and transformation
language, hence lacking the update facility. In this paper, we propose an
extension of the XSPARQL language with update facility. We present
the syntax and semantics for this extension, and we use the real world
scenario of semantic presence in CISCOs Unified Presence Systems to
demonstrate the requirement of update facility. Preliminary evaluation
of the XSPARQL Update Facility is also presented.

Keywords: XSPARQL, SPARQL , XQuery, XMPP, Updates, DML.

1 Introduction

Relational databases (RDB) are the most common way of storing and managing
data in majority of the enterprise environments [1]. Particularly in the enter-
prises where integrity and confidentiality of the data are of the utmost impor-
tance, relational data model is still the most preferred option. However, with the
growing popularity of the Web and other Web related technologies (e.g. Semantic
Web), various data models have been introduced. Extensible Markup Language
(XML) is a very popular data model to store and specify the semi-structured
data over the Web [7]. It is also considered as a de-facto data model for data
exchange. XPath and XQuery are designed to access and query data in XML

This work has been funded by Science Foundation Ireland, Grant No.
SFI/08/CE/l1380 (Lion2) and CISCO Systems Galway, Ireland.

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 4350, 2013.
c Springer-Verlag Berlin Heidelberg 2013

44 M.I. Ali et al.

format[10,5]. Semantic Web is built upon data represented in Resource Descrip-

tion Format (RDF) [4]. RDF is a standard data format for establishing semantic
interpretability and machine readability among the various data sources scat-
tered over the Web. SPARQL is a query language designed to query linked data
in RDF data format [17]. All of the above mentioned data models were designed
to perform a specific task and their importance in their own domains can not
be denied. Several query languages including SQL, XQuery and SPARQL have
been designed to access and query RDB, XML and RDF data respectively. It is
inevitable for modern data integration applications to avoid such heterogeneity
of data formats and query languages. Therefore, often these applications require
integrated access over distributed heterogeneous data sources.
XSPARQL is a W3C Member Submission (http://www.w3.org/
Submission/2009/01/) which combines XQuery, SPARQL and SQL to
provide an integrated access over XML, RDF and RDB data sources [2].
However, in practice the integrated applications not only require the access of
integrated data, but in many scenarios an update in the existing data is also
desired. XSPARQL in its present state only provides Data Query Language
(DQL) while lacking Data Manipulation Language (DML). In this paper we
define syntax and semantics of XSPARQL Update Facility which extends the
capabilities of the XSPARQL language to perform update operations over RDB,
XML and RDF data on the fly while keeping the data sources in their original
data formats. XSPARQL Update Facility is fully compatible with the update
semantics defined individually for SQL, XQuery and SPARQL [16,18,12]. Our
main motivation for this extension was the real world scenario of semantic
presence in the CISCO Unified Presence (CUP) systems, based on processing
and integrating information retrieved from XMPP(http://xmpp.org) messages
and update the underlying data in RDB and RDF data sources to keep track
of up-to-date semantic presence history of registered users. In this scenario we
require not only the simultaneous access to heterogeneous data represented in
RDB, XML and RDF, but also the update operations are desired to update the
underlying RDB and RDF data on the fly [8].
Our main contributions in this paper can be summarised as follows: (i) we
extend XSPARQL with updates, providing what we call XSPARQL Update Fa-
cility, (ii) we define formal semantics of the XSPRQL Update Facility, (iii) we
demonstrate the need for XSPARQL Update Facility by presenting a use case
scenario for semantic presence in the CUP systems, and (iv) we successfully im-
plement and evaluate XSPARQL Update Facility for dealing with group chats
history in the same scenario.

2 Semantic Presence in Ciscos Unified Presence System

Cisco Unified Presence (CUP) systems is a standards-based enterprise platform

that aims to provide an eective way of communication among people in and
across the organisation using instant messaging (IM) over XMPP standard pro-
tocol. A general architecture to collect several presence information from various
 Updates in XSPARQL 45

heterogeneous and dynamic

sources has been introduced
in [14],which provides some
insights on how richer seman-
tic presence services and ap-
plications can be deployed. As
a first step towards the en-
hancement of CUP systems
with richer semantic presence,
a mapping of the core XMPP
messages into RDF is for-
Fig. 1. Semantic Presence in CUP Systems
mally defined in [8]. Figure 1
depicts a simplified high-level
version of the architecture of semantic presence in the CUP systems focus-
ing only on heterogeneous data formats that need to be manipulated during
communication. All XMPP messages are mapped into RDF using various on-
tologies e.g SIOCC (http://rdfs.org/sioc/chat#), Nepomuk (http://www.
semanticdesktop.org/ontologies/2007/03/22/nmo#) and SIOC (http://
rdfs.org/sioc/ns#). CUP server stores history of all the group chat mes-
sages and other communications in an external PostgreSQL (http://www.
postgresql.org/) database, while semantic enhancement component stores
the semantic information in a graph database using Virtuoso Server (http://
openlinksw.com/virtuoso). Listing 1 shows a sample XMPP message gen-
erated by IM client while requesting for the creation of a chat room. List-
ing 2 depicts mapping of XMPP messages into RDF using SIOC and SIOCC
ontologies. We use the XSPARQL Update Facility to interpret the XMPP
message and generate/update RDB and RDF data on the fly. This enables
a semantically enriched CUP system to update/generate data from one data
model to any other data model. Later, in section 4 we evaluate our XSPARQL
Update facility by using the use case of semantic presence in CUP systems.

@base <h t t p : / / l i o n 2 p c i s c o . com

#i n t i z a r @ d e r i . o r g s e e k s t o e n t e r / example#>.
#i n t h e DERICollab room : DERICollab a
s i o c t : ChatChannel ;
<p r e s e n c e s i o c : h a s o wn e r
from= : i n t i z a r @ c o r p . com .
i n t i z a r @ d e r i . o r g / i n t i z a r n o t e b o o k : D E R I Co l l a b Ch a t S e ssi o n a
t o= s i o c c : ChatSession ;
D E RI Co l l a b @ c o n f e r e nc e . c o r p . com/ sioc : has container
i n t i z a r > : DERICollab ;
<x xmlns= h t t p : / / j a b b e r . o r g / p r o t o c o l / siocc : has publishing source
muc/> : intizarNotebook
</ p r e s e n c e > s i o c c : has nickname ali

Listing 1. XMPP request message for chat Listing 2. Mapping XMPP Message of
room creation Listing 1 into RDF [8]
46 M.I. Ali et al.

Prolog: declare namespace prefix=namespace-URI or

prefix prefix: <namespace-URI>
Body: for var in FLWOR expression
let var:= FLWOR expression
where FLWOR expression
order by FLWOR expression or
for varlist
from/ from named ( <dataset-URI> or FLWOR expr)
where pattern
order by expression
limit integer > 0
oset integer > 0 or
for SelectSpec
from RelationList
where WhereSpecList
Update: insert node expression ((as(first|last))?into)|after|before) expression|
delete node expression |
replace (value of) node expression with expression or
insert data Quads |
( ( delete data Quads ) | ( delete where QuadPattern ) ) |
[with <IRI>] ( ( delete QuadPattern [ insert QuadPattern] ) | (insert Quad- or
Pattern ) )(using [named] <IRI>)* where GroupGraphPattern
insert into RelationList values ValueList |
delete from RelationList where WhereSpecList |
update RelationList set ValueList where WhereSpecList
Head: construct template (with nested FLWOR expressions) or
return XML + nested FLWORs expression

Fig. 2. Schematic View of XSPARQL with Updates

3 XSPARQL Update Facility

In this section we specify the XSPARQL Update Facility by giving a detailed
description of its syntax and semantics.

3.1 Syntax
XSPARQL is built on XQuery semantics, XSPARQL Update Facility is also ex-
tended using formal semantics of XQuery Update Facility. It merges the subset
of SPARQL, XQuery and SQL update clauses. We limit ourself to three major
update operations of INSERT: to insert new records in the data set, DELETE:
to delete already existing records from data set and UPDATE: to replace the ex-
isting records or their values with the new records or values. Figure 2 presents a
schematic view of the XSPARQL Update Facility which allows its users to select
data from one data source of any of its data format and update the results into
another data source of dierent format within a single XSPARQL query while
preserving the bindings of the variables defined within query. Contrary to the
select queries, the update queries have no return type. On successful execution
of the update queries no results are returned, only a response is generated which
can be either successful upon successful execution of the query or an appropriate
error is raised if for some reasons the XSPARQL query processor is unable to
update the records successfully. However if a valid updating or delete query with
or without where clause does not match any relevant tuple in the data source,
the XSPARQL query processor will still response with successful execution, but
 Updates in XSPARQL 47

there will be no eect on the existing data source. The basic building block of
XQuery is the expression, an XQuery expression takes zero or more XDM in-
stances and returns an XDM instance. Following new expressions are introduced
in XSPARQL Update Facility
A new category of expression called updating expression is introduced to
make persistent changes in the existing data.
A basic updating expression can be any of the insert, delete or update.
for, let, where or order by clause can not contain an updating expression.
return clause can contain an updating expression which will be evaluated for
each tuple generated by its for clause.

RevalidationDecl::= "declare" "revalidation" ("strict" | "lax" | "skip" )

XSPARQLExpr ::= (FLWORExpr | SPARQLForClause | SQLForClause) |
(InsertClause | DeleteClause | UpdateClause)
(ReturnClause | ConstructClause)
InsertClause ::= XQueryInsert | SPARQLInsert | SQLInsert
DeleteClause ::= XQueryDelete | SPARQLDelete | SQLDelete
UpdateClause ::= XQueryUpdate | SPARQLUpdate | SQLUpdate
...

Fig. 3. XSPARQL Update Facility Grammar Overview

Figure 3 presents an overview of the basic syntax rules for XSPARQL Update
Facility, for complete grammar of XSPARQL Update Facility, we refer our reader
to the latest version of the XSPARQL available at http://sourceforge.net/
projects/xsparql.

3.2 Semantics
Similar to the semantics of the XSPARQL language [6], for the semantics of
updates in XSPARQL we rely on the semantics of the original languages: SQL,
XQuery, and SPARQL. Notably, the semantics for SPARQL updates are pre-
sented only in the upcoming W3C recommendation: SPARQL 1.1 [12]. We start
by presenting a brief overview of the semantics of update languages for the dif-
ferent data models: relational databases, XML, and RDF.
RDB: updates in the SQL language rely on a procedural language that specify
the operations to be performed, respectively: insrdb (r, t), delrdb (r, C), and
modrdb (r, C, C ), where r is a relation name, t is a relational tuple, and C, C
are a set of conditions of the form A = c or A = c (A is an attribute
name and c is a constant). Following [1], insrdb (r, t) inserts the relational
tuple t into the relation r, delrdb (r, C) deletes the relational tuples from
relation r that match the condition, and modrdb (r, C, C ) updates the tuples
in relation r that match conditions C to the values specified by C .
XML: For XML data, the XQuery language is adapted such that an expression
can return a sequence of nodes (as per the XQuery semantics specifica-
tion [9]) or pending update list [18], which consists of a collection of update
48 M.I. Ali et al.

primitives, representing state changes to XML nodes. For this paper we

are focusing on insert, delete, and replace operations, whose semantics are
procedurally defined in [18] and in this paper we represent these procedural
semantics by the functions insxml (SourceExpr , InTargetChoice, TargetExpr ),
delxml (TargetExpr ), and modxml (TargetExpr , ExprSingle).
RDF: SPARQL Updates semantics are similarly defined in terms of changes
to the underlying triple store [12], where the result of an update opera-
tion is a new triple store. We rely on the semantics functions insrdf , delrdf ,
and modrdf (shorthand for the functions OpInsertData, OpDeleteData,
and OpDeleteInsert, respectively). The functions signatures are as per the
SPARQL specification: insrdf (GS , QuadPattern), delrdf (GS , QuadPattern),
and modrdf (GS , DS , QuadPatterndel , QuadPatternins , P ) presented in [12].
The functions insrdf and delrdf insert and delete triples match-
ing QuadPattern from the triple store DS, respectively. The modrdf eval-
uates the graph pattern P against the dataset DS and apply these bindings
to QuadPatterndel in order to determine the triples to be removed from the
graph store GS and to QuadPatternins for the inserted triples.

4 Implementation and Evaluation

Implementation: We use java platform for the implementation of the XS-

PARQL Update Facility. XSPARQL grammar rules are defined using ANTLR
(http://www.antlr.org). XSPARQL uses Saxon (http://www.saxonica.com)
as a query processor for XQuery, Jena ARQ (http://jena.apache.org) for
SPARQL queries and for SQL queries we provide option to connect to multi-
ple relational database management systems including PostgreSQL and MySQL
(http://www.mysql.com).
Evaluation of XSPARQL Update Facility: In order to demonstrate and
evaluate performance and practicality of XSPARQL Update Facility, we consider
a generic use case of group chat of the semantic presence in CUP systems as
described in Section 2. However XSPARQL Update Facility can be used by any
application which requires access and updates in distributed heterogeneous data
stored in any of the RDB, XML or RDF data models.

f o r $x i n / / [ name= p r e s e n c e ]
i f $x / s t a t u s [ @code =201] t h e n
i n s e r t data
{Graph <S e m a n t i c P r e s e n c e>
{
: DERICollab a s i o c t : ChatChannel ;
s i o c : h a s o wn e r : i n t i z a r a l i @ c o r p . com .
: D E R I Co l l a b Ch a t S e ssi o n a s i o c c : ChatSession ;
sioc : has container : DERICollab ;
siocc : has publishing source : intizaraliNotebook
s i o c c : has nickname ali
}}

Listing 3. A Sample XSPARQL Update Query

 Updates in XSPARQL 49

A User Creates a Chat Group: Consider a scenario where a user creates a

persistent chat room over a specific topic as shown in Listing 1 in the Section 2.
When an IM client sends a request to create a chat room an XMPP message will
be generated by IM client and sent to the CUP Server. CUP Server will process
the message. The semantic presence component of CUP systems wants to update
the semantic data stored in Virtuoso Server. Listing 3 shows an XSPARQL query
that interprets XMPP messages using XQuery and if certain condition is met,
the XSPARQL update query will insert the relevant information into RDF data
store using SPARQL update query.

5 Related Work
Several eorts are made to integrate distributed XML, RDF and RDB data
on the fly. These approaches can be divided into two types, (i) Transformation
Based: Data stored in various format is transformed into one format and can be
queried using a single query language [11]. The W3C GRDDL working group
addresses the issues of extracting RDF from XML documents. (ii) Query Re-
writing Based: Query languages are used to transform or query data from one
format to another format. SPARQL queries are embedded into XQuery/XSLT to
pose against pure XML data [13]. XSPARQL was initially designed to integrate
data from XML and RDF and later extended to RDB as well. DeXIN is another
approach to integrate XML, RDF and RDB data on the fly with more focus
on distributed computing of heterogeneous data sources scattered over the Web
[3]. Realising the importance of the updates, the W3C has recommendations for
XQuery and SPARQL updates [18,12]. In [15], SPARQL is used to update the
relational data. However, to the best of our knowledge their is no work available
which can provide simultaneous access over the distributed heterogeneous data
source with updates.

6 Conclusion and Future Work

In this paper we have extended XSPARQL to provide the Update Facility which
further strengthens the capabilities of XSPARQL by enabling simultaneous ac-
cess and update of heterogeneous data sources. We have defined syntax and
semantics of XSPARQL Update Facility, which merges update operations of
SQL, XQuery and SPARQL. Using XSPARQL Update Facility user not only
can query, integrate and transform heterogeneous data on the fly, but can also
update the already stored data in the data sets or in-memory data generated
as a result of a query before an update operation. XSPARQL Update Facility
will lead to the wide adoption of XSPARQL in many applications (e.g. simi-
lar to semantic presence in CUP systems) which require integrated access over
distributed heterogeneous data sources with updates.
In the future, we plan to further investigate query optimisation techniques that
would make it possible to evaluate even complex XSPARQL queries in tractable
time, that would make XSPARQL usable for large-scale applications. In this
50 M.I. Ali et al.

paper, we elaborated the practicality of XSPARQL Update Facility using real

world scenario, however in future, we plan to define detailed formal semantics of
the language, evaluation of static and dynamic rules and experimental evaluation
considering various parameters including scalability and query executaon time.
We also plan to extend XSPARQL Update Facility to include Data Definition
Language (DDL) for its subsequent languages.

References
1. Abiteboul, S., Hull, R., Vianu, V.: Foundations of Databases. Addison-Wesley
(1995)
2. Akhtar, W., Kopeck y, J., Krennwallner, T., Polleres, A.: XSPARQL: Traveling
between the XML and RDF Worlds and Avoiding the XSLT Pilgrimage. In:
Bechhofer, S., Hauswirth, M., Homann, J., Koubarakis, M. (eds.) ESWC 2008.
LNCS, vol. 5021, pp. 432447. Springer, Heidelberg (2008)
3. Ali, M.I., Pichler, R., Truong, H.-L., Dustdar, S.: DeXIN: An Extensible Framework
for Distributed XQuery over Heterogeneous Data Sources. In: Filipe, J., Cordeiro,
J. (eds.) ICEIS 2009. LNBIP, vol. 24, pp. 172183. Springer, Heidelberg (2009)
4. Beckett, D., McBride, B.: RDF/XML Syntax Specification. W3C Proposed Rec-
ommendation (February 2004) (revised)
5. Berglund, A., Boag, S., Chamberlin, D., Fernandez, M.F., Kay, M., Robie, J., Simeon,
J.: XML Path Language (XPath) 2.0. W3C Recommendation (December 2010)
6. Bischof, S., Decker, S., Krennwallner, T., Lopes, N., Polleres, A.: Mapping between
RDF and XML with XSPARQL. Journal on Data Semantics 1, 147185 (2012)
7. Bray, T., Paoli, J., Maler, E., Yergeau, F., Sperberg-McQueen, C.M.: Extensible
Markup Language (XML) 1.0 (5th edn.). W3C Recommendation (November 2008)
8. Dabrowski, M., Scerri, S., Rivera, I., Leggieri, M.: Dx Initial Mappings for the
Semantic Presence Based Ontology Definition (November 2012),
http://www.deri.ie/publications/technical-reports/
9. Draper, D., Fankhauser, P., Fern andez, M., Malhotra, A., Rose, K., Rys, M.,
Simeon, J., Wadler, P.: XQuery 1.0 and XPath 2.0 Formal Semantics. W3C Rec-
ommendation (January 2007)
10. Fernandez, M.F., Florescu, D., Boag, S., Robie, J., Chamberlin, D., Simeon, J.:
XQuery1.0: An XML query language. W3C Proposed Recommendation (April 2009)
11. Gandon, F.: GRDDL Use Cases: Scenarios of extracting RDF data from XML
documents. W3C Proposed Recommendation (April 2007)
12. Gearon, P., Passant, A., Polleres, A.: SPARQL 1.1 Update. W3C Working Draft
(January 2012)
13. Groppe, S., Groppe, J., Linnemann, V., Kukulenz, D., Hoeller, N., Reinke, C.:
Embedding SPARQL into XQuery/XSLT. In: Proc. of SAC (2008)
14. Hauswirth, M., Euzenat, J., Friel, O., Grin, K., Hession, P., Jennings, B., Groza,
T., Handschuh, S., Zarko, I.P., Polleres, A., Zimmermann, A.: Towards Consoli-
dated Presence. In: Proc. of CollaborateCom 2010, pp. 110 (2010)
15. Hert, M., Reif, G., Gall, H.: Updating relational data via SPARQL/update. In:
Proc. of EDBT/ICDT Workshops (2010)
16. Negri, M., Pelagatti, G., Sbattella, L.: Formal Semantics of SQL Queries. ACM
Trans. Database Syst. 16(3), 513534 (1991)
17. Prudhommeaux, E., Seaborne, A.: SPARQL Query Language for RDF. W3C Pro-
posed Recommendation (January 2008)
18. Robie, J., Chamberlin, D., Dyck, M., Florescu, D., Melton, J., Simeon, J.: XQuery
Update Facility 1.0. W3C Recommendation (March 2011)
 GPU-Accelerated Bidirected De Bruijn Graph
Construction for Genome Assembly

Mian Lu1 , Qiong Luo2 , Bingqiang Wang3 , Junkai Wu2 , and Jiuxin Zhao2
1
A*STAR Institute of High Performance Computing, Singapore
[email protected]
2
Hong Kong University of Science and Technology
{luo,jwuac,zhaojx}@cse.ust.hk
3
BGI-Shenzhen, China
[email protected]

Abstract. De Bruijn graph construction is a basic component in de novo

genome assembly for short reads generated from the second-generation
sequencing machines. As this component processes a large amount of
data and performs intensive computation, we propose to use the GPU
(Graphics Processing Unit) for acceleration. Specifically, we propose a
staged algorithm to utilize the GPU for computation over large data sets
that do not fit into the GPU memory. We also pipeline the I/O, GPU,
and CPU processing to further improve the overall performance. Our
preliminary results show that our GPU-accelerated graph construction
on an NVIDIA S1070 server achieves a speedup of around two times over
previous performance results on a 1024-node IBM Blue Gene/L.

1 Introduction

Genome assembly refers to the process of reconstructing a genome sequence from

a large number of sequence fragments (known as reads). These reads are gener-
ated by sequencing machines through randomly sampling the original sequence.
The De Bruijn graph based genome assembly algorithms have been shown ef-
fective for assembling a large number of short reads and have been adopted in
state-of-the-art assemblers [14]. In this paper, we focus on the bidirected De
Bruijn graph construction, which is the first as well as one of the most expensive
steps in genome assembly. We propose to utilize the GPU (Graphics Processing
Units) to accelerate this process, in particular, the construction of a bidirected
De Bruijn graph from a large set of short reads.
The bidirected De Bruijn graph construction is expensive in both memory
consumption and running time. A previous study [2] has shown that the graph
construction for human genome took around 8 hours on a 16-core machine with
2.3GHz AMD quadcore CPUs and consumed 140GB main memory. With lim-
ited GPU memory (up to 6 GB per GPU in the market), the first challenge
in GPU-based De Bruijin graph construction is to develop algorithms that can
handle data larger than the GPU memory. Second, given the superb computa-
tion power of the GPU, the overall performance is likely to be dominated by disk

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 5162, 2013.
c Springer-Verlag Berlin Heidelberg 2013

52 M. Lu et al.

I/O and CPU processing. Therefore, we study how to utilize the hierarchy of
disk, main memory, and GPU memory to pipeline processing and to involve the
CPU for co-processing. These issues are essential for the feasibility and overall
performance on practical applications; unfortunately, they are seldom studied in
current GPGPU research.
Specifically, we address the GPU memory limit by developing a staged al-
gorithm for GPU-based graph construction. We divide the reads into chunks
and load the data chunk by chunk from disk through the main memory to the
GPU memory. To estimate the chunk size that can fit into the GPU memory,
we develop a memory cost model for each processing step. We further utilize the
CPU and main memory to perform result merging each time the GPU finishes
processing a chunk. Finally, we pipeline the disk I/O, GPU processing, and CPU
merging (DGC) to improve the overall performance. We expect this pipelined-
DGC processing model to be useful for a wide range of GPGPU applications
that handle large data.
We have implemented the GPU-based bidirected De Bruijn graph construction
and evaluated it on an NVIDIA Tesla S1070 GPU device with 4 GB memory. Our
initial results show that this implementation doubles the performance reported
on a 1024-node IBM Blue Gene/L and is orders of magnitude faster than state-
of-the-art CPU-based sequential implementations.
The remainder of this paper is organized as follows. In Section 2, we briefly
introduce the graph construction algorithm and related work. We present our de-
sign of the staged algorithm in Section 3. We describe the details of the pipelined-
DGC model in Section 4. The experimental results are reported in Section 5. We
conclude in Section 6.

2 Preliminary
2.1 Genome Assembly
The second generation of sequencing produces very short reads at a high through-
put. Popular algorithms to reconstruct the original sequence for such a large
number of short reads are based on the De Bruijn [5] or bidirected De Bruijn
graph [6]. The graph is constructed through generating k-mers from reads as
graph nodes. For example, suppose a short read is ACCTGC and k = 4, then
this read can generate three 4-mers, which are ACCT, CCTG, and CTGC. The
major dierence between the two graph models is that for each k-mer, its re-
verse complement is represented by a separate node (De Bruijn) or the same
node (bidirected De Bruijn).
At the beginning of assembly, each l-length read generates (l k + 1) k-
mers. Then the De Bruijn graph is constructed using information about overlap
between k-mers. Next, the graph is simplified and corrected by some heuris-
tic algorithms. After the simplification and correction, several long contigs are
generated. Finally, if reads are generated through paired-end sequencing, these
contigs are joined to produce scaolds using relevant information. Another al-
ternative of joining contigs is through Eulerian paths. The detailed algorithms
 GPU-Accelerated Bidirected De Bruijn Graph Construction 53

of these steps can be found in previous work [2, 4]. In this work, we focus on the
De Bruijn graph construction.

2.2 Bidirected De Bruijn Graph

The double-stranded structure of DNA sequences can be naturally mapped to
the bidirected De Bruijn graph [6]. For a birdirected graph, each edge has inde-
pendent directions at two ends. A valid path from node v1 to vk is represented
as a sequence v1 e1 v2 e2 ...ek1 vk , where ei is the edge connecting two node vi and
vi+1 , and ei1 and ei have dierent directions at the node vi for 1 < i < k.
Figure 1 shows an example of bidirected graph. A De Bruijn graph is a directed
graph, where each node represents an ordered sequence, and all sequences have
the same length. An edge from node A representing a1 a2 ...an to node B repre-
senting b1 b2 ...bn exists when a2 a3 ...an is identical to b1 b2 ...bn1 .

K K
' ( )
K

Fig. 1. An example of bidirected graph

For genome assembly, each node in a bidirected De Bruijn graph represents

a k-mer containing k bases. Given the double-stranded structure of DNA se-
quences, each node also implicitly represents its reverse complementary k-mer,
i.e., the base on each position in one sequence is the complement (A to T, C to
G) of another sequence in the reverse order. For a given node v representing two
reverse complementary k-mers, we denote the lexicographically greater one as v +
to be the canonical form of these two k-mers as well as the node representative,
and v the reverse complement of v + . Note that, when k is an odd number, v +
and v will never be identical.
Suppose an edge exists between nodes A and B. There are four cases when A
overlaps B.
Case 1. A+ overlaps B + A B
Case 2. A overlaps B A B
Case 3. A overlaps B + A B
Case 4. A+ overlaps B A B
Note that duplicate edges are expected due to the high coverage of input reads.
However, only distinct edges and their frequency counts (known as multiplici-
ties) need to be maintained. The multiplicity is useful when removing errors from
graphs. After constructing all distinct edges and recording their multiplicities,
we generate an ordered pair <cX , cY > corresponding to an edge representing X
overlaps Y . cX and cY are the label to output when walking from node X to
Y and from node Y to X, respectively. The label is generated as follows corre-
sponding to the four cases of overlap. X + [1] and X [1] denote the first character
of the canonical form and its reverse complement of node X, respectively.
54 M. Lu et al.

Case 1. A+ overlaps B + <A+ [1], B [1]>

Case 2. A overlaps B <A [1], B + [1]>
Case 3. A overlaps B + <A [1], B [1]>
Case 4. A+ overlaps B <A+ [1], B + [1]>

Distinct canonical k-mers are collected as representatives for graph nodes, and
adjacency information is built for the graph representation for further manipu-
lation.

2.3 Related Work

There are several genome assembly tools for short reads on the CPU. Velvet
[4] is a pioneer software for short reads assembly, which is still one of the most
popular short reads assemblers today. SOAPdenovo [2] has been used successfully
for human genome projects. There are other similar assemblers, such as Euler-SR
[7], Shorty [8], Edena [9], ALLPATHS [10], SSAKE [11], SHARCGS [12], ABySS
[3], and YAGA [1]. These assemblers are dierent in implementation details, but
are all based on the (bidirected) De Bruijn graph model.
Among parallel short reads assemblers, ABySS [3] was implemented using
MPI. SOAPdenovo [2] has parallelized several key time-consuming steps. Jack-
son et al. [13] have reported their bidirected De Bruijn graph construction on
an IBM Blue Gene/L. Later, the authors implemented the complete genome as-
sembly procedure on the Blue Gene/L machine and reported the performance
scalability [1]. Kundeti et al. [14] have further improved the parallel graph con-
struction algorithm presented by Jackson et al. [13] to avoid a large amount
of message passing. However, to the best of our knowledge, the only work on
GPU-accelerated genome assembly is GPU-Euler [15], which implements a simi-
lar construction algorithm. The major issue of their work is that it cannot handle
large genome data, e.g., human genome, which cannot fit into the GPU memory.

3 Design and Implementation

In this section, we first present the in-memory implementation of GPU-based
bidirected De Bruijn graph construction assuming sucient memory. Then, we
propose a staged algorithm for out-of-core processing.

3.1 In-Memory Implementation

We consider the graph construction as a component in de novo assembly, where
the program input is a plain text file storing short reads and the output is the
adjacency list representation of the bidirected De Bruijn graph stored in the
main memory. Overall, there are three steps for the graph construction.
 GPU-Accelerated Bidirected De Bruijn Graph Construction 55

Step 1. Encoding. This step loads short reads from the disk to the main
memory and uses two bits per character to represent the reads.

Step 2. Canonical Edge Generation. We adopt an edge-oriented building

method to avoid validating edges [1]. For an l -length short read, we generate
(k+1)-mers, and the number of (k+1)-mers is (l k). Each (k+1)-mer corre-
sponds to an edge connecting two overlapping k-mers. The two k-mers represent
two nodes connected by the edge. We do not extract the two k-mers in this step
since there are many redundant edges and nodes at this time. A 64-bit data
type, such as long, is sucient to hold a (k+1)-mer since (k + 1) is usually set to
less than 32 in practice. Similarly, we do not extract edge directions and output
labels in this step due to the redundancy.
In the GPU-based implementation, generating edges is done through a map
primitive eciently. Each thread takes charge of one encoded read and generates
corresponding (l k) edges. In our implementation, the encoding and edge gen-
eration are done in one kernel program. We use the fast shared memory on the
GPU to hold encoded reads, and bitwise operations to generate all (k+1)-mers.
Duplicate elimination can be implemented through either sorting, or hashing.
In our evaluation, the two methods have a similar performance. Considering
the merging step in out-of-core processing, we have adopted the sorting-based
approach to perform the duplicate elimination.

Step 3. Adjacency List Representation After edges are generated, we can

extract two corresponding k-mers, edge directions, and output labels from each
(k+1)-mer. This step can be implemented as a map on GPUs. To keep the
adjacency information, we use a pair (k-mer, edge id) to represent a node. An
additional duplicate elimination step on the field k-mer is performed for all pairs
to obtain distinct nodes as well as associated edges.

3.2 The Staged Graph Construction Algorithm

The in-memory implementation assumes unlimited GPU memory for the graph
construction. In practice, a staged graph construction algorithm is necessary
for the GPU-based implementation to handle data that cannot fit into the GPU
memory. In common cases, the most memory consumption occurs in Step 2, when
edges are generated from encoded reads before the duplicates are eliminated.
Suppose there are n short reads of length l each, for a user-defined k, the total
size of (k+1)-mers is 8 n (l k) bytes, e.g., several hundreds of gigabytes for
the human genome. Step 3 is executed using a similar staged algorithm when
all distinct edges are produced. Therefore, we focus on introducing the staged
algorithm for Step 1 and 2.
The basic idea in our staged algorithm is to load and process the input reads
through multiple passes. In each pass, only a subset of input reads, denoted
as a chunk of reads, is loaded from the disk to the memory for processing.
Figure 2 shows the workflow of chunk-based processing. The chunk size is set so
56 M. Lu et al.

read chunk i from disk

start C-2: GPU processing
C-3: CPU merge
C-1: I/O to adjacency list
... ...
chunk i n(l-K) edges m distinct edges m distinct edges

n reads

processed distinct edges

Fig. 2. The workflow of one chunk based processing of GPU-accelerated bidirected De

Bruijn graph construction. The color coding of black, grey, and white indicates the
data is stored in the disk, GPU memory, and CPU main memory, respectively

that the memory required for the chunk-based processing does not exceed the
GPU memory size. Specifically, there are three processing steps for one chunk:
1. C-1: I/O. n (n n) short reads are loaded from the disk to the main
memory as a chunk of data.
2. C-2: GPU processing. We transfer the data chunk to the GPU memory.
The GPU performs the encoding and generates n (l k) edges. Duplicates
are eliminated through a sorting-based method for these n (l k) edges.
After the duplicate elimination, there are m distinct edges for this chunk.
3. C-3: CPU merge. These m distinct edges are copied from the GPU mem-
ory to the main memory, and merged with the distinct edges generated from
previously processed data chunks. Since both the newly generated edges and
the existing edges are ordered, the merge step is ecient. The m distinct
edges after merging will be used for the next chunk. The multiplicity infor-
mation are also updated in this step. Due to the high coverage of input reads,
the number of distinct edges is around tens of times smaller than that of all
generated edges. Therefore we assume that the main memory is sucient to
hold these distinct edges.

3.3 The Memory Cost Model

Given the GPU memory size Mg we estimate the memory consumption for each
step to decide the suitable chunk size. Suppose each chunk contains n reads,
and each read is l-length. Then the memory size of input reads is n l bytes
(one byte per character). After encoding, the memory size for encoded reads is
n l
4 bytes. Thus the total GPU memory consumption for this encoding step is:

(n l + n 4l ) bytes. Next we use encoded reads in the memory for generating

edges. The number of edges generated for one chunk is n (l k), and each
edge is represented using a 64-bit word type. Thus the memory size required for
all generated (k+1)-mers in one chunk is 8 n (l k) bytes. Therefore the
total GPU memory consumption
$ % of this edge generation step for one chunk is:
n l
4 + 8 n (l k) bytes. We perform the sorting-based duplicate elimina-

tion algorithm on n (l k) edges. The memory consumption of the GPU-based
radix sort is an output buer with the same size of the input array. Thus the total
memory consumption for the radix sort can be estimated as (2 8 n (l k))
 GPU-Accelerated Bidirected De Bruijn Graph Construction 57

bytes. Finally, we remove duplicates from these sorted edges. The memory con-
sumption for this operation is at most (2 8 n (l k)) bytes. Therefore,
the maximum value of n can be estimated as follows:
& '
4Mg Mg Mg
min , ,
5l 8.25l 8k 16 (l k)

4 The Pipelined Processing Model

4.1 The DGC Model
Our staged algorithm can be generalized as a processing model, which can be
adopted in many GPGPU applications to handle the data set that cannot entirely
fit into the GPU memory or of which some steps are not suitable to be processed
on the GPU. We call it the DGC (Disk-GPU-CPU) model.
The DGC model is defined as follows. There are three components: I/O, the
computation on GPUs, and the processing on CPUs. Suppose the total size of
input data is D, the chunk size d, there are Dd passes. For each pass: (1) A
chunk of data is loaded from the disk to the main memory. (2) The in-memory
chunk is transferred from the main memory to the GPU memory, and processed
by the GPU. (3) The result is transferred back to the main memory, and a merge
(or other post-processing) step is performed on the CPU.
In the DGC model, at least three memory buers are required: bi is the
CPU memory buer to hold the data from the disk, bg is the GPU memory
buer used to store the input data, and bc is the CPU memory buer used to
hold the computation result transferred from the GPU memory. To simplify the
presentation, we assume the computation result on the GPU is also stored in
the input buer, that is bg .

4.2 The Pipelined-DGC Model

The DGC model can be improved using pipelining. Since bg is the GPU memory
buer independently accessed from bi and bc , the DGC model can be pipelined
without additional memory allocated for data exchange. Without loss of gener-
ality, we assume that during processing, bg and bc will not be released by the
GPU and CPU programs, respectively.
We maintain three threads to independently take charge of the I/O, GPU,
and CPU processing. The I/O thread will be blocked when bi is full and also
when the memory copy to bg is ongoing. The GPU thread may be blocked in
two cases. First, the data is not ready, i.e., bg is not full. Second, the processing
on the GPU has been done, but the data is being copied to bc . The CPU thread
may be blocked only when the data is not ready in bc .
At the beginning, we load the first data chunk to bi . When bi is full, a memory
copy from bi to bg is executed. The I/O is blocked when performing the memory
copy. As long as the data has been uploaded to bg , bi becomes available to
load the next chunk of data until it is full again. However, the second memory
58 M. Lu et al.

copy from bi to bg (and all following memory copies) should wait until the GPU
processing for the previous chunk is done to release the use of bg . Similarly,
we pipeline the GPU and CPU processing. A memory copy from the GPU to
CPU is performed when the computation on the GPU is finished. The GPU
processing is blocked when coping the data to bc . After the memory copy, the
CPU can perform the merge step for the data in bc , and the the next chunk of
data can be copied from bi to bg if the data is ready in bi , otherwise the GPU
thread is blocked to wait for the data in bi becoming ready. We can see that
with pipelining, the I/O, GPU processing, and CPU merge can overlap in time.

5 Experiments
5.1 Experimental Setup
Hardware Configuration. We conduct our experiments on a server machine
with two Intel Xeon E5520 CPUs (16 threads in total) and an NVIDIA Tesla
S1070 GPU. The NVIDIA Tesla S1070 has four GPU devices. In our current
implementation, we only utilize one GPU device, and the CPU merge step also
only uses one core on the CPU. This computation capability is sucient for the
overall performance as the I/O is the bottleneck. The main memory size of the
server is 16 GB.
Data Sets. We use a small and a medium sized data set for the evaluation. As
the previous research [13], we randomly sample known chromosomes to simulate
short reads. Two data sets are sampled from Arabidopsis chromosome 1 (denoted
as Arab.) and human chromosome 1 (denoted as Human), which can be accessed
from the NCBI genome databases [16]. The details of the two data sets are shown
in Table 1. Additionally, we fix the k-mer length k to 21, which is a common
value for most genome assemblers.

Table 1. Data sets

Arabidopsis chromosome 1 Human chromosome 1

#nucleotide 30 million 247 million
Read length 36 36
Coverage 17 36
#read 12 million 230 million
File size (FASTA) 800 MB 13 GB

5.2 Performance Results

Time Breakdown without Pipelining. We first study the performance bot-
tleneck of our GPU-based graph construction without the pipeline optimization.
Figure 3 shows the time breakdown of our GPU-accelerated bidirected De Bruijn
graph construction. In this figure, the I/O, the encoding and edge generation on
the GPU, the duplicate elimination for each chunk on the GPU, and the merge
 GPU-Accelerated Bidirected De Bruijn Graph Construction 59

Fig. 3. The time breakdown of GPU-based bidirected De Bruijn graph construction

without pipelining

step on the CPU together take around 90% of the execution time. The DGC
model is applied to these four components. Particularly, among these four com-
ponents, I/O takes around 50%, and both the GPU processing (including encod-
ing and generating edges) and CPU processing take around 25% of the elapsed
time. Since the eciency of the pipeline is limited by the most time-consuming
component, we expect that the overall performance of these four components
can be improved by around two times through pipelining.

I/O
GPU
CPU
0 25 50 75 100 125 150 175 200 225 250
Time (second)

(a) Without pipelining

I/O
GPU
CPU
0 10 20 30 40 50 60 70 80 90 100 110 120 130
Time (second)

(b) With pipelining

Fig. 4. Executions of the staged algorithm without and with pipelining for GPU-based
bidirected De Bruijn graph construction. The solid black rectangles represent the busy
time of a component.

The Pipelining Optimization. We focus on the four steps (I/O, Encod-

ing+Edge gen., Duplicate elimination, Merge) which as shown in Figure 3 that
consume around 90% of the overall elapsed time. We take the larger data set
human chromosome 1 to demonstrate the result. There are 33 passes in the
staged algorithm for this medium size data set. Figure 4(a) shows the execution
of the staged algorithm without pipelining. The I/O, GPU and CPU process-
ing are done sequentially in each pass. As a result, the disk is idle around half
of the overall time, and the utilization of the GPU and CPU is only around
25%. Figure 4(b) shows the execution status of the staged algorithm when
pipelining optimization is adopted. With pipelining, I/O is nearly non-blocking.
60 M. Lu et al.

Even though the GPU and CPU processing cannot completely overlap, their
utilization is improved considerably. As a result, the overall execution time for
these four steps are reduced from around 250 seconds to 128 seconds.
Comparison to Existing Implementations. There are four CPU-based im-
plementations for comparison: Velvet [4], SOAPdenovo [2], ParBidirected [14],
and Jacksons implementation [13]. We only compare the performance of graph
construction. We perform the evaluation on the same machine for all these soft-
ware except the Jacksons implementation. Note that SOAPdenovo, ParBidi-
rected, and Jacksons implementation are parallel implementations.

Table 2. Comparison of the elapsed time in seconds

Arab. Human 1
GPU-accelerated 18 177
Velvet [4] 86 -
SOAPdenovo [2] 78 1,245
ParBidirected [14] 1,740 32,400
Jackson2008 [13] 15 327

Table 2 shows the comparison result for the running time of four dierent im-
plementations. The memory required by preprocessing in Velvet for the Human
data set exceeds our main memory limit and takes excessively long time, thus
we do not report the performance number for the Human data set in the table.
SOAPdenovo has parallelized the hash table building. In the evaluation, it con-
sumes a similar main memory size (around 8 GB for the Human data set) to our
implementation. However, the GPU-accelerated graph construction is around
4-7x faster than SOAPdenovo for the similar functionality. Since ParBidirected
adopts a more conservative method to handle the out-of-core processing, inter-
mediate results need to be written into the disk in most steps, which results in
a slow execution time, but the memory consumption is stable and very low. In
our experiments, we have already modified the default buer size used in the
external sorting to improve the performance. The published performance results
from Jacksons implementation [13] is based on a 1024-node IBM Blue Gene/L.
The input and output are the same as our program. However, they have adopted
a node-oriented graph building approach, and the message passing is very ex-
pensive. In summary, compared with existing implementations, our GPU-based
graph construction is significantly faster. Specifically, compared with the mas-
sively parallel implementation on the IBM Blue Gene/L, our implementation is
still around two times faster.

6 Conclusion
In this paper, we have presented the design and implementation of our GPU-
accelerated bidirected De Bruijn graph construction, which is a first step to build
 GPU-Accelerated Bidirected De Bruijn Graph Construction 61

a complete GPU-accelerated genome assembler. We have addressed the GPU

memory limit issue through a staged algorithm, and the overall performance can
be further improved by around two times through pipelining I/O, GPU and CPU
processing. Furthermore, such optimized processing flow can be generalized as
a pipelined-DGC model, which can be applied to other GPGPU applications to
handle large data sets. Compared with existing implementations on CPUs, the
performance of our implementation is up to two orders of magnitude faster. In
particular, our GPU-accelerated graph construction is around 2X faster than the
published performance numbers for a parallel implementation on a 1024-node
IBM Blue Gene/L.
Acknowledgement. This work was supported by grants 617509 and 616012
from the Research Grants Council of Hong Kong, and grant MRA11EG01 from
Microsoft China.

References

1. Jackson, B., Regennitter, M., Yang, X., Schnable, P., Aluru, S.: Parallel de novo
assembly of large genomes from high-throughput short reads. In: IPDPS 2010:
Proceedings of the 2010 IEEE International Symposium on Parallel&Distributed
Processing, pp. 110 (April 2010)
2. Li, R., Zhu, H., Ruan, J., Qian, W., Fang, X., Shi, Z., Li, Y., Li, S., Shan, G.,
Kristiansen, K., Li, S., Yang, H., Wang, J., Wang, J.: De novo assembly of human
genomes with massively parallel short read sequencing. Genome Research 20(2),
265272 (2010)
3. Simpson, J.T., Wong, K., Jackman, S.D., Schein, J.E., Jones, S.J., Birol, I.: Abyss:
a parallel assembler for short read sequence data. Genome Research 19(6), 1117
1123 (2009)
4. Zerbino, D.R., Birney, E.: Velvet: algorithms for de novo short read assembly using
de bruijn graphs. Genome Research 18(5), 821829 (2008)
5. Pevzner, P.A., Tang, H.: Fragment assembly with double-barreled data. Bioinfor-
matics 17(suppl. 1), S225S233 (2001)
6. Medvedev, P., Georgiou, K., Myers, G., Brudno, M.: Computability of models for
sequence assembly. In: Giancarlo, R., Hannenhalli, S. (eds.) WABI 2007. LNCS
(LNBI), vol. 4645, pp. 289301. Springer, Heidelberg (2007)
7. Chaisson, M.J., Pevzner, P.A.: Short read fragment assembly of bacterial genomes.
Genome Research 18(2), 324330 (2008)
8. Hossain, M.S.S., Azimi, N., Skiena, S.: Crystallizing short-read assemblies around
seeds. BMC Bioinformatics 10(suppl. 1) (2009)
9. Hernandez, D., Francois, P., Farinelli, L., ster
as, M., Schrenzel, J.: De novo bac-
terial genome sequencing: Millions of very short reads assembled on a desktop
computer. Genome Research 18(5), 802809 (2008)
10. Butler, J., MacCallum, I., Kleber, M., Shlyakhter, I.A., Belmonte, M.K., Lander,
E.S., Nusbaum, C., Jae, D.B.: Allpaths: De novo assembly of whole-genome shot-
gun microreads. Genome Research 18(5), 810820 (2008)
11. Warren, R.L., Sutton, G.G., Jones, S.J., Holt, R.A.: Assembling millions of short
dna sequences using ssake. Bioinformatics 23(4), 500501 (2007)
62 M. Lu et al.

12. Dohm, J.C., Lottaz, C., Borodina, T., Himmelbauer, H.: Sharcgs, a fast and highly
accurate short-read assembly algorithm for de novo genomic sequencing. Genome
Research 17(11), 16971706 (2007)
13. Jackson, B.G., Aluru, S.: Parallel construction of bidirected string graphs for
genome assembly. In: International Conference on Parallel Processing, pp. 346
353 (2008)
14. Kundeti, V., Rajasekaran, S., Dinh, H.: Ecient parallel and out of core algorithms
for constructing large bi-directed de bruijn graphs. CoRR abs/1003.1940 (2010)
15. Mahmood, S.F., Rangwala, H.: Gpu-euler: Sequence assembly using gpgpu. In:
Proceedings of the 2011 IEEE International Conference on High Performance Com-
puting and Communications, HPCC 2011, pp. 153160. IEEE Computer Society
(2011)
16. National Center for Biotechnology Information, http://www.ncbi.nlm.nih.gov/
 K Hops Frequent Subgraphs Mining
for Large Attribute Graph

Haiwei Zhang, Simeng Jin, Xiangyu Hu, Ying Zhang,

Yanlong Wen , and Xiaojie Yuan

Department of Computer Science and Information Security, Nankai University.

94, Weijin Road, Tianjin, China
{zhanghaiwei,jinsimeng,huxiangyu,zhangying,wenyanlong,
yuanxiaojie}@dbis.nankai.edu.cn
http://dbis.nankai.edu.cn

Abstract. Attribute Graphs are widely used to describe complex data

in many applications such as bio-informatics and social network. With
the rapid growth of scale for graph data, traditional solutions for mining
frequent subgraphs cannot performed well in large attribute graph be-
cause of time-consuming candidates generation and isomorphism testing.
In this paper, we investigate the problem for k hops subgraph mining in
large attribute graph. The attribute graph is transformed into labeled
graphs by projection for each attribute. K hops frequent subgraph min-
ing algorithm FSGen consists of three procedures is performed. Firstly,
frequent vertices and edges will be extended to frequent subgraphs from
root vertices. Secondly, frequent edges joining frequent vertices will be
added into extended subgraphs. Thirdly, if necessary, isomorphism test-
ing will be used to summarize frequent subgraphs based on Graph Edit
Distance. Then, frequent labeled subgraphs will be merged into attribute
subgraphs by integration according to designated attributes. The com-
plexity of our mechanism is approximately O(2n ), which is more ecient
than existing algorithms. Real data sets are applied in experiments to
demonstrate the eciency and eectiveness of our technique.

Keywords: Attribute graph, k hops frequent subgraph, mining, projec-

tion, integration.

1 Introduction

With the rapidly increasing amount of data in Web and related applications,
more and more data can be represented by graphs. Most mechanism and algo-
rithms can only work on labeled graphs, e.g. a vertex has only one label, while
vertices in many real graph data have more than one label, and this kind of
graphs are named multi-labeled graphs [1], such as graphs describing social net-
work and proteins. Attribute graph is characterized as a kind of multi-labeled

Corresponding author.

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 6374, 2013.
c Springer-Verlag Berlin Heidelberg 2013

64 H. Zhang et al.

graph. There are a set of attributes associated with each vertex in attribute
graph, and each attribute has an exacted name.
Mining frequent subgraphs from graph databases is of great importance in
graph data mining. Frequent subgraphs are useful in data analysis and data
mining tasks, such as similarity search, graph classification, clustering and in-
dexing [2]. In similarity search and classification, frequent subgraphs are usually
used as feature leading to exact results and high scalability. In graph clustering,
frequent subgraphs can be used as a solution for traditional algorithms such as k-
means. Similarity searching is an important issue in many applications of graph
data. The task is considered NP-complete and is inecient for large graph.
Existing research on frequent subgraphs mining is conducted mainly on two
types of graph database. The first one is transaction graph databases [3] that
consist of a set of relatively smaller graphs. Transaction graph databases are
prevalently used in scientific domains such as chemistry, bioinformatics, etc. The
Second one is a graph with a large number of vertices such as social networks [4].
This paper focuses on the latter type of graph database with several attributes
in each vertex.
Mining frequent subgraphs is a iterative processing consists of two main steps
[5]: candidates generation and subgraphs counting. In the first step, candidates
can be mined extending. In the second step, subgraph isomorphism will be in-
vestigated for counting. Subgraph isomorphism is an NP-complete problem [6]
and it is time-consuming for large graphs.
Considering the increasing importance of frequent subgraphs mining, this
paper will present a mechanism for mining frequent subgraphs from a large
attribute graph. Attribute graph will be transformed into labeled graphs by
projection firstly. Then algorithm FSGen will be performed to mining k hops
frequent subgraphs by extending, joining and isomorphism testing. Finally, fre-
quent labeled subgraphs will be merged into attribute subgraphs by integration
according to designated attributes. There is no candidates generation in algo-
rithm FSGen, and isomorphism testing is optional. Thus, the time occupation
of our mechanism for mining frequent labeled subgraphs is better than existing
works.
The rest of this paper is organized as follows. Related works are summa-
rized in section 2 on subgraph mining algorithms for graph databases consist
of multiple small labeled graphs. Section 3 is dedicated to basic concepts and
general definitions of frequent subgraphs mining. We present the k hops frequent
subgraph mining algorithm FSGen consisting of extending, joining and isomor-
phism testing procedures in section 4. Section 5 shows experimental studies for
our technique and section 6 concludes our work.

2 Related Works
Frequent subgraph mining has been widely studied for graph databases consist
of multiple small labeled graphs. To the best of our knowledge, no mechanisms
 K Hops Frequent Subgraphs Mining for Large Attribute Graph 65

can be directly performed in frequent subgraphs mining from graph databases

consist of one large attribute graph.
One popular approach for frequent subgraphs mining is raised in Apriori algo-
rithm. AGM [7] defines induced subgraph G = (V , E ) which satisfies V V
and E E where E includes all edges connecting vertices in V . In the algo-
rithm, all frequent induced subgraphs were mined according to designed support
by using vertices for candidates. FSG [8] algorithm stores labels of all (k-1)-
subgraphs (subgraphs including k-1 edges), then the common structure between
two k-graphs can be found by computing intersection of the label sets of their
subgraphs. As a result, a large amount of common subgraphs are generated as
candidates, which reduces eciency of the algorithm.
The other popular approach for frequent subgraphs mining is based on extend-
ing. Extending-based approaches usually use depth first search to find frequent
subgraphs. Initially, frequent subgraphs are frequent vertices. Then, a new vertex
extending from each subgraphs by another edge is added to the subgraph in order
to generate a new subgraph. Isomorphism testing is used to determine whether
the new subgraph is frequent. All subgraphs can be mined by repeating these
steps. But this case creates many redundant candidates. So all existing methods
based on extending attempt to reduce the generation of redundant candidates.
gSpan [9] defines DFS (Depth First Search) Lexicographic Order and DFS code
of a graph. GraphGen [6] improves gSpan by mining frequent subtree instead of
frequent subgraph in order to reduce the complexity of time. When preforming a
depth first search in a graph, a DFS tree is constructed. One graph can have sev-
eral DFS trees and respective DFS codes. Because the minimum DFS code of a
graph is unique, a DFS tree can be denoted by this minimum DFS code. Isomor-
phism testing can be performed based on minimum DFS codes of each subgraph.
gSpan and GraphGen can only extend vertices in the right path of DFS tree.
Candidates can only generate from finite vertices or edges for reducing redun-
dant subgraphs. FFSM [10] designs canonical adjacent matrix (CAM) as formal
data model for graphs and guarantee all frequent subgraphs are enumerated
unambiguously. FFSM uses two operations: FFSM-Join and FFSM-Extension
to generate candidates eciently and completely avoids subgraph isomorphism
testing by maintaining an embedding set for each frequent subgraph. Indeed,
in FFSM, edges extending and isomorphism testing has been transformed into
operations on CAMs. And the complexity of FFSM concentrates on operations
of CAMs. CAMs will be larger together with the increasing scale of graphs and
mining frequent subgraphs in large graph is inecient.
We can summary from existing works that candidates generation and isomor-
phism testing is essential stages for mining subgraphs. Since the complexity of
isomorphism testing cannot be actually reduced (including substitution of iso-
morphism testing stage in FFSM), all researches focus on reducing creation of
candidates. We will present a new algorithm FSGen for k hops frequent sub-
graphs mining from a large labeled graph. Under the restriction of k, candidates
generation and isomorphism testing can be avoided by the mechanism.
66 H. Zhang et al.

3 Preliminaries

For simplicity of presentation, we restrict our discussion to undirected, multi-

attributes graphs. And graph database consists of only one large attribute graph.
However, our approach can be applied to directed graphs and labeled graphs with
minor changes.
A graph G is defined as a binary (V, E), where V is the set of vertices, E is
the set of edges. A labeled graph GL is defined as a triple (V, E, L), where V
and E is the same as G, L is a labeling function that assigns a label to each
vertex and edge. L(v) = l denotes that the label of vertex v is l. An attribute
graph GA is extending the function L to multiple attributes, as definition 1
interprets.

Definition 1. Attribute Graph. An Attribute Graph GA is a triple (V, E, FA )

where V is a set of (m and E V V is a set of edges. A denotes the set of
(nvertices
attributes. FA = j=1 i=1 fai (vj ) (Assuming there are n vertices in V and m
attributes in each vertex.) is a set of functions , where fai (vj ) (1im, 1jn)
denotes to the value of attribute ai of vertex vj (ai A, vj V ).
Fig. 1 shows examples of graph, labeled graph and attribute graph. In the
attribute graph, each vertex has three attributes: id, name and major.

Y Y D Y Y D $QQ$, Y Y 'DQ'%

Y Y E Y Y F %LOO%LR Y Y 5RVH0HG

D*UDSK* 9(! E/DEHOHGJUDSK F$WWULEXWHJUDSK*$ 9()$!

*/ 9(/!/Y DHJ )DY )DY $QQHJ

Fig. 1. Graph, Labeled Graph and Attribute Graph

Definition 2. K Hops Neighbor. vV , N (v, k) denotes to a set of vertices

that are at most k hops from v. Each vertex uN (v, k) is a k hops neighbor of
vertex v. For example, in Fig. 1(c) N (v1 , 1) = {v3 , v4 }.
In this paper, attribute graph will be transformed into labeled graphs by
projection according to attributes. Thus, following definitions in this section are
designed for labeled graphs.

Definition 3. Support. The support of a subgraph GS in labeled graph GL

is denoted by Sup(GS ) and is given as the total quantity of its occurrence in
GL . Sup(v) denotes the support of vertex v by counting the occurrences of L(v).
Sup(e) denotes the support of edge e by counting the occurrences of e whose
labels of begin and end vertices are the same respectively.
 K Hops Frequent Subgraphs Mining for Large Attribute Graph 67

Definition 4. Frequent. Given a minimum support MinSup and a labeled

graph GL , a vertex vV is a frequent vertex i. Sup(v)M inSup. An edge
eE is a frequent edge i. Sup(e)M inSup. A subgraph gGL is frequent i.
Sup(g)M inSup.

Definition 5. K Hops Frequent Subgraph. A k hops frequent subgraph

Gf = (Vf , Ef , Lf ) of labeled graph GL satisfies: (1)Vf V , Ef E, (2)quantity
of Gf in G is not less than a given M inSup, (3)minimum hops (distance) between
all vertices and a designated root vertex are not more than a given constant k,
where the root vertex is one of frequent vertices in GL . For example, in Fig. 1(b),
assuming k=1 and M inSup=2, 1 hop frequent subgraphs in the labeled graph
are edges (a, b) and (a, c).

4 K Hops Frequent Subgraph Mining

In this section, we discuss algorithms for mining k hops frequent subgraphs in
a large attribute graph. Attribute graph is transformed into labeled graphs by
projection firstly. Then, algorithm FSGen for k hops frequent subgraph mining
in a large labeled graph will be proposed. Finally, integration will be described
in order to merge frequent labeled subgraphs into frequent attribute subgraphs
by designated attributes.

4.1 Projection
Attribute graph will be transformed into labeled graphs by projection before
frequent subgraphs mining. Projection can simplify the complexity of attribute
graph. And mechanisms for processing labeled graphs can be used to work on
attribute graph after projection.

Definition 6. Projection. Given attribute graph GA =(V, E, FA ), a labeled

graph Gp = (Vp , Ep , Lp ) is a projection of GA , i. Vp V , Ep E, and vVp ,
aA then L(v)fa (v).
Fig. 2 shows labeled graphs transformed from the attribute graph shown in
Fig. 1(c) by projection. For each attribute, a projection gpi can be used as a
labeled graph for frequent subgraph mining algorithm presented in next subsec-
tion. And integration will be performed to combine mining results as frequent
attribute subgraphs.
For each projection of an attribute, a linear scanning will be performed to
get each value of the attribute as label of related vertex in labeled graph. Then
a double iteration will be executed to get edges of labeled graph according to
the attribute. Assuming there are m attributes in each vertex of n vertices in
the attribute graph GA . The complexity of time in projection stage is O(n2 ) as
the number of attributes m is a constant.
68 H. Zhang et al.

Fig. 2. Labeled graphs of attribute graph in Fig. 1(c) by projection

4.2 K Hops Frequent Subgraph Mining Algorithm

In this subsection, k hops frequent subgraph mining algorithm FSGen will be
presented. In FSGen, frequent subgraphs are extended from root vertices until
k steps. In each extending step, the subgraphs extended from previous steps are
frequent. Thus, the ultimate extending structures after k steps are frequent. This
idea is based on property 1.

Property 1. Assuming a set of frequent vertices VR ={v1 , v2 , ..., vn }(VR V )

where L(v1 ) = L(v2 ) = ... = L(vn ) and Sup(vi ) M inSup(1 i n). If
uhj N (vi , h)(1 h k) and occurrences of L(uhj ) is no less than given
M inSup, then vertex uhj is frequent and subgraph gi = (Vi , Ei , Li ) is frequent
(k ( (k1
where Vi = {v1 } h=1,j uhj , Ei = { j (v1 , u1j )} h=1,j (uhj , u(h+1)j ).
Proof : The occurrence of L(uhj ) is no less than given M inSup, so in all N (vi , h),
Sup(uhj ) M inSup, and vertex uhj is frequent. e Ei , the start and end
vertices of e is frequent, so edge e is frequent according to definition 3 and 4.
In labeled graph gi , v Vi and e Ei , Sup(vi ) M inSup and Sup(ei )
M inSup, then all components of gi are frequent. As a result, gi is frequent.
VR is considered as the set of root vertices. An iterated processing will be
invoked to get root vertices. If a frequent vertex is not contained in any frequent
subgraph, it will be extracted as new root vertex until all frequent vertices are
contained in frequent subgraphs. Frequency of vertices in a labeled graph will
be calculated by a polynomial time scanning, and the worst complexity of time
is O(n2 ).
In order to brief mining processing, we attempt to avoid overlaps among fre-
quent subgraphs mined by each stage in the algorithm. Overlap is a repeated
structure in subgraph, such as paths, subgraphs and so on. We discuss the prob-
ability of overlap basing on property 2.

Property 2. Assuming frequent vertices v, u, w, if u,wN (v, 1) and L(u)=L(w),

then the frequent subgraphs extended from v and u are probably overlap.
Proof : Given a constant k as the number of hops, assuming u1 N (u, 1),
u2 N (u, 2), ..., u(k1) N (u, k 1) and w1 N (w, 1), w2 N (w, 2), ..., w(k1)
N (w, k 1), where L(ui ) = L(wi ), 1 i k 1. Subgraphs consist of paths
u, u1 , u2 , ..., u(k1) and w, w1 , w2 , ..., w(k1) are overlap.
 K Hops Frequent Subgraphs Mining for Large Attribute Graph 69

Based on above properties, our algorithm for mining frequent subgraph has
following features: (1) Vertices and edges in each extending step are frequent in
all N (vr , h)(1 h k), where vr VR is a root vertex. (2) The same edge
(labels of start and end vertices are the same respectively) occurs once in each
extending step.
K hops frequent subgraph mining algorithm FSGen consists of three stages: ex-
tending, joining and isomorphism testing. Frequent subgraphs are initially
denoted as frequent vertices, and then these vertices will be extended by other
frequent vertices and edges until k hops in extending stage. Frequent subgraphs
obtained from extending stage are constructed as tree model. So in joining stage,
frequent edges in labeled graph will be added into extended k hops frequent sub-
graphs where they do not occur. After above two stages, k hops frequent subgraphs
can be enumerated. If we want to summarize all k hops frequent subgraphs, iso-
morphism testing will be performed and this stage is optional according to actual
requirements. In order to complete our algorithm, isomorphism testing will be pre-
sented along with the other two stages in this subsection. Algorithm 1 shows FS-
Gen performed on labeled graphs by projection. The algorithm consists of three
main procedures: Extending, Joining and Isomorphism Testing.

Algorithm 1. FSGen
Input: (1) k, (2) labeled graph Gp by projection, (3) frequent vertices set Vf , (4)
frequent edges set Ef
Output: k hops frequent subgraphs of Gp
1: Designing Vmf denotes the set of frequent vertices as root vertices set
2: int h = 0;
3: Designing Gf denotes to set of k hops frequent subgraphs and Gf = ;
4: for all vi in Vmf and vi does not belong any extended frequent subgraph do
5: Designing g as a k hops frequent subgraph in Gf and g.V = vi , g.E = ;
6: Extending(vi , g, k, ++h)
7: Joining(g)
8: add g to Gf
9: end for
10: Isomorphism Testing(Gf )
11: return Gf

As each Gp is a labeled graph, an inverted table will be built as index for each
label and sorted by frequency. Then root vertices can be chosen by the inverted
table. In extending stage, frequent vertices that do not occur in any extended
frequent subgraphs can be used as new root vertices in the iteration. Extending
is a recursive procedure as Algorithm 2 shows and the complexity of time is
O(2n ). We design h as the current number of hops. In each steps of extending, a
frequent vertex and a related frequent edge will be added into frequent subgraph
extended in previous steps as line 6 to line 8 shows. In order to avoid overlap, we
use a list of vertices for N (v, 1) which include non-repeated neighbors of vertex
v. While h is equal to k, the procedure will return and g will be extended in
vertices set and edges set.
70 H. Zhang et al.

Algorithm 2. Extending
Input: (1) k, (2)current number of hops h, (3) frequent vertex v, (4) a reference of
frequent subgraph g
Output: reference g as a frequent subgraph
1: if h<k then
2: add v to g.V
3: Designing list L of frequent vertices in N (v, 1), and L includes dierent vertex
only once
4: for all vni in L do
5: if vni is frequent and edge (v,vni ) is frequent then
6: add vni to g.V
7: add (v,vni ) to g.E
8: Extending(vni , g, k, ++h)
9: end if
10: end for
11: end if

Joining procedure is shown in Algorithm 3 and double iteration will be per-

formed in the procedure. Assuming n is the number of frequent vertices in the
labeled graph which is a projection of attribute graph GA and m is the number
of edges in the edges set of frequent subgraph Gf . Time occupation is O(n2 m).
Since m is approximately a constant. The complexity of time in Joining proce-
dure is O(n2 ).

Algorithm 3. Joining
Input: (1) k, (2) k hops frequent subgraph g, (3) frequent edges set Ef
Output: Joined k hops frequent subgraphs g
1: for all vertex vi in g.V do
2: for all vertex vj in g.V do
3: if edge (vi ,vj )Ef and (vi ,vj )
/ g.E then
4: add (vi ,vj ) to g.E
5: end if
6: end for
7: end for
8: return g

Algorithm 4 shows Isomorphism Testing procedure. All k hops frequent sub-

graphs in labeled graph can be enumerated after executing Extending and Join-
ing procedure. If we want to summarize the set of k hops frequent subgraphs,
repeated subgraphs will be reduced, and Isomorphism Testing procedure will be
performed. In existing works, isomorphism testing is considered as NP-complete
problem. In our mechanism, the procedure will be replaced by a approximated
solution named Graph Edit Distance(GED) [11]. GED can be computed in poly-
nomial time O(m3 ) (m denotes to the larger number of vertices in two graphs)
by the idea of dynamic programming. If GED between two frequent subgraphs
 K Hops Frequent Subgraphs Mining for Large Attribute Graph 71

is no more than a given threshold of similarity, these two frequent subgraphs are
considered similar subgraph and one of them will be removed from Gf . The total
complexity of time in isomorphism testing is O(m3 n2 ) where m is the number
of vertices in each frequent labeled subgraph and n denotes the number of all
frequent labeled subgraphs. By the restriction of k, the number of vertices in a
frequent subgraph can be considered as a constant, so the complexity of time in
Isomorphism Testing procedure is O(n2 ).

Algorithm 4. Isomorphism Testing

Input: (1)set of k hops frequent subgraphs Gf , (2)similar threshold s
Output: reduced Gf as summary
1: for all gi in Gf do
2: for all gj in Gf do
3: if Graph Edit Distance(gi , gj )s then
4: remove gj from Gf
5: end if
6: end for
7: end for
8: return Gf

K hops frequent subgraphs mining algorithm FSGen consists of three proce-

dures. The complexities of time in each procedure are: O(2n ), O(n2 ) and O(n2 ).
Therefore, the total complexity of the algorithm FSGen is O(2n +n2 +n2 ), which
can be merged into O(2n ) no matter what including the stage of isomorphism
testing.

4.3 Integration

Attribute graph is transformed into labeled graphs by projection. Algorithms

are performed on labeled graphs for mining k hops frequent subgraphs. Each
frequent subgraph is also a labeled subgraph. In order to merge frequent la-
beled subgraphs into attribute subgraphs, integration will be used. Integration
described in definition 7 is the opposite operation of projection.

Definition 7. Integration. Given n labeled graphs gp1 , gp2 , ..., gpn , where gpi =
(Vpi , Epi , L)
pi ). Integration
)nis a aggregated (
attribute graph GA = (V, E, FA ),
n n
where V = i=1 Vpi , E = i=1 Epi and FA = i=1 Lpi .
According to the definition of integration, many operations of intersection
computing will be carried out. Common solution as definition 7 shows is time-
consuming. On the other hand, there may be no frequent subgraphs mined from
labeled subgraphs for attributes whose values are unique. While integrating all
frequent labeled subgraphs, no frequent attribute subgraphs will be obtained.
Thus, the mining results are not significant. In order to solve problems presented
above, we adopt a flexible method for integration. We add a flag for each attribute
72 H. Zhang et al.

of a vertex, if value of the attribute is frequent, the flag will be set to 1, and
otherwise set to 0. Then we can designate attributes and its related frequent
labeled subgraphs for integration. In each frequent labeled subgraph, scanning
the set of vertices, if flags of the values in designated attributes of a vertex are
equal to 1, the vertex will be selected as a frequent vertex and added to vertices
set of frequent attribute subgraph. This processing is a linear scanning and the
complexity of time is O(m), where m is the number of vertices in the scanned
frequent subgraph. For all frequent subgraphs, the total complexity of time for
integration is O(m n) where n is the number of frequent labeled subgraphs
mined by FSGen.

5 Experimental Studies

The mechanism presented in this paper is written in C++ using Microsoft Visual
Studio.net 2008. All experiments are carried out on a PC with Intel core i7 CPU
2.67GHz, 4GB memory and Microsoft Windows 7.
The experiment consists of two parts to demonstrate the eectiveness of our
work. Firstly, we perform our mechanism for mining frequent subgraphs in at-
tribute graph with dierent k, dierent size of data and dierent support. And
then testing time occupation in each case. Secondly, we perform algorithm FSGen
for mining frequent subgraphs in labeled graphs together with other traditional
algorithms including gSpan, FFSM and GraphGen, which must be rewritten for
large labeled graph instead of multiple small graphs, to compare the time occu-
pations of these algorithms. We obtain large attribute graph data by crawling
from YouTube API and arrange each item to a vertex of graph including 9 at-
tributes. We choose three attributes: category, length and views for frequent
subgraphs mining.

5.1 Mining in Attribute Graph

K is the number of hops denoting distance from root vertex to another in each
frequent subgraph. In this experiment, we firstly discuss the impact of k in time
occupation while running the k hops frequent subgraph mining algorithm FS-
Gen on labeled graphs transformed from the attribute graph based on YouTube
dataset. And then testing the eectiveness with dierent size of data and dier-
ent support.
Fig. 3 shows time occupation according to dierent k. We can see from the
figure that the time occupation is increasing along with the growth of k while
the number of vertices in attribute graph is 500K. The same tendency occurs in
Fig. 4 while the number of hops is set to 3, with the growing amount of vertices,
time occupation of our work will increase. The time occupation is also related to
the support as Fig. 5 shows. The values of x axis is denoted by the ratio between
support (given by definition 3) and total number of vertices in the labeled graph.
More time will be occupied while the support is low because low support will
bring more subgraphs.
 K Hops Frequent Subgraphs Mining for Large Attribute Graph 73

3000 2500 200

2500

Time Occupation (s)

Time Occupation (s)

2000

Time Occupation (s)

150
2000
1500
1500 100
1000
1000
50
500
500

0 0 0
2 3 4 5 6 100000 200000 500000 1000000 2000000 0.1 0.2 0.4 0.6 0.8
Number of hops k Number of vertices support

Fig. 3. Eectiveness of k Fig. 4. Eectiveness of Fig. 5. Eectiveness of

sizes supports

5.2 Mining in Labeled Graph

In this experiment, we compare algorithm FSGen presented in this paper with

gSpan, FFSM and GraphGen. All the algorithms comparing to FSGen can be
used to mine frequent subgraphs in a large labeled graph with few modification.
As a result, these algorithms can be used to mining frequent subgraphs from
labeled graph transformed from attribute graph by projection. The same pro-
cessing of projection and integration will be performed while executing gSpan,
FSSM, GraphGen and FSGen. We only discuss the eectiveness of mining fre-
quent subgraphs from large labeled graph for these algorithms.
Designing the size of attribute graph (assuming 500K vertices in the experi-
ment) and the number of hops (assuming k = 3). We can see from Fig. 6, FSGen
is more ecient than any other algorithms. With the size of graph data grow-
ing, this advantage becomes more and more apparent. And the same tendency
will occur when we compare the time occupation with given supports as Fig. 7
shows.

1000 1400
FSGen
FSGen
gSpan 1200
Time Occupation (s)

800 gSpan
FFSM
Time Occupation (s)

1000
600 GraphGen FFSM
800
GraphGen
400 600
400
200 200
0 0
50000 100000 200000 500000 1000000 2000000 0.1 0.2 0.4 0.6 0.8
Number of vertices support

Fig. 6. Eectiveness of sizes Fig. 7. Eectiveness of supports

6 Conclusion

We introduced a new mechanism for mining k hops frequent subgraphs in large

attribute graph. In our work, large attribute graph was transformed into labeled
graphs by projection, and then algorithm FSGen for mining k hops frequent sub-
graphs in large labeled graph was presented. Finally, integration was performed
to merge frequent labeled subgraphs into frequent attribute subgraphs. Experi-
mental studies show that our mechanism can eciently mining k hops frequent
74 H. Zhang et al.

attribute subgraphs. Comparing with existing algorithms for mining frequent

subgraphs in large labeled graph, algorithm FSGen presented in this paper is
more ecient than gSpan, FFSM and GraphGen on real datasets.

Acknowledgments. This work is partly supported by the National Natural

Science Foundation of China under Grant Nos. 61170184, 61070089; the Na-
tional High Technology Development 863 Programs of China under Grant No.
2013AA013204; the Ph.D. Programs Foundation of Ministry of Education of
China under Grand No. 20120031120038.

References
1. Yang, J., Zhang, S., Jin, W.: DELTA: Indexing and Querying Multi-labeled Graphs.
In: The 20th ACM International Conference on Information and Knowledge Man-
agement (CIKM), pp. 17651774 (2011)
2. Keyvanpour, M.R., Azizani, F.: Classification and Analysis of Frequent Subgraphs
Mining Algorithms. Journal of Software 7(1), 220227 (2012)
3. James, C., Yiping, K., Ada, W.C.F., Jerey, X.Y.: Fast graph query processing
with a low-cost index. The International Journal on Very Large Data Bases 20(4),
521539 (2011)
4. Koren, Y., North, S.C., Volinsky, C.: Measuring and extracting proximity in net-
works. In: The 12th ACM SIGKDD International Conference on Knowledge Dis-
covery and Data Mining (KDD), pp. 245255 (2006)
5. Meinl, T., Borgelt, C., Berthold, M.R.: Discriminative Closed Fragment Mining
and Perfect Extensions in MoFa. In: The 2nd Starting AI Researchers Symposium
(STAIRS), pp. 314 (2004)
6. Li, X.T., Li, J.Z., Gao, H.: An Ecient Frequent Subgraph Mining Algorithm.
Journal of Software 18(10), 24692480 (2007)
7. Inokuchi, A., Washio, T., Motoda, H.: Complete Mining of Frequent Patterns From
Graphs: Mining graph data. Machine Learning 50(3), 321354 (2003)
8. Kuramochi, M., Karypis, G.: Frequent Subgraph Discovery. In: The 1st IEEE In-
ternational Conference on Data Mining (ICDM), pp. 313320 (2001)
9. Yan, X., Han, J.: GSpan: Graph-Based Substructure Pattern Mining. In: The 2nd
IEEE International Conference on Data Mining (ICDM), pp. 721724 (2002)
10. Huan, J., Wang, W., Prins, J.: Ecient mining of frequent subgraph in the pres-
ence of isomorphism. In: The 3rd IEEE International Conference on Data Mining
(ICDM), pp. 549552 (2003)
11. Riesen, K., Bunke, H.: Approximate graph edit distance computation by means of
bipartite graph matching. Image and Vision Computing 27(4), 950959 (2009)
 Privacy Preserving Graph Publication
in a Distributed Environment

Mingxuan Yuan1,2, Lei Chen2 , Philip S. Yu3 , and Hong Mei4

1
Huawei Noah Ark Lab, Hong Kong
2
The Hong Kong University of Science & Technology
3
University of Illinois at Chicago
4
Peking University

Abstract. Recently, many works studied how to publish privacy preserving so-
cial networks for safely data mining or analysis. These works all assume that
there exists a single publisher who holds the complete graph. While, in real life,
people join different social networks for different purposes. As a result, there are
a group of publishers and each of them holds only a subgraph. Since no one has
the complete graph, it is a challenging problem to generate the published graph in
a distributed environment without releasing any publishers local content. In this
paper, we propose a SMC (Secure Multi-Party Computation) based protocol to
publish a privacy preserving graph in a distributed environment. Our scheme can
publish a privacy preserving graph without leaking the local content information
and meanwhile achieve the maximum graph utility. We show the effectiveness of
the protocol on a real social network under different distributed storage cases.

1 Introduction
Recently, privacy preserving graph publication has become a hot research topic and
many graph protecting models have been proposed [9,7,3,12,11,13,14,1]. All these
models assume there is a trustable centralized data source, which has a complete origi-
nal graph G, and can directly generate the privacy preserving graph G from G.

blog1 --- subscribed on 1/9/08

Africa, 20 v1
blog2 --- subscribed on 2/10/08
American, 31 v2
game1 --- score 1-2
Africa, 35 v3 blog1 v2 game1 v5
friend1 --- from 5/6/08 v1 v3
American, 26 v4 v2
blog2 friend1 v6
email1
game2 --- score 4-1 v4 v7
Asia, 28 v5 v3 v5
game2 email2
game3 --- score 8-3-6 friend1 game3 v8
Asia, 42 v6 v4 v6 v9 friend2
email1 --- 1024 byte on 1/3/08
European, 50 v7 v7 v10 email3
email2 --- 716 byte on 4/3/08 v8
Asia, 30 v8 friend2
email4

v9 friend2 --- from 9/9/08

Asia, 33 A1 A2 A3
Africa, 33 v10 email3 --- 512 byte on 7/5/08
email4 --- 1982 byte on 3/15/08

Fig. 1. The complete graph Fig. 2. The distributed storage

However, in reality, people join different social networks due to different interests
or purposes. For example, a person uses Facebook to share his information with his
classmates and coworkers. At the same time, he may also build his blog on a blog
server to share his interests with others. As a result, his connection information is stored
in two social networks. A consequence of this joining preference is that each social
network website only holds partial information (a subgraph) of the complete social

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 7587, 2013.
c Springer-Verlag Berlin Heidelberg 2013
76 M. Yuan et al.

network G. We call such a social network website as a data agent. Consider a social
graph as shown in Figure 1 where each person has two labels and people are involved
in different interactions, the graph can be stored on three data agents, A1 , A2 and A3
separately as shown in Figure 2.
Although people join different networks, it is often necessary to obtain a privacy pre-
serving graph generated from the complete graph for criminal investigation [8] or min-
ing useful patterns/influential persons [4,8,6]. This requests that the distributed agents
should cooperatively generate a privacy preserving graph G . There are three potential
approaches to do this (Figure 3).

Data Agents Data Agents Data Agents

... ... ...

Information without releasing any local content

Local contents Privacy Preserved Graphs (local) SMC

...

A Trustable Agent
Virtual Whole
The Whole Original Graph Virtual Server original graph
...
...

...
...
...

...
Privacy preserved Graph Privacy preserved Graph Privacy preserved Graph

(a) Third-Party Approach (b) Naive Approach (c) SMC Approach

Fig. 3. Architectures for privacy preserving data publishing

A Third-Party approach (Figure 3(a)) is to let all the data agents send their local con-
tents to a trustable third party agent, where the published graph by integrating all the
data. However, since the data of each site is its most valuable asset, no one is willing to
share its data with others, finding such a trustable third party data agent is not feasible in
the real world. The naive approach (Figure 3(b)) is to let each data agent generate a pri-
vacy preserving graph based on its own content and securely combine them into a large
privacy preserving graph 1 . Secure combination means no local content is released dur-
ing the process. However, this approach encounters two problems: (1) It results in the
low quality of the published data since the graph is constructed only based on local in-
formation instead of the complete graph; (2) A privacy preserving graph generated only
by local information may violate the privacy requirement when globally more connec-
tion information is provided considering the connections between nodes [9,13,14,2,1].
The secure combination needs to delete some connections to ensure the final published
graph satisfies the privacy requirement. Therefore, the published graph will have incom-
plete information. Another approach (Figure 3(c)) is that each data agent participates
in a protocol to produce a privacy preserving graph. The privacy preserving graph is
generated just as there virtually exists an agent who has the integrated data. During
the computation, except the content derived from the final published graph, the pro-
tocol controls no additional local content of a data agent is released. This solution is
1
The overlapping between subgraphs should be considered. [5] showed it is not safe even when
each publisher publishes privacy preserved data independently. Here we use Fig. 3(b) to show
the basic workflow of the naive approach.
 Privacy Preserving Graph Publication in a Distributed Environment 77

known as the famous Secure Multi-Party Computation (SMC). SMC deals with a prob-
lem where a set of parties with private data wish to jointly compute a function of their
data. A lot of SMC protocols have been proposed for different computation problems,
but not for privacy preserving graph publishing in a distributed environment. SMC has
two requirements: 1) Correctness requirement: the computation is performed in a dis-
tributed environment just the same as doing it on an agent who holds the integrated
data; 2) Security requirement: each data agent should not know the local information of
other agents even with the intermediate results passing through each other.
In this paper, we follow the SMC approach and design a secure protocol SP to al-
low the data agents to cooperately generate a graph G based on a recently proposed
protection model [1], called S-Clustering. Through clustering methods, S-Clustering
publishes a graph G that only contains super nodes (clusters) where each super node
represents multiple nodes in G. We call a published graph which satisfies S-Clustering
as the S-Clustering graph. We refer the algorithm which generates the S-Clustering
graph in a centralized environment as the centralized algorithm. The distributed ver-
sion of the centralized algorithm, SP , should work the same as running the centralized
algorithm on the complete original graph. For any data agent, SP protects its local
information when generating the published graph. We propose novel solutions in SP
based on random lock, permutation, and Millionaire Protocol [10].

2 Problem Description

2.1 Graph, Storage Model and Problem Definition

An online social network with rich interactions can be represented as a bipartite graph
G(V, I, E) [1], where each node in V represents an entity in the network and each node
in I stands for an interaction between a subset of entities in V . An edge e(u, i) E
(u V , i I) means entity u participants in the interaction i. Each entity has an
identity and a group of attributes. Without loss of generality, each entitys identifier can
be represented as a unique id within [1, |V |]. The commutative encryption scheme [8]
can assign each name a unique number id without releasing the real name values in a
distributed environment. In the rest part of this paper, we refer entity u as the same as
the entity with id u. Each interaction also has an identity and a set of properties as shown
in Figure 1. An interaction can involve more than two entities, such as the interaction
game3. Two entities can also be involved in different interactions at the same time.
For example, in Figure 1, v1 and v2 participate in blog1 and blog2 simultaneously.
In a distributed environment, the graph G is distributively stored on l different data
agents. That is, each agent Ai holds a portion of the graph Gi (Vi , Ii , Ei ) such that
V = li=1 Vi , I = li=1 Ii and E = li=1 Ei . The interactions that an entity participants
may be stored on different agents. The Gi held by different data agents may overlap,
such that the intersection between two graphs stored on different agents might not be
an empty set. 2 That is, li=1 Vi = and li=1 Ii = .
2
When Vi s do not have overlap (li=1 Vi = ), G is a disconnected graph since each data agent
only holds a subgraph that is isolated from other parts.
78 M. Yuan et al.

In the rest of this paper, we use node to specifically represent an entity in V and use
interaction to represent an item in I. When we say two nodes u, v have a connection,
it means u and v are involved in a same interaction i in I (In the graph, we have two
edges e(u, i) and e(v, i)).
We propose a protocol which securely generates a S-Clustering graph [1] for in-
teraction based graphs in a distributed environment. S-Clustering [1] assumes that an
attacker can know the id, label and any connection information of a node. An attacker
uses the information he knows about some users to analyze the published graph in order
to learn more about these users. Given a constant k, S-Clustering publishes a clustered
graph (S-Clustering graph) which guarantees the following three Privacy objectives:
Objective 1: For any node u, an attacker has at most k1 probability to recognize it.
Objective 2: For any interaction i, an attacker has at most k1 probability to know a node
u involves in it.
Objective 3: For any two nodes u and v, an attacker has at most k1 probability to know
they have a connection (u and v participate in a same interaction).
We assume that all participants joined in the computation, including all data agents and
necessary extra servers are semi-honest. The local information beyonds the Privacy
objectives of S-Clustering is protected. The problem we solve in this paper is:
1. Input: A graph G distributed stored on l data agents and a privacy parameter k.
2. Output: A S-Clustering graph G of G under k.
3. Constraints: (a) Correctness: G is computed the same as generating it on an agent
who holds G; (b) Security: The three Privacy objectives of S-Clustering are guaran-
teed even when a participant gets the intermediate results during the computation.

2.2 S-Clustering Model

Given a graph G(V, I, E), to satisfy the three privacy objectives, a S-Clustering graph
G (CV, I, CE) is published, where CV is a super node (cluster) set in which each
super node represents a group of entities in V . We use c to denote a cluster in CV and
|c| to denote the cluster size, which is the number of entities that c represents. For each
interaction i in I, if c contains an entity which participates in i, there is an edge e(c, i)
in G . Figure 4 is a clustered graph of Figure 1.

(1) Generate random vector D with size |V| [4]

blog1 (2) D = D + (degrees on A1)
blog2 (Africa, 20)
{v1,v5} game1 {v1,v5} (Asia, 28) [2] D [3] D
[1] D ...
friend1 (American, 31)
{v2,v7} {v2,v7} (European, 50) (3) D = D + (degrees on A2)
game2 A1 [6] A2 Al
(D) [5] (D)
game3 (Africa, 35) (4) D = D + (degrees on Al)
{v3,v9} {v3,v9} (Asia, 33) (5) Create a random
email1 permutation on [1, |V|]
(Amrican, 26) The Millionaires' Protocol
{v4,v8} email2 {v4,v8} (Asia, 30)
friend2 (Asia, 42) S2 S1
{v6,v10} {v6,v10} (Africa, 41) (6) S1 and S2 cooperately compute the sorting L on (id)s on [ (D) - (D) ]
email3 [ (D)[i] - (D)[i] ] < [ (D)[j] - (D)[j] ] ? == [ (D)[i] - (D)[j] ] < [ (D)[i] - (D)[j] ] ?

email4 Securely Compare

Fig. 4. A clustered graph Fig. 5. Stage 1: SSSP

 Privacy Preserving Graph Publication in a Distributed Environment 79

To guarantee the three privacy objectives, the S-Clustering graph G must satisfy:

Each cluster represents at least k entities. This guarantees the Objective (1).
The Clustering Safety Condition (CSC) [1]:

Definition 1. A division (clustering) of nodes V into clusters satisfies the Clustering

Safety Condition (CSC) if for any node v V , v participates in interactions with at
most one node in any cluster S V , that is:

e(v, i), e(w, i), e(v, j), e(z, j) E : w S z S z = w;

e(v, i), e(w, i) E : v S w S v = w;

This condition guarantees Objectives (2) and (3).

The CSC requires that any two nodes in a cluster cannot be connected or connect to
the same node. Figure 4 is a clustered graph which satisfies CSC. For any two nodes
that appear in the same cluster of Figure 4, they do not have a connection or connected
to the same third node. For any two nodes u and v, if they satisfy (or do not satisfy) the
CSC, we denote this as CSC(u, v) (or CSC(u, v)). Similarly, for a cluster c and a
node u, if u satisfies (or does not satisfy) CSC with all the nodes in c, we denote this
as CSC(c, u) (or CSC(c, u)).

3 Secure Protocol (SP)

3.1 Overview

We proposed SP to generate a S-Clustering graph in a distributed environment as

shown in Algorithm 1. We call a computation that satisfies the Security requirement
as a Secure Computation. Thus, SP contains four stages such that each stage conducts
one Secure Computation in lines 1, 7, 15, and 16, respectively. Stage 1 sorts nodes
without releasing any degree information. Stage 2 clusters nodes by correctly checking

Algorithm 1. Clustering with CSC in distributed environment

1 Securely Sort(V) by degree; /* Stage1 */
2 CV=null; /* Begin Stage2 */
3 for v V do
4 flag = true;
5 for cluster c CV do
6 if SIZE(c)<k then
7 Distributed computing r = CSC(c, v);
8 if r then
9 Insert(c,v) without generating/updating edges between clusters;
10 flag = false;
11 break;

12 if flag then
13 Create a New Cluster c and add c to CV ;
14 Insert(c, v); /* End Stage2 */

15 Compute interactions from distributed agents; /* Stage3 */

16 Publish a graph with node attributes only links to clusters; /* Stage4 */
80 M. Yuan et al.

CSC(c, v) without releasing any connection information. Stage 3 and Stage 4 gener-
ate the interactions on clusters and attributes for clusters respectively without disclos-
ing any interaction-node mapping and node-attribute mapping. To summarize, we must
avoid the releasing of the following information during computation: (1) Degree infor-
mation (from Stage 1); (2) Specific node-attribute mapping (from Stage 4); (3) Specific
connection information, including node-interaction mapping and connection between
nodes (from Stage 2 & 3); (4) Any CSC result and the computation order of nodes
(from Stage 1 & 2).
Next, we introduce our design of SP for each stage in detail. Due to space limitation,
we ignore all the proofs. Before presenting SP , we assume each node/interaction has a
weight that represents how many duplicated copies of this node/interaction have been
stored in the system. We call these weights duplicated weights. Details of the method
to generate duplicated weights is shown in Appendix A.

3.2 Stage 1: Secure Sorting Sub-Protocol (SSSP)

Protocol Design. In this stage, we sort the nodes without revealing any degree infor-
mation and the computation order of nodes (the sorting order of nodes on real ids)
information. The basic idea is to do the sorting on permuted ids with their correspond-
ing locked degree values. We use the random number to lock the real degree of each
node. That is, we make one agent hold a key (a random number) and another agent hold
the locked degree value (real degree plus this random number). Then, we sort all nodes
on permuted ids through the cooperation of these two agents. During the sorting, the
agent who has the locked degree values cannot learn any key value and the agent who
holds the keys cannot learn any locked value. The Secure Sorting Sub-Protocol (SSSP)
works as follows (Figure 5):
1. A1 generates a random real number vector D with size |V |.! A1 constructs a variable vector
D = D. Then, for each node u stored on A1 , A1 adds e(u,i) w1i to D [u] (wi is the
duplicated weight of the interaction i). Finally, A1 sends D to A2 ;
!
2. When A2 receives D , for each node u stored on A2 , we add e(u,i) w1i to D [u]. Then, A2
sends D to A3 ;
3. Each Ai (i > 1) does the same operation as A2 and sends D to Ai+1 if Ai+1 exists;
4. Al generates a random permutation function , passes to A1 and sends (D ) to S1
5. A1 sends (D) to S2 .
6. S1 and S2 cooperately sort all the nodes. During the sorting, each time when S1 needs
to compare two values (D )[i] (D)[i] and (D )[j] (D)[j], S1 and S2 uses the
Millionaires Protocol [10]3 to securely get the result by comparing the value ((D )[i]
(D )[j]) on S1 and ((D)[i] (D)[j]) on S2 .

Theorem 1. SSSP sorts all nodes on (id)s under the Security requirement.
3
The Millionaires Problem is: suppose there are two numbers a and b hold by two people, they
want to know the inequality a > b or a < b without revealing the actual values of a and b.
The Millionaires Protocol [10] can securely compare a and b based on the techniques such
as Homomorphic Encryption. When we want to compare two degree values d1 and d2 in case
that Sx knows r1 , r2 and Sy knows (r1 + d1 ), (r2 + d2 ), we can use the Millionaires Protocol
to compare the value ((r1 + d1 ) (r2 + d2 )) on Sx and the value (r1 r2 ) on Sy .
 Privacy Preserving Graph Publication in a Distributed Environment 81

3.3 Stage 2: Secure Clustering Sub-Protocol (SCSP)

Protocol Design. In Stage 2, we need to cluster nodes based on the sorted order L
without revealing any connection information.
Since S1 holds L after Stage 1, in SCSP , we continue to make S1 generate clusters
on (id)s. Before S1 can do clustering on (id)s, the connection information between
nodes should be mapped on (id)s firstly. So, SCSP contains the following three steps:

1. Generate a noise matrix M R on A1 and a noised adjacent matrix M R of G on Al . For any

two nodes u and v, M R [u][v] M R[u][v] > 0 is equivalent to u, v has a connection;
2. Create a matrix N M R on A1 and N M R on Al based on M R and M R , respectively. For
any two nodes u and v, N M R [u][v]N M R[u][v] is a [0, 1] matrix where N M R [u][v]
N M R[u][v] = 1 indicates that u, v have a connection;
3. Do clustering on (id)s with (N M R ), (N M R) through the cooperations among S1 ,
S2 and S3 .
The first two steps map the connection information on (id)s (The [0, 1] character of
N M R N M R will be used for securely clustering). The last step does the clustering
on the connection information between (id)s. During the computation, to satisfy the
Security requirement, we make the agent who holds M R, N M R or (N M R) cannot
get the noised content M R , N M R or (N M R ) and vice versa.
The first step computes M R and M R securely as follows:

1. A1 generates a random real number matrix M R with size |V | |V |. A1 constructs a vari-

able matrix M R = M R. Then, a self-edge is added for each node (The self-edge will be
used to determine the CSC). For each u, A1 sets M R [u][u] = M R [u][u] + a1 (a1 is a
random positive number). For any interaction i and any two nodes u and v that participate
in i, A1 sets M R [u][v] = M R [u][v] + a2 (a2 is a positive random number), and then A1
sends M R to A2 ;
2. After A2 getting M R , for any interaction i and any two nodes u and v that participate in
i, A2 sets M R [u][v] = M R [u][v] + a3 (a3 is a positive random number), and then A2
sends M R to A3 ;
3. Each Ai (i > 1) does the same operation as A2 and sends M R to Ai+1 if Ai+1 exists;
After the first step, Al gets a noised matrix M R . For any nodes u and v, M R [u][v]
M R[u][v] > 0 is equivalent to u and v have a connection. The second step securely
computes the matrices N M R and N M R as shown in Figure 6:
1. Al generates a random permutation function 2 on |V | |V | numbers and sends 2 to A1 .
Al sorts all the numbers in M R row by row to obtain a vector L2 whose length is |V ||V |;
2. A1 sorts M R row by row to obtain a vector L2 ;
3. A1 sends 2 (L2 ) to S3 and Al sends 2 (L2 ) to S3 ;
4. S3 computes two random number vectors LN2 and LN2 which satisfy
"
0 2 (L2 )[i] 2 (L2 )[i] = 0
LN2 [i] LN2 [i] =
1 2 (L2 )[i] 2 (L2 )[i] > 0

Then, S3 passes LN2 to A1 and LN2 to Al ;

5. A1 computes 21 (LN2 ) and converts the results to a |V | |V | matrix N M R. A1 passes
(N M R) to S2 ; Similarly, Al generates the corresponding matrix N M R based on
21 (LN2 ) and passes (N M R ) to S1 ;
82 M. Yuan et al.

(2) L2 = sort MR row by row

(4) generate matrix NMR
based on 2 -1(LN2)
[4] (NMR)
S1
[1] 2

[2] 2 (L2')
A1 [2]
Al
2 (L2)
(1) generate a random
[4] (NMR) [3] LN2 permutation 2 (|V|*|V|)
[3] LN2'
(2) L2' = sort MR row by row
(4) generate matrix NMR
based on 2-1 (LN2')

(3) Compute random vectors LN2 and LN2'

based on 2 (L2')- 2 (L2)

S2 S3

Fig. 6. Stage 2: SCSP 2

The third step is to do the clustering on (id)s. Before introducing how S1 conducts
the clustering, we prove a property we use in the next computation. For a node u, we
call Eu (|Eu | = |V |) the connection vector of u. Eu is a [0, 1] vector and Eu [u] = 1. If
nodes u and v have a connection, Eu [i] = 1, otherwise Eu [i] = 0. For example, node
v1 s connection vector# in Figure 1 E1 = [1, 1, 0, 0, 0, 0, 0, 0, 0, 0]. For a cluster of nodes
C, we call EC = ( uC Eu ) as Cs connection vector. The connection vector has
the following property:
Theorem 2. For any cluster of nodes C, if there exists a t where EC [t] > 1, then C
does not satisfy CSC; otherwise, C satisfies the CSC.

Knows: (NMR) (1) Generate random AT (|V|*m matrix) [4] ''

(3) Compute Rc '( ( NMR)[u] )
u C
(2) AT = AT + (attributes on A1)
(4) Generate a random permutation vector R (with size |V|)
(5) Compute RC = RC + R
[1] AT [2] AT ... [3] AT
[3] RC [1] (C, )
S2
[2] R
A1 A2(3) AT = AT + (attributes on A2) Al
Knows: (NMR) [5] ''(AT) (4) AT = AT + (attributes on Al)
[5] ''(CV) (5) Generate a random
(1) Compute ERc ( NMR' )[u] whereC c {v}
permutation pattern
[5] T/F u C

(2) Generate a random permutation [6] CVatr (7) Compute CVatr = ''-1(CVatr)
(5) Compute ERC = (ERC) + R
[4] ERC
S3 S1 (6) Compute the attributes cluster mapping CVatr on
(6) Compute (ERC - RC ) to determine if CSC(c,u) ''(id)s using [ ''(AT) - ''(AT) ]
T : CSC(c, v), F : CSC(c, v)
S1

Fig. 7. Stage 2: SCSP 3 Fig. 8. Stage 4: SLSP

#
For a cluster of nodes C on (id)s, it is obvious EC = uC ((N M R )[u]
(N M R)[u]). Based on the property of connection vector, when S1 needs to check
whether v can join a cluster c based on CSC, the protocol works as follows (Figure 7):
!
1. S1 computes a connection vector ERC = uC (N M R )[u]. Where C = c {v}. The
size of this vector is |V |.;

2. S1 generates a random permutation ! pattern and sends (C, ) to S2 ;
3. S2 computes a vector RC = ( uC (N M R)[u]). A2 continues to generate a random
vector R with size |V | and sends R to S1 . S2 sends RC
= RC + R to S3 ;

4. S1 computes ERC = (ERC ) + R and passes ERC to S3 ;

5. S3 computes EC = ERC RC . If EC contains a number bigger than 1, it returns cannot
cluster to S1 . Otherwise, it returns can cluster to S1 .
S1 does the clustering on L and uses the above method to test CSC. S1 finally gets the
cluster set CV on (id)s and passes CV to A1 . A1 computes the cluster set on real
ids CV = 1 (CV ).
 Privacy Preserving Graph Publication in a Distributed Environment 83

Theorem 3. SCSP clusters all nodes without violating the Security requirement.

3.4 Stage 3: Secure Edge Generation Sub-Protocol (SESP)

Protocol Design. In Stage 3, we generate the interactions among clusters. After A1
gets the cluster set CV on real ids, A1 reports CV to all Ai s one by one. Each Ai gen-
erates the interactions between clusters and sends the results to Ai+1 . Finally Al gets a
clustered graph with correct interactions. The above process only passes the connection
information through interactions without the attribute information. Each Ai directly
sends the attribute of each interaction to Al since all the interactions will be clearly
published. Since the interactions between super nodes reported by each Ai is a sub-
set of the final result, the middle results are already included in the published graph.
Therefore the Security requirement is satisfied.

3.5 Stage 4: Secure Label Generation Sub-Protocol (SLSP)

Protocol Design. In Stage 4, we generate the node attributes for each cluster without
disclosing any specific node-attribute mapping. Suppose there are m attributes for each
node. The Secure Label Generation Protocol (SLGP) works as shown in Figure 8. This
is similar to Stage 1. Due to page limitation, we ignore the detailed description.
Theorem 4. SLSP assigns node attributes to each cluster without violating the Secu-
rity requirement.
Based on the above step by step illustration of Algorithm, we conclude that the whole
SP satisfies the Correctness and Security requirements. The intermediate results of
one stage do not influence other stages since computation contents are different. Since
each stage satisfies the Security requirement, the whole SP also satisfies the Security
requirement.

4 Experiment
To demonstrate the effectiveness of SP protocol, we implement a Relaxed Secure Pro-
tocol (RSP) which is based on the naive approach (Fig. 3(b)). The brief introduction
to the design of RSP is shown in Appendix B. We compare the graph generated We
compare the graph generated by RSP and SP .

4.1 Criteria
We focus on two aspects to show the benefit of designing a SM C protocol: the infor-
mation loss and the utilities. We estimate the information loss of RSP comparing with
SP (SP does not delete any interaction). Assume del is the number of interactions that
are deleted by RSP and the original complete graph contains |I| interactions, we use
the ratio of deleted interactions ( 100del
|I| %) to represent the information loss.
Utility is used to estimate the quality of the published graph. For the clustering-based
protection models [7,12,1], the utility testing is operated by drawing sample graphs
from the published one, measuring the utility of each sample and aggregating utilities
across samples. We test the following measures:
84 M. Yuan et al.

1. Degree distribution: Suppose the sorted degree sequence of the original graph G is
DG and the one of a sampled graph Gc is DGc . The different !
between the degree dis-
tributions of G and Gc is represented as: EDDD (DG , DGc ) = |V1 | "|V
i=1 (DG [i] DGc [i]) .
| 2

We compute the average EDDD of 30 sampled graphs for each protocol. Suppose
the SP s result is EDDD,SP and RSP s result is EDDD,RSP , we use
EDDD,RSP EDDD,SP
EDDD,SP 100% to compare these two protocols.
2. A group of randomly selected queries: We test three aggregate queries as the same
as Paper [1]. We compute the average query error between the sampling graphs and
the original graph. Suppose SP s result is errorSP and RSP s result is errorRSP ,
errorSP
we use errorRSP
errorSP 100% to compare them. The three aggregate queries
include: (1) Pair Queries: how many nodes with certain attribute interact with nodes
with another attribute; (2) Trio Queries: how many two hop neighbors. (3) Trian-
gle Queries: how many triangles given three attributes. We select a random set of
queries of each type to do the testing. For each type, we select 20 queries and com-
pute the average query error between the sampling graphs and original graph.

4.2 Data Set and Distributed Storage

Our experiment is operated on a real life social network we crawled from
ArXiv (arXiv.org). ArXiv(arXiv.org) is an e-print service system in Physics, Computer
Science etc. We extract the co-author graphs in Computer Science. Each node denotes
an author, and each interaction means a unique coauthor paper. The ArXiv provides 37
categories in Computer Science to search the papers. Since most people work in multi-
ple sub-fields and each paper also belongs to multiple sub-fields, the authors of papers
in different categories are overlapped and finally form a large graph. We set each au-
thors research field as his/her node label. We divide the 37 categories to 6 sets and store
the crawled graph of each category set on one data agent. This can be seen as different
data agents maintain the relationship of people in different fields. The integrated graph
contains 28868 nodes and 23290 interactions. On average, each interaction is involved
by 2.4 nodes.
We consider the following two distributed storage cases:
1. Real Distributed Case: The 6 subgraphs we crawled in different research fields can
be seen as 6 data agents maintain the relationship of people in different research
fields. The number of nodes stored in the 6 data agents are 6371, 6370, 7705, 5923,
4676 and 7876 respectively. There are r = 34% nodes stored multiple times on
different agents. We call r the node overlapping ratio.
2. Simulated Distributed Case: We manually divide the integrated graph G to 6 parts
that have similar node numbers with different node overlapping ratios (rs).

4.3 Result
Result on Real Distributed Case
We compare the performance of SP and RSP under different privacy parameter ks.
Figure 9(a) shows the result of information loss. RSP deletes 13% to 17% interac-
tions. This is a fairly significant information loss. The published graph by RSP fails
 Privacy Preserving Graph Publication in a Distributed Environment 85

18 30
RSP vs. SP 80 RSP vs. SP RSP vs. SP

Increased query error (%)

Deleted interactions (%)

Increased Distance (%)

75 25
17
70
20
16 65
15
60
15
55 10
50
14
5
45
13 40 0
4 6 8 10 12 14 16 4 6 8 10 12 14 16 4 6 8 10 12 14
k k k

(a) Information Loss (b) Distance of DD (c) Error of queries

Fig. 9. Real Distributed Case

15 42 40
RSP vs. SP RSP vs. SP RSP vs. SP

Increased query error (%)

40 35
Deleted Interactions (%)

14
Increased distance (%)

38 30
13
36 25
12 34 20

11 32 15
30 10
10
28 5
9
26 0
8 24 -5
0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80
Node overlapping ratio (%) Node overlapping ratio (%) Node overlapping ratio (%)

(a) Information Loss (b) Distance of DD (c) Error of queries

Fig. 10. Simulated Distributed Case

to correctly represent the original graph. Figure 9(b) shows the comparison of degree
sequence distance. RSP performs 45% to 80% worse than SP . The published graph
by RSP is much worse than SP when estimating by the degree distribution. Figure
9(c) shows the results of queries. In most cases, the RSP performs 10% to 25% worse
than SP . In most cases, SP gets a much better graph than RSP .
Result on Simulated Distributed Case
For the Simulated Distributed Case, we set k = 5 and compare the performance of SP
and RSP with different node overlapping ratio rs. Figure 10(a) shows the number of
deleted interactions. Larger r means more overlapping between data agents and more
chances to delete an interaction. In most cases, RSP deletes 10% to 20% interactions.
Figure 10(b) shows the comparison of degree sequence distance. RSP performs 25%
to 45% worse than SP . Figure 9(c) shows the results of queries. In most cases, the
RSP performs 5% to 35% worse than SP . SP generates a graph with higher utilities
than RSP .
From the testing results, we can find SP exhibits benefits on both information com-
pleteness and the utilities of the published graphs. Firstly, the RSP deletes roughly
13%-19% interactions, publishing a graph which loses such a large portion of infor-
mation is not acceptable. While SP guarantees no information loss in the published
graph. Secondly, the utilities of the graph generated by SP are much better than RSP .
It is necessary to design a SMC protocol such as SP for the privacy preserving graph
publication problem. Actually, since SP has the same effect as running the state-of-art
centralized graph construction algorithm on the complete original graph, the published
graph generated by SP can be seen as the best.
86 M. Yuan et al.

5 Related Work
Two models are proposed to publish a privacy preserving graph: the edge editing based
model [9,13,14,2,11] and the clustering based model [7,12,1]. The edge editing based
model is to add/delete edges to make the published graph satisfy certain properties ac-
cording to the privacy requirements. For example, Liu[9] defined and implemented the
k-Degree anonymous model on network structure, with which there exists at least other
k 1 nodes having the same degree as any node in a published graph. The clustering
based model is to cluster similar nodes into super nodes. Since a clustered graph only
contains super nodes, the probability of re-identifying any user can be bounded to k1 by
making each clusters size at least k.
Frikken [4] designed a protocol which allows a group of agents to generate an in-
tegrated graph. However, the graph generated by this protocol cannot provide the pro-
tections against the complex attacks proposed in recent works [9,7,3,12,11,13,1,14].
Kerschbaum [8] designed a protocol to generate an anonymized graph from multiple
data agents. Each edge in the anonymized graph has a digital signature which can help
trace where this edge comes from. While, simple removing the identifiers of nodes in
a graph cannot resist an attack which aims to re-identify the nodes/links [7]. There-
fore, it is essential to investigate protocols that support the stronger protection models
[9,7,3,12,11,13,1,14] in a distributed environment. Our work generates the published
graph which follows the recently proposed S-Clustering protection model. This model
protects nodes and links against the strongest attack which may use any information
around nodes. Moreover, our work supports a more flexible graph model while the pro-
tocols of Keith [4] and Florian [8] only support the homogeneous graphs.

6 Conclusion
In this paper, we target on the secure multi-party privacy preserving social network
publication problem. We design a SMC protocol SP for the latest clustering based
graph protection model. SP can securely generate a published graph with the same
quality as the one generated in the centralized environment. As far as our knowledge,
this is the first work on SMC privacy preserving graph publication against the structure
attack. In the future, one interesting direction is to study how to enhance the current
solution to handle the malicious or accessory agents. How to design SMC protocols for
the editing based graph protection models will be another interesting direction.

Acknowledgments. This work is supported in part by Hong Kong RGC grants

N HKU-ST612/09, National Grand Fundamental Research 973 Program of China under Grant
2012CB316200 and NSFC 619308.

References
1. Bhagat, S., Cormode, G., Krishnamurthy, B., Srivastava, D.: Class-based graph anonymiza-
tion for social network data. Proc. VLDB Endow. 2, 766777 (2009)
2. Cheng, J., Fu, A.W.-C., Liu, J.: K-isomorphism: privacy preserving network publication
against structural attacks. In: SIGMOD 2010, pp. 459470. ACM, New York (2010)
 Privacy Preserving Graph Publication in a Distributed Environment 87

3. Cormode, G., Srivastava, D., Yu, T., Zhang, Q.: Anonymizing bipartite graph data using safe
groupings. Proc. VLDB Endow. 1, 833844 (2008)
4. Frikken, K.B., Golle, P.: Private social network analysis: how to assemble pieces of a graph
privately. In: WPES 2006, pp. 8998. ACM, New York (2006)
5. Ganta, S.R., Kasiviswanathan, S., Smith, A.: Composition attacks and auxiliary information
in data privacy. CoRR (2008)
6. Garg, S., Gupta, T., Carlsson, N., Mahanti, A.: Evolution of an online social aggregation
network: an empirical study. In: IMC 2009, pp. 315321. ACM, New York (2009)
7. Hay, M., Miklau, G., Jensen, D., Towsley, D., Weis, P.: Resisting structural re-identification
in anonymized social networks. Proc. VLDB Endow. 1, 102114 (2008)
8. Kerschbaum, F., Schaad, A.: Privacy-preserving social network analysis for criminal investi-
gations. In: WPES 2008, pp. 914. ACM, New York (2008)
9. Liu, K., Terzi, E.: Towards identity anonymization on graphs. In: SIGMOD 2008, pp. 93106
(2008)
10. Yao, A.C.: Protocols for secure computations. In: SFCS 1982, pp. 160164. IEEE Computer
Society, Washington, DC (1982)
11. Ying, X., Wu, X.: Randomizing social networks: a spectrum preserving approach. In: SDM
2008 (2008)
12. Zheleva, E., Getoor, L.: Preserving the privacy of sensitive relationships in graph data. In:
Bonchi, F., Malin, B., Saygn, Y. (eds.) PinKDD 2007. LNCS, vol. 4890, pp. 153171.
Springer, Heidelberg (2008)
13. Zhou, B., Pei, J.: Preserving privacy in social networks against neighborhood attacks. In:
ICDE 2008, pp. 506515 (2008)

14. Zou, L., Chen, L., Ozsu, M.T.: k-automorphism: a general framework for privacy preserving
network publication. Proc. VLDB Endow. 2, 946957 (2009)

A Method for Duplicated Weights Computation

Since each node and interaction has a unique id, A1 generates two random vectors RVdup and
RIdup . RVdup has size |V | and RIdup has size |I|. Then A1 generates two variable vectors Vdup
and Idup with Vdup = RVdup and Idup = RIdup . Vdup [i] RVdup [i] will be used to represent
the duplicated weight of node i. Idup [i] RIdup [i] will be used to represent the duplicated
weight of interaction i. Each Ai does the following operations: (1) For each node i stored by
Ai , Vdup [i] = Vdup [i] + 1; (2) For each interaction i stored by Ai , Idup [i] = Idup [i] + 1; (3)
Passes Vdup and Idup to Ai+1 if Ai+1 exists. After doing this, Al sends the Vdup and Idup to A1 .
Finally, A1 gets the duplicated weights of all nodes and interactions by computing Vdup RVdup
and Idup RIdup . A1 passes the result to each Ai and Ai stores the duplicated weights for the
nodes and interactions in Gi .

B Relaxed Secure Protocol (RSP)

We implement RSP by clustering nodes on local data agents and deleting some cross agent inter-
actions. By doing this, the Security can be guaranteed. A cross agent interaction is an interaction
whose participants contain at least two nodes which have duplicates in the system. We only delete
the cross agent interactions which break or have the potential risk to break the CSC. A1 gen-
erates a clustered graph and passes it to the next agent A2 . Each Ai generates a larger clustered
graph by adding its local content into the received clustered graph. The cross agent interactions,
which have potential to break the CSC, are deleted.
 Correlation Mining in Graph Databases
with a New Measure

Md. Samiullah1 , Chowdhury Farhan Ahmed1 , Manziba Akanda Nishi1 ,

Anna Fariha1 , S M Abdullah2 , and Md. Rafiqul Islam3
1
Department of Computer Science and Engineering, University of Dhaka, Bangladesh
sami cse [email protected], [email protected], [email protected],
[email protected]
2
Department of Computer Science and Engineering,
United International University, Bangladesh
[email protected]
3
School of Computing and Mathematics, Charles Sturt University, Australia
[email protected]

Abstract. Correlation mining is recognized as one of the most impor-

tant data mining tasks for its capability to identify underlying dependen-
cies between objects. Nowadays, data mining techniques are increasingly
applied to such non-traditional domains, where existing approaches to
obtain knowledge from large volume of data cannot be used, as they
are not capable to model the requirement of the domains. In particu-
lar, the graph modeling based data mining techniques are advantageous
in modeling various real life complex scenarios. However, existing graph
based data mining techniques cannot eciently capture actual corre-
lations and behave like a searching algorithm based on user provided
query. Eventually, for extracting some very useful knowledge from large
amount of spurious patterns, correlation measures are used. Hence, we
have focused on correlation mining in graph databases and this paper
proposed a new graph correlation measure, gConf idence, to eciently
extract useful graph patterns along with a method CGM (Correlated
Graph M ining), to find the underlying correlations among graphs in
graph databases using the proposed measure. Finally, extensive perfor-
mance analysis of our scheme proved two times improvement on speed
and eciency in mining correlation compared to existing algorithms.

Keywords: Correlation mining, knowledge discovery, correlated graph

patterns, graph mining, graph correlation.

1 Introduction
Data mining extracts useful implicit patterns within data along with important
correlations/anities between the patterns to discover knowledge from databases.
Moreover, in data mining, correlation analysis is a special measure which finds the
underlying dependencies between objects. The frequent itemset can be mined but
correlation calculation among items can define the dependencies and mutual cor-
relation among items within an itemset. The hidden information within databases,

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 8895, 2013.
c Springer-Verlag Berlin Heidelberg 2013

 Correlation Mining in Graph Databases with a New Measure 89

and mainly the interesting association relationships among sets of objects, may
disclose useful patterns for decision support, financial forecast, marketing poli-
cies, even medical diagnosis and many other applications.
Nowadays, data mining techniques are applied to non-traditional domains e.g.
the most complicated real life scenarios where objects are interacting with their
surrounding other objects. To model such scenarios graph can be used, where
vertices of the graph will correspond to entities and edges will correspond to
relations among entities. Because of combinatorially explosive search for sub-
graphs which includes subgraph isomorphism testing, the graph structured data
mining is dicult. Moreover, in mining graph data, the emphasis is on frequent
labels and common topologies unlike traditional data. Here mining can be di-
vided into level-by-level generate-and-test method and pattern growth-based
approach. AGM[1] and FSG[2] are of the former type, where as gSpan[3] and
graph pattern [4] are of later type, which require no candidate generation. In
order to discover correlations, several measures are used [5] and to mine corre-
lation in graph databases existing works such as [6] mainly focus on structural
similarity search. However, graphs those are structurally dissimilar but always
appear together within the database, may be more interesting , such as, the
chemical properties of isomers [7].
For capturing eective correlations, authors of [7] proposed CGS algorithm
which mines graph correlation by adopting P earson s correlation coef f icient by
taking into account the occurrence distributions of graphs. However, CGS works
for searching correlation of a specific query graph with the database. Therefore,
it has some limitations in describing inherent correlation among graphs and the
domain knowledge is obligatory in using CGS, otherwise lots of queries would be
meaningless. These facts motivated us in developing a new measure which can
prune a large number of un-correlated graphs and designing an algorithm to ef-
ficiently mine graph correlation. Our contributions are: a new graph correlation
measure gConf idence to mine inherent correlation in graph databases, pruning a
large number of candidates using the downward closure property of the measure
and an algorithm CGM (Correlated Graph M ining) to eciently mine correla-
tion by constructing a hierarchical reduced search space.
Rest of the paper is organized as follows: Section 2 contains our proposed
scheme and Section 3 focuses on the performance analysis of our proposed algo-
rithm. Finally, we concluded our work in Section 4.

2 Our Proposed Approach

Correlated graph mining is one of the most important graph mining tasks. So, we
have proposed a new measure, gConf idence and a new method, CGM (Correlated
Graph M ining), to search correlation among graphs within a graph database. We
have mined correlated graphs by constructing a hierarchical search space using our
proposed measure and algorithm.
90 Md. Samiullah et al.

In graph domains, transactions can be represented by G = {V (G), E(G),

L(V (G)), L(E(G))} that is the set of vertices, set of edges and set of labels for
vertices and edges respectively. Size of a graph varies in various algorithms and
can be the number of nodes or edges or disjoint paths. To tackle graph isomor-
phism problem, various labeling or coding is introduced in many algorithms.
For a graph database, GD, with Gs , Gb and Gh are sub-graphs of any graph
G GD, the support of Gs is the ratio among the frequency of Gs s supergraph
and total number of transaction graphs. Moreover, confidence of Gb with respect
to Gh is the ratio between the joint-occurrence frequency of Gb along with Gh
and frequency of the super graph for Gb . The canonical labeling for the normal
form representation, X of a graph G can be defined as[3], the minimum DFS
code among all possible DFS codes of G . The proposed measure gConf idence,
is defined as follows where the correlation is defined based on the co-occurrence
probability of edges of the graphs within the database.

Definition 1. (gConfidence) Given a graph database GD and one of its trans-

action graph G, then we have defined the gConf idence of Gs , a subgraph of G
as
{N o. of graphs G | Gs G GD}
gConf idence(Gs ) = (1)
max({N o. of graphs Gi G GD | Gi Gs })

max({N o. of graphs Gi G GD | Gi Gs }) is the maximum support of

any sub-graph of Gs . This prescribed clarity merely entails that gConf idence
is the smallest correlation of any graph, Gs and fall within the range [0, 1].
The measure gConf icence, has numerous crucial properties those made it ef-
ficient in mining graph correlation more eectively. These are the downward
closure property, null invariance, lower bound marker of correlation and co
occurrence observer.

Lemma 1. Given a graph database GD and one of its transaction graph G,

then we can define the gConf idence of Gs G derived in Equation 1 as

{N o. of Graphs G | Gs G GD}
gConf idence(Gs ) = (2)
max({ei E(Gs ), N o. of graphs Gj | ei E(Gj ), Gj GD})

Consider a scenario shown in Figure 1, where two frequent closed graphs found
from a set of graphs representing a group of people. Each graph in the set repre-
sents friend circle of an individual where nodes represent individuals and edges
represent interaction among individual pairs. The circles around closed graphs
represent the interaction of a group of people all together. Labels of edges and
circles represent the frequencies of the edges and closed graphs respectively. How-
ever, in mining the most correlated group, frequency of closed-frequent graphs
cant help due a tie i.e. 30 for each. As a consequence, Group G2 will be sug-
gested as most correlated by our proposed measure. Since, maximum interaction
30
of any pair in G1 is 100 and in G2 is 60. Therefore, gConf idence(G1 ) = 100 =
30
0.3 <gConf idence(G2) = 60 = 0.5.
 Correlation Mining in Graph Databases with a New Measure 91

Fig. 1. Motivating Scenario Fig. 2. gConfidence Tree : A search space

Definition 2. (Correlated Graph Mining) Given a graph database GD =

{G1 , G2 , ..., GN }, a user specified minimum support threshold, and a user
specified minimum correlation threshold, . We have to search for interesting
graph/subgraphs that is, we have to search for the set of graphs GI ={Gi |
supp(Gi ) ; gConf idence(Gi ) }. It means the problem is to search for
graphs having support count greater than or equal to and gConf idence value
greater than or equal to .

We have created a hierarchical tree like structure for eciently searching the
correlation among graphs within graph databases. The tree is defined as follows:
Definition 3. (gConfidence Tree) A tree, where each node represents a graph
or subgraph by storing corresponding DFS code and represents correlation by
storing gConf idence value gC. Moreover, the relation between parent node
and child node complies with the relation that a parent is one edge shorter in
size than its child and a child is one edge larger than its parent and no child
has gC greater than its parent. The relation between siblings is consistent with
the DFS lexicographic order. That is, the pre-order search of gConf idence tree
follows the DFS lexicographic order.

In Figure 2, we have shown a gConf idence code tree where the (n+1)-th level of
the tree has nodes which contain DFS codes of n-edge graphs and a value gC
which represents the inherent correlation among the nodes, edges and subgraphs
of that particular graph. Any node in the gConf idence tree contains a valid DFS
code. Certainly some of the nodes contain a minimum DFS code while others
do not. And there could be some nodes having gC values smaller than the
minimum correlation threshold. The value for gC also maintains a parent and
child relationship that is gC() gC() where = (a0 , a1 , ..., am ) and =
(a0 , a1 , ..., am , b) that is is s parent.
Now we will describe our algorithm for mining graph correlation using our
proposed measure gConf idence within graph databases. Since correlation is
searched based on user specified minimum correlation threshold then the search
space can be pruned based on two values, one for minimum support threshold
and another one is minimum confidence threshold. Proposed CGM is an edge
based correlation mining algorithm and we also used the concept of Projected
Database to reduce the costly searching operation for counting occurrences of
any graph/subgraph.
92 Md. Samiullah et al.

Fig. 3. gConfidence mining Illustration

For illustrating the working procedure of our CGM algorithm, we can consider
the graph database for the chemical dataset shown in Figure 3. We have assumed
the support threshold, = 13 and correlation threshold, = 23 . According to our
algorithm we have calculated support and gConf idence of each edge and vertex
and then selected the frequent and correlated vertices and edges. Now, we have
to construct single edge graph for each frequent-correlated edge and the edge set
is also used to construct GLOBAL EDGE M AT RIX. This matrix is sorted
based on support count and DFS code and can be used in a chronological order
for constructing potential children (smallest DFS code oriented child first) and
also helps in counting maximum elementary edge support count of a graph.
We have created a Null-rooted gConf idence tree and then started mining for
the first 1-edge graph C1 . Since it satisfies both the threshold values, we have
added it in the search space and start mining the correlation of its potential
children recursively. Its first child C1.1 is not frequent, hence we have pruned
it. The second child C1.2 is frequent and correlated, hence added to the search
space. However, recursive calculation of its children shows that C1.2.1 and C1.2.4
are frequent but not correlated and C1.2.2 and C1.2.3 are not frequent.
 Correlation Mining in Graph Databases with a New Measure 93

As a consequence, the fifth child of C1.2 that is C1.2.5 is found correlated and
added to gConf idence tree. The eighth child of C1.2.5 is found correlated and
added in the search space but first seven children of C1.2.5 are not frequent hence
will not be correlated and been pruned from the search space. Therefore, we have
to further search for the correlation of all possible children of C1.2.5.8 . Since no
children of C1.2.5.8 found frequent-correlated, we can backtrack to C1.2.5 . But
already we have checked all possible children of C1.2.5 hence we can trace back
again to C1.2 for checking correlation of its remaining children.
In this way we have calculated the correlation for the graph database of Figure
3 and found a complete gConf idence tree, where nodes of the tree contain
correlated graphs along with the amount of correlation. All the above discussed
steps are illustrated in Figure 3.

3 Experimental Results

We have performed a comprehensive performance study in our experiments on

both synthetic and real world datasets. We have used our own synthetic graph
generator to construct synthetic dataset. The real life datasets we have tested are
cancer dataset, found from [8], namely MOLT-4 and NCI-H23. On an average,
the real datasets contain about 40K graphs with average 25 nodes and 30 edges
along with an average of 17 distinct labels for vertices and edges. The synthetic
dataset can be identified by four properties, | D | representing numbers of graphs,
| N | indicating number of distinct labels for vertices and edges, | T | and | V |
for average graph size wrt. edges and vertices respectively.
In both types of data (real and synthetic) CGM algorithm is proved to be
sound and ecient as well as found scalable and faster enough that it can mine
correlation among various graphs with any size and any level of complexity in
comparison to CGS[7] and gSpan[3]. All experiments of CGM, CGS and gSpan
have been performed on a 2.1 GHz Intel(R) Core(TM) Duo PC with 1GB RAM,
running Windows 7 operating system, using C/C++ programming language. We
have kept the support threshold fixed at 5% and varied the correlation threshold
from 45% to 85% unless stated otherwise. In some cases, we have varied the size
of database by adding or removing graph transactions randomly from the actual
database.
In Figure 4 and Figure 5, we have shown the performance of our proposed
CGM algorithm for Processing time vs. Graph Density with varying Correlation
Threshold, when run on the MOLT-4 graph database and a synthetic database
characterized by D200kN 30T 80V 50 respectively. Since we are assessing the per-
formance of our algorithm against graph density, we have varied the size of the
graph database. It can be noticed that maximum 150 seconds were needed and
minimum required time was 25 seconds in mining real life database, where most
of the time processing completes within 1000 seconds for synthetic dataset with
any confidence threshold within range.
Figure 6 and Figure 7 contain the performance analysis of our proposed algo-
rithm for Scalability wrt Time with varying Data Size, when run on the NCI-H23
94 Md. Samiullah et al.

Fig. 4. Processing time wrt Graph Fig. 5. Processing time wrt Graph
Density on (MOLT-4) Density on (D200kN30T80v50)

Fig. 6. Scalability wrt processing Fig. 7. Scalability of CGM wrt processing

time on (NCI-H23) time for Synthetic Data(D200kN20T40v30)

Fig. 8. CGM vs CGS Fig. 9. CGM vs gSpan Fig. 10. Performance in Fil-
on (D200kN20T40v30) on (D200kN20T40v30) tering(%) on (MOLT-4)

graph database and a synthtic database characterized by D200kN 20T 40v30 re-
spectively. We have found that 50 to 100 seconds are required in mining NCI-H23
database and 300 to 1200 seconds are required for the synthetic dataset with any
confidence threshold within range.
To compare the performance of CGS with CGM, once again we have used the
denser synthetic dataset used in the scalability assessment earlier. However, the
support threshold is considered 5% and confidence threshold 50%. The compar-
ison is shown in the Figure 8, which illustrates a significant performance gain.
We have also provided the performance of our proposed algorithm in filtering
less significant graphs. By comparing it with the well known frequent subgraph
mining algorithm gSpan, using denser synthetic dataset, used in the scalability
assessment, we found our algorithm ecient enough in comparison with existing
algorithms. Here, we have considered the support threshold 5% and confidence
 Correlation Mining in Graph Databases with a New Measure 95

threshold 50%. The comparison can be found in Figure 9. Figure 10 contains the
percentage of graphs, those are un-correlated, filtered by CGM with respect to
the graphs selected by gSpan. Figure 10, shows that the filtering percentage of
CGM can be from 10% up to 40%, on a real life data set MOLT-4, for various
gConf idence threshold.

4 Conclusions
Mining frequent patterns or sub-patterns with larger support threshold could
miss some interesting patterns. At the same time if the threshold considered
is small enough to capture such rare but interesting items could generate lots
of spurious patterns. Therefore, association and correlation analysis is used in
association of frequent pattern mining for mining frequent-interesting patterns
from a collection of itemsets, but correlation searching is a challenging task. In
a graph database, correlation searching is more challenging due to the fact that
searching frequent subgraphs faces the graph isomorphism problem. Therefore, a
new measure gConf idence and an algorithm CGM are proposed for correlation
mining in graph databases which can capture more interesting inherent correla-
tion among graphs. gConf idence has downward closure property, which helps in
pruning descendants of non-correlated candidates. Proposed method constructs
a tree-like search space named gConf idence tree to eciently mine the corre-
lation. We have performed extensive performance analysis of CGM and found
it ecient enough which outperforms existing works in correlation search based
on speed. The proposed algorithm can be applied in both traditional and non-
traditional domains such as bio-informatics, computer vision, various networks,
machine learning, chemical domain and various other real life domains.

References
1. Inokuchi, A., Washio, T., Motoda, H.: An apriori-based algorithm for mining fre-

quent substructures from graph data. In: Zighed, D.A., Komorowski, J., Zytkow,
J.M. (eds.) PKDD 2000. LNCS (LNAI), vol. 1910, pp. 1323. Springer, Heidelberg
(2000)
2. Kuramochi, M., Karypis, G.: Frequent subgraph discovery. In: ICDM, pp. 313320.
IEEE Computer Society (2001)
3. Yan, X., Han, J.: gSpan: Graph-based substructure pattern mining. In: ICDM, pp.
721724. IEEE Computer Society (2002)
4. Li, J., Liu, Y., Gao, H.: Ecient algorithms for summarizing graph patterns. IEEE
Trans. Knowl. Data Eng. 23(9), 13881405 (2011)
5. Tan, P.N., Kumar, V., Srivastava, J.: Selecting the right interestingness measure for
association patterns. In: KDD, pp. 3241. ACM (2002)
6. Yan, X., Zhu, F., Yu, P.S., Han, J.: Feature-based similarity search in graph struc-
tures. ACM Trans. Database Syst. 31(4), 14181453 (2006)
7. Ke, Y., Cheng, J., Ng, W.: Ecient correlation search from graph databases. IEEE
Trans. Knowl. Data Eng. 20(12), 16011615 (2008)
8. Pubchem web site for information on biological activities of small molecules (2011),
http://pubchem.ncbi.nlm.nih.gov
 Improved Parallel Processing
of Massive De Bruijn Graph for Genome Assembly

Li Zeng1, Jiefeng Cheng1,, Jintao Meng1,2,3, Bingqiang Wang4,

and Shengzhong Feng1
1
Shenzhen Institutes of Advanced Technology, CAS, Shenzhen, 518055, P.R. China
2
Institute of Computing Technology, CAS, Beijing, 100190, P.R. China
3
Graduate University of Chinese Academy of Sciences, Beijing, 100049, China
{li.zeng,jf.cheng,jt.meng,sz.feng}@siat.ac.cn
4
Beijing Genomics Institute, Shenzhen, 518083, China
[email protected]

Abstract. De Bruijn graph is a vastly used technique for developing genome

assembly software nowadays. The scale of this kind of graph can reach billions of
vertices and edges which poses great challenges to the genome assembly task. It is
of great importance to study scalable genome assembly algorithms in order to
cope with this situation. Despite some recent works which begin to address the
scalability problem with parallel assembly algorithms, massive De Bruijn graph
processing is still very time consuming which needs optimized operations. In this
paper, we aim to significantly improve the efficiency of massive De Bruijn graph
processing. Specifically, the time consuming and memory intensive processing are
the De Bruijn graph construction phase and the simplification phase. We observe
that the existing list ranking approach repeatedly performs parallel global sorting
over all De Bruijn graph vertices, which results in a huge amount of
communications between computing nodes. Therefore, we propose to use depth-
first traversal over the underlying De Bruijn graph once to achieve the same
objective as the existing list ranking approach. The new method is fast, effective
and can be executed in parallel. It has a computing complexity of and
communication complexity of , which is smaller than the existing list
ranking approach, here g is the length of genome reference, p is the number of
processors. Our experimental results using error-free data show that, when the
number of processors scales from 8 to 128, our algorithm has a speedup of 10
times on processing simulated data of Yeast and C.elegans.

Keywords: Parallelized Graph Algorithm, De Bruijn Graph, Genome

Assembler.

1 Introduction

Whole Genome sequencing is one of the most fundamental problems in Biology.

Since 1977, when Sanger sequencing method appeared, tens of thousands of genome

Corresponding author.

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 96107, 2013.
Springer-Verlag Berlin Heidelberg 2013
 Improved Parallel Processsing of Massive De Bruijn Graph for Genome Assembly 97

sequence has been sequencced. Due to the diversity of species, there are still a laarge
number of species that neeed to be sequenced. Moreover, modern medical reseaarch
shows that most of the diiseases and gene have connections. Genome sequenccing
results can lead to reveal th
he mystery of genetic variation, and is now widely usedd in
genetic diagnosis, gene therrapy and drug design, etc.[4].
Whole Genome sequenccing has two main steps. The first step is to obtain the
sequence of short reads. It needs amplifying the DNA molecule (multiples of
amplification is referred to o as coverage) first, and then these DNA molecules are
randomly broken into many y small fragments (each small fragment is referred to aas a
read), the short read is then
n determined by sequencing devices, which is illustratedd in
Figure 1. The second step is to reconstruct genome sequence from these short reaads,
which is usually referred asa genome assembly. Genome assembly problem is proved
to be is NP-hard [1], reduceed from the Shortest Common Superstring(SCS) problem m.

break the DNA into short fragments randomly

read, sequenced by sequencing instrument

Fig. 1. DNA and its reads

There are two types of sequence

s assembly algorithms. one is called overlap grraph
model [2]. On an overlap grraph, each vertex corresponds to a read. There is a direccted
edge between two reads iff overlap exists between them and the overlapping lenngth
exceeds a certain threshold d. Therefore, sequence assembly problem is converted iinto
finding a Hamilton path th hrough each node in the overlap graph, which is NP-hhard
[2]. Another model is based d on De Bruijn graph [3]. In a De Bruijn graph, each reaad is
cut into small fragments off length k, called k-mer. Each k-mer contributes a nodee in
the graph. If there are k-1 characters
c overlapping between two adjacent k-mers, theen a
directed edge exists betweeen the two corresponding nodes. In this way, each read can
produce a corresponding pathp in the graph. So the sequence assembly problem has
been transformed into findiing the shortest tour of the De Bruijn graph that includee all
the read path [4]. The pro oblem is extremely difficult because repeat fragmentss of
various lengths exist in the t original sequence. Also, sequencing machines w will
introduce errors into thesee reads, where the error rate is about 1%~3% in moddern
sequencing machines. Th hese two issues make the sequence assembly m more
complicated.
98 L. Zeng et al.

Many optimization criteria and heuristics have been presented to deal with
assembly problem. Early, the length of fragments obtained by the Sanger sequencing
method can be achieved 200bp (base pairs). Overlap graph model algorithm is more
effective, but the cost of sequencing is relatively high. For example, the Human
Genome Project, by the completion of the first-generation sequencing technology,
already spent $3 billion in 3 years. More recently, genome sequencing entered in a era
of large-scale applications due to the second-generation sequencing technology (also
known as the next-generation sequencing technology), such as Solexa, 454, SOLID
[5]. There are three prominent features of the second-generation sequencing
technology, namely, high throughput, short sequence and high coverage. For high-
throughput sequencing, a sequencing machine can simultaneously sequence millions
of reads, which greatly reduces the whole cost. For short sequence, it stands for that
the sequence length is between 25-80 bp in general. In this regard, genome assembly
has to greatly increase DNA coverage [5] to ensure the integrity of the information.
With the increasing of the coverage, the total number of reads also increases greatly.
For overlap graph, then the size of the graph will significantly increase accordingly.
But for De Bruijn graph, the size of the graph is only linear to the length of the
reference genome. Therefore, in the face of the second generation of gene sequencing
technology, De Bruijn graph model has a great advantage and become much popular.
The proposed algorithms based n De Bruijn graph includes Euler[3], ALLPATHS[6],
Velvet[7], IDBA[8], SOAPdenovo[9], ABySS[10] and YAGA[11]. Euler,
ALLPATHS, Velvet and IDBA algorithms are serial algorithms, only suitable for
small data sets. SOAPdenovo is multi-threaded SMP mainframe-based algorithm,
which can process a large data set. But it also spends more than 40 hours on human
genome [9] and the required infrastructure is very expensive.
The main phases of genome assembly based on De Bruijn graph is as follows: First,
De Bruijn graph construction. Second, errors correction. To remove the known
erroneous structures in the De Bruijn graph. Third, graph simplification. It is to
compress a single chain in the De Bruijn graph into a single compact node, which also
generates preliminary contig information. Fourth, generating scaffold. It is to merge
contigs from the previous step with pair-end information obtained by sequencing,
then to derive longer contigs (known as scaffold). Fifth, to output the obtained
scaffold. Generally, the first phase has the largest memory consumption, and the
second phase and the third phase has the largest time expense. Initially, the graph
after constructed is extremely sparse, chain nodes ratio is more than 99% [12].
YAGA[11] is a distributed algorithm with well parallelism and scalability which
represent state of the art. In this paper, we improve the parallel processing of De
Bruijn graph by constructing a De Bruijn graph in distributed manner for the
construction phase and then a parallel graph traversal to obtain all chains for the
simplification phase. After assuming that the error has been removed, our work
focuses on improving the efficiency of parallel graph simplification. The main
contribution of this paper is to formulate the graph simplification problem into graph
traversal problem. In YAGA algorithm, Graph simplification is approached as a list
 Improved Parallel Processing of Massive De Bruijn Graph for Genome Assembly 99

Fig. 2. k-mer and k-molecule

ranking problem with a total of four times global sorting and a multi-round recursion.
Our algorithm directly traverses from end node to access all chains and visit each
node only once. Our algorithm has a computing complexity of and
communication complexity of , which is smaller than YAGA algorithm, here
g is the length of genome reference, p is the number of processors.
Section 2 gives the definition of De Bruijn and the detailed description on graph
construction. Section 3 describes parallel graph simplification. We show experimental
results in section 4. Section 5 concludes this paper.

2 Parallel De Bruijn Graph Construction

We first explain the definition of De Bruijn graph. Then, in order to achieve efficient
parallel De Bruijn graph construction, we discuss a compact representation of De
Bruijn graph. Finally, a parallel De Bruijn graph construction is given, which
distributes the storage of the graph across multiple computational nodes.

2.1 Defining De Bruijn Graph

Let s be a DNA sequence of length n, which is a sequence consists of the four bases,
namely A C T and G. A substring derived from s with length k, represented as
, is called as a k-mer. The set of all k-mers
derived from s is referred to as the k-spectrum of s. For a k-mer represents the
reverse sequence of s complementary sequence (for example , then
). can be obtained by first complementing each character in alpha and
then reversing and the resulted sequence. The complementary rules are
.
A k-molecule represents a pair of k-mers { }, and and is reverse
complementary. The notation is the ordering relation between the two k-mers with
length k, such that indicates that is larger than lexicographically. We
designate the lexicographically larger one as the positive k-mer denoted , and
the other one as negative k-mer denoted (as shown in Figure 2), here there is
. When the context is clear, stands for a k-molecule , that is
. Fig.2 depicts the relationship between a k-mer and its
corresponding k-molecule with an example.
100 L. Zeng et al.

Fig. 3. k-molecule from read

Fig. 4. The structure of De Bruijn graph node

2.2 Structure of De Bruijn Graph

The node structure in our De Bruijn graph is different from the existing one of YAGA
algorithm [11]. Specifically, we design a compacted representation of De Bruijn
graph, where an ID field and an arc field are devised for every node. And this node
structure not only uniquely identifies a node by the ID field, it also encodes all
incident edges for that node in the arc field. Therefore, our De Bruijn graph
representation does not store edges explicitly.
Lets consider a De Bruijn graph node, which is a k-molecule (as shown in Figure 3).
We compute a 64 bit unsigned integer based on the sequence of all characters in the
positive k-mer using the rule {A: 00, C: 01 G: 10 T: 11} (shown in Figure 4). In turn,
the 64 bit unsigned integer forms the ID field (or simply ID) of the node. Since we have
a maximum value for k as 31, therefore, all characters in the positive k-mer can at most
encode the low 2k bits of the ID. And the remaining high bits are filled with 0.
All edges are stored in the arc field of the node. Because DNA consists of four
bases {A, C, G, T}, a node can connect to eight adjacent nodes at most, corresponding
to eight edges. Since the two adjacent nodes have k-1 characters overlapped in their
 Improved Parallel Processing of Massive De Bruijn Graph for Genome Assembly 101

Fig. 5. Obtaining all k-mers

k-mers, positive or negative, only one different character can be found over the two
adjacent nodes. So we can simply use this character to represent the corresponding
edge. We use an 8-bit char type variable to store these edges. The bit with value of 1
indicates that the existence of this edge, otherwise, it means there is no corresponding
edge. The rule is: the top 4 bits represent the edges of positive k-mer, 0 to 3
corresponding to the A C G T; the after 4 bits represent the edges of negative k-mer, 4
to 7 corresponding to A C G T. For example, the node TAG owns three edges
connected to three nodes (TTA AGT and TAG), as defined above, we use the positive
k-mer TAG to represent the node. We set the third bit of arc field to 1. It means that
an edge exist from the positive k-mer(TAG) to the negative k-mer(AGG) of node
CCT, for the number of the bit is the third and in the top 4. The edges record of node
TAG is shown in Figure 4. The bits of the shaded part in Figure 4 with value of 1
represent the existence of these edges. Accordingly, only 9 bytes is used to store one
node in our De Bruijn graph.

2.3 Parallel De Bruijn Graph Construction Algorithm

The parallel De Bruijn graph construction algorithm has three main steps. The first
step is to read all short sequences in parallel; the second step is to obtain all k-mers
and the edges between any two consecutive k-mers, and to obtain all k-molecules
based on those k-mers, repeated edges in a k-molecule are merged. The third step to
102 L. Zeng et al.

Fig. 6. Three types of erroneous connections

distribute those k-molecules in a number of computational nodes. The details are

given below.
Step 1: Read short sequences in parallel.
Each processor is assigned with different blocks of short sequences according to
the rank of processor and the file size of short sequences. Assume that the length
of the file is L and there are n processors, then the i-th processor get short
sequences from the to .
Step 2: Parallel segmentation of k-mers (see Figure 5).
Each processor cuts all short reads into k-mers by a sliding window with length k
from the first k-mer to the last k-mer. Then, each k-mer is handled as follows:
a) By the rules of complementation, we get all reverse complementary k-mers
of the existing k-mers. Moreover, we obtain all k-molecule based
on those k-mers. The lexicographically larger k-mer is encoded by the rule
{A: 00, C: 01 G: 10 T: 11} as the ID of the k-molecule.
b) Record edges between adjacent k-molecules. Form the compact De Bruijn
graph node representation for all k-molecules.
Step 3: Distribute all k-molecules.
Each processor stores a partition of the De Bruijn graph. All processors are
assigned with a certain portion of De Bruijn graph nodes with a hash function
over the 64-bit node representation. Specifically, the location of a De Bruijn
graph is computed by taking the ID of the k-molecule mod of the number of
processors. So, all k-molecule are mapped to different processors by an all-to-all
communication operation. After this step, any processor can computed a node
directly by the same hash function. This is to boost the performance of the
following graph simplification phase.

3 Parallel Simplification of De Bruijn Graph

In this section, we first explain the problem of De Bruijn simplification. Then, we

discuss our parallel De Bruijn graph simplification algorithm based on concurrent
graph traversal starting from multiple sources in the graph.
 Improved Parallel Processing of Massive De Bruijn Graph for Genome Assembly 103

Fig. 7. De Bruijn graph in our algorithm

3.1 The De Bruijn Graph Simplification Problem

There can be a lot of sequencing errors introduced by gene sequencing devices.

Therefore, the De Bruijn graph obtained by the construction algorithm can have a
large number of erroneous paths, which give rise to three types of erroneous
connections in the graph, namely, tip, bubble and spurious link, which are depicted in
Figure 6. Therefore, after the De Bruijn graph construction, the immediate processing
is to remove those erroneous connections using the knowledge from Biology study
[12]. After this, the De Bruijn graph is error-free and very sparse such that it contains
a large number of chains (as shown in Figure 7) [13]. Therefore, an important
processing is to simplify the De Bruijn graph by finding all chains in the graph. After
simplification, the size of the graph can be reduced sharply, almost about 90% off and
the graph can be processed for latter tasks such as processing branches and repeats in
a single machine easily.
Since our focus in this paper is an efficient and effective parallel De Bruijn graph
simplification algorithm, we assume the processing to remove erroneous connections
is completed. So, in the following of this paper, our discussion is on the De Bruijn
graph without erroneous connections, that is, the error-free data. First, we give a few
notations and a detailed description of the problem. We call a node with a degree of 1
as the 1 degree node. A node with degree of 2, by an incoming edge and an outgoing
edge, is called as the chain node. The chain node can lead to a chain. all nodes with a
degree of 2, by two incoming edges or two outgoing edges, as well as all nodes with
the degree larger than 3 is bifurcation node. A chain is a simple path in the graph
containing chain nodes only. Nodes at both ends of the chain are called as end node.
Although an end node is not part of the chain, the chain can be obtained by traversal
from it.
In Figure 7, the green nodes are chain nodes, which are the target of simplification.
The red node is bifurcation node and the yellow is 1 degree node, both of them are
end node. Although the degree of the circled node in Figure 7 is 2, it cannot be
traversed for it has two incoming edges or two outgoing edges.
104 L. Zeng et al.

Fig. 8. The merging thread and the communication thread

Fig. 9. Traversing De Bruijn graph

3.2 Parallel De Bruijn Graph Simplification Algorithm

Overview. Our algorithm directly traverses from an end node of a chain to the other
end node to simplify the chain.
Therefore, we use two threads for each processor, which collectively performs the
concurrent graph traversal to identify all chains. As shown in Figure 8, thread A is
responsible for the simplification task as to merge all chain nodes identified over a
chain, on the other hand, thread B is dedicated for inter-process communications. The
communication is for resolving local requests and dealing with other processors
requests on traversing end nodes.
Data Structure. Each processor has two Maps. One is locationMap, which store
nodes before simplification. Another is subGraphMap which stores nodes after
simplification. During initialization, the subGraphMap is empty and the
subGraphMap is inserted with nodes during the simplification by the process of
identifying chains.
 Improved Parallel Processing of Massive De Bruijn Graph for Genome Assembly 105

Algorithm. Based on the above data structure, we outline the algorithm as below:
Step 1: Initialization: De Bruijn graph is distributed on all processor and each
processor has a locationMap and a subGraphMap.
Step 2: Traversal of the graph (as shown in Figure 9): Every processor starting to
traverse from an end node in local locationMap. Since a chain has two ends
nodes, traversal can be divided into the following three cases: 1) a processor
visits a same chain twice from the two end nodes; 2) two different processors
visit the same chain twice, each one from a different end node; 3) two processors
visit the same chain simultaneously by starting from the two end node at almost
the same time. In order to reduce repeated access to the same chain, we add a
endNodeID variable to store the ID of end node in the direction of the end.
According to the ID number, the merging thread can determine whether to
continue or exit.
Step 3: Simplification: Delete all chain nodes and append their information to the
corresponding end node, and then insert them into subGraphMap.
Step 4: Select another end node to repeat the processing from Step 2 until no end
node can be found.

3.3 Analysis

YAGA algorithm transforms the problem of simplifying all chains into list ranking
problem which requires four times of global sorting over all graph nodes. In this
procedure, a large number of nodes are moved four times. The time and
communication cost is very large. The computational complexity of four times global
sorting is , and the communication complexity is . In addition to
the cost of four times of global sort, the cost of other involved processing is also very
large.
In our algorithm, each processor directly traverses from the local end node to get all
chains. A chain only holds two end nodes, so each chain node is accessed twice at
most. Also our algorithm successfully prevents a chain from being visited twice. So,
in the worst case, all nodes are calculated twice, the time complexity is .And
each node is requested twice at most, the communication complexity is .

4 Performance Evaluation

Our assembly software is written in C++ and MPI. The experimental platforms
includes 10 servers interconnected by InfiniBand network, which is configured as 16-
core, 32G shared memory. We choose Yeast and C.elegans genomes, using Perl
scripts automatically generated yeast test data with 17, 007, 362 reads and C.elegans
test data, which consists of 140, 396, 108 reads. The length of reads ranges from 36bp
to 50bp with error rate of 0, and the coverage for both data is 50X. Our goal is to
demonstrate the scalability of our assembler to graph construction and graph
compaction on a large distributed De Bruijn graph.
106 L. Zeng et al.

Fig. 10. The time used in three phases on Yeast dataset

Fig. 11. The time used in three phases on C.elegans dataset

Firstly, we tested assembler on the Yeast data set. The runtime is displayed in
Figure 10 and the time is divided into three phrases. First, parallel read file from a
distributed file system. Second, graph construction. Third, graph compaction. The
most part time on the third phase is network transmission cost. But, in our algorithms
most nodes only be moved once, that is an efficient optimization. Figure 10 indicates
that the algorithm has good scalability. When the number of processors is increased
from 8 to 128, the total time is reduced from 490s to 63s that means the speedup is
about 8 times.
The C.elegans dataset is 10 times larger than Yeast dataset. The running time is
shown in Figure 11. When the number of processors is increased from 8 to 128, the
total time is reduced from 3844s to 375s with the speedup being 10 times.
 Improved Parallel Processing of Massive De Bruijn Graph for Genome Assembly 107

5 Conclusion

We propose to use depth-first traversal over the underlying De Bruijn graph once
to simplify it. The new method is fast, effective and can be executed in parallel.
By testing the two data sets, the experimental results show that the algorithm has great
scalability. After the completion of the graph simplification, the scale of the graph is
sharply reduced, thus the graph simplification is of great value. The latter stage
includes resolving branches and repeats using pair-end information in a single
machine. We will study these problems in our future work.

6 Conclusion

This work is supported by NSFC of China (Grant No. 61103049) and Shenzhen
Internet Industry Development Fund (Grant No. JC201005270342A). The authors
would like to thank the anonymous reviewers for their helpful comments.

References
[1] Kundeti, V.K., Rajasekaran, S., Dinh, H., et al.: Efficient parallel and out of core
algorithms for constructing large bi-directed de Bruijn graphs. BMC Bioinformatics
11(560) (2010)
[2] Kececioglu, J.D., Myers, E.W.: Combinatorial algorithms for DNA sequence assembly.
Algorithmica 13(1), 751 (1995)
[3] Pevzner, P.A., Tang, H., Waterman, M.S.: An Eulerian path approach to DNA fragment
assembly. Proceedings of the National Academy of Sciences of the United States of
America 98(17), 97489753 (2001)
[4] Medvedev, P., Georgiou, K., Myers, G., Brudno, M.: Computability of models for
sequence assembly. In: Giancarlo, R., Hannenhalli, S. (eds.) WABI 2007. LNCS
(LNBI), vol. 4645, pp. 289301. Springer, Heidelberg (2007)
[5] Jackson, B.G., Aluru, S.: Parallel Construction of Bidirected String Graphs for Genome
Assembly, 346353 (2008)
[6] Butler, J., Maccallum, I., Kleber, M., et al.: ALLPATHS: de novo assembly of whole-
genome shotgun microreads. Genome Res. 18(5), 810820 (2008)
[7] Zerbino, D.R., Birney, E.: Velvet: algorithms for de novo short read assembly using de
Bruijn graphs. Genome Res. 18(5), 821829 (2008)
[8] Peng, Y., Leung, H.C.M., Yiu, S.M., Chin, F.Y.L.: IDBAA Practical Iterative de
Bruijn Graph De Novo Assembler. In: Berger, B. (ed.) RECOMB 2010. LNCS,
vol. 6044, pp. 426440. Springer, Heidelberg (2010)
[9] Li, R., Zhu, H., Ruan, J., et al.: De novo assembly of human genomes with massively
parallel short read sequencing. Genome Res. 20(2), 265272 (2010)
[10] Simpson, J.T., Wong, K., Jackman, S.D., et al.: ABySS: a parallel assembler for short
read sequence data. Genome Res. 19(6), 11171123 (2009)
[11] Jackson, B., Regennitter, M., Yang, X., et al.: Parallel de novo assembly of large
genomes from high-throughput short reads. IEEE (2010)
 B3Clustering: Identifying Protein Complexes
from Protein-Protein Interaction Network

Eunjung Chin and Jia Zhu

School of ITEE, The University of Queensland, Australia

{eunjung,jiazhu}@itee.uq.edu.au

Abstract. Cluster analysis is one of most important challenges for data mining
in the modern Biology. The advance of experimental technologies have produced
large amount of binary protein-protein interaction data, but it is hard to find pro-
tein complexes in vitro.We introduce new algorithm called B3Clustering which
detects densely connected subgraphs from the complicated and noisy graph.
B3Clustering finds clusters by adjusting the density of subgraphs to be flexi-
ble according to its size, because the more vertices the cluster has, the less dense
it becomes. B3Clustering bisects the paths with distance of 3 into two groups
to select vertices from each group.We experiment B3Clustering and two other
clustering methods in three different PPI networks. Then, we compare the re-
sultant clusters from each method with benchmark complexes called CYC2008.
The experimental result supports the efficiency and robustness of B3Clustering
for protein complex prediction in PPI networks.

1 Introduction
Protein complexes play essential roles in biological process and they are also important
to discover drugs in pharmaceutical process. Recent progress in experimental technolo-
gies has led to an unprecedented growth in binary protein-protein interaction (PPI) data.
Despite this accumulation of PPI data, considerably small amount of protein complexes
are identified in vitro. Thus, computational approaches are used to overcome the tech-
nological limit for detecting protein complexes. It becomes one of the most important
challenges for modern biology to identify protein complexes from binary PPI data. Pro-
teins are represented as vertices and their interactions are represented as edges in the
PPI network. Since protein complexes are shown as densely connected subgraphs in
PPI network, clustering methods are applied to find them from the network.
For example, some used clique finding algorithms to predict protein complexes from
PPI network (Spirin and Mirny [23] and Liu et al.[18]). They devised their own methods
to merge overlapping cliques.
Besides, Van Dongen (2000)[27] introduced MCL (Markov Clustering) as graph par-
titioning method by simulating random walks. It used two operators called expansion
and inflation, which boosts strong connections and demotes weak connections. Brohee
et al. (2006)[6] shows the robustness of MCL with comparison to three other clustering
algorithms (RNSC[12], SPC[5], MCODE[4]) for detecting protein complexes.
One of the recent emerging techniques is to use protein core attachments methods,
which identify protein-complex cores and add attachments into these cores to form pro-
tein complexes (CORE[15] and COACH[28]). Li et al.(2010)[17] shows that COACH

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 108119, 2013.
c Springer-Verlag Berlin Heidelberg 2013

 Identifying Protein Complexes from Protein-Protein Interaction Network 109

Fig. 1. The graph comes from the Biogrid database using Cytoscape

performs better than seven other clustering algorithms (MCODE[4], RNSC[12],

MCL[27], DPClus[3], CFinder[1], DECAFF[16], and CORE[15]) when they are ap-
plied to two PPI datasets (DIP[30] and Krogan et al.[13] data).
PPI data has increased by about forty times for the last six years from 50,726(July
2006) to 2,013,176(April 2012) in Biogrid database[25] for Sacchromyces Cerevisiae
called bakers yeast, while number of yeast proteins has grown by just 1.2 times from
5,300 to 6,226 for the same period. It means that PPI networks become more com-
plicated and tightly interconnected. For example, DIP and Krogan data have average
degrees of 6.98 and 7.86, respectively, but Biogrid 3.1.85 has an average degree of
62.22 for Sacchromyces Cerevisiae.
Even though the clustering methods which we mentioned before prove their compe-
tency in the small size of PPI data (DIP and Krogan data), they show poor performance
in the large size of tightly interconnected PPI network (Biogrid 3.1.85). The existing
clustering methods are not suitable to detect protein complexes from the tightly inter-
connected network such as Biogrid 3.1.85. Normally, they has been experimented with
the yeast datasets produced before 2006.
Since Biogrid 3.1.85 is densely connected and hard to partition as shown in Fig.1,
the existing algorithms do not perform very well. MCL has low F-measure of 0.0261
by producing 52 clusters for Biogrid 3.1.85 and one of the clusters has included 5,786
proteins out of 6,194 total proteins in PPI network. COACH has predicted a large num-
ber of clusters as many as 3,138 and it also shows low F-measure of 0.055 for Biogrid
3.1.85 Scchromyces Cerevisiae.
To detect clusters in such a tightly interconnected network, we introduce new algo-
rithm called B3Clustering which finds densely connected subgraphs from tightly inter-
connected PPI network. B3Clustering tries to get protein complexes from each pair of
proteins by bisecting vertices of paths with distance of 3 into two groups and calculat-
ing the probability of interaction between them. It chooses vertices with the probability
of interaction and check how much two groups have common vertices. If there remain
vertices in each group and two groups overlap more than the given threshold, then we
identify the union of two groups as a complex.
110 E. Chin and J. Zhu

We evaluated B3Clustering using three PPI datasets and compared the results with
MCL and COACH. We used CYC2008[22] as a benchmark of protein complexes.
B3Clustering shows capability to identify protein complexes in the tightly intercon-
nected PPI network as well as partially missing PPI network. We found that it achieved
higher F-measure than MCL and COACH.
All the proteins that we mentioned in this paper are those of Sacchromyces Cere-
visiae, so to speak, a bakers yeast. We introduce the basic concepts and explain the
methods which B3Clustering uses in the next section. In section 3, we give full details
of PPI data and complex benchmark data which we use in this paper. In section 4, we
explain the evaluation methods we hire here. We show the results of our experiments in
section 5. In the last section, we summarize our project and suggest future improvement
of B3Clustering.

2 Method
Protein-protein interaction(PPI) data come in the form of connections between proteins,
which is easily described as a graph model. Proteins are represented as vertices and their
interactions are represented as edges in the graph. Since protein complexes are shown
as densely connected subgraphs in PPI network, clustering methods are applied to find
them from the PPI network. We introduce B3Clustering to identify clusters in the graph
and we apply it to detect protein complexes in the PPI network.
B3Clustering proceeds clustering in three steps. Firstly, it selects vertices and we call
it vertex selection step. Secondly, it decides if the vertices can form clusters or not and
it is called cluster selection step. Lastly, it merges clusters, which is called trimming
step. We introduce terminology we use before we explain three steps.

2.1 Terminology
For graph G = (V, E), V (G) and E(G) denote the sets of vertices and edges of G, re-
spectively. We assume that graph G is an undirected and simple graph. One of important
objects in B3Clustering is a path with distance of 3. A path in a graph is a sequence of
vertices which are connected to the next vertex in the sequence. We use paths with
distinct vertices in B3Clustering.

Definition 1 (3-Path Reachable). Let a graph G on a vertex set V (G) be an undirected

and simple graph with an edge set E(G). Given two vertices u, v V (G), we say that
u is 3-path reachable to v if there exists a path with two distinct internal vertices and
three different edges between u and v. We also say that the path is 3-path reachable
from u to v.

That is, u is 3-path reachable to v if there exists a path whose distance is 3.

Definition 2 (3-Path Set). Path3(u, v) is defined as the set of paths 3-path reachable
from u to v.
Path3 (u,v)={ {u,i, j,v}|(u,i),(i, j),( j,v)E(G),u=i= j=v, for i, jV (G)}
 Identifying Protein Complexes from Protein-Protein Interaction Network 111

We assume that u, i, j, v are member of V (G) and they should be different from each
other. Path3(u, v) could be empty set if u is not 3-path reachable to v. Otherwise, it will
have paths with distance of 3 as members.

Definition 3 (Bi-sets). Du (u, v) is defined as the set of u, v, and vertices which are
linked to u in paths Path3(u, v). Dv (u, v) is defined as the set of u, v, and vertices
which are linked to v in paths Path3(u, v). Du (u, v) and Dv (u, v) are called Bi-sets of
Path3(u, v).
Du (u,v)={ u,v,i|(u,i)E(G),{u,v,i}z, for any zPath3 (u,v) }

Dv (u,v)={ u,v, j|( j,v)E(G),{u,v, j}z, for any zPath3 (u,v) }

We extracts two sets of vertices from 3-path set and call them Du (u, v) and Dv (u, v).
Each set includes u and v. Additionally, Du (u, v) includes the set of vertices which are
linked to u in paths Path3(u, v) and Dv (u, v) includes the set of vertices which are
linked to v in paths Path3(u, v).

2.2 Vertex Selection Step

B3Clustering detects clusters from every pair of vertices in PPI network. It has three
steps to proceed clustering. The first step is vertex selection. We assume that a vertex

YDR397C

YGR047C YDR087C
YBR123C

YOR110W

YLP007C
YDR362C

YAL001C

YDL097C

Fig. 2. Sample subgraph from krogan data

pair (u, v) is given. We can get Du (u, v) and Dv (u, v) from Definition 3. Then, we calcu-
late the probability to interact with members of other set for each vertex. We define the
probability of interaction pu (i) and pv ( j) as follows.

jDv (u,v) w(i, j)

pu (i) =
|Dv (u, v)|

iDu (u,v) w(i, j)

pv ( j) =
|Du (u, v)|
112 E. Chin and J. Zhu

w(i, j) is an edge weight between i and j. If there exists an edge between i and j, then
w(i, j) should be 1. If i is the same vertex with j, then w(i, j) should 1. Otherwise,
w(i, j) should be 0. !
1, if (i, j) E(G) or i = j;
w(i, j) =
0, otherwise.
We calculate pu (i) for every vertex i Du (u, v) and we calculate pv ( j) for every vertex
j Dv (u, v). If pu (i) is higher than , then i remains in Du (u, v). Otherwise, we remove i
from Du (u, v). Likewise, if pv ( j) is higher than , then v remains in Dv (u, v). Otherwise,
we remove j from Dv (u, v).
is the threshold which decides lowest margin for probability of interaction. As the
protein complex consist of more proteins, they show lower connectivity, Therefore, we
set a lower value when there are more paths with distance of 3, and vice versa.
|4Path3 (u,v)|
= 2 m

where m = max i, jV (G) |Path3 (u,v)| and is given. We set = 0.7 in this paper.
Lets see Fig.2 as a example. If we start from (Y OR110W,Y DR362C), then we get
bi-sets as shown Table 1. From now on, we call DY OR110W (Y OR110W,Y DR362C) D1 and
DY DR362C (Y OR110W,Y DR362C) D2 in short.

Table 1. Bi-sets of YOR110W and YDR362C before removing vertices

D1 YOR110W, YDR362C, YAL001C,

YBR123C, YGR047C, YPL007C,
YDR087C
D2 YOR110W, YDR362C, YAL001C,
YBR123C, YGR047C, YDR397C,
YDL097C

If we calculate the probability of interaction for each vertex in D1 and D2, each
vertex has a value in Table 2.
If We set that is equal to 0.5, then YDR087C is removed from D1 and YDL397C
and YDL097C are removed from D2. Finally, we get the D1 and D2 as shown in Table 3.

2.3 Cluster Selection Step

After we get D1 and D2 after removing vertices with lower probability of interaction
value than , we decide whether the union of D1 and D2 can be a cluster or not. We use
similarity value s(D1, D2) between D1 and D2 and define s(D1, D2) as follows.
| D1 D2 |
s(D1, D2) =
| D1 D2 |
If s(D1, D2) is higher than , D1 D2 is selected as a cluster. We set that is equal to
0.6 in this paper.
 Identifying Protein Complexes from Protein-Protein Interaction Network 113

Table 2. Probability of interaction

pYOR110W pY DR362C
p1(YOR110W)=1 p2(YOR110W)=0.714
p1(YDR362C)=0.714 p2(YDR362C)=1
p1(YAL001C)=0.857 p2(YAL001C)=0.857
p1(YBR123C)=0.857 p2(YBR123C)=0.857
p1(YGR047C)=0.714 p2(YGR047C)=1
p1(YPL007C)=0.571 p2(YDR397C)=0.281
p1(YDR087C)=0.281 p2(YDL097C)=0.281

Table 3. Bi-sets of YOR110W and YDR362C after removing vertices

D1 YOR110W, YDR362C, YAL001C,

YBR123C, YGR047C, YPL007C
D2 YOR110W, YDR362C, YAL001C,
YBR123C, YGR047C

Lets look at example in Table 3. Then the union of D1 and D2 has 6 members. Since
D1 and D2 has 5 common vertices, s(Y OR110W,Y DR362C) = 0.833. Hence, D1 D2
is considered as a cluster because it has higher similarity value than .

2.4 Merging Step

Since B3Clustering calculates similarity value from every pair of vertices, it can pro-
duce a large number of overlapping clusters. Therefore, we need to merge similar clus-
ters together. We calculate the overlapping ratio between clusters and merge similar
clusters. We define overlapping ratio l(C1,C2)between cluster C1 and cluster C2 as
follows.
| C1 C2 |
l(C1,C2) =
| C1 C2 |
where is a marginal overlapping threshold. We set = 0.9 in this paper.
After merging step, we get final clusters.

3 Evaluation
3.1 Data
We use three PPI datasets for Saccharomyces cerevisiae. One is a dataset from Krogan
et al.(2006)[25] and another comes from Dip database[30]. The last dataset is biogrid
3.1.85 downloaded from BIOGRID[25] database 1 .
1 http://www.thebiogrid.org/
114 E. Chin and J. Zhu

Start

PPI data  ,

Make a graph G=(V,E)

Get Path3(u,v)

Get Du(u,v) and Dv(u,v)

Select vertices in Select vertices in

Du(u,v) Dv(u,v)

Get C=Du(u,v) U Dv(u,v)

Get next vertex pair (u,v)

Get l(u,v)

Yes l(u,v) >

Add C to cluster set

No

More (u,v) ? Yes

No

Merge clusters

Display results

End

Fig. 3. Overall algorithm of B3Clustering

Krogan and dip datasets are used by Li et al.(2010) to evaluate the performance of
several clustering algorithms[17]. As shown in Table 4, Krogan and dip dataset has
similar number of average degree, but biogrid has much higher average degree than
them. Biogrid has about ten times higher average degree then dip dataset.
PPI data have a high rate of false positives, which has been estimated to be about 50%
[24]. The noise of the data disturbs clustering methods to detect protein complexes from
PPI data.
We use CYC2008 complexes as a benchmark set. CYC2008 is published by Pu et
al.(2008)[22] and it can be downloaded from (http://wodaklab.org/cyc2008/). CYC2008
provides a comprehensive catelogue of manually curated 408 protein complexes in Sac-
charoyces cerevisiae.
CYC2008 complexes consists of 1627 proteins, which is about one quarter of pro-
teins in Biogrid 3.1.85. Among the 408 protein complexes, 259 protein complexes are
composed of less than four proteins with 97% of connectivity. As the protein complex
consist of more proteins, they have lower connectivity.
There are 172 protein complexes consisting of two proteins out of 408 total protein
complexes in CYC2008. Two-member complexes exist as a form of edge in a graph so
that they should be chosen from edges in PPI networks. For example, we try to detect
172 protein complexes from 193,679 edges in Biogrid 3.1.85.
 Identifying Protein Complexes from Protein-Protein Interaction Network 115

Table 4. Features of PPI datasets

Dataset # Nodes # Edges Average degree

Biogrid(3.1.85) 6194 193976 62.22
Dip 4930 17201 6.98
Krogan et al. 3181 14076 7.86

3.2 F-Measure
Precision, recall, and F-measure have been used to evaluate the performance in infor-
mation retrieval and data mining. We calculate affinity score used by Li et al.[17] to get
F-measure. The neighbourhood affinity score NA(p,b) is defined as follows.

|Vp Vb |2
NA(p,b) =
|Vp | |Vb |

where P = (Vp , E p ) is a predicted complex and B = (Vb , Eb ) is a benchmark complex.

Precision. Precision is calculated as follows.

Ncp
Precision =
|P|

where Ncp =|{p|pP,NA(p,b), forbB}|.

Recall. Recall is calculated as follows.

Ncb
Recall =
|B|

where Ncb =|{b|bB,NA(p,b), for pP}|.

F-measure. F-measure is the geometric mean of Precision and recall as follows.

F-measure = Precision Recall

4 Result
We experiment B3Clustering with three datasets. Three databases are Biogrid 3.1.85,
Dip, and Krogan as describes in section 3. We experiment MCL and COACH with the
same datasets to compare the performance of B3Clustering. We set to be 0.7, to be
0.6, and to be 0.9.
116 E. Chin and J. Zhu

Table 5. Result from Krogan et al. dataset

Algorithms # Clusters Ncp Ncb

MCL 624 135 159
COACH 570 269 155
B3Clustering 629 398 144

Table 6. Result of Dip dataset

Algorithms # Clusters Ncp Ncb

MCL 835 171 186
COACH 746 311 215
B3Clustering 786 306 211

From the results, B3Clustering has the highest F-measure in Krogan data and Biogrid
data, but it has lower F-measure than COACH in Dip data. COACH and B3Clustering
score F-measure twice higher than MCL in Krogan and Dip datasets as shown in Fig.
4. MCL supposes to have the lowest F-measure because it partitions PPI graph. On
the contrary, COACH and B3Clustering consider some of vertices and overlap clusters
instead of partitioning.
For Biogrid 3.1.85 data, MCL has a F-measure of 0.0261 by producing 52 clusters,
one of which includes 5,786 proteins out of 6,194 total proteins. COACH predicts a
large number of clusters as many as 3,138 and only 105 are similar to protein complexes
in CYC2008. Since COACH has low precision for Biogrid 3.1.85, it shows low F-
measure of 0.055 as well.
B3Clustering has the similar F-measure to COACH in Krogan and Dip data, but it
has six times higher F-measure than COACH in Biogrid 3.1.85.
The experimental results show that B3Clustering has the capability to detect clusters
from the noisy graph as well as tightly interconnected graph. They support the efficiency
and robustness of B3Clustering for protein complex prediction in PPI networks.

5 Related Works
There exist several clustering algorithms applied to biological networks in order to find
protein complexes. One of traditional clustering approaches are maximal clique finding
algorithm, which induces complete subgraphs that is not a subset of any larger sub-
graph. Clique finding algorithms have been used to predict protein complexes from
PPI network (Spirin and Mirny [23] and Liu et al.[18]). They devised their own meth-
ods to merge overlapping cliques. Liu et al.[18] proposed an algorithm called CMC
(Clistering-based on Maximal Cliques), which used a variant of Czekanowski-Dice dis-
tance to rank cliques and merged highly overlapped cliques. If the interaction networks
 Identifying Protein Complexes from Protein-Protein Interaction Network 117

Table 7. Result of Biogrid 3.1.85

Algorithms # Clusters Ncp Ncb

MCL 52 2 2
COACH 3,138 105 63
B3Clustering 467 117 52

0.3
0.6 0.6
MCL MCL 0.25 MCL
0.5 0.5
0.2
0.4 0.4
0.15
0.3 COACH 0.3 COACH COACH

0.2 0.2 0.1

0.1 0.1 0.05

B3Clustering B3Clustering B3Clustering
0 0 0
Recall Precision F-measure Recall Precision F-measure Recall Precision F-measure

Fig. 4. Result from Krogan et al., Dip and Biogird 3.1.85 datasets

are accuate and complete, then maximal clique finding algorithm [26] can be ideal for
detecting protein complexes from the networks. However, since biological data is highly
noisy and incomplete, protein complexes are not formed as shapes of complete sub-
graphs in the biological networks. To reduce the impact of the noise in the network, Pei
et al.[21] proposed using quasi-cliques instead of maximal cliques. Other approaches
employ vertex similarity measures to improve confidence of noisy networks [14].
Besides, Van Dongen (2000)[27] introduced MCL (Markov Clustering) as graph par-
titioning method by simulating random walks. It used two operators called expansion
and inflation, which boosts strong connections and demotes weak connections. Iter-
ative expansion and inflation separate the graphs into many subgraphs. Brohee et al.
(2006)[6] shows the robustness of MCL with comparison to three other clustering algo-
rithms (RNSC[12], SPC[5], MCODE[4]) for detecting protein complexes. Each clus-
tering algorithm was applied to binary PPI data in order to test the ability to extract
complexes from the networks and the clusters were compared with the annotated MIPS
complexes [20]
MCODE [4] uses a vertex-weighting method based on the clustering coefficient
which assesses the local neighbourhood density of a vertex. MCODE is the overlap-
ping clustering approach which allows to overlap clusters.
One of the recent emerging techniques is to use protein core attachments methods,
which identify protein-complex cores and add attachments into these cores to form pro-
tein complexes (CORE[15] and COACH[28]). Li et al.(2010)[17] shows that COACH
performs better than seven other clustering algorithms (MCODE[4], RNSC[12],
MCL[27], DPClus[3], CFinder[1], DECAFF[16], and CORE[15]) when they are ap-
plied to two PPI datasets (DIP[30] and Krogan et al.[13] data).
118 E. Chin and J. Zhu

6 Conclusion
As the advances of technologies has produced a wealth of data, data becomes compli-
cated and erroneous. The noise of the data makes it hard to find desirable results from
the overflow of data. We introduce B3Clustering to detect clusters in the complicated
and noisy graph. B3Clustering controls the connectivity of clusters according to the
size of clusters by two thresholds. Since protein complexes has lower connectivity with
larger size, it is reasonable to control threshold according to the size of protein com-
plexes. As we can see in the experimental results, B3Clustering performs better than
the existing state-of-art clustering methods in the densely connected graph as well as
noisy graph.
In the future, data size will be bigger and data may be more complicated. For saving
operation time, we may consider B3Clustering detect clusters from the selected vertex
pairs instead of all the vertex pairs.

References
1. Adamcsek, B., Palla, G., Farkas, I.J., Derenyi, I., Vicsek, T.: CFinder: locating cliques and
overlapping modules in biological networks. Bioinformatics 22(8), 10211023 (2006)
2. Aloy, P., et al.: Structure-based assembly of protein complexes in yeast. Science 303(5666),
20262029 (2004)
3. Altaf-Ul-Amin, M., Shinbo, Y., Mihara, K., Kurokawa, K., Kanaya, S.: Development and
implementation of an algorithm for detection of protein complexes in large interaction net-
works. BMC Bioinformatics 7, 207 (2006)
4. Bader, G., Hogue, C.: An automated method for finding molecular complexes in large protein
interaction networks. MBC Bioinformatics 4, 2 (2003)
5. Blatt, M., Wiseman, S., Domany, E.: Superparamagnetic clustering of data. Phys. Rev.
Lett. 76(18), 32513254 (1996)
6. Brohee, S., van Helen, J.: Evaluation of clustering algorithms for protein-protein interaction
networks. BMC Bioinformatics 7, 488 (2006)
7. Cho, Y., Hwang, W., Ramanthan, M., Zhang, A.: Semantic integration to identify overlapping
functional modules inprotein interaction networks. BMC Bioinfotmatics 8, 265 (2007)
8. Dwight, S.S., et al.: Saccharomyces Genome Database provides secondary gene annotation
using the Gene Ontology. Nucleic Acids Research 30(1), 6972 (2002)
9. Friedel, C.C., Krumsiek, J., Zimmer, R.: Boostrapping the interactome: Unsupervised Iden-
tification of Protein Complexes in Yeast. In: Vingron, M., Wong, L. (eds.) RECOMB 2008.
LNCS (LNBI), vol. 4955, pp. 316. Springer, Heidelberg (2008)
10. Gavin, A., Aloy, P., Grandi, P., Krause, R., Boesche, M., Marzioch, M., Rau, C., Jensen,
L.J., Bastuck, S., Dumpelfeld, B., et al.: Proteome survey reveals modularity of the yeast cell
machinery. Nature 440(7084), 631636 (2006)
11. Gentleman, R., Huber, W.: Making the most of high-throughput protein-interaction data.
Genome Biology 8(10), 112 (2007)
12. King, A., Przulj, N., Jurisica, I.: Protein complexes prediction via cost-based clustering.
Bioinformatics 20(17), 30133020 (2004)
13. Krogan, N., Cagney, G., Yu, H., Zhong, G., Guo, X., et al.: Global landscape of protein
complexes in the yeast Saccharomyces cerevisiae. Nature 440(7082), 637643 (2006)
14. Leight, E., Holme, P., Newman, J.: Vertex similarity in networks. Physical Review E 73,
026120 (2006)
 Identifying Protein Complexes from Protein-Protein Interaction Network 119

15. Leung, H.C., Yiu, S.M., Xiang, Q., Chin, F.Y.: Predicting Protein Complexes from PPI Data:
A Core-Attachment Approach. Journal of Computational Biology 16(2), 133144 (2009)
16. Li, X., Foo, C., Ng, S.K.: Discovering protein complexes in dense reliable neighborhoods of
protein interaction networks. In: Comput. Sys. Bioinformatics Conf., pp. 157168 (2007)
17. Li, X., Wu, M., Kwoh, C.K., Ng, S.K.: Computational approaches for detecting protein com-
plexes from protein interaction networks: a survey. BMC Bioinformatics 11(suppl. 1), S3
(2010)
18. Liu, G.M., Chua, H.N., Wong, L.: Complex discovery from weighted PPI networks. Bioin-
formatics 25(15), 18911897 (2009)
19. Mete, M., Tang, F., Xu, X.D., Yuruk, N.: A structural approach for finding functional modules
from large biological networks. BMC Bioinformatics 9(suppl. 9), S19 (2008)
20. Mewes, H.W., et al.: MIPS: Analysis and annotation of proteins from whole genomes. Nu-
cleic Acids Res. 32(Database Issue), D41D44 (2004)
21. Pei, J., Jiang, D., Zhang, A.: On mining cross-grph quasi-cliques. In: Proceedings of the
ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (2005)
22. Pu, S., Wong, J., Turner, B., Cho, E., Wodak, S.J.: Up-to-date catalogues of yeast protein
complexes. Nucleic Acids Res. 37(3), 825831 (2009)
23. Spirin, V., Mirny, L.: Protein complexes and functional modules in molecular networks.
PNAS 100(21), 1212312128 (2003)
24. Sprinzak, E., Sattah, S., Magalit, H.: How reliable are experimental protein-protein interac-
tion data. Journal of Molecular Biology 327(5), 919923 (2003)
25. Stark, C., Breitkreutz, B.J., Reguly, T., Boucher, L., Breitkreutz, A., Tyers, M.: Biogrid: A
general Repository for Interaction Datasets. Nucleic Acids Res. 34, D535D539 (2006)
26. Tomita, E., Tanaka, A., Takahashi, H.: The worst-case time complexity for generating all
maximal cliques and computational experiments. Theoretical Computer Sciene 363, 2842
(2006)
27. Van Dongen, S.: Graph Clustering by Flow Stimulation. University of Utrecht (2000)
28. Wu, D.D., Hu, X.: An efficient approach to detect a protein community from a seed. In:
2005 IEEE Symposium on Computational Intelligence in Bioinformatics and Computational
Biology (CIBCB 2005), pp. 135141. IEEE, La Jolla (2005)
29. Wu, M., Li, X., Kwoh, C.K., Ng, S.K.: A core-attachment based method to detect protein
complexes in PPI networks. BMC Bioinformatics 10, 169 (2009)
30. Xenarios, I., Salwinski, L., Duan, X., Higney, P., Kim, S., Eisenberg, D.: DIP, the Database
of Interacting Proteins: a research tool for studying cellular networks of ptoein interactions.
Nucleic Acids Research 30, 303305 (2002)
 Detecting Event Rumors on Sina Weibo Automatically

Shengyun Sun1, Hongyan Liu2,*, Jun He1,*, and Xiaoyong Du1

1 Key Labs of Data Engineering and Knowledge Engineering, Ministry of Education, China
School of Information, Renmin University of China, 100872, China
{sun_shengyun,hejun,duyong}@ruc.edu.cn
2
Department of Management Science and Engineering, Tsinghua University, 100084, China
[email protected]

Abstract. Sina Weibo has become one of the most popular social networks in
China. In the meantime, it also becomes a good place to spread various spams.
Unlike previous studies on detecting spams such as ads, pornographic messages
and phishing, we focus on identifying event rumors (rumors about social
events), which are more harmful than other kinds of spams especially in China.
To detect event rumors from enormous posts, we studied the characteristics of
event rumors and extracted features which can distinguish rumors from ordi-
nary posts. The experiments conducted on real dataset show that the new fea-
tures are effective to improve the rumor classifier. Further analysis of the event
rumors reveals that they can be classified into 4 different types. We propose an
approach for detecting one major type, text-picture unmatched event rumors.
The experiment demonstrates that this approach is well-performed.

Keywords: rumor, Sina Weibo, rumor detection, social network.

1 Introduction

Over the past several years, online social networks such as Facebook, Twitter, Re-
nren, and Weibo have become more and more popular. Among all these sites, Sina
Weibo is one of the leading micro-blogging service providers with eight times more
users than Twitter [1]. It is designed as a micro-blog platform for users to communi-
cate with their friends and keep track of hot trends. It allows users to publish micro-
blogs including short text messages with no more than 140 Chinese characters,
attached videos, audios and pictures to show their opinions and interests. All these
micro-blogs published by users will appear in their followers pages. Also users can
make comments on or retweet others' micro-blogs to express their view.
Weibo is a kind of convenient sites for users to show themselves and communicate
with others. It is put into service in 2009 by Sina Corporation. A recent official report
shows that Sina Weibo has more than 300 million registered users by the end of Feb-
ruary 29, 2012 and posts 100 million micro-blogs per day. Unfortunately, the wealth
of information also attract interests of spammers who attempt to send kinds of spams

*
Corresponding authors.

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 120131, 2013.
Springer-Verlag Berlin Heidelberg 2013
 Detecting Event Rumors on Sina Weibo Automatically 121

such as rumors, ads, pornographic messages or links to phishing or malicious web

sites. Some of these malicious spams may cause significant economic loss to our so-
ciety and even threaten national security and harmony.
In this paper, we focus on a kind of spam, and we call it event rumor, which is
more harmful than other kinds of spams especially in China. Event rumors are spams
that make up untruthful social events and theyre harmful to society security and har-
mony in potential. Figure 1 shows an instance of event rumor which says that the
picture is not a scene of a movie or a war, but shows that an officer was on his way to
investigate peoples life under the protection of several soldiers. It was posted by a
Sina weibo user (for the purpose of privacy protection we hide the users name) at
08:31 on May 11.

Fig. 1. Instance of event rumor on Sina Weibo Fig. 2. Instance of Sina Weibo Rumor-
Busting

To detect spams, Sina Weibo has tried several ways such as adding a "report as a
spam" button to each micro-blog as shown in the left bottom of Figure 1. Any users
considering a micro-blog to be suspicious can report it to Sina official Weibo account
(with the translated English user name of Weibo Rumor-Busting [1]) who specia-
lizes in identifying fake micro-blogs manually. Figure 2 shows a micro-blog
published by the Sina official account, which points out that the message showed in
Figure 1 is a rumor and gives detailed and authentic reasons. Though artificial method
is accurate, it costs a lot of human efforts and financial resources. What's more, there
is a certain delay by this approach. It cannot identify a spam quickly once it appears.
The rumor may have been spread widely and produced bad effects on society before it
is identified manually. So an automatic approach to detecting event rumors is essen-
tial. To the best of our knowledge, [1] is the first paper that detect rumor automatical-
ly on Sina Weibo, but it aims to detect general rumors not event rumors. And the
accuracy has big room to be improved.
In this paper we study how to detect the special kind of rumor, event rumor, be-
cause it is more harmful to our society than others. We consider the problem of rumor
detection as a classification problem firstly. To distinguish rumors from ordinary
posts, we design special features about event rumors from Sina Weibo and build clas-
sifiers based on these features. Then we divide event rumors into 4 types and propose
a method to identifying one major type among them. The experimental results indi-
cate that our approaches are effective.
The rest of the paper is organized as follows. In section 2 we give a review of re-
lated work. In section 3 we describe our two approaches to identifying event rumors
and provide a detailed description of the new features. In section 4 we describe how
to collect data and present the experimental results. Finally, we conclude this paper in
section 5.
122 S. Sun et al.

2 Related Work
The previous research about spam detection mainly focuses on detecting spams re-
lated to ads, pornography, viruses, phishing and so on. These works can be divided
into two types. One is detecting spammers who publish social spams [5-10], while the
other is detecting spams directly [1-4].
Lee et al. [6] proposed a honeypot-based approach to uncover social spammers in
online social systems including MySpace and Twitter. They deployed social honey-
pots to harvest deceptive spam profiles from social networking communities firstly.
Then they analyzed the properties of these spam profiles and built spam classifiers to
filter out spammers. Benevento et al. [11] also used honeypots to collect spam pro-
files. Benevento et al. [5] extracted attributes related to content of tweets and
attributes about user behavior and built a SVM classifier to detect spammers of Twit-
ter. This approach is well-performed in filtering out spammers who focus on posting
ads, especially link deception. Stringhini et al. [7] did detailed analysis of spammer on
Facebook, MySpace and Twitter. They created honey profiles accepting all requests
but sending none request to make friends in order to collect data about spamming
activity. Spam bots were divided into four categories Displayer, Bragger, Poster and
Whisperer. Random forest algorithm was used to identify spammers.
Wang [3] extracted three graph-based features and three content-based features,
and formulated the problem of spam bots detection as classification problem. And
then he applied different classification algorithms such as decision tree, neural net-
work, support vector machines, and k-nearest neighbor. Finally they found that Bayes
classifier is best well-performed. The three graph-based features were extracted from
Twitters social network graph which represents the following relationship among
users. The content-based features contained the number of duplicate tweets, the num-
ber of HTTP links, and the number of replies or mentions. Similar to this work [3], in
Wangs another work [4], a new content-based feature named trending topics
representing the number of tweets that contains the hash tag # in a users 20 most
recent tweets was added to features set. Besides the features studied in previous
works, Yang, et al. [1] extracted two new features, client-based feature and location-
based feature. The former refers to the client program that user has used to post a
micro-blog while the later refers to the actual place where the event mentioned in the
micro-blog has happened. They trained a Support Vector Machine classifier to meas-
ure the impact of proposed features on the classification performance. A noteworthy
point is that the dataset they used is from Sina Weibo. To the best of our knowledge,
this is the first paper that studies rumor detection of Sina Weibo. The difference be-
tween this work and ours lies in that we focus on event rumors. Gao, et al. [2] studied
the wall posts in Facebook which are usually used to spread malicious content by
spammers. They built wall post similarity graph, clustered similar wall posts into a
group and then identified spam clusters. The difference from previous works is using
clustering method.
As mentioned above, majority of the previous works are about detecting spams
such as ads, viruses, phishing, pornography and so on. Unlike previous studies, we
focus on identifying event rumors (rumors about social events), which are more harm-
ful to national security and society harmony than other kinds of spams especially in
China.
 Detecting Event Rumors on Sina Weibo Automatically 123

3 Event Rumors Detection

3.1 Problem Statement

Driven by different purposes, spammers spread various kinds of spams such as adver-
tise, pornography, viruses, phishing websites and rumors on Sina Weibo. Most of the
previous works focus on detecting spams like ads and pornography as mentioned above.
In this paper, we focus on the detection of event rumor which is more harmful than
other kinds of spams in China. Other rumors like gossips about stars are not covered by
our study. Event rumors are messages which describe fake social events. Usually, an
event is described by attributes such as time, location, character and content.
In the following we list the notations to be used in the subsequent sections.

: set of users
: set of micro-blogs
: a set of s followers
: a set of s friends
: the posting time of micro-blog
: the text content of micro-blog
: the picture attached to micro-blog

3.2 Detecting Event Rumors

In this paper, we consider event rumors detection as a classification problem. To iden-
tify the event rumors on Sina Weibo automatically, we need to train classifiers and the
key point is to extract powerful features. By analyzing micro-blogs characters we
extract 15 features, which are divided into three types: content-based features,
user-based features and multimedia-based features. Some of these features have been
studied in previous works [1,3-7], and others are proposed in this paper. Here we
introduce them briefly.

Content-Based Features. Content-based features are 8 attributes related to the mes-

sage content including the number of comments, the number of retweets, the number
of @ tags, the number of URLs, the number of duplications, the client program
type, the number of verbs used to describe event, and the message contains strong
negative words or not. The former 6 features have been studied already while the
latter two are new features proposed by us.
The number of comments and the number of retweets. Event rumors are usually trend-
ing topics that attract more comments and retweets than common micro-blogs.

The number of @ tags. @ is considered as a mention tag in Sina Weibo similar

to that in Twitter. In Sina Weibo, users use @+username anywhere in the micro-
blog messages as a reply or mention of another user. A user can mention anyone in a
micro-blog as long as the user follows him or her. If a user is mentioned in a micro-
blog by the mention tag, he or she will see the notification message. The reply and
mention function is often utilized by the rumor accounts to draw users attention to
spread rumors.
124 S. Sun et al.

The number of URLs. Since Sina Weibo only allows users to post a short message
within 140 Chinese characters, the URLs are always shown in a shorten format [3].
Shorten URLs can hide the URLs source and users cant get direct information from
the URLs. So URL is an attribute that is often utilized by rumor publisher.
The number of duplications. To spread rumors widely, users may post the same mi-
cro-blog several times. So the number of duplications is also a noteworthy feature.
Given two micro-blogs named and , we measure the similarity based on Jaccard
coefficient as shown in Equation 1.
(1)

Where is represented by a set of keywords in the text content of . Keywords

are the words of the micro-blog text after removing stopping words. When the value
of similarity is bigger than 0.8, we consider the two micro-blogs are duplicates.
Client. It refers to the client program used to post micro-blogs by users. Clients are
divided into two types [1]: non-mobile client and mobile client. The non-mobile client
program includes Sina Weibo web-app, times posting tools and embedded Sina Wei-
bos third party applications, while the mobile client program includes mobile-based
client and Tablet Personal Computer client. Due to the convenience of non-mobile
clients, they are often utilized to post and spread event rumors by users.
The number of event verbs. In order to discriminate event rumors from other micro-
blogs, we regard the number of event verbs as one of the features. We create an event
verb dataset using the following 2-step method. First, we extract all the verbs from
news of Sina and Fenghuang. Second, we filter out the common verbs usually used in
daily life. Figure 3 compares the number of event verbs in event rumor and non-rumor
micro-blogs. In this figure, each point represents a micro-blog message.
The message contains strong negative words or not. We find that most of event ru-
mors are negative social events and contain strong negative sentiment words and opi-
nion words. Figure 4 shows the comparison result. As we expect, event rumors are
more likely to contain strong negative words than normal micro-bogs.

rumor non-rumor
20
No. of event verbs

15

10

0
0 10 20 30 40 50 60 70 80 90 100110

Fig. 3. The number of event verbs used in each micro-blog

 Detecting Event Rumors on Sina Weibo Automatically 125

including nagetive words

unincluding negative words

no
on-rumor
rumor

0% 50% 100%

Fig. 4. The distributtion of micro-blogs that contains strong negative words

Fig. 5. An example social network graph on Sina Weibo

User-Based Features. Useer-based features are 6 attributes about users, including the
number of users followerss, the number of users friends, reputation of user, whetther
the user has VIP authenticaation, ratio of micro-blogs containing event verbs, and raatio
of micro-blogs containing strong negative words. The former 4 features have bbeen
studied already, while the later 2 features are new features proposed by us. Sina WWei-
bo provides a platform to users
u to create social connections by following others and
hem. Figure 5 shows a simple social network graph and the
allowing others to follow th
arrows depict the following g relationship between users. We can see that user A and
user D are the followers of user B, while user C and user D are friends of user B. U
User
B and user D are mutual folllowers.
The number of followers. The
T micro-blogs posted by a user will be seen by all hiss or
her followers. The more folllowers the user has, the wider the rumors spread.
The number of friends. An nyone can follow a user to be this users follower withhout
seeking permission. As Sina Weibo will send a private message to you when someone
follows your account activeely, rumor publishers use the following function to attrract
attention of users attention
n. If you follow other accounts, you can also mention thheir
accounts by using a @ taag in your micro-blogs, then they will receive your miccro-
blogs even though they are not your followers.
Reputation. If a user has a small number of followers but a lot of friends, the poossi-
bility that he or she posts rumors is higher than others. The reputation of a useer is
defined as follows by Equattion 2.

(2)
126 S. Sun et al.

Where means the number of followers and means

the number of friends. The value of reputation is between 0 and 1. Value 1 means a
high reputation while 0 means a low reputation.

VIP authentication. If users identity is verified by Sina Weibo officially, then a VIP
tag will occur after the user name. The verified users include famous actors, expor-
ters, organizations, famous corporations and so on. Usually, VIP users are less likely
to post rumors.

Ratio of micro-blogs containing event verbs. This attribute means the proportion of
micro-blogs that contain event verbs over all micro-blogs posted by the user. The
higher this value, the more likely the users post event rumors.

Ratio of micro-blogs containing strong negative words. This attribute means the pro-
portion of micro-blogs that contain strong negative words over all micro-blogs posted
by the user. The higher this value, the more likely the users post event rumors.

Multimedia-Based Features. Multimedia-based features are attributes about picture.

Sina Weibo allows the micro-blogs containing multimedia like pictures, videos and
audios. The most powerful feature we extracted from picture is time span.

Timespan. To make the rumors look like normal micro-blogs, users usually attach
corresponding pictures to the micro-blogs. These pictures often come from Internet
and were published before. We call them outdated pictures. Users utilize the outdated
pictures and new messages to make up a new micro-blog related to event rumors. The
timespan function is defined by Equation 3.

(3)

Where is the time function, , and is a

timespan threshold. In our experiment we set the threshold value as a week.
The time of picture is calculated by the following steps. First, we use the picture-
to-action function provided by Baidu1 search engine which is used to identify similar
pictures to find out the attached picture on the Internet. If we submit a picture shown
in Figure 6 which is an attached picture of the rumor shown in Figure 1 as a query, the
search engine will output lots of result records and each record includes the corres-
ponding picture, title, description, URL, and time as shown in Figure 7. Second, ac-
cording to the time order, we identify the oldest picture and consider this pictures
time as the attached pictures time. If a micro-blog doesnt contain any picture, then
we consider the value of timespan to be 0. If the timespan between the text and the
picture is bigger than the threshold, then the value is 1. Otherwise the value is 2.

1
http://shitu.baidu.com/
 Detecting Event Rumors on Sina Weibo Automatically 127

Fig. 6. An attached picture Fig. 7. An result record

Figure 8 depicts the distribution of timespan of micro-blogs and Figure 9 depicts

the distribution of timespan of micro-bogs that contain pictures. The two distributions
indicate that rumors are more likely to contain an outdated picture than normal micro-
blogs indeed.

rumor non-rumor rumor non-rumor

1 1
proportion

proportion

0.5 0.5

0 0
0 1 2 1 2
timespan timespan
Fig. 8. The distribution of times pan of mi- Fig. 9. The distribution of timespan of mi-
cro-blogs on Sina Weibo cro-blogs containing pictures on Sina Weibo

Based on these features, classifiers are built and used to predict whether a micro-
blog is an event rumor or not.

3.3 Detecting Text-Picture Unmatched Rumors

Further analysis of sample event rumors reveals that event rumors can be classified
into different types. In the following discussion, we provide our analysis of several
common types of event rumors and propose an effective method to one major type,
detect text-picture unmatched rumors.
Based on our observation, most event rumors belong to one of the following types:

Complete fiction: The social event itself is a pure fabrication.

Time-sensitive: First, time-sensitive event rumors were true messages in the past,
but the user published it again later and made it like just happened, so these mes-
sages are out of date and become time-sensitive event rumors.
Fabricated details: As mentioned in section 3.1, an event contains the following
attributes: time, location, character and content. This kind of rumor is made up by
modifying the time, location or characters of another real event happened before.
128 S. Sun et al.

The difference between fabricated details type and time-sensitive type is that the
latter was true but is out of date while the former is fake.
Text-picture unmatched rumors: As shown in Figure 1, to increase event rumors
credibility, users often attach a picture to the text, but there is no relationship be-
tween the picture and the text.

From Figure 8 we can see that about 80% event rumors contain a picture and a major-
ity of them are text-picture unmatched rumors according to our observation. There-
fore, we focus on study how to detect this kind of event rumors. Figure 1 shows an
instance of text-picture unmatched rumor, whose text describes an officers investiga-
tion of peoples life under the protection of soldiers, but the picture actually describes
the team leader of Chinese Embassy in the Republic of Iraq visiting temporary em-
bassy accompanied by security personnel in fact. The rumor account utilizes the unre-
lated picture and new text to make up a new micro-bog and we call it text-picture
unmatched rumor. We propose a 5-step method to identify this kind of event rumors.
First, we create a list, including 60 news websites consisting of almost all the major
domestic media, and foreign media2. Second, we submit the picture attached to a mi-
cro-blog as a query to search engine to look for similar pictures. If the output result is
nothing, then we consider that the picture matches the text, and the micro-blog is non-
rumor. Otherwise, we order the result records according to their websites reliability
and their posting time descending. The websites in the list created in the first step are
credible and their credibility is the same while others are not credible. Third, we crawl
the main content of the top ranked website. Finally, we use Jaccard coefficient to
calculate the similarity between micro-blogs text and the crawled content after re-
moving stopping words. If they are similar, we consider the micro-blog is non-rumor.
Otherwise, they are text-picture unmatched rumors.

4 Experiments

4.1 Collection of Dataset

In order to evaluate the proposed approach for detecting event rumors, we need a
labeled dataset of micro-blogs. To the best of our knowledge, no such collection is
publicly available, so we need to build one. Labeling manually costs lots of human
labor. Besides, the result may not be accurate by the errors in human judgment. For-
tunately, Sina Weibo has an official account named Weibo Rumor-Busting for pub-
lishing micro-blogs relevant to rumor topics. All the announced rumors are verified
by authoritative sources. Based on this official rumor busting service, we build a high
quality dataset.
The micro-blogs posted by Weibo Rumor-Busting are messages explaining
which micro-blog is rumor and why it is rumor, but they are not the original rumors
themselves, so we need to construct queries manually to find out the original rumors

2
http://news.hao123.com/wangzhi
 Detecting Event Rumors on Sina Weibo Automatically 129

by the search function provided by Sina Weibo. We select the top one micro-blog
from the search results. We collect 104 event rumors matching the keywords of mi-
cro-blogs published by Weibo Rumor-Busting from May 2011 to December 2011.
Also we extract the profiles (number of followers, number of followings and VIP
authentication) of users who published the event rumors correspondingly and all their
micro-blogs. The profiles of the followers and followings and their micro-blogs are
also included in our dataset. Finally we create a high quality dataset that consists of
1943 users and 26972 micro-blogs, which includes 104 event rumors and 26868 non-
rumors.

4.2 Evaluation

To assess the effectiveness of our methods, we use the standard information retrieval
metrics of precision, recall and F1 [12]. The precision is the ratio of the number of
rumors classified correctly to the total number of micro-blogs predicted as rumors.
The recall is the ratio of the number of rumors correctly classified to the total number
of true rumors. The F1 is the harmonic mean between both precision and recall, and is
defined as . We train 4 different classifiers used in previous
works [1, 3-5] including Nave Bayes, Bayesian Network, Neural Networks and Deci-
sion Tree. 10-fold cross validation strategy is used to measure the impact of those
features introduced in Section 3.2 on the classification performance.

Table 1. The comparison of different classifiers using different set of features

Precision Recall F-measure

Classifier
Without With Without With Without With
Nave Bayes 0.375 0.176 0.606 0.827 0.463 0.29
Bayesian Network 0.5 0.850 0.058 0.654 0.103 0.739
Neural Network 0.5 0.783 0.163 0.692 0.246 0.735
Decision Tree 0.553 0.750 0.202 0.663 0.296 0.704

The experimental results are shown in Table 1. The experimental results indicate
that before adding the 5 new features the precision, recall and F-measure are 0.555,
0.202 and 0.269 respectively. There are two reasons why the performance of these old
features in our experiment is worse than that in previous works. One is that the data-
sets are different, and the other is that the purposes are different. Previous studies
focus on identifying spams such as ads, while we focus on detecting event rumors.
After introducing the 5 new features to the classification process, the precision, recall
and F1 are improved to 0.85, 0.654 and 0.739 respectively. As shown in Figure 10,
the precision, recall and f1-measure have increased 0.475, 0.048 and 0.276 separately,
demonstrating the effectiveness of our proposed new features.
130 S. Sun et al.

with new features without new features

F1

recall

precision

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Fig. 10. The comparison between effect of feature set without and with new features

To evaluate the effectiveness of our approach to detecting text-picture unmatched

event rumors, we create a dataset including 40 messages, while 17 of them are text-
picture unmatched event rumors and 23 non-rumors. All these micro-blog messages
contain a picture. Our experiment result shows that the approach is well-performed
and the precision, recall and F-measure are 0.833, 0.823 and 0.857 respectively. The
disadvantage of this method is that the effectiveness of the picture search engine af-
fects the experimental results to some extent.

5 Conclusion

In this paper we focus on detecting event rumors on Sina Weibo, which are more
harmful than other kinds of spams especially in China. To distinguish event rumors
from normal messages, we propose five new features and use classification technique
to predict which one is event rumor. Further, we divide event rumors into four types
and propose a method to identifying text-picture unmatched event rumors with the
help of picture search engine. The experiments show the proposed approaches are
effective. In the future we will focus on study how to detect other three forms of event
rumors.

Acknowledgement. This work was supported by the HGJ PROJECT 2010ZX01042-

002-002-03 and National Natural Science Foundation of China under Grant No.
71272029, 70890083, 71110107027 and 61033010.

References
1. Yang, F., Liu, Y., Yu, X., Yang, M.: Automatic detection of rumor on Sina Weibo. In:
Proceedings of the ACM SIGKDD Workshop on Mining Data Semantics, Beijing, China,
pp. 17. ACM (2012)
2. Gao, H.Y., Hu, J., Wilson, C., Li, Z.C., Chen, Y., Zhao, B.Y.: Detecting and characterizing
social spam campaigns. In: Proceedings of the 10th ACM SIGCOMM Conference on In-
ternet Measurement, Melbourne, Australia, pp. 3547. ACM (2010)
 Detecting Event Rumors on Sina Weibo Automatically 131

3. Wang, A.H.: Dont follow me: Spam detection in Twitter. In: Proceedings of the 2010 In-
ternational Conference on Security and Cryptography (SECRYPT) (2010)
4. Wang, A.H.: Detecting spam bots in online social networking sites: A machine learning
approach. In: Foresti, S., Jajodia, S. (eds.) Data and Applications Security and Privacy
XXIV. LNCS, vol. 6166, pp. 335342. Springer, Heidelberg (2010)
5. Benevenuto, F., Magno, G., Rodrigues, T., Almeida, V.: Detecting Spammers on Twitter.
In: Collaboration, Electronic messaging, Anti-Abuse and Spam Conference, Redmond,
Washington, US (2010)
6. Lee, K., Caverlee, J., Webb, S.: Uncovering social spammers: social honeypots + machine
learning. In: Proceedings of the 33rd International ACM SIGIR Conference on Research
and Development in Information Retrieval, Geneva, Switzerland, pp. 435442. ACM
(2010)
7. Stringhini, G., Kruegel, C., Vigna, G.: Detecting spammers on social networks. In: Pro-
ceedings of the 26th Annual Computer Security Applications Conference, Austin, Texas,
pp. 19. ACM (2010)
8. Yang, C., Harkreader, R.C., Gu, G.: Die free or live hard? Empirical evaluation and new
design for fighting evolving twitter spammers. In: Sommer, R., Balzarotti, D., Maier, G.
(eds.) RAID 2011. LNCS, vol. 6961, pp. 318337. Springer, Heidelberg (2011)
9. Benevenuto, F., Rodrigues, T., Almeida, V., Almeida, J., Goncalves, M.: Detecting spam-
mers and content promoters in online video social networks. In: Proceedings of the 32nd
International ACM SIGIR Conference on Research and Development in Information Re-
trieval, Boston, MA, USA, pp. 620627. ACM (2009)
10. Benevenuto, F., Rodrigues, T., Almeida, V., Almeida, J., Zhang, C., Ross, K.: Identifying
video spammers in online social networks. In: Proceedings of the 4th International Work-
shop on Adversarial Information Retrieval on the Web, Beijing, China, pp. 4552. ACM
(2008)
11. Webb, S., Caverlee, J., Pu, C.: Social Honeypots: Making Friends with a Spammer near
You. In: Collaboration, Electronic messaging, Anti-Abuse and Spam Conference, Moun-
tain View, California, US (2008)
12. Yang, Y.: An Evaluation of Statistical Approaches to Text Categorization. Journal of In-
formation Retrieval 1(1/2), 6788 (1999)
 Uncertain Subgraph Query Processing
over Uncertain Graphs

Wenjing Ruan1 , Chaokun Wang1,2,4 , Lu Han1 , Zhuo Peng1 , and Yiyuan Bai1,3
1
School of Software, Tsinghua University, Beijing 100084, China
2
Tsinghua National Laboratory for Information Science and Technology
3
Department of Computer Science and Technology, Tsinghua University
4
Key Laboratory of Intelligent Information Processing, Institute of Computing
Technology, Chinese Academy of Sciences
{ruanwj11,hanl09,pengz09,baiyy10}@mails.tsinghua.edu.cn,
[email protected]

Abstract. In some real-world applications, data cannot be measured

accurately. Uncertain graphs emerge when this kind of data is modeled
by graph data structures. When the graph database is uncertain, our
query is highly possibly uncertain too. All of the prior works do not
consider the uncertainty of the query. In this paper, we propose a new
algorithm called Mutual-Match to solve the problem of uncertain sub-
graph query under the situation where both the graph dataset and the
query are uncertain. Considering that the subgraph isomorphism verifi-
cation is an NP-hard problem, and it will be more complex on uncertain
graph data, the Mutual-Match algorithm uses MapReduce processes to
find out all the subgraph matches and can adapt to dynamic graphs well.
Experimental results on both real-world and synthetic datasets show the
correctness and eectiveness of our proposed methods.

Keywords: Uncertain subgraph query, uncertain graphs, mutual-match,

dynamic graphs, mapreduce.

1 Introduction

With the rapid development of web technology, graph has considerable impor-
tance in modeling of data in diverse scenarios, e.g., social networks, mobile
networks and sensor networks. If the graph-structured data is uncertain, it is
called an uncertain graph. For example, each edge of the graph has a probabil-
ity of its existence in this paper. That is to say the uncertainty is embodied in
the relationships between nodes of the graph. A typical example of the uncer-
tain graph is protein-protein interaction networks (PPI networks for short) [5].

This work was supported by the National Natural Science Foundation of China
(No. 61170064, No. 61133002), the National High Technology Research and Devel-
opment Program of China (No. 2013AA013204) and the ICT Key Laboratory of
Intelligent Information Processing, CAS.

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 132139, 2013.
c Springer-Verlag Berlin Heidelberg 2013

 Uncertain Subgraph Query Processing over Uncertain Graphs 133

Fig. 1. An Example of Matching Pairs

In PPI networks, nodes denote protein molecules while edges denote the interac-
tion relationships between protein molecules. Limited by the measuring methods,
we cannot accurately detect whether there is an interaction between two protein
molecules. Therefore, an existence probability is set for each interaction.
Recently, uncertain graphs have become an increasingly attractive focus in
many fields. There are a number of algorithms proposed to solve mining and
query problems in uncertain graph database. Representative mining techniques
in this area include finding top-k maximal cliques [11] and mining frequent sub-
graph patterns [10]. Various query methods have been proposed to solve the
essential problems of uncertain graph database, e.g., KNN query [4], the short-
est path query [9], reliable subgraph query [3], and so on. All of the prior works
haven not considered the uncertainty of the query graph, which is used to search
matches in the uncertain graph database.
In the research field of uncertain graph management, the problem of uncertain
subgraph query emerges when both the dataset and the query are uncertain.
As shown in Figure 1, Q is an uncertain query graph, and D is an uncertain
data graph. The uncertain subgraph query returns all the pairs of isomorphic
existences between Q and D.
There are two challenges to deal with the above-mentioned problem. Firstly,
the subgraph isomorphism verification is an NP-hard problem, and it will be
more complex when the dataset and the query are both uncertain. Secondly,
uncertain subgraph query problem has high complexity, especially when the size
of the uncertain dataset is large.
To address the above challenges, a novel algorithm called Mutual-Match is
proposed in this paper. The intuitive idea of Mutual-Match is using distributed
computing methods to greatly improve the query performance. Thanks to the
MapReduce model [1], Mutual-Match can eectively find out all matching pairs,
each of which is composed by two isomorphic subgraphs derived, respectively,
from the query graph and the data graph.
134 W. Ruan et al.

To the best of our knowledge, this paper is the first one addressing the problem
of uncertain subgraph query. The main contributions of the paper are summa-
rized as follows.
1. The problem of uncertain subgraph query over uncertain graph is formally
presented.
2. An algorithm called Mutual-Match is proposed to process the uncertain sub-
graph query over uncertain graphs with the MapReduce model. It can also exe-
cute on dynamic uncertain graphs.
3. Extensive experiments are conducted to show the correctness and eectiveness
of our proposed method.
The rest of the paper is organized as follows. Section 2 gives the formal definition
of the problem of uncertain subgraph query processing. We present a brute-force
algorithm, and then propose an index structure and the Mutual-Match algorithm
in Section 3. Moreover, we conduct an empirical study on both real-world and
synthetic datasets in Section 4. Finally, we conclude in Section 5.

2 Problem Definition
Definition 1. An uncertain graph is defined as G(V, E, , , ), where (1) V is
a set of nodes; (2) E V V is a set of edges; (3) is a set of labels; (4)
: V is a function that sets a label to each node of V ; and (5) : E [0, 1]
is a probability function for edges.
Given an uncertain graph G(V, E, , , ), we can find out a certain graph
G (V, E, , , ) which is generated from G when all the edges probabilities
of G are 1. We call G the underlining graph of G.
For a clear formalization of the problem of uncertain subgraph query over
uncertain graphs, we use the possible world model. Each possible world of G
corresponds to a subgraph of G and has an existence probability that is com-
puted according to the probability of every edge in G. In this paper, we assume
that all the probabilities of edges distribute independently.
Definition 2. Given an uncertain query graph Q and an uncertain data graph
D, uncertain subgraph query is to find a set of matches. Each match which is
denoted as a pair m(q, d) should satisfy the following two requirements.
1. q and d are certain graphs from the possible worlds of Q and D, respectively.
2. q and d are isomorphic.
As shown in Figure 1, if we limit that the edge number of isomorphic subgraph
is not less than 3, there are 5 matching pairs when searching uncertain subgraph
matches of Q in D. One of the pairs is m(q, d).
Definition 3. For n given uncertain query graph Q, an uncertain data graph D
and one of their matches m(q, d), the matching probability of m(q, d) is defined as

P ro(q, d) = P ro(q) P ro(d), (1)

 Uncertain Subgraph Query Processing over Uncertain Graphs 135

where P ro(q) and P ro(d) are the existence probabilities of the corresponding
possible worlds of Q and D, respectively.

3 Algorithms

In this section, we first present the brute-force algorithm of uncertain subgraph

query and then propose an eective index and the mutual-match algorithm which
transforms the complex uncertain subgraph query on a large uncertain database
to a series of MapReduce operations with the index structure.

3.1 Brute-Force Algorithm

For a given uncertain query Q and an uncertain data graph D, the brute-force
algorithm first transforms them into possible worlds WQ = {g1 , g2 , , gs } and
WD = {g1 , g2 , , gt }, respectively. Let CS = {q1 , q2 , , qm } be a subset of WQ ,
which contains all connected components of WQ . Each element of CS or WD is
a certain graph with an existence probability, so that the brute-force algorithm
evaluates the uncertain subgraph query by processing certain subgraph queries
on WD m times. We build an index structure on all elements of WD with gIndex
[7,8], which performs the best for most queries for sparse datasets in recent
researches about indexes [2]. And then, search subgraph matches against each
graph in CS using the gIndex.
This brute-force algorithm has to deal with exponential times of subgraph
queries within an extremely large search space, if the query and data graphs are
large. Clearly, it can only be applied to small datasets.

3.2 Mutual-Match Algorithm

With the Mutual-Match algorithm, we need not know the possible worlds of
uncertain data and query graphs before the query processing. In order to get
a matching pair m(q, d) from the uncertain query graph Q and the uncertain
data graph D, the Mutual-Match algorithm first splits both Q and D into some
small components which are denoted as bi-edge graphs. Then, with the help of
the edge join operation (EJoin) which joins components together, q and d are
gradually constructed respectively.

Definition 4. Given an uncertain graph G, a subgraph of G is called a bi-edge

graph, denoted as (el (vl , vm ), er (vm , vr )), if it has three nodes vl , vm , vr and
two corresponding edges el , er . In a bi-edge graph, vm is called the middle node.
Bie(G) is the set containing all bi-edge graphs of G.

Definition 5. Given two uncertain graphs G1 and G2 , and e is one of their

common edges. The operator EJoin(G1 , G2 , e) joins two graphs by merging the
common edge e.
136 W. Ruan et al.

Bi-Edge Index Structure. Before the query processing, we build an index struc-
ture of the uncertain graph database based on the bi-edge graphs. Given an uncer-
tain data graph D, we generate Bie(D) by decomposing D into bi-edge graphs. If
(el (vl , vm ), er (vm , vr )) Bie(D), the label of is labell /labelm/labelr ,
where labelm is from vm and labell < labelr (< is the partial order). The value
of is saved as (dbi id, el , er , vm , vl , vr ), where dbi id is the unique identifier of .
We index bi-edge graphs by their labels. This bi-edge graph based index is imple-
mented by a hash chain H. The ids of bi-edge graphs which have the same keyword
are saved in a list denoted by blist(label).
We generate Bie(Q) and store each bi-edge graph in Bie(Q) with the same
method as the data graph. And then, we search the database index for every
label of bi-edge graphs of Bie(Q) and get the input file for MapReduce jobs.
The format of the input file is (el , er , qbi id, vm , posv, pose), where el , er , vm are
from ( Bie(D)) and qbi id is the unique identifier of ( Bie(Q)). posv
and pose are vectors that save the matched nodes and edges, respectively. Every
record of the input file is an initial matching pair.

Algorithm Description. The Mutual-Match algorithm decomposes the query

and data graphs into bi-edge graphs and then joins dierent bi-edge graphs
according to the matching relation of these two graphs by EJoin. It coincides
with the idea of MapReduce. The map and reduce phase can process this kind
of query very well.
A MapReduce job chain assembles these initial matching pairs to find out all
the query answers. The mapper of the Job1 maps each line of the input file into
two key-value pairs. Every edge of a bi-edge graph of the data graph acts as the
key of a pair. Pairs that have the same key may be joined together with the
common edge so as to form a bigger matching pair.
The reducer constructs matching pairs by the EJoin operation. It finds out
the pairs with the same keyword, and then finds the joinable pairs, according
to some join rules. These join rules contain: (1) common edge check, (2) join
manner check, (3) identity check. The first two rules can be obtained from the
process that decomposes the query graph into bi-edge graphs, and the last one
will be obtained through the EJoin process. The common edge check aims to
find out whether the two bi-edge graphs of the query graph, matched by bi-
edge graphs from the data graph respectively, have a common edge. If there is
a common edge, the matched bi-edge graphs from the data graph are joinable.
Apart from this, there are two kinds of join manners about dierent matching
pairs. As shown in Figure 2, when the two bi-edge graphs are joined as the above
manner, they have the same middle node. This is a middle-join. On the contrary,
when joined as the below manner, their middle nodes are dierent, and this is a
non-middle-join. The EJoin operation requires the two bi-edge graphs from data
graph have the same join manner with the two matched bi-edge graphs from the
query graph. If the two key-value pairs are joinable, nodes from Q and D must
satisfy the one-to-one correspondence in their two posv vectors. The identity
check will do the job to ensure that nodes of the two posv vectors satisfy the
one-to-one correspondence.
 Uncertain Subgraph Query Processing over Uncertain Graphs 137

Fig. 2. Two Dierent Join Manners. The above is a middle-join, and the below is a
non-middle-join.

3.3 Improvement

Our method can adapt to dynamic uncertain graphs well. That is the Mutual-
Match algorithm can be improved to address the problem of uncertain subgraph
query over dynamic uncertain graphs.
When new edges are inserted or existing edges are removed and original graph
D becomes the new graph D , we dont need to rebuild the bi-edge index for D
considering all edges of D . Instead, we only take the new edges and the removed
edges into consideration.
Suppose we have run the Mutual Match Algorithm for the data graph D and
the query graph Q, and get the uncertain subgraph query result file RF which
contains all matching pairs of D and Q. When D changes, for each removed edge
e, we find matching pairs which contain e and remove these pairs from RF . For
each new edge e, we build new bi-edge index related with e. Then we get the
new input file F . We process F in Mapper to produce key-value pairs P and
choose records in RF whose keys are the same with the keys in P . Then we
add these records to the file F . Finally, we construct new matching pairs of D
and Q by EJoin operation in the MapReduce job chain (Job2 Job3 . . .
Jobn ) with F as the input file.

4 Experiments

Our experiments are conducted on two dierent environments. The index con-
struction, MapReduce input file construction and the brute-force algorithm are
performed on a 2.4GHZ, 2GB-memory, Intel(R) Core(TM) i5 CPU PC, running
Ubuntu 10.10 operation system. The other is a MapReduce platform with 5
multithreaded cores. All the algorithms are implemented with Java.

4.1 Results on the Real-World Dataset

We use a real-world dataset of CAL road network in our experiments. It con-

tains 21,692 edges and 21,047 nodes. Every edge has a weight that indicates the
uncertainty. We distribute labels to all of the nodes randomly.
Figure 3 shows the results on the real dataset. Firstly, we compare the time
costs of the index construction of both Mutual-Match and brute-force algorithms.
As shown in Figure 3(a), the time cost of the index construction of the brute-force
138 W. Ruan et al.

100 100
90 BF 90 BF
Time cost (s) 80 MRMM 80 MRMM

Time cost (s)

70 70
60 60
50 50
40 40
30 30
20 20
10 10
0 0
4 6 8 10 12 14 16 4 6 8 10 12 14 16
# of edges in the data graph # of edges in the data graph
(a) Index Time Cost (b) Query Time Cost

Fig. 3. Results on the Real-world Dataset

algorithm increases rapidly. Meantime, that of the Mutual-Match algorithm is

satisfactory and nearly constant. For example, the Mutual-Match method can
index a 16-edge uncertain graph within just 2 ms.
The query eciencies of both methods are showed as Figure 3(b). The query
time of Mutual-Match is computed by the sum of time costs to get MapReduce
input file and MapReduce jobs. It shows that the MapReduce-Based Mutual-
Match has a poor performance when the uncertain data size is too small. MapRe-
duce is not designed for small data size. With the increasing of the data size,
Mutual-Match is much more eective than brute-force.

4.2 Results on the Synthetic Dataset

The synthetic dataset is generated using the gengraph win [6]. It is a node-
labeled, undirected, uncertain data graph which contains 50,000 edges, and the
average degree of the nodes is 3. As same as the real-world dataset experiments,
labels of every node are distributed randomly. The probabilities of every edge
that describe the uncertainty are distributed as Gaussian distribution.
Figure 4 shows the eciency of our method on the synthetic dynamic graph
dataset. We use MRMM to stand for the algorithm that computes matches

3000 16
MRMM 14 MRMM
2500
DMRMM
Time cost(min)

DMRMM
Time cost(ms)

12
2000 10
1500 8
1000 6
4
500 2
0 0
10 15 20 25 30 35 40 45 50 10 15 20 25 30 35 40 45 50
#of edges in the data graph(x1000) #of edges in the data graph(x1000)

(a) Index Time Cost (b) Query Time Cost

Fig. 4. Results on the Synthetic Dynamic Dataset

 Uncertain Subgraph Query Processing over Uncertain Graphs 139

from the beginning and DMRMM to stand for the algorithm that only takes
changed edges into consideration. Figure 4(a) shows that the time cost of the
index building of the DMRMM method is much smaller than that of the MRMM
method for DMRMM does not compute all edges of the new graph. Figure 4(b)
shows the eciency of query processing in the MapReduce jobs. Clearly, the
DMRMM method is faster than the MRMM method.

5 Conclusions
Uncertain subgraph query is an essential operation on uncertain graph database.
It is complicated due to the NP-hard isomorphism verification and uncertainty.
MapReduce provides a good solution to this problem. In this paper, we have
explored a Mutual-Match algorithm and an applicative index method to con-
duct the uncertain subgraph query problem within MapReduce framework. Our
MapReduce-Based Mutual-Match algorithm has eective performances on un-
certain subgraph matching processing over static uncertain graphs and dynamic
uncertain graphs.

References
1. Dean, J., Ghemawat, S.: Mapreduce: Simplified data processing on large clusters.
Communications of the ACM 51(1), 107113 (2008)
2. Han, W., Lee, J., Pham, M., Yu, J.: iGraph: A framework for comparisons of disk-
based graph indexing techniques. Proceedings of the VLDB Endowment 3(1-2),
449459 (2010)
3. Hintsanen, P., Toivonen, H.: Finding reliable subgraphs from large probabilistic
graphs. Data Mining and Knowledge Discovery 17(1), 323 (2008)
4. Potamias, M., Bonchi, F., Gionis, A., Kollios, G.: K-nearest neighbors in uncertain
graphs. Proceedings of the VLDB Endowment 3(1-2), 9971008 (2010)
5. Saito, R., Suzuki, H., Hayashizaki, Y.: Interaction generality, a measurement to
assess the reliability of a proteinprotein interaction. Nucleic Acids Research 30(5),
1163 (2002)
6. Viger, F., Latapy, M.: Ecient and simple generation of random simple connected
graphs with prescribed degree sequence. In: Wang, L. (ed.) COCOON 2005. LNCS,
vol. 3595, pp. 440449. Springer, Heidelberg (2005)
7. Yan, X., Yu, P., Han, J.: Graph indexing: A frequent structure-based approach. In:
SIGMOD, pp. 335346. ACM (2004)
8. Yan, X., Yu, P., Han, J.: Graph indexing based on discriminative frequent structure
analysis. ACM Transactions on Database Systems (TODS) 30(4), 960993 (2005)
9. Yuan, Y., Chen, L., Wang, G.: Eciently answering probability threshold-based
shortest path queries over uncertain graphs. In: Kitagawa, H., Ishikawa, Y., Li,
Q., Watanabe, C. (eds.) DASFAA 2010, Part I. LNCS, vol. 5981, pp. 155170.
Springer, Heidelberg (2010)
10. Zou, Z., Gao, H., Li, J.: Discovering frequent subgraphs over uncertain graph
databases under probabilistic semantics. In: SIGKDD, pp. 633642. ACM (2010)
11. Zou, Z., Li, J., Gao, H., Zhang, S.: Finding top-k maximal cliques in an uncertain
graph. In: ICDE, pp. 649652. IEEE (2010)
 Improving Keyphrase Extraction
from Web News by Exploiting Comments Information

Zhunchen Luo, Jintao Tang, and Ting Wang

College of Computer, National University of Defense Technology,

410073 Changsha, Hunan, China
{zhunchenluo,tangjintao,tingwang}@nudt.edu.cn

Abstract. Automatic keyphrase extraction from web news is a fundamental task

for news documents retrieval, summarization, topic detection and tracking, etc.
Most existing work generally treats each web news as an isolated document. With
the rapidly increasing popularity of Web 2.0 technologies, many web news sites
provide various social tools for people to post comments. These comments are
highly related to the web news and can be considered as valuable background
information which can potentially help improve keyphrase extraction. In this pa-
per we propose a novel method to integrate the comment posts into the task of
extracting keyphrases from a web news document. Since comments are typically
more casual, conversational, and full of jargon, we introduce several strategies to
select useful comments for improving this task. The experimental results show
that using comments information can significantly improve keyphrase extraction
from web news, especially our comments selection method, using machine learn-
ing technology, yields the best result.

Keywords: Keyphrase Extraction, Comments, Web News, Machine Learning.

1 Introduction
As thousands of web news are posted on the Internet everyday, it is a challenge to
retrieve, summarize, classify or mine this enormous repository eectively. Keyphrases
can be seen as condensed represents of web news documents which can be used to help
these applications. However only a small minorities of documents have author-assigned
keyphrases and manually assigning keyphrases for each document is very laborious.
Therefore it is highly desirable to extract keyphrases automatically. In this paper we
consider this task in web news scenario.
Most existing work on keyphrase extraction from a web news document only con-
siders the internal information, such as the phrases Tf-idf, position and other syntactic
information, which neglects the external information of the document. Some researches
also utilize background information. For example, Wan and Xiao proposed a method,
which selected similar documents for a given document from a large documents cor-
pus, to improve the performance of keyphrase extraction [14]. Mihalcea and Csomai
used Wikipedia to provide background knowledge for this task [11]. However both of
these work needs to retrieve topic-related documents as external information, which is
a non-trivial problem. Fortunately, nowadays many web news sites provide various so-
cial tools for community interaction, including the possibility to comment on the web

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 140150, 2013.
c Springer-Verlag Berlin Heidelberg 2013

 Improving Keyphrase Extraction from Web News 141

news. These comments are highly related to the web news documents and naturally
bound with them. The motivation of our work is that comments are valuable resource
providing background information for each web news document, especially some use-
ful comments contain rich information which is the focus of the readers. Hence, in this
paper we consider exploiting comments information to improve keyphrase extraction
from the web news documents.
Unfortunately, although the main entry and its comments are certainly related and at
least partially address similar topics, they are markedly dierent in several ways [16].
First of all, their vocabulary is noticeably dierent. Comments are more casual, conver-
sational, and full of jargon. They are less carefully edited and therefore contain more
misspellings and typographical errors. There is more diversity among comments than
within the single-author post, both in style of writing and in what commenters like to
talk about. Simply using all the comments as external information does not improve the
performance of keyphrase extraction (see section 5.2). Therefore, we propose several
methods to select useful comments, based on some criterions such as textual similarity
and helpfulness of the comment. The experimental results show our comment selection
approaches can improve keyphrase extraction for web news. Especially our machine
leaning approach which integrates many factors of comments can significantly improve
this task.
The rest of this paper is organized as follows. Section 2 introduces the related work.
Section 3 describes the framework of our approach to extract keyphrases from a web
news document. In section 4 we propose the strategies for selecting useful comments.
Section 5 presents and discusses the evaluation results. Lastly we conclude the paper.

2 Related Work
Most existing work only uses the internal information of a document for keyphrase ex-
traction. The simplest approach is to use a frequency criterion (or Tf-idf model [12]) to
select keyphrases in a document. This method was generally found to give poor perfor-
mance [3]. However, Hasan recently conducted a systematic evaluation and analysed
some algorithms on a variety of standard evaluation datasets, his results indicated that
Tf-idf remained oering very robust performance across dierent datasets [4]. Another
important clue for keyphrase extraction is the location of phrase in the document. Kea
[2] and GenEx [13] used this clue for their studies. Hulth added more linguistic knowl-
edge, such as syntactic features, to enrich term representation, which were eective for
keyphrase extraction [5]. Mihalcea and Tarau firstly proposed a graph-based ranking
method for keyprase extraction [10] and currently graph-based ranking methods have
become the most widely used approaches [3,8,9]. However, all the approaches above
never consider the external information when extracting keyphrases.
Another important information, which can be used for keyphrase extraction, is ex-
ternal resource. Wikipedia1 has been intensively used to provide background knowl-
edge. Mihalcea and Csomai used the link structure of Wikipedia to derive a novel word
feature for keyphrase extraction [11]. Grineva et al. utilized article titles and the link
structure of Wikipedia to construct a semantic graph to extract keyphrases [3]. Xu et al.
1
http://www.wikipedia.org/
142 Z. Luo, J. Tang, and T. Wang

extracted the in-link, out-link, category, and inbox information of the documents related
the article in Wikipedia for this task [15]. All approaches using Wikipedia significantly
improve the performance of keyphrase extraction. Another external resource is neigh-
bor documents. Wan and Xiao proposed an idea that a user would better understand a
topic expressed in a document if the user reads more documents about the same topic
[14]. He used a textual similarity measure (e.g., cosine similarity) to find documents
topically closed to the document from a large documents corpus. These neighbor doc-
uments are beneficial to evaluate and extract keyphrases from the document. However
using the external resource introduced above needs to retrieve topic-related documents,
which is a non-trivial problem. Unlike these resources, we exploit comments of web
news as external information. This information is topic-related to the original docu-
ment. Moreover, the characteristic of naturally bound with the original news makes the
comments easier to obtain.

3 Framework

A web news document has two parts: an article and comments. The article contains
the text of event written by the author. Comments can be seen as the discussion of
the article including the attitude of readers. Most of comments are topic-related to the
article. Especially some comments contain rich information which is focused by the
readers. We call the comments, which provide additional information for readers to
better understand the content of the article, useful comments. Intuitively, these useful
comments are expected to be beneficial to evaluate and extract the salient keyphrases
from the article. However traditional keyphrase extraction methods only consider the
information of the article and ignore its comments information. This study we propose
an approach exploiting comments information to improve keyphrase extraction for the
web news documents.
Figure 1 gives the framework of our approach for keyphrase extraction. First we
segment a web news document into the article part and the comments part based on
its natural structure. Next, since not all comments are helpful for keyphrase extraction
(see section 5.2), we use a comments selector, based on some strategies, to select useful
comments from the comments part. Then we integrate the selected comments with the
article part to form a new document. At last we use a keyphrase extractor to extract
phrases from the new document. The extracted phrases, occurred in the article, are
taken as keyphrases. Our approach can be easily incorporated with any state-of-the-art
keyphrase extractor to extract keyphrases from web news.
Since some comments contain noise information such as advertisement, irrelevant
texts and texts only containing meaningless word emoticons. For example, a comment
Who cares! just reflects a readers attitude to the news without any useful informa-
tion for other readers to understand the article. Hence, the key point of our approach
is to find the useful comments for keyphrase extraction. We propose three strategies to
select useful comments, which are expected to find the articles keyphrases with higher
accuracy. The first is selecting comments which is similar to the original article, called
Similar Comments Selector. The idea of this approach is the same as Wan and Xiao
[14]. Many web news sites provide ratings setting about the comments for other readers
 Improving Keyphrase Extraction from Web News 143

Keyphrase
Extractor

Article
Article Article

Comments Comments Keyphrases

Extract
Comment Useful Comment
Comment Useful Comment
Comment Comments Comments
Useful Comment






Comment Useful Comment






Comment
Select Useful Comment
Comment Useful Comment

Comments
Selector

Fig. 1. Framework of Keyphrase Extraction from a Web News Document

(e.g, thumb up or thumb down votes). These meta ratings information can help filter
useful comments more eciently. We call the approach based on ratings information
Helpfulness Comments Selector. Furthermore, since existing technologies of natural
language processing and machine learning provide strong supports to mine the poten-
tial knowledge, we use these technologies to select useful comments, called KeyInfo
Comments Selector. Next section we will introduce the three comments selectors in
detail.

4 Selecting Useful Comments

4.1 The Similar Comments Selector

Wan and Xiao proposed a method to select similar documents for a given document to
improve the performance of keyphrase extraction [14]. Their assumption is that multiple
documents within an appropriate cluster context usually have mutual influences and
contain useful clues, which can help to extract keyphrases from each other. For example,
two documents about the same topic earthquake would share a few common phrases,
e.g. earthquake, victim, and they can provide additional knowledge for each other
to better evaluate and extract salient keyphrases from each other.
The similar comments selector adopts the same idea to find useful comments. Firstly
it uses the widely used cosine measure to evaluate document similarity between the web
news article and its comments. The term weight is computed by Tf-idf. Then the selector
chooses top N comments with the highest similarity values as the useful comments for
the web news keyphrase extraction.

4.2 The Helpfulness Comments Selector

Due to the lack of editorial and quality control, comments on web news vary greatly.
The low-quality comments may hurt the performance of keyphrase extraction.
144 Z. Luo, J. Tang, and T. Wang

Many web news sites also allow readers to vote for the helpfulness of each comment and
provide a function to assess the quality. The helpfulness function h is usually defined as
follow:
Numthumbup
h= (1)
Numthumbup + Numthumbdown
where Numthumbup is the number of people, who consider the comment is helpful.
Numthumbdown is the number of people, who consider the comment is unhelpful. This
function actually provides a direct and convenient way to estimate the utility value of a
given comment. Therefore, we use the helpfulness of comments based on this function
to select useful comments. The top N comments with highest scores voted by at least 5
users are selected for the following keyphrase extraction step.

4.3 The KeyInfo Comments Selector

A simple idea is that given an article, if the comments include more keyphrases, they
are more likely to be useful comments than the other comments which have less or
no keyphrase, since these comments contain more information which the readers focus
on. Intuitively, integrating these useful comments into original article is expected to im-
prove the performance of keyphrase extraction. However, we can not obtain the accurate
number of keyphrases in each comment, since it is the final purpose of the task. Fortu-
nately, some factors of comments are related to the probability of comments containing
more keyphrases such as the textual similarity and helpfulness of comments. Moreover,
existing natural language processing technologies can help us find many factors related
to the probability of comments containing keyphrases, and using machine learning tech-
nologies, based on these factors, can also estimate the probability accurately. Therefore,
we propose another comments selector called KeyInfo comments selector using natural
language processing and machine learning technologies.
The principle of the KeyInfo comments selector is that if a shorter comment contains
more keyphrases, it is more likely to improve the performance of keyphrases extraction
from the web news when integrating it into the original article. We define a function k
called keyInfo function as follow:

Numkeyphrase
k= (2)
Numword

where Numkeyphrase is the number of keyphrases in the comment and Numword is the
number of words in the comment. We choose the top N comments with highest score
of k as the useful comments for keyphrase extraction.
The key point of this approach is to predict the value of Numkeyphrase . We take
this prediction as a regression-learning problem. Firstly, some web news and their
comments are collected as training data. These news documents have been originally
annotated keyphrases. Then using these gold standard keyphrases, each comments
Numkeyphrase can be easily obtained. Some features, which can potentially help to pre-
dict the Numkeyphrase , are designed for each comment. Next each comment is repre-
sented by a feature vector. Then a classical regression tool-simple linear regression
 Improving Keyphrase Extraction from Web News 145

(SLR) implemented in WEKA2 is used to train a prediction model, called KeyInfo

model. Lastly, give a new comment represented as a feature vector, this model can pre-
dict its Numkeyphrase value.
To the features for each comment, textual similarity and helpfulness are not enough.
Since the most similar comments are similar to the original entire news article text, the
probability of containing keyphrases in each part of the article text, however, is dier-
ent. For example, the headline of article, the beginning and the end part of article may
have higher probability than another parts of text. If a comment is more similar to these
key parts of the article, it is more likely to be useful comment for keyphrase extrac-
tion. Additionally, for the comments with high score of helpfulness, the phenomenon
of rich-get-richer might hurt the performace of this approach. The reason is that,
how fast a comment accumulate votes depends on the number of votes they already
have [7] and this led to a bias to the comments with more user voting. Therefore, we
develop more features, based on natural language processing technologies, for the Key-
Info model. These features are designed to maintain the advantages and overcome the
disadvantages of two approaches introduced above. Table 1 shows all the features for the
KeyInfo Model. All developed features are classified into five categories: Structural,
Similar, Lexical, Syntactic, Semantic, and Meta-data. We use Stanford Log-linear
Part-Of-Speech Tagger3 to tag the comment and use General Inquirer4 for sentiment
analysis.

5 Experiment
We empirically evaluate our approach for keyphrase extraction from a web news docu-
ment, especially comparing the performance of various strategies to select useful com-
ments for this task. In addition we conduct in-depth analysis of the dierences among
useful comments sets chosen by dierent comments selectors.

5.1 Experimental Setup

The AOL Corpus. We construct our own keyphrase extraction data set by collecting
60 web news posted on the AOL websites5 . All these news are world news posted from
November 16th 2010 to November 26th 2010. Every news document contains at least
20 comments. The main reason we choose AOL web news for the experiment is that
every web news document has its own tags. We take the tagged phrases that occur in the
article as gold standard keyphrases. Table 2 provides an overview of the AOL corpus.

Evaluation Metric. For the evaluation of keyphrase extraction results, the automatic ex-
tracted keyphrases are compared with the gold standard keyphrases. The precision p =
countcorrect /countextrator , recall r = countcorrect /countgold , and F1 score f = 2pr/(p + r)

2
http://www.cs.waikato.ac.nz/ml/weka/
3
http://nlp.stanford.edu/software/tagger.shtml
4
http://www.wjh.harvard.edu/ inquirer/
5
http://www.aol.com/
146 Z. Luo, J. Tang, and T. Wang

Table 1. Features of the KeyInfo Model

Feature category Feature name Description

NumTokens The number of tokens in the comment
NumSentences The number of sentences in the comment
Structural
NumAskMarks The number of ask marks in the comment
NumExclamationMarks The number of exclamation marks in the
comment
ArticleSimilar The similarity between the comment and the
Similar article
TitleSimilar The similarity between the comment and the
title
HelpfulnessCommentsSimilar The similarity between the comment and
other comments with the highest score of
helpfulness (the number of other comments
is 5)
RatioNouns The percentage of nouns in the comment
RatioVerb The percentage of verb in the comment
Syntactic
RatioAdj The percentage of adjective in the comment
RatioAdv The percentage of adverb in the comment
NumPosWord The number of positive words in the com-
Semantic
ment
NumNegWord The number of negative words in the com-
ment
NumThumbUps The number of thumb up votes in the com-
ment
MetaData
NumThumbDowns The number of thumb down votes in the
comment
HelpfulNess The helpfulness score of the comment
TimeLines The number of comments having been
posted before this comment

Table 2. AOL Corpus Statistics

Ave Max Min

#Tokens in Article 570.77 1558 162
#Comments 103.98 316 22
#Gold Standard Keyphrases 7.6 17 3

are used as evaluation metrics, where countcorrect is the total number of correct keyphrases
extracted by the extractor, countextrator is the total number of automatic extracted
keyphrases, and countgold is the total number of gold keyphrases.

The Keyphrase Extractors. Most state-of-the-art keyphrase extractors can be used

in our approach. Hasan conducted a systematic evaluation and analysed many unsu-
pervised keyphrase extraction methods on a variety of standard evaluation datasets.
He found that Tf-idf and SingleRank extractors give very robust performance across
 Improving Keyphrase Extraction from Web News 147

dierent datasets [4]. Therefore, we use these two keyphrase extractors in our experi-
ment. A publicly available implementation of these two extractors can be downloaded6.
Tf-idf Keyphrase Extractor assigns a score to every term in a document based on its
frequency (term frequency, tf) and the number of other documents containing this term
(inverse document frequency, idf). Given a document the extractor first computes the
Tf-idf score of each candidate word, then extracts all the longest n-grams consisting of
candidate words and scores each n-gram by summing the Tf-idf scores of its constituent
unigrams, and finally output the top N n-grams as keyphrases.
SingleRank Keyphrase Extractor is expanded from TextRank Keyphrase Ex-
tractor [10,14]. Both of these two extractors take a text represented by a graph, in
which each vertex corresponds to a word type and the edge connects two vertices if the
two words co-occur within a window of W words in the associated text. The goal is
to compute the score of each vertex, which reflects its importance. The two extractors
use the word types that correspond to the highest-scored vertices to form keyphrases
for the text. Some graph-based ranking algorithms such as Googles PageRank [1] and
HITS [6], which compute the importance of vertices, can be easily integrated into the
TextRank model and SingleRank model. There are some dierences between these two
extractors. First, each edge has a weight equal to the number of times the two corre-
sponding word types co-occur within a window of W words in a SingleRank graph,
while each edge has the same weight in a TextRank graph. Second, SingleRank only
uses high score sequence nouns and adjectives words to form keyphrases, but any high
score sequence words can form keyphrases in TextRank.

5.2 Evaluation of the Proposed Approaches

First we investigate all the comments information can help keyphrase extraction. We
integrate all comments into their related articles in our AOL corpus to form a new
dataset. We extract 20 keyphrases in each news document for evaluation using Tf-idf
Keyphrase Extractor and SingleRank Keyphrase Extractor respectively. Table 3
shows the result. We can see that, compared the approaches extracted keyphrases from
original article (Base Tf-idf and Base SingleRank), the approaches adding all com-
ments into the article (All Tf-idf and All SingleRank) decrease the performance of
extracting keyphrases. We conducted a paired t-test between the results of these meth-
ods and found statistically significant dierences (at p = 0.05). All these means simply
using all comments can not help for keyphrase extraction.

Table 3. Performance of All Comments Testing

Precision Recall F1
Base Tf-idf 0.139 0.366 0.202
All Tf-idf 0.153 0.259 0.192
Base SingleRank 0.120 0.316 0.174
All SingleRank 0.096 0.163 0.121

6
http://www.utdallas.edu/ ksh071000/code.html
148 Z. Luo, J. Tang, and T. Wang

Next we investigate some useful comments can help keyphrase extraction. We ranked
all the comments based on keyInfo function k. We use the gold keyphrases in each
article to calculate the accurate Numkeyphrase value for each comment. For each web
news document, we choose the top 15 comments with high value of k and integrate them
into the original article (called Oracle Tf-idf and Oracle SingleRank), which achieve
the best result of keyphrase extraction in our experiment (see Table 4). Moreover, they
are oracle tests and the results are significantly dierent (at p = 0.05). It means there
is a subset of useful comments, which can help for keyphrase extraction, in comments
part for each web news (see Figure 1).

Table 4. Performance of Useful Comments Testing

Precision Recall F1
Base Tf-idf 0.139 0.366 0.202
Oracle Tf-idf 0.165 0.432 0.239
Base SingleRank 0.120 0.316 0.174
Oracle SingleRank 0.146 0.382 0.211

At last we investigate the performance of our three comments selectors, choosing top
15 high score comments, for keyphrase extraction. The Similar Tf-idf (SingleRank)
and Helpfulness Tf-idf (SingleRank) methods used the measures of comments sim-
ilar and helpfulness score to choose the comments. To KeyInfo Tf-idf (SingleRank)
approach, we used 5 new AOL web news comments excluding AOL corpus to train a
regression model, which can predict the Numkeyphrase of each new comment. The 5 new
documents contain 524 comments and each document has gold standard keyphrases.
We extract all features in Table 1 to represent each comment as a feature vector for
training. The trained model can be used to predict the KeyInfo score of a new com-
ment. Table 5 gives the comparison results of various selecting methods for keyphrase
extraction. We can see that the performance of most methods integrating comments
outperform the Base Tf-idf(SingleRank) for keyphrase extraction. It shows comments
information indeed can help for keyphrase extraction. We conducted a paired t-test be-
tween the results and found all methods significantly improve the keyphrase extraction
(at p = 0.05) except Helpfulness Tf-idf. Especially the KeyInfo Tf-idf (SingleRank)
achieves the best result. It shows similar and helpfulness information are helpful for se-
lecting useful comments. Moreover, using more knowledge, extracted based on natural
language processing and machine learning technologies, can collect more useful com-
ments. All these shows exploiting comments information are eective for keyphrase
extraction.
We are interested in the detail of comments selected by these three selectors. We
compare three selected comment sets with the comments selected based on the accurate
Numkeyphrase value (see Oracle Test). All of comparisons use the top 15 high score
comments for each document. Table 6 gives the result. We can see that the KeyInfo
selector collect the largest most useful comments. It demonstrates our machine learning
approach is eective to find more comments with high function K value for keyphrase
extraction.
 Improving Keyphrase Extraction from Web News 149

Table 5. Performance of Extracting 20 Keyphrases

Precision Recall F1
Base Tf-idf 0.139 0.366 0.202
Similar Tf-idf 0.151 0.393 0.218
Helpfulness Tf-idf 0.144 0.379 0.209
KeyInfo Tf-idf 0.157 0.406 0.226
Base SingleRank 0.120 0.316 0.174
Similar SingleRank 0.134 0.324 0.190
Helpfulness SingleRank 0.133 0.332 0.190
KeyInfo SingleRank 0.144 0.331 0.201

Table 6. Useful Comments Sets Comparing Evaluation

Number
SetOracle SetKeyIn f o 437
SetOracle SetHelp f ulness 331
SetOracle SetS imilar 300

6 Conclusion
In this paper we propose a novel approach to use comments information for keyphrase
extraction from web news documents. Since comments are full of noisy and low-quality
data, simply integrating comment into an original article does not improve the perfor-
mance of this task. Therefore we give three strategies to select useful comments. The ex-
perimental results show that our comments selection approaches can improve keyphrase
extraction from web news. Especially our machine leaning approach which considers
many factors of comments can significantly improve this task.

Acknowledgements. This research is supported by the National Natural Science Foun-

dation of China (Grant No. 61170156 and 61202337) and a CSC scholarship.

References
1. Brin, S., Page, L.: The anatomy of a large-scale hypertextual web search engine. In: Proceed-
ings of the Seventh International Conference on World Wide Web 7, WWW7, pp. 107117.
Elsevier Science Publishers B. V., Amsterdam (1998)
2. Frank, E., Paynter, G.W., Witten, I.H., Gutwin, C., Nevill-Manning, C.G.: Domain-specific
keyphrase extraction. In: Proceedings of the Sixteenth International Joint Conference on Ar-
tificial Intelligence, IJCAI 1999, pp. 668673. Morgan Kaufmann Publishers Inc., San Fran-
cisco (1999)
3. Grineva, M., Grinev, M., Lizorkin, D.: Extracting key terms from noisy and multitheme doc-
uments. In: Proceedings of the 18th International Conference on World Wide Web, WWW
2009, pp. 661670. ACM, New York (2009)
4. Hasan, K.S., Ng, V.: Conundrums in unsupervised keyphrase extraction: making sense of the
state-of-the-art. In: Proceedings of the 23rd International Conference on Computational Lin-
guistics: Posters, COLING 2010, pp. 365373. Association for Computational Linguistics,
Stroudsburg (2010)
150 Z. Luo, J. Tang, and T. Wang

5. Hulth, A.: Improved automatic keyword extraction given more linguistic knowledge. In: Pro-
ceedings of the 2003 Conference on Empirical Methods in Natural Language Processing,
EMNLP 2003, pp. 216223. Association for Computational Linguistics, Stroudsburg (2003)
6. Kleinberg, J.M.: Authoritative sources in a hyperlinked environment. J. ACM 46(5), 604632
(1999)
7. Liu, J., Cao, Y., Lin, C., Huang, Y., Zhou, M.: Low-quality product review detection in opin-
ion summarization. In: Proceedings of the 2007 Joint Conference on Empirical Methods
in Natural Language Processing and Computational Natural Language Learning (EMNLP-
CoNLL), pp. 334342 (2007)
8. Liu, Z., Huang, W., Zheng, Y., Sun, M.: Automatic keyphrase extraction via topic decom-
position. In: Proceedings of the 2010 Conference on Empirical Methods in Natural Lan-
guage Processing, EMNLP 2010, pp. 366376. Association for Computational Linguistics,
Stroudsburg (2010)
9. Liu, Z., Li, P., Zheng, Y., Sun, M.: Clustering to find exemplar terms for keyphrase extrac-
tion. In: Proceedings of the 2009 Conference on Empirical Methods in Natural Language
Processing, EMNLP 2009, vol. 1, pp. 257266. Association for Computational Linguistics,
Stroudsburg (2009)
10. Mihalcea, R., Tarau, P.: Textrank: Bringing order into texts. In: Proceedings of EMNLP,
Barcelona, Spain, vol. 4, pp. 404411 (2004)
11. Mihalcea, R., Csomai, A.: Wikify!: linking documents to encyclopedic knowledge. In: Pro-
ceedings of the Sixteenth ACM Conference on Information and Knowledge Management,
CIKM 2007, pp. 233242. ACM, New York (2007)
12. Salton, G., Buckley, C.: Term-weighting approaches in automatic text retrieval. Inf. Process.
Manage. 24(5), 513523 (1988)
13. Turney, P.D.: Learning to extract keyphrases from text. CoRR cs.LG/0212013 (2002)
14. Wan, X., Xiao, J.: Single document keyphrase extraction using neighborhood knowledge. In:
Proceedings of the 23rd National Conference on Artificial Intelligence, AAAI 2008, vol. 2,
pp. 855860. AAAI Press (2008)
15. Xu, S., Yang, S., Lau, F.: Keyword extraction and headline generation using novel word
features. In: Proc. of the Twenty-Fourth AAAI Conference on Artificial Intelligence (2010)
16. Yano, T., Cohen, W., Smith, N.: Predicting response to political blog posts with topic models.
In: Proceedings of Human Language Technologies: The 2009 Annual Conference of the
North American Chapter of the Association for Computational Linguistics, pp. 477485.
Association for Computational Linguistics (2009)
A Two-Layer Multi-dimensional Trustworthiness
Metric for Web Service Composition

Han Jiao1 , Jixue Liu1 , Jiuyong Li1 , and Chengfei Liu2

1
School of Information Technologies and Math Sciences,
University of South Australia, SA 5095, Australia
2
Faculty of Information and Communication Technologies
Swinburne University of Technology, Melbourne, VIC3122, Australia
{han.jiao,jixue.liu,jiuyong.li}@unisa.edu.au, [email protected]

Abstract. Web service composition is a promising area to revolutionize

the collaboration among heterogeneous applications. Trust is recognized
as an important topic in this research area. In this paper, we propose
a two-layer multi-dimensional metric to represent the trustworthiness of
web service composition programs. We explored several dimensions for
both composition level and service level and finalize the metric, as an
indicator, to represent the trustworthiness of the automatic service com-
position. We also propose a method to transfer the multi-dimensional
metric into a single value for comparison purpose. A simulation experi-
ment shows that the proposed metric shows better results than the em-
pirical judgement based on past experiences.

Keywords: trust, trustworthiness, web service, composition, metric.

1 Introduction
The Web originates for enabling users the information sharing power. As evolv-
ing, it has been enriched to provide more flexible ways to link dierent sources,
one of which is to use web services, a set of related functionalities that can be
programmatically accessed through the web [16], usually via XML-based pro-
tocols. Web service composition, aiming at composing distributed web services
to carry out complex business processes, is considered to be a promising area
to revolutionize the collaboration among heterogeneous applications[3]. Due to
the constantly changing environment and the large scalability of the Web, the
research focus on web service composition is moving from static to dynamic
technology, known as Automatic Service Composition (ASC) [15]. The software
programs that provide ASC functionalities are termed ASC agents. An ASC
agent has to compose several web services to deliver results, which brings them
a few particular characteristics that are dierent from single-functionality web
services.
Researchers has realized the significance of trust in the web service environ-
ment. Methods, models and frameworks are proposed to help service consumers
identify trustworthy web services [17]. However, determining the trustworthiness

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 151162, 2013.
c Springer-Verlag Berlin Heidelberg 2013

152 H. Jiao et al.

for ASC agents is left as an open issue. We adopt the major consideration in
[6] by defining the trustworthiness of ASC agents to be a two-layer structure,
namely composition-layer and service-layer. This caters to the unique character-
istics owned by ASC agents that their deliveries depend on not only their own
activities, but also the single-functionality web services they compose.
The previous work did not give a complete investigation on how to use QoS
factors to represent the trustworthiness of an ASC agent. The conducted ex-
periment cannot reflect the eectiveness of the proposed model. In this paper,
we will thoroughly explore the relationships between trustworthiness and QoS
factors. A proper set of indicators, based on QoS factors, will be constructed
that can better present ASCs trustworthiness. We will also use experiment to
prove that our proposed metric shows better result than result of the forecast
simply using historical data. The contributions of this paper are:

We explore the relationships between trustworthiness QoS factors for ASC

agents. And compare them with their counterparts in normal web services.
We design a metric of Trustworthiness Status Indicators (TSIs) that can be
used to evaluate the trustworthiness of ASC agents. The metric is open and
customizable. Users can config it to cater for their own preferences.
We provide a mathematic method to transfer the multi-dimensional TSI
metric into single-valued format to ease comparison and evaluate the method
using experiments.

The rest of the paper is structured as follows. In Section 2, we explore the rela-
tionship between the trustworthiness of ASC agent and QoS factors. In Section
3, we present trustworthiness metric in detail. A method to integrate TSIs for
trustworthiness comparison is introduced in Section 4. A simulation experiment
is presented in Section 5. Section 6 gives the recent related works and Section 7
summarizes the paper and looks at the future work.

2 Trust-Related Concepts and QoS Factors

In this section, we clarify the concept of trust and trustworthiness. We will

also discuss the commonly agreed dimensions of trustworthiness and what QoS
factors can be used to represent those trustworthiness dimensions.

2.1 Trust and Trustworthiness

Trust and trustworthiness are dierent. Trust is the willingness of a trustor to

believe that the trustee will behave as expected. This willingness makes the
trustor confident to be vulnerable to some trustees behaviors. Dierent from
trust, trustworthiness is a trustees attribute to represent the extent that it can
be trusted [5]. In other words, a trustor gives its trust to a trustee based on the
trustworthiness of the trustee. Strictly speaking, trust is a mental activity. People
are not necessarily to give their trust based on trustworthiness. In reality, there
 A Two-Layer Multi-dimensional Trustworthiness Metric 153

are people that irrationally trust others without any reasons. However, rational
people give their attitude based on gathered information on trustworthiness.
In this paper, the ASC requesters are trustors and the ASC agents are the
trustees. Instead of directly modeling trust, we design a metric to represent the
trustworthiness of ASC agents, which has impact on the trust from requesters.

2.2 Trustworthiness Dimensions and QoS Factors

In [11,9], three dimensions of trustworthiness were studied. They are ability,
benevolence and integrity. This taxonomy is agreed by most researcher since
proposed, for example in [4,5,12]. Since the process of automatic service com-
position is entirely computer-controlled and a computer software does not have
emotions to act altruistically or kindheartedly, it makes no senses to measure
the dimension of benevolence in this paper. The concept of trustworthiness only
covers the first two dimensions: ability and integrity.
As we mentioned, ASC agents own the speciality that their deliverables are
based on their composition ability as well as the single-functionality web service
they use. When we investigate the trustworthiness of an ASC agent, we have
dierent concerns on these two sources. At composition layer, people implicitly
make an assumption that they believe the ASC agent has the ability to complete
the required tasks. What they concern is that whether their actual performance
can be really as good as what they claim to be. In other words, people concerns
the deviation between the actual action with the advertised one. At single service
layer, people know that even the ASC agent cannot control the single services,
because they are provided by third parties. So people concern whether an ASC
agent have enough qualified resources to compose. The selected QoS factor at
service layer should be able to reflect the general conditions of the resource pool
maintained by the ASC agents.
Based on the above discussion, we design the following trustworthiness metric
consisting of 4 aspect at each layer. Each aspect is represented by a Trustworthi-
ness Status Indicator (TSI) to distinguish it from the basic corresponding QoS.
At composition layer, we have: 1)Functional Eectiveness Deviation, 2)Execu-
tion Eciency Deviation, 3)Availability Deviation, and 4)Reputation. At service
layer, another four TSIs are taken into account: 1)Average Execution Duration,
2)Average Availability, 3)Average Reliability, and 4)Service Monitoring. Please
note that at composition layer, we concern deviations between the promises
and the actual performance. At service layer, we concern the general quality of
the maintained resource pool. In the next section, we present all the TSIs in
detail.

3 Trustworthiness Indicators
3.1 Composition-Layer TSIs
Functional Eectiveness Deviation (Tf ed ). This TSI represents the sta-
tus of functional eectiveness. To define the functional eectiveness deviation,
154 H. Jiao et al.

we first need to define functional error rate. Functional error rate is the pro-
portion of requests answered with functional errors over the total number of
requests. Given an ASC agent c and a certain period du, it is calculated as
(c, du) = NNtotal
err
, where Nerr is the total number of requests answered in error
and Ntotal is the total number of composition requests. The functional eective-
ness deviation Tf ed (c, du) represents the dierence between the actual functional
error rate a (c, du) and the claimed one c (c, du) within that duration. It is cal-
culated as Tf ed (c, du) = max{0, a (c, du) c (c, du)}. In practice, a (c, du) can
be monitored by consumers or a third party and c (c, du) is claimed by the ASC
agent itself. The lower the Tf ed (c, du) is, the more trustworthy the ASC agent
is. In general, this TSI reflects how trustworthy the ASC agent is in the aspect
of functional capability.

Execution Eciency Deviation (Teed ). This TSI indicates the status of the
execution eciency of an ASC agent. For an ASC agent c within duration du, the
execution eciency deviation Teed (c, du) is the dierence between the claimed
on-time response rate c (c, du) and the actual on-time response rate a (c, du).
On-time response rate is the proportion of on-time response times over the to-
tal times an ASC agent should response, i.e., = NNontime
total
. Therefore, Teed (c, du)
is calculated as Teed (c, du) = max{0, c (c, du) a (c, du)}. In general, execution
eciency deviation stands for how trustworthy the ASC agent is in the aspect
of on-time response.

Availability Deviation (Tad ). In [14,17], availability is defined to be the pro-

portion of the available duration in the total duration. Based on those defini-
tions, we define availability deviation Tad (c, du) to be the dierence between
the claimed availability and the actual availability for ASC agent c in duration
du. It is dicult to make real-time availability monitoring of an ASC agent.
One possible solution is to change this real-time continuous monitoring to
randomly discrete monitoring. For every time the ASC agent is called by con-
sumers, its availability information is recorded. After a certain period of time
du, the actual availability of an ASC agent c can be represented by the propor-
tion of the available times in the total accessing times, i.e., = Navailable
Ntotal . Then,
we use the following formula to calculate availability deviation: Tad (c, du) =
max{0, c (c, du) a (c, du)}. Same as the above two TSIs, the lower the avail-
ability deviation is, the more trustworthy the ASC agent is.

Public Reputation (Tpr ). The public reputation of an ASC agent is a mea-

sure of its trustworthiness from public opinions. There is no doubt that in an
open environment, dierent peoples opinions will impact each other. Dier-
ent consumers may have dierent opinions on the same ASC agent. Our public
reputation is defined as the average ranking scored"by consumers. This can be
n
i
represented by the following formula: Tpr (c, du) = i=1
n , where i is the score
of ranking given by consumer i in the last duration du and n is total number of
consumers who give scores in that duration. To simplify the calculation in our
 A Two-Layer Multi-dimensional Trustworthiness Metric 155

framework, we define the range of ranking score is [0,1]. Therefore, the range of
Tpr (c, du) is also [0,1]. Dierent from the above three TSIs, the higher the public
reputation is, the more trustworthy the ASC agent is.
The above four TSIs are composition-layer TSIs that reflect the trustwor-
thiness of an ASC agent itself. In the next sub-section, we will introduce sev-
eral service-layer TSIs which reflects the status of candidate elementary services
maintained by an ASC agent. Their properties also impacts the final deliverables
of ASC agents and therefore, impact the trust consumers give.

3.2 Service-Layer TSIs

Average Execution Duration (Taed ). Given an ASC agent c, the average

execution duration Taed (c, du) is the average processing time of the elementary
services
"n that ASC agent c calls during the last duration du, i.e., Taed (c, du) =
i=1 N orm(ti )
n , where ti is the execution duration of the ith call to an elemen-
tary service and n is the total number of calls in duration du. To simplify the
future calculation, we would like to normalize ti into range [0,1] by giving a
value of maximum tolerable time tmax , which is the largest execution time that
consumers would like to tolerate, and a minimum satisfactory time tmin , which
is the execution time that consumers are 100% satisfied with (consumers will
not get more satisfaction even the web service is faster than that). These two
reference values can be determined by consumer surveys. Here they are treated
as environmental constants. If the average execution duration exceeds tmax , the
score from this aspect deserves 1 point. If it is lower than tmin , the score is 0. If it
ti tmin
is in between, we use the following formula to calculate ti : N orm(ti ) = tmax tmin
and the final value is ranging from [0,1]. Therefore, Taed will be in the-less-the
better family. In general, this TSI represents how fast generally the elementary
services can complete their tasks. The less it is, the better the overall quality of
elementary services is.

Average Service Availability (Tasa ). Average service availability Tasa (c, du)
is the average value of the availability of each elementary"nservice maintained in
Avaii
ASC agent c in the last duration du, i.e., Tasa (c, du) = i=1n , where Avaii
is the availability of the ith elementary service and n is the total number of ele-
mentary service. The definition of the availability of a single elementary service
can be referred to [14]. The higher the average service availability is, the better
the quality of elementary services is. Generally, average service availability re-
flects the overall quality of elementary services in the aspect of on-line available
duration.

Average Service Reliability (Tasr ). Reliability is defined as the probability

that a request is correctly responded within the maximum expected time frame
indicated in the web service description. Therefore, the reliability of service s in
nresp
duration du can be calculated as Rel(s, du) = ntotal , where nresp is the num-
ber of correct responses and ntotal is the number in total. Our average service
156 H. Jiao et al.

reliability Tasr (c, du) is just the average value

"n
of all the maintained elementary
Rel(s ,du)
services in ASC agent c, i.e., Tasr (c, du) = i=1 n i , where n is the total
number the elementary services in the service pool of ASC agent c. Same as av-
erage service availability, the higher the average service reliability is, the better
the quality of elementary services is. Average service reliability stands for the
overall quality in the aspect of correctly on-time responding.

Monitoring Strength (Tms ). Monitoring strength stands for the strength that
elementary services are monitored by ASC agents or a third party. It is reason-
able to think that if elementary services are monitored regularly, their services
will be of consistent quality. Two factors are considered to have impact on the
monitoring strength: monitoring coverage mc and monitoring frequency mf .
Monitoring coverage is the proportion of the monitored services over all the ser-
vices. Its range is [0,1]. Monitoring frequency stands for how often the monitoring
session is. To simplify the calculation, we normalize monitoring frequency by giv-
ing satisfactory monitoring frequency smf as a reference so that the range of
mf becomes [0,1]. The satisfactory monitoring frequency is a frequency in which
if elementary services are monitored, consumers will consider the monitoring to
be satisfactory. This reference value can be decided by conducting consumer sur-
vey. Once it is decided, it keeps constant and seldom changes. For a certain ASC
agent c in duration du, " we use the following formula to calculate Monitoring
n
mc mf
Strength: Tms (c, du) = i=1n i M in{ smf , 1}, where mci is the monitoring
coverage of the ith monitoring process. The higher the monitoring strength is,
the better the quality that it represents is.
In summary, we proposed several composition-layer and service-layer TSIs,
which form a comprehensive metric, reflecting the trustworthiness status of an
ASC agent. As mentioned before, it is impossible to give complete metric that
shows every aspects of a composer, since the properties are countless and dierent
contexts may concern dierent things. Our metric is extensible so that consumers
can add or remove TSIs as they preferred, as far as the new added ones can be
scaled into the range [0,1]. In the next part, we will integrate all the TSIs together
to form a single value through which the trustworthiness of ASC agents can be
comprehensively distinguished.

4 Trustworthiness Comparison

The trustworthiness of ASC agent c in the last duration du can be represented

by the follow vector: (Tf ed (c, du), Teed (c, du), Tad (c, du), Trp (c, du), Taed (c, du),
Tasa (c, du), Tasr (c, du), Tms (c, du)). To simplify the representation, the eight
TSIs are numbered from 1 to 8 in the order as shown. Then, the above vec-
tor can be re-written as {Tj (c, du)|j = 1, 2, ..., 8}. Given an ASC agent ci , the
jth TSI T SIi,j can be represented as T SIi,j = Tj (ci , du). We adopt simple addi-
tive weighting technique to help transfer the multi-dimensional TSI metric into
a single value. This technique has two phases: normalization and weighting.
 A Two-Layer Multi-dimensional Trustworthiness Metric 157

Table 1. Settings for ASC agents

ASC agent Avai. Exec.Drtn Func.Err.Rate Moni.Cov. Moni.Freq.

Agent01 75% 40ms 15% 30% 1/4
Agent02 77% 36ms 14% 40% 1/4
Agent03 79% 32ms 13% 40% 1/4
Agent04 81% 28ms 12% 50% 1/2
Agent05 83% 24ms 11% 60% 1/2
Agent06 85% 20ms 10% 70% 1/2
Agent07 87% 16ms 09% 80% 2/3
Agent08 89% 12ms 08% 80% 2/3
Agent09 91% 8ms 07% 90% 4/5
Agent10 93% 4ms 06% 100% 1

4.1 Normalization Phase

The eight TSIs can be grouped into two, as four of them (No. 4,6,7,8) are the
bigger the better and the rest are entirely opposite. The normalization processes
for these two types of TSIs are dierent. Given n ASC agents in total, the ith
TSI of the jth ASC agent can be normalized by the following formula:
T SI max T SIi,j


i
T SIimax T SIimin
if j=1,2,3,5



and T SIimax = T SIimin ;




N
vi,j = T SIi,j T SIimin (1)
T SIimax T SIimin
if j=4,6,7,8



and T SIimax = T SIimin ;





0 if T SIimax = T SIimin

In the above formula, T SIimax and T SIimin are respectively the highest and low-
est values of the ith TSI of all the ASC agents, i.e., T SIimax = M ax(T SIi,j ), 1
j n and T SIimin = M in(T SIi,j ), 1 j n. By applying the above formula,
we obtain a 8 n matrix V N ,

v1,1 v1,2 ... v1,8
v2,1 v2,2 ... v2,8

v3,1 v3,2 ... v3,8 (2)

... ... ... ...
vn,1 vn,2 ... vn,8
where each column stands for one TSI dimension and each row represents an
ASC agent.

4.2 Weighting Phase

Even though we can normalize the TSI values of all the ASC agents into one
matrix, we still cannot judge which ASC agent is more trustworthy, since one
158 H. Jiao et al.

ASC agent may be better in some TSIs and the other may be better in other
TSIs. This problem is related to how important each TSI is in the context that
the comparison occurs. There are no fixed rules saying what TSI is the most
important one and what is the second one. The significance of each TSI depends
on user preference.
#8 method, users are given an 1 8 matrix
In our trustworthiness comparison
W = (wj |1 j 8, 0 wj 1, j=1 wj = 1) to allocate weight values for each
TSI, where element wj is the weight value for the jth TSI we described before.
Given an 8 n normalized TSI matrix V N = (vi,j N
|1 i n, 1 j 8), by
applying the following formula, an 1 n matrix T will be achieved, in which
each value Ti stands for trustworthiness of an ASC agent.

v1,1 v1,2 ... v1,8 w1
v2,1 v2,2 ... v2,8 w2

T =V W =
N
v3,1 v3,2 ... v3,8 w3 (3)
... ... ... ... ...
vn,1 vn,2 ... vn,8 wn

The above formula results in n Tj , each of which is the final normalized trust-
worthiness for the jth ASC agent. That value is ranging in [0,1]. The higher
the value is, the more trustworthy the ASC agent is. Please note that if users
only want to calculate an absolute trustworthiness value for an ASC agent (no
comparison), they can directly multiple weight values with the result calculated
from the formulas given in TSI description. In that case, normalization is not
necessary.
As we mentioned, it is very dicult to know which TSI is more important.
However, there are some general rules that can be applied to any contexts. The
first rule is eort-oriented division, which means that we put more weight to
where the major computation jobs are done. If you know most of your work
will be done by elementary services and the composer is just for integrating
the results, you should put more weight on service-layer TSIs, and vice versa.
The second rule is diculty-driven. We put more weight to the TSIs that are
hard to achieve. For example, if you have a massive computation task in which
computation logic is very simple but the amount is quite large, rather than
putting too much weight on the TSIs reflecting the ability factors, you should
give more attention on the quality factors such as execution eciency. If the
computation logic is very complex, you should focus on ability factors such as
Functional Eectiveness Deviation. The third rule is concern-driven. If you are
quite keen in certain aspects, e.g., in some time-crucial cases where even one
microsecond determines everything, you should definitely put the highest weight
to the corresponding factors. Again, our metric is open and it accepts any new
TSIs as its member, as far as they can be scaled to [0,1]. Users are allowed to
create their own TSIs based on their contexts.
 A Two-Layer Multi-dimensional Trustworthiness Metric 159

Fig. 1. Trustworthiness Evolution

5 Experiments and Simulation

In this section, we present a simulation experiment to illustrate the eectiveness
of our proposed TSI metric. We want to test that given several ASC agents
with the same claimed capabilities status and dierent actual capabilities, the
higher their calculated trustworthiness values are, the better the actual Eligible
Return Rate (ERR) they give. Here ERR stands for the proportion of the eli-
gible returned results over all the results. The eligible returns are the on-time
return with valid composition results. An ASC agent giving an eligible return
means it successfully delivered its job without any flaws. We use this indicator
to represent the actual level of ASC agents. The whole experiment is conducted
under Microsoft Oce Excel 2007 Professional with Visual Basic for Application
(VBA) as the programming language.
Suppose that there is a business process that needs to composing three types of
elementary web services. For each types of function, there are 500 elementary web
services available on the web. Their availability follows the normal distribution
with mean value of 90% and variance of 10%. Same probability distributions are
also followed by their reliability with mean value of 85% and variance of 10% and
by actual execution duration with mean value of 20 milliseconds and variance of
5 milliseconds. ASC agents regard web services that can return results in 10ms
as perfect and the web services that exceed 35ms to return as non-tolerable.
There are 10 available ASC agents which claim to oer the corresponding
composition service. They claim to provide results with 95% score for TSI 1, 2
and 3 and 100% for TSI 8. However, their actual capabilities are shown in the
160 H. Jiao et al.

above table. Except the MC and MF data, the rest data in the table are just
the average values. The performance of ASC agents follows a normal distribu-
tion with the mean values given in the table and 1% variance for availability, 5
milliseconds variance for execution duration and 0.5% for error rate. 200ms is
the longest duration consumers would like to wait for one composition job. Each
ASC agent maintain their own elementary service pool that consists of 60 web
services in total with 20 for each function. Those services are randomly selected
from the web. The ideal monitoring strength in our experiment is suppose to
check service pool every 10 compositions with coverage 100%.
We simulate the composition processes for each ASC agents for 300 rounds.
In each round, for a certain ASC agent, we first check whether it is available by
comparing a even-distribution random number with its availability that follows
normal distribution. Then we use same method to check whether the composition
has errors. After that, the ASC agent has to select three elementary web service
with dierent functionalities. If a selected web service cannot work properly
(not available, not reliable or execution duration is too long), ASC agent has to
select another elementary service with the same# functionality. The final execution
n
duration of ASC agent is calculated as Dasc = i=1 Tiws + (n 2) Tasc, where
n is the total times that the WSC interact with dierent web services, Tiws
is the execution duration of ith web service and Tasc is the duration. If the
final total execution duration is greater than 200ms, the composition is regarded
as not ecient. Until now, we finish calculation of the composition-layer TSIs.
The calculation of the first three service-layer TSIs is done based on the current
services maintained in each ASC agent. Then based on the monitoring frequency,
we decide whether a monitoring process must be started after the current round.
If a monitoring process has to be conducted, we will randomly select several
web services for each function type according to monitoring coverage and check
their availability, reliability and execution duration. The web services that give
ineligible results will be removed. In those cases, new services whose quality is
better than the removed ones will be added in from the web. A web service that
was once selected by a certain ASC agent is not allowed to be added again. After
each round, a trustworthiness value (equal weight for TSIs) and a ERR value of
each ASC agent are calculated based on the previous composition records. The
results are shown in Fig.1.
Fig.1(a) and (b) give the general evolution curves of trustworthiness values
and eligible returning rate (ERR) for each ASC agent. In general, as we ex-
pected, the better actual capability an ASC agent has, the better its ERR and
trustworthiness are. This shows that our calculated trustworthiness result eec-
tively reflect the actual performance. We note that at the first several rounds,
the trustworthiness curves and ERR curves fluctuate drastically and can not
reflect the actual capability status. This is because the performance of the ASC
agents and elementary services are of random nature (normal distributions in
the settings). As the experiment goes, our calculated trustworthiness values be-
come stable much faster than ERR. It tells that our selected TSIs and trust-
worthiness calculation method have better sensitivity of reflecting the actual
 A Two-Layer Multi-dimensional Trustworthiness Metric 161

performance than ERR which purely reflects capability by following previous

historical records. This is more clear in Fig.1(c) and Fig.1(d). (c) give the trust-
worthiness and ERR status after 50 rounds. At that time ERR was still fluc-
tuating, while trustworthiness values had already correctly shown the actual
capability comparison. In Fig.1(d) after 150 rounds when ERR gives general ca-
pability status, our trustworthiness values are able to present clear and precise
dierences among the ASC agents.

6 Related Works
Research on trust in the online environment has been studied mainly in the
information systems domain and computer science domain.
In the information systems domain, people concerns the fundamental concep-
tualization and the implication of each concepts. The seminal work that clarify
the subtilty includes [9,4,12,5]. There is a branch of trust research which focuses
on analyzing the institutional trust elements and their functions [18,10,13]. Re-
cently, some brain visualization technologies are used to analyze trust reactions
[8,2]. Basically, the trust research in information systems focuses on exploring
the nature and roles of trust and use dierent methods to analyze the impact of
peoples trustworthy and untrustworthy behaviors. Their research is valuable in
guiding people and organizations to form trust relationships.
Researchers from computer science paid more attention on trust modeling and
trust framework construction. There have been some very good survey papers in
this field [17,7,1]. Details cannot be listed due to the limit of the required page
size.

7 Conclusions and Future Work

In this paper, we clarified several trust-related concepts and explore the dimen-
sions, manifestations and formation of trustworthiness. Based on the special
characteristics of automatic web service composition, we proposed a combined
composition-layer and service-layer trust status indicator metric to calculate the
trustworthiness of ASC agents. For comparison purposes, we use simple additive
weight method to transfer the multi-dimensional TSI metric into a single trust-
worthiness value. We use experimental simulation to prove that our proposed
metric and the comparison method can eectively dierentiate the trustworthi-
ness of ASC agents. They also present a better performance than the method
that purely use historical data to judge trustworthiness.
Our next step will focus on implementing the proposed structure into a soft-
ware framework and doing experiments with real data to determine the reason-
able weights for TSIs under dierent situations. Exploring new service-layer and
composition-layer trustworthiness factors to enrich the framework is also a job
on our schedule.
162 H. Jiao et al.

References
1. Artz, D., Gil, Y.: A survey of trust in computer science and the Semantic Web.
Journal of Web Semantics: Science, Services and Agents on the World Wide
Web 5(2), 5871 (2007)
2. Dimoka, A.: What Does the Brain Tell Us about Trust and Distrust? Evidence
From A Functional Neuroimaging Study. MIS Quarterly 34(2), 373396 (2010)
3. Dustdar, S., Schreiner, W.: A Survey on Web Service Composition. International
Journal of Web and Grid Services 1(1), 130 (2005)
4. Gefen, D.: Reflections on the Dimensions of Trust and Trustworthiness among
Online Consumers. SIGMIS Database 33(3), 3853 (2002)
5. Gefen, D., Benbasat, I., Pavlou, P.: A Research Agenda for Trust in Online Envi-
ronments. Journal of Management Information Systems 24(4), 275286 (2008)
6. Jiao, H., Liu, J., Li, J., Liu, C.: Trust for Web Service Composers. In: The
12th Global Information Technology Management Association World Conference,
GITMA, pp. 180184 (2011)
7. Jsang, A., Ismail, R., Boyd, C.A.: A survey of trust and reputation systems for
online service provision. Decision Support Systems 43(2), 618644 (2007)
8. King-Casas, B., Tomlin, D., Anen, C., Camerer, C.F., Quartz, S.R., Montague,
P.R.: Getting to Know You: Reputation and Trust in a Two-Person Economic
Exchange. Science 308(5718), 7883 (2005)
9. Mayer, R.C., Davis, J.H., Schoorman, F.D.: An Integrative Model of Organizational
Trust. The Academy of Management Review 20(3), 709734 (1995)
10. McKnight, D.H., Chervany, N.L.: What Trust Means in E-Commerce Customer
Relationships: An Interdisciplinary Conceptual Typology. International Journal
on Electronic Commerce 6(2), 3559 (2001)
11. McKnight, D.H., Choudhury, V., Kacmar, C.: Developing and Validating Trust
Measures for e-Commerce: An Integrative Typology. Info. Sys. Research 13(3),
334359 (2002)
12. Pavlou, P.A., Dimoka, A.: The Nature and Role of Feedback Text Comments in
Online Marketplaces: Implications for Trust Building, Price Premiums, and Seller
Dierentiation. Information Systems Research 17(4), 392414 (2006)
13. Pavlou, P.A., Tan, Y.-H., Gefen, D.: The Transitional Role of Institutional Trust in
Online Interorganizational Relationships. In: 36th Hawaii International Conference
on Sytem Sciences, HICSS 2003, p. 215a (2003)
14. Ran, S.: A Model for Web Service Discovery with QoS. ACM SIGecom Ex-
changes 4(1), 110 (2003)
15. Rao, J., Su, X.: A Survey of Automated Web Service Composition Methods.
In: Cardoso, J., Sheth, A.P. (eds.) SWSWPC 2004. LNCS, vol. 3387, pp. 4354.
Springer, Heidelberg (2005)
16. Tsur, S., Abiteboul, S., Agrawal, R., Dayal, U., Klein, J., Weikum, G.: Are Web
Services the Next Revolution in e-Commerce. Paper Presented at the The 27th
International Conference on Very Large Data Bases, VLDB 2001 (2001)
17. Wang, Y., Vassileva, J.: A Review on Trust and Reputation for Web Service Selec-
tion. In: 27th International Conference on Distributed Computing Systems Work-
shops, ICDCSW 2007, pp. 2533 (2007)
18. Zucker, L.G.: Production of trust: Institutional sources of economic structure, 1840-
1920. Research in Organizational Behavior 8(1), 53111 (1986)
 An Integrated Approach
for Large-Scale Relation Extraction from the Web

Naimdjon Takhirov1, Fabien Duchateau2 , Trond Aalberg1, and Ingeborg Slvberg1

1
Norwegian University of Science and Technology, 7491 Trondheim, Norway
{takhirov,trondaal,ingeborg}@idi.ntnu.no
2
Universit Lyon 1, LIRIS, UMR5205, Lyon, France
[email protected]

Abstract. Deriving knowledge from information stored in unstructured docu-

ments is a major challenge. More specifically, binary relationships representing
facts between entities can be extracted to populate semantic triple stores or large
knowledge bases. The main constraint of all knowledge extraction approaches is
to find a trade-off between quality and scalability. Thus, we propose in this pa-
per SPIDER, a novel integrated system for extracting binary relationships at large
scale. Through series of experiments, we show the benefit of our approach, which
in general, outperforms existing systems both in terms of quality (precision and
the number of discovered facts) and scalability.

Keywords: Relation Extraction, Knowledge Bases, Web Mining.

1 Introduction

Information available on the Web has the potential to be a great source of structured
knowledge. However this potential is far from being realized. The main benefit of ob-
taining exploitable facts such as relationships between entities from natural language
texts is that machines can automatically interpret them. The automatic processing en-
ables advanced applications such as semantic search, question answering and various
other applications and services. The Linked Open Data (LOD) aims at making the
vision of creating a large structured database a reality. In this domain, the building
of semantic knowledge bases such as DBpedia, MusicBrainz or Geonames is (semi-
)automatically performed by adding new facts which are usually represented by triples.
However, most of these triples express a simple relationship between an entity and one
of its properties, such as the birthplace of a person or the author of a book. By mining
structured and unstructured documents from the Web, one can provide more complex
relationships such as parodies. A different vision known as the web of concepts shares
similar objectives with LOD [12]. As a consequence, knowledge harvesting [9,10,14]
and more generally open-domain information extraction [5] are emerging fields with
the goal of acquiring knowledge from textual content.
In this paper, we propose a relation extraction approach named SPIDER1 . It aims at
addressing the previously mentioned issues by integrating the most relevant techniques
1
Semantic and Provenance-based Integration for Detecting and Extracting Relations.

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 163175, 2013.
c Springer-Verlag Berlin Heidelberg 2013
164 N. Takhirov et al.

to generate trustworthy patterns and relationships. SPIDER is based on a perpetual pro-

cess of generating patterns and examples either in supervised or unsupervised mode.
The main intuition is that generic patterns which are derived from similar sentences and
which are discovered in trustworthy documents are useful for detecting relationships in
other documents. In summary, our contributions in this paper are:
Extracting relationships from a Web-scale source is a major bottleneck. To the best
of our knowledge, SPIDER is the first approach which does not require several days
to perform a single iteration.
Contrary to most knowledge extraction tools, we tackle the problem of uniquely
identifying entities to extend their list of spelling forms and to facilitate the match-
ing to LOD.
SPIDER includes a flexible pattern definition scheme. This scheme is used to merge
similar patterns for efficiency purposes. In addition, we introduce the notion of
confidence score that controls the ranking of patterns. The confidence score evolves
over time as the system runs continuously.
Experiments confirm the benefits of our approach w.r.t. similar tools (ReadTheWeb,
Prospera) in terms of quality and performance.

2 Overview
2.1 Problem Definition
The overall goal of SPIDER is to continuously generate relationships and patterns. Let
us first define a relationship. It is a triplet <e1 , , e2 > where e1 and e2 represent entities
and stands for a type of relationship. An entity might be represented with different
labels in natural language text, and we note le L one of the mentions for the entity
e. An example of a relationship is createdBy between the work entity The Lord of the
Rings and the person entity J.R.R. Tolkien. Note that both the entities and the type
of relationship are uniquely identified using an URI. An example denotes a pair of
entities (e1 , e2 ) which satisfies a type of relationship.
The patterns are extracted from a collection of documents D = {d1 , d2 , ..., dn }.
Although not limited to, these documents are webpages in our context. Each document
is composed of sentences, which may contain mentions of the two entities. In that case,
the sentence is extracted as a candidate pattern. We note CP the set of candidate
patterns given a collection of documents and a set of initial examples. Each candidate
pattern is defined as a tuple cp = {tb , e1 , tm , e2 , ta } with tb , tm and ta respectively
standing for the text before, the text in the middle and the text after the entities. A
sentence Bored of the Rings is a parody of Lord of the Rings is transformed to a
corresponding candidate pattern {, e1 , is a parody of, e2 , }.
From the set of candidate patterns CP, we derive a set P of generic patterns by ap-
plying a strategy s. A strategy is defined as a sequence of operations s =< o1 , o2 , ...,
ok >, each operation aiming at generalizing the candidate patterns. Namely, this gen-
eralization implies the detection of frequent terms and the POS-tagging of the other
terms of the candidate patterns. Thus, a strategy is a function such that s(CP) P.
All generated patterns are associated to a specific type of relationship . For instance,
 An Integrated Approach for Large-Scale Relation Extraction from the Web 165

a generic pattern for the parody type is illustrated with {e1 } is/VBZ a/DT {parody,
illusion, spoof} of/IN {e2 } 2 .
Finally, a pattern and an example both have a confidence score noted confp and
confe respectively. This score is based on the support, the provenance, the number of
occurrences, the number of strategies and the iteration. A pattern similarity metric in-
dicates the proximity between two (candidate) patterns. The notion of confidence score,
as well as the one for operation, strategy, pattern similarity and (candidate) pattern, are
further detailed in the next sections.

2.2 Workflow

Given two labels, SPIDER generates patterns and derives to a relationship. This pattern
generation capability is guaranteed by the following two processes:
Pattern Generation is in charge of detecting candidate patterns by using examples
or provided entities and of generalizing these candidate patterns to obtain patterns
for a given type of relationship (see Figure 1).
Example Generation exploits the previously generated patterns in order to discover
new examples which satisfy the type of relationship.
The knowledge base stores all generated examples and patterns. These examples can
be used to maintain the system continuously running, but they can be exploited from a
user perspective too.

3 Pattern Generation

The pattern generation process either requires a few examples for a given type of re-
lationship so that patterns for this type of relationship can be automatically generated
(supervised), or it directly tries to guess the type of relationship for two given labels
(unsupervised). The process is similar in both modes and it is composed of three main
steps: extension of entities, extraction of candidate patterns from the collection of doc-
uments and their refinement into patterns.

3.1 Extending Entities

In a document, entities are not uniquely identified by a label but they have alternative la-
bels or spelling forms. Therefore, extending these entities with their alternative labels is
a crucial step and it requires the correct identification of the entity. For instance, the en-
tity Lord of the Rings can be labeled LOTR or The Lord of the Rings. To avoid
missing potentially interesting relationships, we search for these alternative forms of
spelling in the documents. Given an entity e represented by a label l, the goal is to dis-
cover its set of alternative labels Le = {l, l1, l2 , ..., ln }. The idea is to match the entity
against LOD semantic knowledge bases to obtain this list of alternative labels. Namely,
2
VBZ=Verb, 3rd person sing. present, DT=Determiner, IN=Preposition or subordinating con-
junction.
166 N. Takhirov et al.

LOTR {e2} is a parody of {e1}

LotR a spoof of {e1}, entitled{e2}
.
Lord_of_the_Rings {e2} is a short satirical novel by
lord of the rings
l11 TLotR parodying {e1}
l12
e1 ...
l1m
DOCUMENT candidate generic ranked
EXTENSION GENERALIZATION SELECTION
EXTRACTION patterns patterns patterns
e2 l21
l22
...
l2n {e2} is/VBZ a/DT parody of/IN {e1}
.
Simple strategy a/DT spoof of/IN {e1}, entitled/VBD {e2}
DBpedia Freebase
Contextual
OpenCyc strategy
collection of documents

Fig. 1. The Pattern Generation Process

we build various queries by decomposing the initial label and we query in the aliases at-
tributes of knowledge bases (i.e., common.topic.alias for Freebase, wikiPageRedirects
for DBpedia, etc.). In most cases, several candidate entities are returned and the tool
tries to automatically select the correct one.
The process of automatically selecting the correct entity is achieved as follows. First,
an AND query is constructed with the two labels. Clusters of documents are built rep-
resenting documents belonging to a set of specific type of entities. The n number of
words around labels are extracted and stemming performed on words. Our assumption
is based on the fact that documents mentioning the same entities tend to have similar
words. Therefore, a graph of semantically related words is built. The most important
documents in the cluster are then compared against the abstract of the automatically
selected entities. Next, we extract frequent terms from the most important documents
in the result set and use these frequent words as extensions. Note that if disambiguation
is not possible, we discard the example and we do not use it for subsequent pattern gen-
eration. The result of the extension process is a list of alternative labels as illustrated in
Figure 1.
The main issue in this step deals with the absence of the two labels in any knowl-
edge base, which means that the entities cannot be extended. The number of retrieved
documents in that case could not be sufficient to extract good candidate patterns. The
first solution consists of analyzing these retrieved documents to detect potential alter-
native spellings by applying metrics such as tf-idf and Named Entity Recognition tech-
niques. Another possibility is to relax the similarity constraint when searching a label
in a knowledge base. In other words, a strict equality measure would not be applied
between the label and the candidate spelling forms from a knowledge base. Rather n-
grams or Levenshtein similarity metrics with a high threshold would be a better choice.

3.2 Extraction of Candidate Patterns

As depicted in Figure 1, the outcome of document extraction process is candidate pat-
terns. Given the lists of extended labels for both entities, our tool associates all alterna-
tive labels of the first entity to all labels of the second entity (Cartesian product) to build
different queries. The documents resulting from these queries are ranked according to
 An Integrated Approach for Large-Scale Relation Extraction from the Web 167

their relevance score. The candidate patterns are extracted by parsing these documents
and locating the sentences with co-occurrence of both entities (defined by a maximum
number of words between them, currently 15 words). Note that we include in the can-
didate patterns the text before and after the entities to obtain full sentences. The final
step aims at refining the candidate patterns to obtain patterns.

3.3 Selection of Patterns

The last issue deals with the selection or ranking of the generic patterns. Thus, a confi-
dence score noted confp is computed for each pattern p with Formula 1. Our intuition
is to exploit all information which allowed the discovery of the patterns and to compare
a pattern with the ones of the same type of relationship.
& '
supp + occp + provp
conf (p) = (1)
++
The support supp is defined as the ratio between the number of examples exp that
this pattern is able to discover and the total number of examples ex discovered by all
patterns of the same type of relationship . Note that the support cannot be computed
at the first iteration.
Similarly, the occurrency occp stands for the number of candidate patterns which
led to the generation of the pattern p. It is normalized by the total number of candidate
patterns used to generalize all patterns of the same type of relationship .
exp occp
suppp = occp =
ex occ
The provenance provp refers to the relevance of the documents from which the can-
didate patterns which generalize a given pattern have been extracted. The relevance is
evaluated given three metrics: the relevance score namely applies tf-idf on the content
of the document and its values are bounded by a maximal value which depends on the
query. PageRank3 is widely known due to the Google search engine. The PageRank
scores of our collection are in the range [0.15, 10]. Finally, SpamScore indicates the
probability that a document is a spam or not [8]. The idea is to average the scores re-
turned by these three metrics for all documents from which patterns have been derived.
We note Dp this set of documents for the pattern p, and dip a document of this set. The
following formula computes a score in the range [0, 1] to evaluate the average relevance
of this set of documents, and thus the provenance of the pattern:
#dip 1 $ relevance(dip ) spamscore(dip ) prank(dip )
%
3 max(relevance(Dp )) + 100 + 10
provp =
|Dp |

4 Relationship and Example Discovery

In the previous step, we have generated patterns for a given type of relationship. SPI-
DER aims at discovering relationships between entities and discovering new examples.
3
http://j.mp/Clueweb09-Pagerank
168 N. Takhirov et al.

4.1 Challenges for Exploiting Patterns

Using the generated patterns to discover knowledge from plain texts involves the ad-
dressing of the following two challenges: pattern similarity and NER.
Pattern Similarity. When analyzing sentences in a document, SPIDER needs to eval-
uate the similarity between a sentence and a pattern. Thus, we have designed a pattern
similarity metric. The intuition which underlies our metric is twofold: (i) the presence
of frequent terms in the sentence is crucial while there is more flexibility for less im-
portant terms and (ii) the position of the words should be taken into account. First, the
sentence is transformed (cleaning, POS-tagging and replacing the mentions of the enti-
ties) so that both the sentence s and the pattern p are composed of POS-tagged terms.
The pattern may also include a list of frequent terms, which is noted F T p . The idea
is to compute , the minimal total distance to transform the sentence into the pattern.
Thus, our metric is an adaptation of the Levenshtein distance [7]. However, we do not
compute a number of operations (delete, transform or add a term) between two words
or characters, but rather we evaluate the semantic distance between two words. Given
the ith word wpi associated to the tag tip in the pattern p and the word wsj with the tag
tjs in the sentence s, we compute their semantic distance semdist(wpi , wsj ) using the
following formula:

0.0 if wsj F T p
i j
semdist(wp , ws ) = resnik(wp , ws ) if wsj
i j
/ F T p and tip = tjs (2)

1.0 otherwise

Namely, the distance is equal to 0 if the word in the sentence is a frequent term. POS-
tagged terms whose tags are identical (e.g., two verbs) have a similarity obtained by
applying the Resnik distance in Wordnet [13]. Else, words with different tags have
the maximal distance. The minimal total distance (p, s) between a pattern p and a
sentence s is computed by the matrix algorithm of the Levenshtein distance4 using the
semantic distance for all pairs of words. The distance is normalized in the range [0, 1]
with Formula 3 which assesses the similarity pattsim.
1
pattsim(p, s) = (3)
1 + (p, s)
Our metric is flexible because extra or missing words in a sentence do not significantly
affect the similarity value. Similarly, less important words are mainly compared on their
nature (POS-tag). Finally, we can select the sentences which are modeled by a pattern
according to a threshold.
NER. When a sentence in a document corresponds to a pattern, our approach needs to
identify and extract entities contained in the sentence. Thus, the NER issue is crucial
as it determines (part of) the output. Indeed, it is necessary to correctly identify the
(labels of) entities in the sentence based on a pattern. To solve this issue, we rely on
the formalism of our patterns: since they have been POS-tagged, the tags serve as a
delimiter and may constraint the candidate entities.
4
http://j.mp/Levenschtein
 An Integrated Approach for Large-Scale Relation Extraction from the Web 169

4.2 Discovering the Type of Relationship

Given two labels (representing an entity), the goal is to determine the possible type(s)
of relationships between these two entities. The two entities are first extended to ob-
tain their alternative labels (see Section 3.1). A set of documents is analyzed to extract
all sentences which contain one label of each entity, and these sentences are then com-
pared to all patterns stored in SPIDERs knowledge base (see Section 4.1). If a sentence
is similar to a trustable pattern (with a sufficient confidence score), then the type of re-
lationship corresponding to this pattern is proposed to the user. Note that if there is no
trustable pattern in the knowledge base, the user has the possibility to provide training
data to the system.

4.3 Discovering New Examples

The discovery of new examples is at the basis of the never-ending feature which en-
ables the feeding of the knowledge base with additional training data for generating
new patterns. SPIDER selects in the knowledge base the patterns of a given type of
relationship. It retrieves a set of documents by querying each de-tagged pattern (or only
their frequent terms). We compute the similarity between a pattern and the sentences of
the documents which include a frequent term of this pattern. If the sentence is modeled
by the pattern, then we apply the NER techniques for discovering the two entities. Note
that each discovered example has a confidence score, which is computed with the same
formula as in Section 3.3 for the confidence of a pattern, except that supe replaces supp
and it indicates the number of patterns which discovered this example. Examples with
a very low confidence are discarded while others are stored in the knowledge base.

5 Scalability
In order to efficiently process documents, we distribute the jobs into several machines.
MapReduce inspired techniques have been popular to tackle such tasks. Therefore, the
collection is indexed with Hadoop enabling efficient indexing and searching. To com-
pute statistics, SPIDER makes use of Pig5 which is a high level platform for analyzing
a large collection of data.
Additionally, we propose a document partition to incrementally provide results on
a subset of the collection. The general idea is that highly ranked documents should
be a better source for obtaining patterns than those with lower PageRank, SpamScore
and relevance score values. The size of a partition depends on the quality of the ob-
tained patterns and examples. SPIDER is able to automatically tune the ideal size of
a partition. Initially, the documents are sorted by their PageRank, SpamScore and the
relevance score as described above. The top k documents are selected for analysis from
the head of the ranked list of documents. For the i-th round, the cursor is moved to
the range [k, i k] and the documents in that range are picked out for analysis. Fur-
thermore, when there are too few patterns discovered for two given labels out of these
initial set of documents, the partition size is subsequently adjusted to a higher number.
5
http://pig.apache.org
170 N. Takhirov et al.

The combination of scores as well as the partitioning mechanism makes obtaining the
URLs of the documents and their content for a given query fast, i.e., it only requires a
few seconds. This efficiency is demonstrated in Section 7.2.

6 Related Work

Relationship extraction has been widely studied in the last decade. Supervised systems
for relationship extraction are mainly based on kernel methods, but they suffer from the
processing costs and the difficulty for annotating new training data. One of the earliest
semi-supervised system, DIPRE, uses a few pairs of entities for a given type of rela-
tionship as initial seeds [3]. It searches the Web for these pairs to extract patterns repre-
senting a relationship, and use the patterns to discover new pairs of entities. These new
entities are integrated in the loop to generate more patterns, and then find new pairs of
entities. Snowball [1] enhances DIPRE in two different directions. First, a verification
step is performed so that generated pairs of examples are checked with MITRE named
entity tagger. Secondly, the patterns are more flexible because they are represented as
vectors of weighted terms, thus enabling the clustering of similar patterns.
Espresso [11] is a weakly-supervised relation extraction system that also makes use
of generic patterns. The system is efficient in terms of reliability and precision. How-
ever, experiments were performed on smaller datasets and it is not known how the
system performs at Web-scale.
TextRunner brought further perspective in the field of Open Information Extraction,
for which the types of relationships are not predefined [2]. A self-supervised learner is
in charge of labeling the training data as positive or negative, and a classifier is then
trained. Relations are extracted with the classifier while a redundancy-based probabilis-
tic model assigns confidence scores to each new examples. The system was further
developed into a ReVerb framework [6] which improves the precision-recall curve of
the TextRunner.
ReadTheWeb / NELL [4] is another project that aims at continuously extracting cat-
egories (e.g., the type of an entity) and relationships from web documents and improv-
ing the extraction step by means of learning techniques. Four components including a
classifier and two learners are in charge of deriving the facts with a confidence score.
According to the content of the online knowledge base, more iterations provide high
confidence scores (almost 100%) for irrelevant relationships. Contrary to NELL, we do
not assume that one document returning a high confidence score for a given relation-
ship is sufficient for approving this relationship. In addition, NELL is mainly dedicated
to the discovery of categories (95% of the discovered facts) rather than relationships
between entities.
A recent work reconciles three main issues in terms of precision, recall and perfor-
mance [10]. Indeed, Prospera utilizes both pattern analysis with n-gram item sets to
ensure a good recall and rule-based reasoning to guarantee an acceptable precision. The
performance aspect is handled by partitioning and parallelizing the tasks in a distributed
architecture. A restriction of this work deals with the pattern which only covers the mid-
dle text between the two entities. This limitation affects the recall, as shown with the
example Lord Of The Rings, which Tolkien has written.
 An Integrated Approach for Large-Scale Relation Extraction from the Web 171

Contrary to the oldest systems which include hard representations of patterns, Pros-
pera and SPIDER includes a more flexible definition of the patterns, so that similar
patterns can be merged. In addition, the patterns are at the sentence level, which means
that the texts before, after and between the entities are considered. Our confidence
score for a pattern or an example takes into account crucial criteria such as provenance.
In addition, the support and the occurrence scores are correlated within the same type of
relationship, thus the confidence in a pattern or an example may decrease over time. To
the best of our knowledge, none of these approaches deal with the issue of identifying
the alternative labels of an entity. In the future, we plan to demonstrate the impact of
these alternative labels on the recall. Finally, our approach is scalable with document
partitioning based on smart sorting using the SpamScore, PageRank and relevance score
of the documents.

7 Experimental Results

Our collection of documents is the English portion of the ClueWeb09 dataset (around
500 million documents). For the components, we have used the contextual strategy,
with the Maxent POS-tagger provided with StandfordNLP6. The NER component is
based on the OpenNLP toolkit7 . Evaluation is performed using well-known metrics
such as precision (number of correct discovered results divided by the total number of
discovered results). The recall can only be estimated since we cannot manually parse
the 500 million documents to check if some results have been forgotten. However, we
show that the number of correct results increases over time.

7.1 Quality Results

In this section, we present our results in terms of quality of label extension, relationship
discovery and comparison with state of the art baseline knowledge extraction tools. The
evaluation of relationship discovery depends on the quality of generated patterns and
hence we present this evaluation rather than the pattern generation itself.
Relationship Discovery. We evaluate the quality obtained by SPIDER when running
the first use case (Section 4.2). Given two labels (representing entities), we search for
the correct type of relationship which links them. To fulfill this goal, we have man-
ually designed a set of 200 relationships, available at this URL8 . Note that the type
of relationship associated to each example is the most expected one, but several types
of relationship are possible for the same example. Table 1 provides a sample of ex-
amples (e.g., Obama, Hawai) and some candidate types of relationship discovered by
SPIDER (e.g., birthplace). A bolded type of relationship indicates that it is correct for
this example. A second remark about our set of relationships deals with the complexity
of some relationships (e.g., <cockatoo, tail, yellow>). The last column shows the ini-
tial confidence score computed for the candidate relationship. The quality is measured
6
http://nlp.stanford.edu/software/CRF-NER.shtml
7
http://opennlp.apache.org/
8
http://j.mp/apweb2013
172 N. Takhirov et al.

Table 1. Examples of Discovered Types of Relationship and Confidence Scores

Example Discovered type of relationship Confidence score

birthplace 0.42
Obama, Hawai senator 0.31
president-elect 0.18
amazon 0.32
cockatoo, yellow parrot 0.31
tail 0.16
plant 0.51
eucalyptus, myrtaceae family 0.43
specie 0.27
inventor 0.60
Bartolomeo Cristofori, instrument 0.43
piano maker 0.19

in terms of precision at different top-k. Indeed, SPIDER outputs a ranked list of rela-
tionship types according to their confidence scores. In addition, our approach is able to
run with or without training data. Thus, we have tested the system when a few training
data have been provided. Using 1 training example means that the system has randomly
selected 1 correct example for bootstrapping the system. Experiments with the training
data are based on cross-validation and 5 runs reduce the impact of randomness. The
manual validation of the discovered relationships has been performed by 8 people. This
manual validation includes around 3000 invalid relationships and 600 correct ones, and
it facilitates the automatic computation of precision. In addition, we are able to estimate
the recall, i.e. to evaluate the number of correct types discovered during a run w.r.t. all
validated types. This is an estimation because there may exist more correct types of
relationship than the ones which have been validated. Besides, a discovered type may
have a different spelling from a validated type while both have the same meaning, thus
decreasing the recall.
Figures 2(a) and 2(b) respectively depict the average precision and the average
recall for the 200 relationships by top-k and by number of provided training data. We
first notice that SPIDER achieves low quality without training data (precision from
40% at top-1 to 30% at top-10). The estimated recall values are also quite low at top-1
because there is an average of 3 correct types of relationships for each example. The
top-3 results are interesting with 5 training data: the precision is acceptable (more than
80%) while the recall value (32%) indicates that one type of relationship out of three is
correctly identified. Since our dataset contains complex relationships, this configuration
is promising for bootstrapping the system. Precision strongly decreases at top-5 and
top-10, mainly because each example roughly includes 3 types. However, the top-5 and
top-10 recall values indicate that we discover more correct examples. Finally, we notice
that providing a few training data (5 examples) enables at least a 10% improvement both
for precision and recall. This remark is important since our approach aims at running
perpetually by reusing previously discovered examples and patterns.
The quality results are subject to the complexity of the set of relationships, since we
have selected some complex ones to discover, such as <cockatoo, tail, yellow>.
Other problems of disambiguation occurred, for instance the example Chelsea, Lon-
don mainly returns types of relationships about accommodations because Chelsea is
identified as a district of London and not as the football team.
 An Integrated Approach for Large-Scale Relation Extraction from the Web 173

100 100
no training no training
1 example 1 example
5 examples 5 examples

Estimated recall in percentage

80 80
Precision in percentage

60 60

40 40

20 20

0 0
top-1 top-3 top-5 top-10 top-1 top-3 top-5 top-10

(a) Precision (b) Estimated Recall

Fig. 2. Quality results according to top-k and Training Data

Baseline Comparison. A final experiment aims at comparing our system with two other
approaches, ReadTheWeb (NELL) [4] and Prospera [10], both described in Section 6.
These two approaches have been chosen as baseline because the dataset along with the
results are available online. An evaluation of these tools is described online9 . Since the
seed examples are available, we have used them as training data. Table 2 summarizes
the comparison between the three systems in terms of estimated precision, as ex-
plained in the experiments reported in [4,10]. Similarly to Prospera and ReadTheWeb,
our precision is an estimation due to the amount of relationships to validate. Namely,
1000 random types have been validated for each relationship. The average precision of
the three systems is the same (around 0.91). However, the total number of facts discov-
ered by SPIDER (71,921) is 36 times higher than ReadTheWeb (2,112) and 1.3 times
higher than Prospera (57,070), outperforming both baselines.
Prospera provides slightly better quality results than our approach on
AthletePlaysForTeam relation. However, several factors have an influence on the pre-
cision results between Prospera, ReadTheWeb and SPIDER. First, Prospera is able to
use seeds and counter seeds while we only rely on positive examples. On the other side,
Prospera includes a rule-based reasoner combined with the YAGO ontology. Although
SPIDER does not support this feature, the combination of POS-tagged patterns and
NER techniques achieves outstanding precision values.

7.2 Performance

Since knowledge extraction systems deal with large collections of documents, they need
to be scalable. Figure 3(a) depicts the performance of SPIDER for retrieving and pre-
processing (i.e., clean up the header, remove html tags) the documents. The total time
(sum of retrieval and preprocessing) is also indicated. Although there is no caching,
the total time is not significant for collecting and preprocessing one million documents
(around 40 seconds). Note that in real cases, a conjunctive query composed of two
labels rarely returns more than 20, 000 documents. The peak for retrieval at 600, 000
documents is due to an overhead processing from the thread manager. Increasing the
9
http://www.mpi-inf.mpg.de/yago-naga/prospera/
174 N. Takhirov et al.

Table 2. Estimated Precision (with Number of Discovered Facts) values obtained by

ReadTheWeb (RTW), Prospera and SPIDER

Relation RTW Prospera SPIDER

AthletePlaysForTeam 1.00 (456) 0.82 (14,685) 0.80 (15,234)
TeamWonTrophy 0.68(397) 0.94 (98) 0.96 (92)
CoachCoachesTeam 1.00 (329) 0.88 (1,013) 0.90 (1,629)
AthleteWonTrophy n/a 0.92(10) 0.94 (124)
AthletePlaysInLeague n/a 0.94 (3,920) 0.95 (4,211)
CoachCoachesInLeague n/a 0.99 (676) 0.89 (741)
TeamPlaysAgainstTeam 0.99 (1,068) 0.89 (15,170) 0.93 (15,729)
TeamPlaysInLeague n/a 0.89 (1,920) 0.95 (2,409)
TeamMate n/a 0.86 (19,578) 0.84 (31,752)

400 500000
80 candidate patterns
retrieval-time patterns
preprocessing-time 350 total time
70 total-time
400000
60 300

Number of (candidate) patterns

Time in seconds

50 250
Time in seconds

300000

40 200

30 200000
150

20 100
100000
10 50

0 0 0
200000 400000 600000 800000 1M
1000 5000 10000 20000 50000
Number of documents Number of documents

(a) Retrieval and Preprocessing Performance (b) Query Performance

Fig. 3. Performance results

number of threads above 400 leads to higher thread switching latency while decreasing
this number only reports the peak earlier during the process. This issue could be simply
solved by dispatching this task on different servers.
ReadTheWeb and Prospera expose their knowledge base but not their tools. Thus, it
is not possible to evaluate the three systems on the same hardware and the following
comparison is based on the performance described in the original research papers. It
mainly aims at showing the significant improvement of SPIDER over ReadTheWeb and
Prospera, for a better average quality. To produce the results shown in Table 2, Prospera
needed more than 2 days using 10 servers with 48 GB RAM [10]. In a similar fashion,
ReadTheWeb has generated an average of 3618 facts per day during the course of 67
days using the Yahoo M 45 supercomputing cluster [4]. On the contrary, our approach
performed the same experiment in a few hours. The generation of facts for each type
of relationship took between 20 to 60 minutes with four servers equipped with 24 GB
RAM. Although these values are mainly indicative, SPIDER is more efficient than the
two other systems when dealing with dynamic and large-scale environments.

8 Conclusion
In this paper, we have presented SPIDER, an approach to automatic extraction of
binary relationships from large text corpora. The main advantage of our system is to
 An Integrated Approach for Large-Scale Relation Extraction from the Web 175

guarantee both a better quality and a strong improvement in terms of performance

over similar approaches, thus providing new opportunities for discovering relationships
at large scale. Finally, we have demonstrated the feasibility of SPIDER at Web-scale.

References
1. Agichtein, E., Gravano, L.: Snowball: Extracting relations from large plain-text collections.
In: Proc. of DL, pp. 8594. ACM (2000)
2. Banko, M., Cafarella, M.J., Soderland, S., Broadhead, M., Etzioni, O.: Open information
extraction from the web. In: Proc. of IJCAI, pp. 26702676. Morgan Kaufmann (2007)
3. Brin, S.: Extracting patterns and relations from the world wide web. In: Atzeni, P., Mendel-
zon, A.O., Mecca, G. (eds.) WebDB 1998. LNCS, vol. 1590, pp. 172183. Springer, Heidel-
berg (1999)
4. Carlson, A., Betteridge, J., Kisiel, B., Settles, B., Hruschka Jr., E., Mitchell, T.M.: Toward an
architecture for never-ending language learning. In: Proc. of AAAI. AAAI Press (2010)
5. Etzioni, O., Banko, M., Soderland, S., Weld, D.S.: Open information extraction from the
web. Communication of ACM 51, 6874 (2008)
6. Fader, A., Soderland, S., Etzioni, O.: Identifying relations for open information extraction.
In: Proc. of EMNLP, pp. 15351545. ACL (2011)
7. Levenshtein, V.: Binary Codes Capable of Correcting Deletions, Insertions and Reversals.
Journal of Soviet Physics Doklady 10, 707 (1966)
8. Lynam, T.R., Cormack, G.V., Cheriton, D.R.: On-line spam filter fusion. In: Proc. of SIGIR,
pp. 123130. ACM (2006)
9. Mausam, Schmitz, M., Soderland, S., Bart, R., Etzioni, O.: Open language learning for in-
formation extraction. In: Proc. of EMNLP, pp. 523534. ACL (2012)
10. Nakashole, N., Theobald, M., Weikum, G.: Scalable knowledge harvesting with high preci-
sion and high recall. In: Proc. of WSDM, pp. 227236. ACM (2011)
11. Pantel, P., Pennacchiotti, M.: Espresso: leveraging generic patterns for automatically harvest-
ing semantic relations. In: Proc. of ACL, pp. 113120. ACL (2006)
12. Parameswaran, A., Garcia-Molina, H., Rajaraman, A.: Towards the web of concepts: extract-
ing concepts from large datasets. VLDB Endowment 3, 566577 (2010)
13. Resnik, P.: Semantic similarity in a taxonomy: An information-based measure and its appli-
cation to problems of ambiguity in natural language. Journal of Artificial Intelligence Re-
search 11, 95130 (1999)
14. Takhirov, N., Duchateau, F., Aalberg, T.: An evidence-based verification approach to extract
entities and relations for knowledge base population. In: Cudr-Mauroux, P., Heflin, J., Sirin,
E., Tudorache, T., Euzenat, J., Hauswirth, M., Parreira, J.X., Hendler, J., Schreiber, G., Bern-
stein, A., Blomqvist, E. (eds.) ISWC 2012, Part I. LNCS, vol. 7649, pp. 575590. Springer,
Heidelberg (2012)
 Multi-QoS Eective Prediction
in Web Service Selection

Zhongjun Liang1, , Hua Zou1 ,

Jing Guo , Fangchun Yang1 , and Rongheng Lin1
2

1
State Key Laboratory of Networking and Switching Technology
2
School of Computer Science
Beijing University of Posts and Telecommunications, Beijing, China
{kkliangjun,guo.apple8}@gmail.com, {zouhua,fcyang,rhlin}@bupt.edu.cn

Abstract. The rising development of service-oriented architecture

makes Web service selection a hot research topic. However, there still
remains challenges to design accurate personalized QoS prediction ap-
proaches for Web service selection, as existing algorithms are all fo-
cused on predicting individual QoS, without considering the relationship
between them. In this paper, we propose a novel Multi-QoS Eective Pre-
diction (MQEP for short) problem, which aims to make eective Multi-
QoS prediction based on Multi-QoS attributes and their relationships. To
address this problem, we design a novel prediction framework Multi-QoS
Eective Prediction Approach (MQEPA for short). MQEPA first takes
use of Gaussian method to normalize the QoS attribute values, then ex-
ploits Non-negative Matrix Factorization to extract the feature of Web
services from Multi-QoS attributes, and last predicts the Multi-QoS of
unused services via Multi-output Support Vector Regression algorithm.
Comprehensive empirical studies demonstrate the utility of the proposed
method.

Keywords: Web service, Multi-QoS, QoS Prediction.

1 Introduction
In recent years, Web services are designed as computational components to
build service-oriented distributed systems [1]. The rising development of service-
oriented architecture makes more and more alternative Web services oered by
dierent providers with equivalent or similar functions. So how to select Web
services from a large number of functionally-equivalent candidates become an
important issue.
In an intuitive way, Web service can be selected from all candidates by their
QoS values. However, it is hard to obtain all their QoS observed by a target user,
as some Web service invocations may be charged. Hence QoS prediction [2-5]
become a challenging research topic. For example, Jieming Zhu et al [2] proposes
a Web service positioning framework for response time prediction. Zibin Zheng

Corresponding author.

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 176183, 2013.
c Springer-Verlag Berlin Heidelberg 2013

 Multi-QoS Eective Prediction in Web Service Selection 177

and Michael R. Lyu [3] propose a collaborative reliability prediction approach.

Yilei Zhang et al [4] propose a neighborhood-based approach for collaborative
and personalized quality prediction of cloud components.
However, there remains challenges to design accurate personalized QoS pre-
diction approaches for Web service selection.
How to predict Multi-QoS simultaneously? In Web service selection, the
user wants to obtain Multi-QoS at same time. Existing algorithms are all
focused on predicting individual QoS one by one, which cannot scale to
Multi-QoS prediction problem. So how to eective predict the Multi-QoS is
an interesting problem.
How to improve the accuration of Multi-QoS prediction? Existing algorithms
make QoS prediction without considering the relationship between QoS at-
tributes. In fact, there are certain relationships among the QoS attributes
of a same Web service. For instance, the response time and the through-
put obtained by users will become worse when the network performance
is poor. In addition, probable QoS prediction should be made based on
many aspects, and multi-dimensional QoS information is helped for rational
decision-making. So introducing Multi-QoS attributes and their relationships
for Multi-QoS prediction deserves further research.
In this paper, we propose a novel Multi-QoS Eective Prediction problem, which
aims to make eective Multi-QoS prediction based on Multi-QoS attributes and
their relationships. To address this problem, we design a novel prediction frame-
work Multi-QoS Eective Prediction Approach. It first take use of Gaussian
method to normalize the QoS attribute values, then exploits Non-negative Ma-
trix Factorization to extract the latent feature of users and services from Multi-
QoS attributes, and last predicts the Multi-QoS value of unused services via
Multi-output Support Vector Regression algorithm. Comprehensive empirical
studies demonstrate the utility of the proposed method.
The contribution is as follows.
First, we propose a novel Multi-QoS Eective Prediction problem, which aims
to make eective Multi-QoS predict based on Multi-QoS attributes and their
relationships.
Second, we propose a novel prediction framework Multi-QoS Eective Pre-
diction Approach to address this MQEP problem. Unlike existing algorithms
that all focused on predicting QoS independently one by one, and taking no ac-
count of the relationship between them, MQEPA taking use of these Multi-QoS
attributes and their relationships to predict QoS.
Last, we conduct several experiments on real-world Web service dataset, com-
prehensive empirical studies demonstrate the utility of the proposed method.

2 Problem Statement
In this section, we formally describe the Multi-QoS eective prediction problem
as follow.
178 Z. Liang et al.

Given a set of users and a set of Web services, based on the existing Multi-
QoS values from dierent users, predict the missing Multi-QoS values of Web
service when invoked by a target user at same time. For example, a bipartite
graph G = (U S, E) represents the user-Web services invocation (shown as
Fig.1), where U denotes the users set, and S denotes the Web services set, E
denotes the invocations set between U and S. If user ui has invoked Web service
1 2 m
si , where ui U , sj S, and eij E, the weight Qij = {qij , qij , ..., qij } on edge
eij corresponds to the Multi-QoS values (e.g., throughput, response-time in this
example) of that invocation. Our task is to eectively predict the Multi-QoS
values of potential invocations (the broken lines) for a targeted user.

0.973 18.86 9.66

User-1 User-2 User-3 User-4 Q
oS
0.334 17.54 9.049

5.982 0.228 0.221 11.19

2.13 0.273 0.254

User
8.72
0.366 0.375
0.693 0.223 0.22 0.344
Web service

0.973 18.86 27.67 9.66 10.0 6.849

oS 7.172
Q 0.334 17.54 25.31 9.049 9.009 7.025

5.982 0.228 0.237 0.221 0.222 0.527 11.19

2.13 0.273 0.254
User

Web Web Web Web Web Web 0.366 0.375

8.72
service-1 service-2 service-3 service-4 service-5 service-6 0.693 0.223 0.22 0.344
Web service

Fig. 1. Bipartite Graph of User-Web service Invocation

In fact, the Multi-QoS eective prediction problem can be considered as how

to eective predict the missing entries in the user-Web service matrix based on
the existing entries, where the existing entries are Multi-QoS values between U
and S, and can be observed in Fig.1.

3 Multi-QoS Eective Prediction Approach

The personalized QoS experience, obtained from a user after the invocation,
is co-determined by the feature of corresponding invoked Web service and the
user. Therefore, for a target user, he always obtains dierent QoS experience
after invoking the dierent Web services, as the feature of these Web services is
dierent. As a result, we predict the Multi-QoS based on the extracted feature
of Web services and historical QoS experience of a target user.
In this paper, we propose a novel three-step prediction framework Multi-QoS
Eective Prediction Approach (shown as Figure.2).
First, preprocess step, which aims to normalize the QoS attribute values by
Gaussian method. This step can map the QoS values to the interval [0,1] for
feature extraction.
Second, extraction step, which aims to extract the feature of Web services
from existing Multi-QoS values. This step can handle the acquired sparse
Multi-QoS information.
 Multi-QoS Eective Prediction in Web Service Selection 179

Third, prediction step, which aims to predict the missing Multi-QoS val-
ues via Multi-output Support Vector Regression algorithm. This step finally
solves our problem based on the extraction feature of Web services and the
targeted users historical QoS experience.

Convert the
normalization result
Normalization Web service feature
of multi-QoS values extraction
Output QoS result
0.973 18.86 27.67 9.66 10.0 6.849

oS
7.172

Q
0.334 17.54 25.31 9.049 9.009 7.025
Predict Multi-QoS
5.982 0.228 0.237 0.221 0.222 0.527
with MSVR 2.13 0.273 0.254
11.19

User
8.72
0.366 0.375

QoS experience 0.693 0.223 0.22 0.344

Web service

Fig. 2. The Whole Framework of Proposed Approach

3.1 The QoS Normalization

Since dierent QoS attributes have dierent range of values, we employ Gaussian
method[6] to normalize the QoS attribute values for feature extraction. It can
not only map the QoS values to the interval [0,1], but also can better avoid the
influence of abnormal values (such as high value or low value) compared with
other normalization methods (such as Extremum Regularization Method). The
detail can be described as follow.

First, exploit formulation (1) to normalize the Multi-QoS attribute values in

each dimension. It has been proved that 99% data can be directly mapped
to the [0,1] interval[6].
k
qij qij
k
Qkij = 0.5 + (1)
2 3
Where Qk is normalized value of k-th QoS attribute, qk denote the aver-
ij ij
age value of k-th QoS attribute obtained from all users, is the standard
deviation of the k-th QoS attribute.
Second, assign the Qkij with 0 or 1 based on the distance with 0 or 1, if it is
out of the [0,1] interval.

3.2 Feature Extraction Based on Multi-QoS Attributes

To better reflect the feature of Web service, we extract it based on Multi-QoS
attributes. Due to the existing QoS matrix is always sparse, we employ ma-
trix factorization to extract the features of Web services. The main idea of
this method is to derive a high-quality low-dimensional matrix to approximate
the historical QoS data, which represent the feature of Web service and user
respectively.
180 Z. Liang et al.

Consider Multi-QoS matrix Q = {Q1 , Q2 , ..., Qm }(Q Rlmn ) is consisting

of l users, n Web services and m QoS attributes. Our method is to factorize
Q into two matrices C Rlmr and W Rrn , which satisfy the following
equation:
r
" r
" r
"
Qij = {Q1ij , Q2ij , ..., Qm
ij } { c1ik wjk , c2ik wjk , ..., cm
ik wjk } (2)
k=1 k=1 k=1

Where Qij is the Multi-QoS performance of Web service sj observed by user

ui , vector wj = [wj1 , wj2 , ..., wjr ] reflects the feature of Web service sj , matrix
ci = [c1i , c2i , ..., cm d d d d
i ] reflects the feature of user ui , vector ci = [ci1 , ci2 , ..., cir ] is its
sub feature, r is the rank of matrix C and W , it generally satisfies (lm+n)r <
m l n.
In this paper, we put Multi-QoS attribute values together and use NMF al-
gorithm[7] to learn the feature of Web services. The detail can be described as
follow:
First, we put normalized Multi-QoS attribute values together to form a super
matrix Q = [Q1 , Q2 , ..., Qm ]T .

Second, we decompose the Q to extract the matrix C and W with following

objective function min f (C, W ).
lm
"" n
min f (C, W ) = Iij [Qhj log(CW )hj (CW )hj ] (3)
h=1 j=1

where Iij is the indicator function that is equal to 1, if user ui invoked

service sj and 0 otherwise. Then, an incremental gradient descent method is
employed to extract the feature of Web service by following iterative process.
" Qhj
Cha = Cha Waj (4)
j
(CW )hj
Cha
Cha = # (5)
k Cka
" Qhj
Waj = Waj Cha (6)
(CW )hj
h

3.3 Predict QoS with Multi-output Support Vector Regression

Based on the extracted feature of Web services and users historical QoS expe-
rience, we predict the missing Multi-QoS values for a targeted user. Consider
the relationship between QoS attributes, we employ the Multi-output Support
 Multi-QoS Eective Prediction in Web Service Selection 181

Vector Regression (MSVR for short) algorithm[8] to predict the missing Multi-
QoS values at same time by following steps:
First, we select the training data according to the invoked experience of
the targeted user. Suppose a targeted user ue has invoked d Web services,
then the training data can be represented by T = {(x1 , y1 ), ..., (xd , yd )},
where xd W is the obtained feature of Web service sd , yd represents QoS
performance of Web service sd observed by ue .
Second, we employ MSVR to build a function f , which represents the relation
between the feature of Web service and Multi-QoS values obtained from a
target user. Consider Rr is the feature space of Web service, Rm is the QoS
space. A function f can be described as formulation (7).
4
f : Rr Rm
(7)
xk W T (xk ) + b

where W T (xk ) + b reflects a map from the feature space of Web service to
the QoS space, which is learning from the training data by MSVR.

Last, we predict the miss Multi-QoS values for a targeted user, based on the
feature of unused Web service. let xu is the feature of unused Web service su, the
predicted Multi-QoS Qeu for a targeted user ue can be calculated as following:

Qeu = W T (xu ) + b (8)

4 Experiment
In this section, we conduct several experiments to show the prediction quality of
our proposed approach on real-world Web service dataset[9]. In the experiments,
we randomly remove 90% entries from the original QoS matrix, only use the
remaining 10% entries to predict the removed entries of a targeted user (which is
chosen by random). The parameter settings are r = 4, = 0.02, C = 678, = 0.6.
Figure.3 and Figure.4 show the predictions of response-time and throughput
for dierent users. The experimental results show that our approach have good
performance on predicting the QoS value of dierent Web services.

4.1 Performance Comparison

In order to evaluate the performance of our approach, we compare our prediction
precision with following methods on Mean Absolute Error (MAE) and Root
Mean Squared Error (RMSE).

UserMean : predict the missing QoS values based on the mean QoS value of
each user.
UPCC: predict the missing QoS values based on similar users[10].
IPCC: predict the missing QoS values based on similar items (item refers to
Web service in this paper)[11].
182 Z. Liang et al.

1.6 150

1.4 actual value actual value

predicted value predicted value
1.2
Response time(s)

Throughput(kbps)
100
1

0.8

0.6
50
0.4

0.2

0 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Different Web services Different Web services

Fig. 3. Multi-QoS Prediction for Random Chosen User-1

3.5 100

actual value 90 actual value

3
predicted value predicted value
80
2.5
70
Response time(s)

Throughput(kbps)
2 60

50
1.5
40

1 30

20
0.5
10

0
0 5 10 15 20 25 30 35 40 0
0 5 10 15 20 25 30 35 40
Different Web services
Different Web services

Fig. 4. Multi-QoS Prediction for Random Chosen User-2

Table 1. Accuracy Comparison with Dierent Users

Dierent User RandomUser-1 RandomUser-2

QoS Response time Throughput Response time Throughput
Metrics MAE RMSE MAE RMSE MAE RMSE MAE RMSE
UserMean 0.5783 0.8173 21.2979 27.6767 0.3792 0.4279 18.2875 28.7697
IPCC 0.4466 0.6941 21.6252 26.9089 0.2730 0.3890 15.9715 19.1147
UPCC 0.3069 0.5800 5.8532 10.2054 0.3872 0.5604 8.2635 12.5219
MQEPA 0.0809 0.1444 4.1930 6.1993 0.0396 0.0653 3.8689 5.6997

Table.1 shows the experimental results of dierent approaches, and the result of
MQEPA is specially highlighted. From Table.1, we can observe that our approach
can obtain smaller MAE and RMSE, shows that MQEPA has better performance
on prediction accuracy on both response-time and throughput for dierent users.
This indicates that our approach suits better for Web service selection.

5 Conclusion
In this paper, we propose a novel Multi-QoS Eective Prediction problem, which
aims to make eective Multi-QoS prediction based on Multi-QoS attributes and
their relationships. To address this problem, we propose a novel prediction frame-
work Multi-QoS Eective Prediction Approach. Comprehensive empirical studies
demonstrate the utility of the proposed method.
 Multi-QoS Eective Prediction in Web Service Selection 183

Acknowledgments. This research is supported by the 973 Program of China

No. 2009CB320406; 863 Program of China No.2011AA01A102; the CPSF
No.2011M500226; the RFDP No.20110005130001.

References
1. Zhang, L., Zhang, J., Cai, H.: Services Computing: Core Enabling Technology of
the Modern Services Industry. Tsinghua University Press (2007)
2. Zhu, J., Kang, Y., Zheng, Z., Lyu, M.R.: WSP: A Network Coordinate Based
Web Service Positioning Framework for Response Time Prediction. In: Proc. of
the IEEE ICWS 2012, pp. 9097 (2012)
3. Zheng, Z., Lyu, M.R.: Collaborative reliability prediction of service-oriented sys-
tems. In: Proc. of 32nd International Conference on Software Engineering, pp.
3544 (2010)
4. Zhang, Y., Zheng, Z., Lyu, M.R.: Exploring Latent Features for Memory-Based
QoS Prediction in Cloud Computing. In: Proc. of the 30th IEEE Symposium on
Reliable Distributed Systems, pp. 110 (2011)
5. Jiang, Y., Liu, J., Tang, M., Liu, X.F.: An eective web service recommendation
method based on personalized collaborative filtering. In: Proc. of the IEEE ICWS
2011, pp. 211218 (2011)
6. Ortega, M., Rui, Y., Chakrabarti, K., Mehrotra, S., Huang, T.S.: Supporting sim-
ilarity queries in MARS. In: Proc. of the ACM Multimedia, pp. 403413 (1997)
7. Lee, D.D., Seung, H.S.: Learning the parts of objects by non-negative matrix fac-
torization. Nature 401(6755), 788791 (1999)
8. Yu, S.P., Yu, K., Volker, T.: Multi-Output Regularized Feature Projection. IEEE
Transactions on Knowledge and Data Engineering, 16001613 (2006)
9. Zheng, Z., Zhang, Y., Lyu, M.: Distributed QoS Evaluation for Real-World Web
Services. In: Proc. of ICWS 2010, pp. 8390 (2010)
10. Shao, L., Zhang, J., Wei, Y., Zhao, J., Xie, B., Mei, H.: Personalized qos prediction
for web services via collaborative filtering. In: Proc. of ICWS 2007, pp. 439446
(2007)
11. Resnick, P., Iacovou, N., Suchak, M., Bergstrom, P., Riedl, J.: GroupLens: an open
architecture for collaborative filtering of netnews. In: Proc. of CSCW 1994, pp.
175186 (1994)
Accelerating Topic Model Training on a Single Machine

Mian Lu2 , Ge Bai2 , Qiong Luo2 , Jie Tang3 , and Jiuxin Zhao2
1
A*STAR Institute of High Performance Computing, Singapore
[email protected]
2
Hong Kong University of Science and Technology
{luo,gbai,zhaojx}@cse.ust.hk
3
Tsinghua University
[email protected]

Abstract. We present the design and implementation of GLDA, a library that

utilizes the GPU (Graphics Processing Unit) to perform Gibbs sampling of La-
tent Dirichlet Allocation (LDA) on a single machine. LDA is an effective topic
model used in many applications, e.g., classification, feature selection, and infor-
mation retrieval. However, training an LDA model on large data sets takes hours,
even days, due to the heavy computation and intensive memory access. Therefore,
we explore the use of the GPU to accelerate LDA training on a single machine.
Specifically, we propose three memory-efficient techniques to handle large data
sets on the GPU: (1) generating document-topic counts as needed instead of stor-
ing all of them, (2) adopting a compact storage scheme for sparse matrices, and
(3) partitioning word tokens. Through these techniques, the LDA training which
would take 10 GB memory originally, can be performed on a commodity GPU
card with only 1 GB GPU memory. Furthermore, our GLDA achieves a speedup
of 15X over the original CPU-based LDA for large data sets.

1 Introduction
Statistical topic models have recently been successfully applied to text mining tasks
such as classification, topic modeling, and recommendation. Latent Dirichlet Alloca-
tion (LDA) [1], one of the recent major developments in statistical topical modeling,
immediately attracted a considerable amount of interest from both research and indus-
try. However, due to its high computational cost, training an LDA model may take
multiple days [2]. Such a long running time hinders the use of LDA in applications that
require online performance. Therefore, there is a clear need for methods and techniques
to efficiently learn a topic model.
In this paper, we present a solution to the efficiency problem of training the LDA
model: acceleration with graphics processing units (GPUs). GPUs have recently be-
come a promising high-performance parallel architecture for a wide range of appli-
cations [3]. The current generation GPU provides a 10 higher computation capability
and a 10 higher memory bandwidth than a commodity multi-core CPU. Earlier studies
of using GPUs to accelerate the LDA have shown significant speedups compared with
CPU counterparts [4,5]. However, due the limited size of GPU memory, these methods
are not scalable to large data sets. The GPU memory size is relatively small compared
with the main memory size, e.g., up to 6GB for GPUs in the market. This limits the

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 184195, 2013.
c Springer-Verlag Berlin Heidelberg 2013
 Accelerating Topic Model Training on a Single Machine 185

use of GPU-based LDA for practical applications since the memory size required of
many real-world data sets clearly exceeds the capability of the GPU memory. More-
over, some parallel implementations of LDA, such as AD-LDA [2], cannot be directly
adapted on the GPU either since the overall memory consumption would easily exceed
the available GPU memory size due to a full copy of word-topic counts maintained for
each processor.
To address this issue, we study how to utilize limited hardware resources for large-
scale GPU-based LDA training. Compared with the existing work on GPU-base LDA,
our implementation is able to scale up to millions of documents. This scalability is
achieved through the following three techniques:
1. On-the-fly generation of document-topic counts. We do not precompute and store
the global document-topic matrix. Instead, the counts for specific documents are
generated only when necessary.
2. Sparse storage scheme. We investigate a more compact storage scheme to store the
document-topic and word-topic count matrices since they are usually sparse.
3. Word token partitioning. Word tokens in the data set are divided into a few disjoint
partitions based on several criteria. Then for each partition, only a part of document-
topic and word-topic matrices are needed in each iteration.
Based on these techniques, we have successfully trained LDA models on data sets orig-
inally requiring up to 10 GB memory on a GPU with only 1 GB memory. Furthermore,
our GPU-based LDA training achieved significant performance speedups. These results
show that our approach is practical and cost-effective. Also, our techniques can be ex-
tended to solutions that involve multiple GPUs.
The remainder of this paper is organized as follows. In Section 2, we briefly introduce
the LDA algorithm and review GPGPU. We present the design and implementation of
our GLDA in detail in Section 3. In Section 4, we experimentally evaluate our GLDA
on four real-world data sets. We conclude in Section 5.

2 Preliminary
2.1 Gibbs Sampling of Latent Dirchlet Allocation
We develop GLDA based on Gibbs sampling of LDA [6] due to its simplicity and
high effectiveness for real-world text processing. In the LDA model, D documents are
modeled as a mixture over K latent topics. Given a training data set, we use a set x
to represent the words in documents, where xij is the jth word in the ith document.
Furthermore, a corresponding topic assignments set z is maintained, where zij is the
assigned topic for the word xij . The total number of words in the data set is N . Two
matrices, njk and nwk are used in LDA. Specifically, njk is the document-topic count
the number of words in document j assigned to topic k, and nwk is the word-topic
count the number of occurrences of word w assigned to topic k. The sizes of matrices
njk and nwk are D K and W K, respectively, where W is the number of words in
the vocabulary for the given data set.
The original Gibbs sampling is inherently sequential, as each iteration in the training
process depends on the results from the previous iteration. To implement it in parallel,
186 M. Lu et al.

the previous study has proposed a parallel approximate LDA algorithm [2], which has
been shown similar accuracy to the sequential algorithm. Therefore, in this work, we
adopt a similar parallel approximate LDA algorithm for our GPU implementation.

2.2 Graphics Processing Units (GPUs)

GPUs are widely available as commodity components in modern machines. We adopt

the massive thread parallelism programming model from NVIDIA CUDA to design
our algorithm. The GPU is modeled as an architecture including multiple SIMD multi-
processors. All processors on the GPU share the device memory, which has both a high
bandwidth and a high access latency. Each multi-processor has a fast on-chip local
memory (called shared memory in NVIDIAs term) shared by all the scalar processors
in the multi-processor. The size of this local memory is small and the access latency
is low. The threads on each multi-processor are organized into thread blocks. Threads
in a thread block share the computation resources on a multi-processor. Thread blocks
are dynamically scheduled on the multi-processors by the runtime system. Moreover,
when the addresses of the memory accesses of the multiple threads in a thread block are
consecutive, these memory accesses are grouped into one access. This feature is called
coalesced access.

2.3 Parallel and Distributed LDAs

Two parallel LDA algorithms, named AD-LDA and HD-LDA, are proposed by New-
man et al. [2]. The speedup of AD-LDA on a 16-processor computer was up to 8 times.
Chen et al. [7] have implemented AD-LDA using MPI and MapReduce for large-scale
LDA applications. Their implementations achieved a speedup of 10 when 32 ma-
chines were used. Another asynchronous distributed LDA algorithm was proposed by
Asuncion et al. [8], which had a speedup of 15-25 when there were 32 processors.
Among existing studies on GPU-based LDAs, Masada et al. [4] have accelerated
collapsed variational Bayesian inference for LDA using GPUs and achieved a speedup
of up to 7 times compared with the standard LDA on the CPU. Moreover, Yan et al.
[5] have implemented both collapsed Gibbs sampling (CGS) and variational Bayesian
(CVB) methods of LDA on the GPU. Compared with the sequential implementation on
the CPU, the speedups for CGS and CVB are around 26x and 196x, respectively. How-
ever, none of these GPU-based LDAs has effectively addressed the scalability issue. In
contrast, our GLDA can train very large data sets on the GPU efficiently.

3 GLDA Design and Implementation

3.1 Parallel Gibbs Sampling of LDA on GPUs

A traditional parallel approach [2] for LDA (AD-LDA) is that both D documents and
document-topic counts njk are distributed evenly over p processors. However, each
processor needs to maintain a local copy of all word-topic counts. Then each proces-
sor performs LDA training based on its own local documents, document-topic, and
 Accelerating Topic Model Training on a Single Machine 187

word-topic counts independently. At the end of each iteration, a reduction operation

is performed to update all copies of local nwk counts. The major difficulty in apply-
ing this approach to GPUs is that multiple copies of the word-topic count matrix may
not fit into the limited GPU memory. Therefore, Yan et al. [5] have proposed a data
partitioning method to avoid access conflicts while maintaining only one copy of nwk
matrix. The major issue of their implementation is it may cause workload imbalance
since the partitioning scheme is based on both documents and vocabulary, and different
partitions are processed in parallel. The approximate algorithm they have proposed may
not work well for non-uniform data distributions. Moreover, it still cannot handle the
data set when either njk or nwk matrix cannot be entirely stored in the GPU memory.
Instead, we have adopted a different algorithm without such 2-dimension partitioning
and solved the scalability issue.

Algorithm 1. GLDA algorithm for one iteration.

Input:
xp : word tokens assigned to the pth processor
Output: njk , nwk , zij
1 for all processors in parallel do
2 foreach xij xp do
3 Sample zij with global counts njk and nwk
4 /* Global synchronization */
5 Update njk
6 Atomic update nwk

Our parallel algorithm is based on the following fact: With atomic increment and
decrement operations, concurrent execution of these operations is guaranteed to pro-
duce a correct result. Therefore, in our implementation, we also maintain only one copy
of nwk matrix, and adopt two modifications. (1) We use atomic increment and decre-
ment operations, which are supported by recent generation GPUs, in count updates.
(2) We serialize the computation and update on the nwk matrix. With these modifica-
tions, Algorithm 1 outlines our parallel LDA algorithm for one iteration. Compared
with AD-LDA, which does not perform global updates until the end of each iteration,
our algorithm performs global updates after each p words are sampled in p processors
in parallel and guarantees the updated results are correct. Since the multi-processors
on the GPU are tightly coupled, the communication overhead is insignificant in this
implementation. Specifically, we map each thread block in CUDA as a processor, and
the inherent data parallelism of the sampling algorithm is implemented using multiple
threads within each thread block.

3.2 Memory Consumption of Original LDA

We estimate the total memory consumption of the original LDA algorithm in the fol-
lowing four components: the memory used by the word token set x = {xij } and topic
188 M. Lu et al.

assignment set z = {zij } are both sizeof (int) N bytes. The document-topic and
word-topic count matrix consume sizeof (int)DK bytes and sizeof (int)W K
bytes, respectively. Therefore, the estimated total memory size is at least M bytes:

M = sizeof (int) (2 N + (D + W ) K)

Where N , D, and W are fixed for a given data set, and the number of topics K is
assigned by users.
A large K is usually required for large data sets, for example, a
typical value is D. Figure 1 shows the estimated memory consumption for NYTimes
and PubMed data sets that are used in our evaluations (see Table 1 for experimental
setup). Obviously, current generations of GPUs cannot support such large amounts of
required memory space. A straightforward solution is to send the data set xij and topic
assignment set zij to the GPU on-the-fly. However, this method does not solve the
problem when either njk or nwk matrix cannot be entirely stored in the GPU memory.
Therefore, we study more advanced techniques to enable our GLDA to handle large
data sets.

(a) NYTimes (b) PubMed

Fig. 1. Estimated memory consumptions for NYTimes and PubMed data sets (Table 1) with num-
ber of topics varied

3.3 Discarding Global Document-Topic Matrix

We have observed that to generate the njk counts for a given document j, only the word
tokens in that document are necessary to scan. Moreover, the sequential sampling of
words in a single document can share an array with K elements storing the document-
topic counts corresponding to that document. Therefore, we consider discarding the
storage for the global document-topic counts. We only produce the necessary njk counts
for specific documents when these documents are processed (the technique is denoted
as no njk ). Since we process p words in p different documents in parallel, the estimated
memory size required for this method is Mno njk bytes.

Mno njk = sizeof (int) (2 N + W K + p K)

Moreover, the additional cost introduced is equivalent to one scan on the topic assign-
ment set z for each iteration. Due to the high memory bandwidth of the GPU, the gen-
eration of document-topic counts can be implemented efficiently through the coalesced
access pattern.
 Accelerating Topic Model Training on a Single Machine 189

Algorithm 2. Generating document-topic counts for a specific document through

the coalesced access.
Input:
z: the topic assignment array for a given document in the device memory
l: the array with K elements in the local memory
n: the number of elements in z
tx: the thread index in a thread block
nt: the number of threads in a thread block
Output:
njk : the document-topic count array with length K in the device memory for the given
document
1 l is initialized to 0;
2 for i = tx; i < n; i = i + nt do
3 atomicInc(l[z[i]]);
4 /*synchronization within the thread block*/
5 for k = tx; k < K; k = k + nt do
6 njk [k] = l[k];

We use a thread block to generate the document-topic counts for a given document.
Algorithm 2 demonstrates the GPU code for a thread to produce document-topic counts
based on the coalesced access. We have adopted a temporary array stored in the local
memory to improve the memory access efficiency for counting as well as to facilitate
the coalesced access for both reads and writes on the device memory. Without storing
all njk counts, the memory saving is significant with a low computation overhead.
In contrast, the nwk counts are stored. This design choice is made because
re-calculating nwk is expensive - as words are distributed across documents, gener-
ating nwk for each iteration would require approximately N scans of the original data
set. Fortunately, the number of topics and the size of vocabulary are both much smaller
than the data set size; therefore, it is feasible to store the nwk counts even for large data
sets. In the following, we further discuss how to store these matrices efficiently and how
to partition them if they do not fit into memory.

3.4 Sparse Storage Scheme for Count Matrices

We observe that a considerable portion of the values in the nwk and njk matrices be-
come zeros after a burn-in period. For example, after about 100 iterations of training on
the NIPS data set, which is used in our evaluation, the percentages of non-zero values
are around 45% and 10% for njk and nwk counts, respectively. This observation has
motivated us to study the sparse matrix storage scheme for these two counts.
We only store the topic-count pair for non-zero elements for the njk or nwk count
matrix. An array of K topic counts is generated from the sparse storage decompressed
for a given document or word token. The decompression can also be implemented effi-
ciently through the coalesced access as well as the local memory usage, which is similar
to Algorithm 2. Suppose we use a data structure including two elements to store a topic-
count pair, namely id and count. Then to generate an array including all topic counts
for a specific document or word, the value of count will be written to the idth element
190 M. Lu et al.

in the array for specific pairs. Since p word tokens in p different documents are pro-
cessed in parallel, we need to decompress those corresponding njk and nwk counts.
The estimated memory required is Msparse bytes.

Msparse = sizeof (int) (2 N + 2 p K+

2 (D K nzjk % + W K nzwk %))

Where nzjk % and nzwk % refer to the percentage of non-zero elements for njk and
nwk counts, respectively. Note that, when nzjk % or nzwk % is over 50%, this storage
scheme will consume more memory than the original storage format for the njk or nwk
matrix, respectively. Therefore, to decide whether adopting the sparse matrix storage,
we should first estimate the memory space saving.
Although the introduced computation overhead is inexpensive, there is a critical
problem for the GPU-based implementation. Since the percentage of non-zero elements
is not guaranteed to monotonically decrease with iterations, a new topic-count pair may
be inserted when an original zero element becomes non-zero. Thus an additional check-
ing step is necessary after each word sampling process to examine whether new pairs
should be inserted. Unfortunately, our evaluated GPUs do not support dynamical mem-
ory allocation within the GPU kernel code, and the kernel code cannot invoke the CPU
code at runtime. To perform the checking step and also the GPU memory allocation,
some information must be copied from the device memory to the main memory. The
number of such memory copies is around N/p. The overhead of such a large number
of memory copies may be expensive. Our evaluation results confirm this concern. With
the newer generation GPUs, this concern may be partially or fully addressed.

3.5 Data Set Partitioning

Either no njk or sparse storage scheme only partially addresses the memory limitation
issue. To perform LDA for an arbitrary size of data set on the GPU, we consider a
partitioning scheme for word tokens. The basic idea is that the GPU only holds the
necessary information for a subset of word tokens.
Specifically, we divided the set of documents and the vocabulary into n and m dis-
joint subsets by ranges of document IDs or word IDs, respectively. Then there are totally
(n m) partitions and corresponding (n m) computation phases in each iteration. For
the (i m + j)th phase, where i {0, 1, ..., (n 1)} and j {0, 1, ..., (m 1)}, only
the word tokens belonging to the ith document subset and jth vocabulary subset are
processed on the GPU, and corresponding document-topic and word-topic counts are
generated on the GPU. The topic assignments set zij is processed exactly the same as
the word token set xij . Through this partitioning scheme, the GPU can process an arbi-
trary size of data set since we can always decompose a large data set to several smaller
partitions that can fit into the GPU memory.
The algorithms of generating necessary njk and nwk counts are similar to Algorithm 2.
The computation overhead depends on the order of partition processing. Since generating
nwk counts is more expensive, we process the partitions in a vocabulary-oriented style.
Specifically, for the given jth vocabulary subset, the partitions {i m + j|0 i < n}
are processed sequentially. This way, each n partitions can share nwk partitioning counts
 Accelerating Topic Model Training on a Single Machine 191

for the same word token subset. Therefore the overhead is (1 + m) more scans on the
original data set. We denote this global partitioning scheme as doc-voc-part. Partitioning
schemes with m = 1 and n = 1 are denoted as doc-part and voc-part, respectively. doc-
part and voc-part are suitable when the word-topic and document-topic matrix can be
entirely stored in the GPU memory, respectively. For a uniform data distribution with
range partitioning, the memory size requirement is estimated as Mpart bytes:
2N DK W K
Mpart = sizeof (int) ( + + )
nm n m
Note that the workload difference between different partitions will not hurt the overall
performance significantly since these partitions are processed in different computation
phases sequentially rather than in parallel on the GPU. Moreover, if the entire word to-
ken set, the topic assignment set, and all njk and nwk counts can be stored in the main
memory, we can directly copy the required words and counts from the main memory to
the GPU memory and avoid the additional computation for the corresponding njk and
nwk counts generation in each iteration. This can further reduce the overhead but re-
quires large main memory. In our implementation, we do not store njk and nwk counts
in the main memory thus our implementation can also be used on a PC with limited
main memory.

3.6 Discussion
The three techniques we have proposed can also be combined, e.g., either no njk or
sparse storage scheme can be used together with the word token partitioning. To choose
appropriate techniques, we give the following suggestions based on our experimental
evaluation results: doc-part should be first considered if the data set and topic assign-
ments cannot be entirely stored in the GPU memory. Then if all njk counts cannot be
held in the GPU memory, no njk or doc-part should be considered. Finally, the voc-
part should be further adopted if the GPU memory is insufficient to store all nwk counts,
and the sparse matrix storage may be used for nwk .

4 Empirical Evaluation
4.1 Experimental Setup
We conducted the experiments on a Windows XP PC with 2.4 GHz Intel Core2 Duo
Quad processor, 2GB main memory and an NVIDAI GeForce GTX 280 GPU (G280).
G280 has 240 1.3 GHz scalar processors, and 1GB device memory. Our GLDA is devel-
oped using NVIDIA CUDA. The CPU counterparts (CPU LDA) are based on a widely

Table 1. Statistics of four real-world data sets

KOS NIPS NYTimes PubMed

D 3,430 1,500 300,000 1,000,000
W 6,906 12,419 102,660 116,324
N 0.47 106 1.9 106 100 106 70 106
192 M. Lu et al.

used open-source package GibbsLDA++1 . Note that CPU LDA is a sequential program
using only one core. Due to the unavailability of the code from the previous work on
GPU-based LDA [5], we could not compare our GLDA with it. However, the speedup
reported by their paper is similar to ours when the data can be entirely stored on the
GPU.
Four real-world data sets are used in our evaluations 2 , which are illustrated in Table
1. We use a subset of the original PubMed data set since the main memory cannot
hold the original PubMed data set. The two small data sets KOS and NIPS are used
to measure the accuracy of GLDA, and the two large data sets NYTimes and PubMed
are used to evaluate the efficiency. Moreover, we also have a randomly sampled subset
with a size half of the original NY Times dataset (denoted as Sub-NYTimes) to study the
overhead of different techniques when all data can be stored in the GPU memory. For
each data set, we randomly split it to two subsets. 90% of the original documents are
used for training, and the remaining 10% are used for test. Throughout our evaluations,
the hyperparameters and are set to 50/K and 0.1, respectively [6].
We use the perplexity [9] to evaluate the accuracy of GLDA. Specifically, after ob-
taining the LDA model from 1000 iteration training, we calculate the perplexity of the
test data.

4.2 Perplexity Performance

Figure 2 shows the perplexity results for the KOS and NIPS data sets with number of
thread blocks varied when K=64 and K=128. Note that when there is only one thread
block, the implementation is equivalent to the standard sequential LDA. Therefore, per-
plexity with multiple thread blocks is expected to be similar to that with one thread
block. This figure shows that there is no significant difference for the perplexity results
between the standard LDA model and our parallel LDA algorithm.

(a) KOS (b) NIPS

Fig. 2. Perplexity of KOS and NIPS data sets with number of thread blocks varied

4.3 Memory Consumption and Efficiency

We first study the performance impact of the thread parallelism when all data is stored
in the GPU using the data set Sub-NYTimes. We first set the number of threads in
1
http://gibbslda.sourceforge.net/
2
http://archive.ics.uci.edu/ml/datasets/Bag+of+Words
 Accelerating Topic Model Training on a Single Machine 193

each thread block to 32 when varying the number of blocks. Then the number of thread
block is fixed to 512 and the number of threads in each thread block is varied. Figure 3
shows the elapsed time of one iteration for the GLDA with the number of thread blocks
and number of threads in each thread block varied. Compared with the CPU-LDAs
performance of 10 minutes per iteration, this figure shows significant speedups when
either number of thread blocks or threads per block is increased. However, it nearly
maintains a constant or slightly reduced when the GPU resource is fully utilized.

(a) Number of blocks varied. (b) Number of threads varied.

Fig. 3. Elapsed time of one iteration on the GPU for Sub-NYTimes with the number of blocks
and threads per block varied, K=256

Next, we study the overhead of techniques used to address the scalability issue. We
first measure the performance when only a single method is adopted. Figure 4(a) shows
the elapsed time of GLDA adopting different techniques. opt refers to the implementa-
tion without any additional techniques and is optimized through the thread parallelism.
There are four partitions for both doc-part and voc-part. doc-voc-part has divided both
documents and vocabulary into two subsets, thus also resulting four partitions. The
figure shows that the sparse storage scheme has a relatively large overhead compared
with the other methods. Through our detailed study, we find the additional overhead is
mainly from the large number of memory copies between the main memory and GPU
memory. Moreover, voc-part is slightly more expensive than doc-part and no njk since
it needs to scan the original data set for each partition. For data sets used in our evalu-
ations, the sparse storage scheme is always more expensive than other techniques, thus
we focus on the other two techniques in the following evaluations. We further study the
performance overhead when the technique no njk and partitioning are used together.
Figure 4(b) shows the elapsed time when these two techniques are combined. The com-
bined approach is slightly more expensive than only one technique adopted since the
overhead from two techniques are both kept for the combined method.
To investigate the memory space saving from different techniques, Figure 5 shows
the GPU memory consumptions corresponding to Figure 4. It demonstrates that the
memory requirement is reduced by around 14%-60% through various techniques. Since
the number of topics is not very large and the partitioning scheme can also reduce the
memory size consumed for the word token set and topic assignment set, partitioning
is more effective than the other two techniques. Moreover, Figure 5(b) shows that the
combined technique can further reduce the memory consumption compared with a sin-
gle technique adopted.
194 M. Lu et al.

(a) Adopting single technique (b) Adopting multiple techniques

Fig. 4. Elapsed time of one iteration on the GPU for Sub-NYTimes when adopting different
techniques, K=256

(a) Adopting single technique (b) Adopting multiple techniques

Fig. 5. Overall memory consumption on the GPU for Sub-NYTimes of adopting different tech-
niques, K=256

(a) NYTimes (b) PubMed

Fig. 6. Performance comparisons for one iteration of NYTimes and PubMed data sets with num-
ber of topics varied

Finally, we show the overall performance comparisons based on the NYTimes and
PubMed data sets with reasonable numbers of topics. The CPU counterparts are im-
plemented using the same set of techniques since the main memory is also insufficient
to handle such data set for the original CPU implementation. The estimated memory
consumptions using a traditional LDA algorithm for these two data sets are presented
in Figure 1, and the actual GPU memory consumption is around 700MB for each data
sets. The technique selection is based on the discussion in the implementation section.
Specifically, the evaluation with NYTimes has adopted the techniques doc-part, while
 Accelerating Topic Model Training on a Single Machine 195

doc-voc-part and no jjk are both adopted for the PubMed data set. Figure 6 shows
that our GLDA implementations are around 10-15x faster than their CPU counterparts.
Such speedup is significant especially for large data sets. For example, suppose 1000
iterations are required for the PubMed with 2048 topics, GLDA could accelerate the
computation from originally more than two months to only around 5 days.

5 Conclusion
We implemented GLDA, a GPU-based LDA library that features high speed and scal-
ability. On a single GPU with 1 GB GPU memory, we have evaluated the performance
using the data sets originally requiring up to 10 GB memory successfully. Our exper-
imental studies demonstrate that GLDA can handle such large data sets and provide a
performance speedup of up to 15X on a G280 over a popular open-source LDA library
on a single PC.

Acknowledgement. This work was supported by grant 617509 from the Research
Grants Council of Hong Kong.

References
1. Blei, D.M., Ng, A.Y., Jordan, M.I.: Latent dirichlet allocation. Journal of Machine Learning
Research 3, 9931022 (2003)
2. Newman, D., Asuncion, A., Smyth, P., Welling, M.: Distributed inference for latent dirichlet
allocation. In: NIPS (2007)
3. Owens, J.D., Luebke, D., Govindaraju, N.K., Harris, M., Kruger, J., Lefohn, A.E., Purcell,
T.J.: A survey of general-purpose computation on graphics hardware. In: Eurographics 2005,
State of the Art Reports (2005)
4. Masada, T., Hamada, T., Shibata, Y., Oguri, K.: Accelerating collapsed variational bayesian
inference for latent dirichlet allocation with nvidia CUDA compatible devices. In: Chien, B.-
C., Hong, T.-P., Chen, S.-M., Ali, M. (eds.) IEA/AIE 2009. LNCS, vol. 5579, pp. 491500.
Springer, Heidelberg (2009)
5. Yan, F., Xu, N., Qi, Y.: Parallel inference for latent dirichlet allocation on graphics processing
units. In: NIPS 2009, pp. 21342142 (2009)
6. Griffiths, T.L., Steyvers, M.: Finding scientific topics. Proceedings of the National Academy
of Sciences, PNAS 2004 (2004)
7. Chen, W.Y., Chu, J.C., Luan, J., Bai, H., Wang, Y., Chang, E.Y.: Collaborative filtering for
orkut communities: discovery of user latent behavior. In: WWW 2009 (2009)
8. Asuncion, A., Smyth, P., Welling, M.: Asynchronous distributed learning of topic models. In:
NIPS (2008)
9. Azzopardi, L., Girolami, M., van Risjbergen, K.: Investigating the relationship between lan-
guage model perplexity and ir precision-recall measures. In: SIGIR 2003, pp. 369370 (2003)
 Collusion Detection in Online Rating Systems

Mohammad Allahbakhsh1, Aleksandar Ignjatovic1, Boualem Benatallah1 ,

Seyed-Mehdi-Reza Beheshti1 , Elisa Bertino2 , and Norman Foo1
1
The University of New South Wales, Sydney, NSW 2052, Australia
{mallahbakhsh,ignjat,boualem,sbeheshti,norman}@cse.unsw.edu.au
2
Department of Computer Science and CERIAS, Purdue University, Indiana, USA
[email protected]

Abstract. Online rating systems are subject to unfair evaluations. Users may
try to individually or collaboratively promote or demote a product. Collabora-
tive unfair rating, i.e., collusion, is more damaging than individual unfair rating.
Detecting massive collusive attacks as well as honest looking intelligent attacks
is still a real challenge for collusion detection systems. In this paper, we study
impact of collusion in online rating systems and asses their susceptibility to col-
lusion attacks. The proposed model uses frequent itemset mining technique to
detect candidate collusion groups and sub-groups. Then, several indicators are
used for identifying collusion groups and to estimate how damaging such collud-
ing groups might be. The model has been implemented and we present results of
experimental evaluation of our methodology.

1 Introduction

Today, in e-commerce systems, buyers usually rely on the feedback posted by others
via online rating systems to decide on purchasing a product [13]. The higher the num-
ber of positive feedback on a product, the higher the possibility that one buys such a
product. This fact motivates people to promote products of their interest or demote the
products in which they are not interested by posting unfair rating scores [7, 6]. Unfair
rating scores are scores which are cast regardless of the quality of a product and usu-
ally are given based on personal vested interests of the users. For example, providers
may try to submit supporting feedback to increase the rating of their product in order
to increase their income [6]. The providers also may attack their competitors by giv-
ing low scores on their competitors products. Also, sometimes sellers in eBay boost
their reputations unfairly by buying or selling feedback [7]. Unfair rating scores may be
given individually or collaboratively [18]. Collaborative unfair ratings which are also
called collusion [17, 18] by their nature are more sophisticated and harder to detect than
individual unfair ratings [17]. For that reason, in this work we focus on studying and
identifying collusion.
Although collusion is widely studied in online collaborative systems [4, 8, 10, 13,
12, 9, 19], such systems still face some serious challenges. One major challenge arises
when a group of raters try to completely take control of a product i.e., when the number
of unfair reviewers is significantly higher than the number of honest users; the existing
models usually can not detect such groups. Also, the existing models do not perform

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 196207, 2013.
c Springer-Verlag Berlin Heidelberg 2013
 Collusion Detection in Online Rating Systems 197

well against intelligent attacks, in which group members try to give an appearance of
honest users. For example, typically they will not deviate from the majoritys ranking on
most of the cast feedback and target only a small fraction of the products. Such attacks
are hard to identify using the existing methods [19].
Moreover, there are cases in which a large group of people have rated the same group
of products, and thus could be considered as a potential collusion group. Such a group
might be subsequently deemed as non collusive. However, in such cases there may exist
some smaller collusive sub-groups inside the large group which are collusive, but when
put along with others they might be undetected. Detection of such sub-groups is not
addressed in the existing collusion detection models.
In this paper we propose a framework for detecting and representing collusion in
online rating systems. Besides indicators used in the already existing work, we also
define two novel indicators. One indicator which we call Suspiciousness of a reviewer,
which is a metric to estimate to what extent ratings posted by such reviewer correspond
to majority consensus. Such indicator is calculated using two distance functions: the Lp
norm and the Uniform distance (see Section 4.1 for more details). The other indicator
which we call spamicity estimates the likelihood that a particular rating score given to
a product by a reviewer is unfair.
We also propose a graph data model for representing rating activities in online rat-
ing system. This model allows representing products, reviewers and the rating scores
reviewers have cast on products and identifying collusive groups. We propose a new
notion and representation for collusion groups called biclique. A biclique is a group of
users and a group of products such that every reviewer in such a group has rated every
product in the corresponding group of products. We use bicliques to detect collusion.
Moreover, we propose an algorithm which employs two clustering algorithms built
upon frequent itemset mining (FIM) technique [1] for finding bicliques and sub-bicliques.
Then the algorithm uses the proposed collusion indicators to find any possible collusion
in online rating systems. We have implemented our model and tested it and the evaluation
results show the efficiency of our model.
The rest of the paper is organized as follows. In section 2 we propose our data model.
In Section 3 we propose our example scenario and data preprocessing details. Our collu-
sion detection method is proposed in Section 4 and the corresponding algorithm appears
in Section 5. In section 6 we propose implementation details and also evaluate results.
We discuss related work in section 7 and conclude in section 8.

2 Representing Online Rating Entities

We define a graph data model (i.e. ORM: Online Rating Model) for organizing a set of
entities (e.g. reviewers and products) and relationships among them in an online rating
system. ORM can be used to distinguish between fair and unfair ratings in an online
rating system and helps to calculate a more realistic rating score for every product. In
ORM, we assume that interactions between reviewers and products are represented by
a directed graph G = (V, E) where V is a set of nodes representing entities and E is a
set of directed edges representing relationships between nodes. Each edge is labeled by
a triplet of numbers, as explained in the followings.
198 M. Allahbakhsh et al.

An entity is an object that exists independently and has a unique identity. ORM
consists of three types of entities: products, reviewers and bicliques. We use the concept
of folder used in our previous work [3] to represent bicliques.
Product. A product is an item which has been rated by system users in terms of quality
or any other possible aspects. Products are described by a set of attributes such as the
unique indicator (i.e. ID), title, and category (e.g. book, cd, track, etc). We assume that
there are Np products P = {pj |1 j Np } in the system to be rated.
Reviewer. A reviewer is a person who has rated at least one product. Reviewers are
described by a set of attributes and are identified by their unique identifier (i.e. ID). We
assume that there are Nu reviewers U = {ui |1 i Nu } rating products.
Rating Relationship. A relationship is a directed link between a pair of entities, which
is associated with a predicate defined on the attributes of entities that characterizes the
relationship. We assume that no reviewer can rate a product more than once. So, in
ORM, there is at most one relation between every pair of products and reviewers.
When a reviewer rates a product, a rating relationship is established between corre-
sponding reviewer and product. We define R = {eij | ui U pj P } as the set of
all rating relationships between reviewers and products i.e. eij R is the rating score
which the ui has given to pj . A rating relationship is weighted with the values of the
following three attributes:

1. The value is the evaluation of the the reviewer from the quality of the product and
is in the range [1, M ], M > 1. M is a system dependent constant. for example in
Amazon, rating scores reside between 1 and 5. We denote the value field of the
rating by eij .v.
2. The time shows the time in which the rating has been posted to the system. The
time field of the rating is an integer number and is denoted by eij .t.
3. As we mentioned earlier we assume that every reviewer can have at most one
relation to each product (i.e., we do not allow multi-graphs). However, in real
world, e.g. in Amazon rating system, one reviewer can rate a product several times.
Some collusion detection models like [12] just eliminate duplicate rating scores.
We rather use them for the purpose of detecting unfair ratings. Spamicity shows
what fraction of all scores cast for this particular product are cast by the reviewer in
the considered relationship with this product. We denote the spamicity of a rating
relation by eij .spam.

2.1 Biclique
A biclique is a sub-graph of ORM containing a set of reviewers R, a set of products
P and their corresponding relationships Rel. All reviewers in R of a biclique have
rated all products in the corresponding P , i.e., there is a rating relationship between
every r R and every p P . A biclique is denoted by BC = {R, P, Rel}. For
p1 , ...}} means that reviewers u1 ,
example, BC = {{r1 , r2 , r3 }, {p1 , p2 , p3 }, {r1
u2 , and u3 all have voted on p1 , p2 and p3 . We will use terms biclique and group
interchangeably throughout this paper.
 Collusion Detection in Online Rating Systems 199

3 Example Scenario
Amazon is one of the well-known online markets. Providers or sellers put products on
the Amazon online market. Buyers go to Amazon and buy products if they find them
of an acceptable quality, price, etc. Users also can share their experiences with others
as reviews or rating scores they cast on products. Rating scores are numbers in range
[1, 5]. Amazon generates an overall rating score for every product based on the rating
scores cast by users. Some evidence like [6] show that the Amazon rating system has
widely been subject to collusion and unfair ratings.
We use the log of Amazon online rating system1 which was collected by Leskovec
et. al. for analyzing dynamics of viral Marketing [11], referred in the following as AM-
ZLog. This log contains more than 7 million ratings cast on the quality of more than
500 thousands of products collected in the summer of 2006.
We preprocess AMZLog to fit our graph data model. We delete inactive reviewers
and unpopular products and only keep reviewers who reviewed at least 10 products and
the products on which at least 10 ratings are cast. We also change the date format from
yyyy-mm-dd to an integer number reflecting the the number of days between the date
on which the first rating score in the system has been posted and the day in which this
rating score has been cast.
We use redundant votes from one rater on same product to calculate spamicity of
their relation. Let E(j) be the set of all ratings given to product pj and E(i, j) be the
set of all ratings given by reviewer ui to the product pj . We calculate spamicity degree
of the relationship between the reviewer ri and the product pj as follows:
!
0 if (|E(i, j)| 2)
eij .spam = |E(i,j)|
|E(j)| otherwise

In the above equation allowing casting two votes instead of one accommodates for the
situations where a genuine mind change has taken place.

4 Collusion Detection Method

To find collusive bicliques, first we have to identify all collaborative rating groups as
collusion biclique candidates, then check them to find real collusion groups. We em-
ploys FIM technique [1] to find bicliques. More details on our biclique detection algo-
rithm can be found here [2].

4.1 Collusion Indicators

It is almost impossible to identify collusion bicliques just by analyzing the behavior of
individual members or even only based on the posted rating scores [12]. Rather, several
indicators must be checked to realize in what extent a group is a collusion group [13].
In the followings, we propose some indicators and show how they indicate to a possible
collusive activity.
1
http://snap.stanford.edu/data/amazon-meta.html
200 M. Allahbakhsh et al.

Group Rating Value Similarity (GVS). A group of reviewers can be considered as

formers of a collusive biclique (denote by g), if they have posted similar rating scores
for similar products. For example, all of them promoted a set of products and demoted
another set of products by similar rating scores. To find this similarity we calculate the
Pairwise Rating Value Similarity between every pair of reviewers in the group. Pair-
wise rating value similarity denoted by V S(i, j) shows to what extent ui and uj have
cast similar rating scores to every product in g.P . We use cosine similarity model, a
well-known model for similarity detection [15], for finding similarities between group
members. Since V S(i, j) is the cosine of the angle between two vectors containing rat-
ings of two reviewers, it is a value in range [0, 1]. The value 1 means completely same
and 0 means completely different. Suppose that ui and uj are two members of group g
i.e. ui and uj g.R . We calculate similarity between them as follows:
#
kg.R vik vjk
V S(i, j) = 5# 5# (1)
2 2
kg.R (vik ) kg.R (vjk )

We then calculate an overall degree of similarity for every group to show how all mem-
bers are similar in terms of values they have posted as rating scores and call it Group
rating Value Similarity (GVS). GVS for every group is the minimum amount of pair-
wise similarity between group members. The GVS of group g is calculated as follows.

GV S(g) = min(V S(i, j)), for all ui and uj g.R (2)

The bigger the GV S, the more similar the group members are in terms of their posted
rating values.
Group Rating Time Similarity (GTS). Another indicator for finding collusive groups
is that they usually cast their ratings in a short period of time. In other words, when a
group of users aim to promote or demote a product they should do it in a short period
of time in order to gain benefit, get paid, etc. We use this fact as an indicator to find
collusive groups and call it Group rating Time Similarity (GTS). To calculate GTS, we
find the the Time Window (TW) in which the rating scores have been posted on each
product p g.P . The beginning of the TW is the time of casting the first rating score
on the products and its end is the time of casting the last rating score on the product by
members of the group. We use the size of the TW in comparison to a constant M axT W
to show how big such a TW is. The parameter M axT W is the maximum size of a TW
which might be suspected to be collusive. So, for every product in the group we have:
4
0 if (M axT (j) M inT (j) > M axT W )
T W (j) = (3)
1 MaxTMaxT
(j)MinT (j)
W Otherwise

Where M inT = min(eij .t), M axT = max(eij .t), j g.P for all i g.R
Now, we choose the largest T W (j) as the degree of time similarity between the
ratings posted by group members on the target products. We say
GT S(g) = max(T W (j)), for all pj g.P (4)
The bigger the GT S, the more similar the group members are in terms of the time of
posting their rating scores.
 Collusion Detection in Online Rating Systems 201

Group Ratings Spamicity (GRS). As we described in Sections 2 and 3, the spamicity

of a rating relationship shows the suspiciousness of the rating score because of high
number of ratings posted by a same user to a same product. We define the Group Rating
Spamicity (GRS) as follows:
#
eg.Rel e.v e.spam
GRS(g) = # (5)
eg.Rel e.v

Group Members Suspiciousness (GMS). We identify suspicious users in four steps.

Step 1: Suppose that E(j) is the set of all ratings posted on the product pj . We find the
median of the members of E(j) and denote it by mj . We then calculate the average dis-
tance of all ratings posted on pj from the median mj and denote it by dj . We calculate
the average distance using equation (6).
6#
2
iE(j) (eij .v mj )
dj = (6)
||E(j)||

Now, we calculate a credibility degree for every eij E(j). We denote this credibility
degree with ij and use it to marginalize outliers. The ij is calculated as follows.
4
1 if (mj dj ) rij (mj + dj )
ij = (7)
0 otherwise

Equation (7) implies that only the ratings which fall in range (mj dj ) are considered
as credible.
Step 2: In this step, we calculate the averages of all credible ratings on the product pj
and assume that it is a dependable guess for the real rating of the product. We show it
by gj and calculate it as follows.
#
iE(j) (eij .v ij )
gj = # (8)
iE(j) ij

Step 3: In the third step, using the ratings we guessed for every product (Equation (8)),
we calculate two error rates for every reviewer. The first error rate is the LP error rate
(here L2 ) which is L2 norm of all differences between the ratings cast by the reviewer
on the quality of pj and the gj . We denote the Lp error rate of reviewer ui by LP (i).
Suppose that Ji is the set of indices of all products have been rated by ui . The LP (i) is
calculated as follows. 6
" 788 8%2
8
LP (i) = 8eij .v gj 8 (9)
jJi

The second error rate we calculate for every reviewer is the uniform norm of differences
between eij .v and gj for all products have been rated by ui . We call this error rate
uniform error rate of ui , denote it by U N (i) and calculate it as follows.

U N (i) = max (|eij .v gj |) , j Ji (10)

202 M. Allahbakhsh et al.

Algorithm 1. Collusion Detection Algorithm

Input: ORM Graph, , M axT W
Output: C as the set of all collusive Bicliques
Find all bicliques and store them in T empC
for all tc T empC do
Calculate all indicators for tc
end for
repeat
tc = first member of T empC
if tc.P OC > then
Move tc to C
else
if tc.DI < then
Remove tc from T empC
else
SC = Sub-Bicliques of tc
Calculate Indicators for all members of SC
T empC = T empC SC
Remove tc from T empC
end if
end if
until (T empC.size > 0)
return C

Step 4: The suspicious reviewers are the people who have have large LP (i) or while
they have normal LP (i) they have high U N (i) values. To identify suspicious review-
ers, we identify the normal range of error rates for all reviewers. Based on the calculated
range, the outliers are considered as suspicious reviewers. Suppose that LP 9 is the me-
dian of all LP (i) and U9 N is the median of all U N (i). Also, we assume that LP and
U N are the standard distance of all LP (i) and U N (i) respectively, calculated similar
to the method being used in Equation (6). The list of suspicious reviewers is denoted by
S and built using Equation (11).
: $ % $7 ; 7 ;
S = u| u U LP (u) < LP 9 LP LP (u) > LP 9 + LP
7 ; 7 ;% (11)
9
U N (u) < U 9
N U N U N (u) > U N + UN .

Now, we define GMS of a group as follows.

|S(g)|
GM S(g) = where S(g) = (g.R) S (12)
|g.R|

4.2 How Destructive a Biclique Might Be (The Damaging Impact)

The damaging impact of a collusion group is its ability to impact normal behavior of the
system and final rating scores of products. There are two important parameters reflect-
ing damaging impact of a group: (i) size of the group and (ii) the number of products
 Collusion Detection in Online Rating Systems 203

have been attacked by the group. The bigger these two parameters are, the more defec-
tive such a collusive group is.
Group Size (GS). Size of a collusive group (GS) is proportional to the number of
reviewers who have collaborated in the group i.e., g.R.size and is calculated as follows.
|g.R|
GS(g) = where g is a group (13)
max(|g.R|)
GS is a parameter between (0, 1] and showing how large is the number of members of
a group in comparison with other groups.
Group Target Products Size (GPS). The size of the target of a group (GPS) is propor-
tional to the number of products which have been attacked by members of a collusion
group. The bigger the GTS, the more defective the group will be. GTS is a number in
range (0, 1] and is calculated as follows.
|g.P |
GP S(g) = where g is a group (14)
max(|g.P |)

4.3 Finding Collusive Sub-bicliques

When finding bicliques, we try to maximize size of the group, i.e. g.R.size, to find the
largest possible collusive bicliques. It is possible that a large group is identified as an
honest group due to large and diverse number of users who have just rated same subset
of products. But possibly, there might exist some smaller collusive sub-groups inside.
To find such sub-groups, we use again the idea of finding candidate bicliques using FIM
technique. But in this stage, we only investigate bicliques which have damaging impact
greater than a threshold called . is the collusion threshold and is specified when
collusion cliques are found using our proposed algorithm. Due to space limitation, both
biclique and sub-biclique detection algorithms are available in [2].

5 Collusion Detection Algorithm

It is quite impossible to certainly find out if a group is collusive or not [13]. There-
fore, we define a metric called Probability of Collusion (POC) to show to what ex-
tent a group seems to be collusive. POC is an aggregation of four collusion indicators.
Since in different environments, the importance of these indicators may be different, our
model enables user to assign weight to every indicator to have adequate impact on POC.
Suppose that WGV S , WGT S , WGRS and WGMS are corresponding weights for GV S,
GT S, GRS and GM S respectively so that WGV S + WGT S + WGRS + WGMS = 1.
The POC is calculated as follows.

P OC(g) = GV S WGV S + GT S WGT S + GRS WGRS + GM S WGMS

Moreover, for every group we calculate its damaging impact (DI) as follows.
GP S + GS
DI(g) =
2
204 M. Allahbakhsh et al.

Finally we propose Algorithm 1 for collusion detection. This Algorithm uses POC, DI
and ORM graph along with constants and M axtT W to detect collusion in online
rating systems.

6 Implementation and Evaluation

We have implemented a framework for
GVS GTS
querying and analyzing collusion in on- 1.2 1.2

line rating systems. Our framework com- 1 1

prises several parts which enable user 0.8 0.8

to easily customize collusion detection 0.6 0.6

process and query and analyze collu-

0.4 0.4

0.2 0.2

sion with the simple query language pro- 0 0

vided for this purpose. More details of 0 0.5 1 1.5 0 0.5 1 1.5

the framework architecture and imple- GRS GMS

1.2 1.2
mentation can be found in [2] and [3]. 1 1

To evaluate our model, we randomly 0.8 0.8

chose 20 groups identified within AM- 0.6 0.6

ZLog. Then we asked domain experts 0.4 0.4

to manually check whether these groups

0.2 0.2

0 0

are collusive or not. Experts using their 0 0.5 1 1.5 0 0.5 1 1.5

experience and information provided in Fig. 1. Cumulative distribution of values of col-

Amazon web site such as rating scores, lusion indicators for collusive(red/dashed) and
corresponding written reviews,etc ana- non-collusive (blue/solid) selected groups
lyzed these 20 selected groups. 7 groups
out of 20 were identified as collusive and
13 as honest.We used this dataset to evaluate our model.
Statistical Evaluation. For each group in dataset, we calculate all four collusion indica-
tors. Then we calculate the cumulative distribution of every indicator for both collusive
and non-collusive groups. The results of these calculations are shown in Figure 1. In
every section of the chart, the red/dashed line shows the cumulative distribution of the
corresponding indicator of collusive groups while the blue/solid line reflects the corre-
sponding value for non-collusive groups. Also, the vertical axis is the cumulative value
and horizontal axis is the percentage.
For every indicator in the chart, the cumulative distribution charts of the non-collusive
groups are on the left side while collusive charts are more closer to one. It simply means
that in average, for every indicator, values calculated for collusive groups are larger than
values calculated for non-collusive groups. Therefor, these indicators and the way they
are calculated are truly reflecting the behavior of the group, and can be used as indica-
tors to collusive behavior.
Evaluating Quality of the Results. To evaluate quality of the results, we use the well-
known measures of precision and recall [16]. Precision measures the quality of the
results and is defined by the ratio of the relevant results to the total number of retrieved
results. Recall measures coverage of the relevant results and is defined by the ratio of
relevant results retrieved to the total number of existing relevant results in database.
 Collusion Detection in Online Rating Systems 205

1.2 Table 1. A Sample Set of Identified

1 bicliques
0.8
Group Probability of Damaging
0.6
ID Collusiveness Impact
0.4
... ... ...
0.2 2 0.598 0.599
0 4 0.37 0.49
0 0.2 0.4 0.6 0.8 1 1.2
 6 0.45 0.42
9 0.056 0.173
Fig. 2. How precision (blue/solid) and recall 11 0.121 0.41
(red/dashed) change with different threshold ... ... ...
values.

An effective model should achieve a high precision and a high recall; but its is not
possible in real world, because these metrics are inversely related [16]. It means that
cost of improvement in precision is reduction of recall and vice versa. We calculate pre-
cision and recall with different thresholds to show how value of impacts the quality
and accuracy of the results. Figure 2 shows the results of running our algorithm with
different threshold values. We do not specify particular value for precision and recall.
We only say that if the user wants to achieve the highest possible values for both preci-
sion and recall metrics, Figure 2 obviously shows that the optimal value for threshold
() is 0.4. In this case 71% of the bicliques are retrieved and 71% of retrieved results are
really collusive. One can change to change quality or coverage of data. Also, for more
accuracy one can run the model with different groups of randomly chosen bicliques.
Our model also calculates a damaging impact (DI) factor for every group to show
how damaging the group is. DI helps users to identify groups that have a high potential
damaging power. These groups are mainly groups with high number of targeted products.
Table 1 shows a sample set of bicliques and their DI and POC. The group 2 has high values
for POC and DI, so it is a damaging group. On the other hand, group 9 has small values
for POC and DI, and is not a really defective group. Looking at group 11 reveals that
although the POC of the group is small (0.121), but it still can be damaging because its
DI is 0.41. DI is also useful when comparing groups, e.g, comparing 4 and 6. Although,
the POC of group 6 is 0.08 higher than POC of group 4, the DI of group 4 is 0.07 higher
than DI of group 6. Therefore we can classify them similar rather than deeming group 4
more damaging than group 6. Without having DI, the damaging potential of groups like
11 or 4 may be underestimated and it can lead to incorrect calculation of rating scores.

7 Related Work

Collusion detection has been widely studied in P2P systems [4, 8, 14]. For instance,
EigenTrust [8] tries to build a robust reputation score for P2P collaborators but a re-
search [14] shows that it is still prone to collusion. A comprehensive survey on collusion
detection in P2P systems can be found here [4]. Models proposed for detecting collud-
ers in P2P systems are not directly applicable to online rating systems, because in P2P
systems models are mostly built on relations and communications between people; but
206 M. Allahbakhsh et al.

in online rating systems there is no direct relation between raters. Reputation manage-
ment systems are also targeted by collusion. Very similar to rating systems, colluders in
reputation management systems try to manipulate reputation scores by collusion. Many
efforts are put on detecting collusion using majority rules, weight of the voter and tem-
poral analysis of the behavior of the users [17] but none of these models is enough strong
to detect all sorts of collusion [17, 18]. Yang et. al. [19] try to identify collusion by em-
ploying both majority rule and temporal behavior analysis. Their model is not still tested
thoroughly and just is applied to a specific dataset and a particular type of attack.
The most similar work to ours is [13]. In this work Mukherjee et.al. propose a model
for spotting fake review groups in online rating systems. The model analyzes textual
feedbacks cast on products in Amazon online market to find collusion groups. They use
8 indicators to identify colluders and propose an algorithm for ranking collusion groups
based on their degree of spamicity. However, our model is different from this model
in terms of proposed indicators, analyzing personal behavior of the raters and dealing
with redundant rating scores. Also a recent survey [5] shows that buyers rely more on
scores and ratings when intend to buy something rather than reading textual items. So,
in contrast with this model we focus on numerical aspect of posted feedback. However,
the model proposed by Mukherjee et.al. is still vulnerable to some attacks. For example,
if the number of attackers is much higher than honest raters on a product the model can
not identify it as a potential case of collusion.
Another major difference between our work and other related work is that, we pro-
pose a graph data model and also a flexible query language [2] for better understanding,
analyzing and querying collusion. This aspect is missing in almost all previous work.

8 Conclusion
In this paper we proposed a novel framework for collusion detection in online rating
systems. We used two algorithms designed using frequent itemset mining technique for
finding candidate collusive groups and sub-groups in our dataset. We propose several
indicators showing the possibility of collusion from different aspects. We used these
indicators to assign every group a rank to show their probability of collusion and also
damaging impact. We evaluated our model first statically and showed the adequacy of
the way we define and calculate collusion indicators. We then used precision and recall
metrics to show quality of output of our method.
As future direction, we plan to identify more possible collusion indicators. We also
plan to extend the implemented query language with more features and enhanced visual
query builder to assist users employing our model. Moreover, we plan to generalize our
model and apply it to other possible areas which are subject to collusive activities.

References
1. Agrawal, R., Srikant, R.: Fast algorithms for mining association rules in large databases. In:
Proceedings of VLDB 1994, pp. 487499 (1994)
2. Allahbakhsh, M., Ignjatovic, A., Benatallah, B., Beheshti, S.-M.-R., Foo, N., Bertino, E.:
Detecting, Representing and Querying Collusion in Online Rating Systems. ArXiv e-prints
(November 2012)
 Collusion Detection in Online Rating Systems 207

3. Beheshti, S.-M.-R., Benatallah, B., Motahari-Nezhad, H.R., Allahbakhsh, M.: A framework

and a language for on-line analytical processing on graphs. In: Wang, X.S., Cruz, I., Delis, A.,
Huang, G. (eds.) WISE 2012. LNCS, vol. 7651, pp. 213227. Springer, Heidelberg (2012)
4. Ciccarelli, G., Cigno, R.L.: Collusion in peer-to-peer systems. Computer Networks 55(15),
35173532 (2011)
5. Flanagin, A., Metzger, M., Pure, R., Markov, A.: User-generated ratings and the evaluation
of credibility and product quality in ecommerce transactions. In: HICSS 2011, pp. 110.
IEEE (2011)
6. Harmon, A.: Amazon glitch unmasks war of reviewers. NY Times (February 14, 2004)
7. Brown, J.M.J.: Reputation in online auctions: The market for trust. California Management
Review 49(1), 6181 (2006)
8. Kamvar, S.D., Schlosser, M.T., Garcia-Molina, H.: The eigentrust algorithm for reputation
management in p2p networks. In: Proceedings of the WWW 2003, pp. 640651 (2003)
9. Kerr, R.: Coalition detection and identification. In: The 9th International Conference on Au-
tonomous Agents and Multiagent Systems, pp. 16571658 (2010)
10. Lee, H., Kim, J., Shin, K.: Simplified clique detection for collusion-resistant reputation man-
agement scheme in p2p networks. In: ISCIT 2010, pp. 273278 (2010)
11. Leskovec, J., Adamic, L.A., Huberman, B.A.: The dynamics of viral marketing, vol. 1. ACM,
New York (2007)
12. Lim, E., et al.: Detecting product review spammers using rating behaviors. In: Proceedings
of the 19th ACM International Conference on Information and Knowledge Management,
CIKM 2010, pp. 939948. ACM, New York (2010)
13. Mukherjee, A., Liu, B., Glance, N.: Spotting fake reviewer groups in consumer reviews. In:
Proceedings of the 21st International Conference on World Wide Web. ACM (2012)
14. Qio, L., et al.: An empirical study of collusion behavior in the maze p2p file-sharing system.
In: Proceedings of the ICDCS 2007, p. 56 (2007)
15. Salton, G., Buckley, C., Fox, E.A.: Automatic query formulations in information retrieval.
Journal of the American Society for Information Science 34(4), 262280 (1983)
16. Salton, G., McGill, M.: Introduction to modern information retrieval. McGraw-Hill computer
science series. McGraw-Hill (1983)
17. Sun, Y., Liu, Y.: Security of online reputation systems: The evolution of attacks and defenses.
IEEE Signal Processing Magazine 29(2), 8797 (2012)
18. Swamynathan, G., Almeroth, K., Zhao, B.: The design of a reliable reputation system. Elec-
tronic Commerce Research 10, 239270 (2010), 10.1007/s10660-010-9064-y
19. Yang, Y., Feng, Q., Sun, Y.L., Dai, Y.: Reptrap: a novel attack on feedback-based reputation
systems. In: Proceedings of SecureComm 2008, pp. 8:18:11 (2008)
 A Recommender System Model Combining Trust
with Topic Maps

Zukun Yu1, William Wei Song2, Xiaolin Zheng1, and Deren Chen1
1
Computer Science College, Zhejiang University, Hangzhou, China
2
School of Technology and Business Studies, Dalarna University, Borlnge, Sweden
{zukunyu,xlzheng,drc}@zju.edu.cn,
[email protected]

Abstract. Recommender Systems (RS) aim to suggest users with items that
they might like based on users opinion on items. In practice, information about
the users opinion on items is usually sparse compared to the vast information
about users and items. Therefore it is hard to analyze and justify users favo-
rites, particularly those of cold start users. In this paper, we propose a trust
model based on the user trust network, which is composed of the trust relation-
ships among users. We also introduce the widely used conceptual model Topic
Map, with which we try to classify items into topics for Recommender analysis.
We novelly combine trust relations among users with Topic Maps to resolve the
sparsity problem and cold start problem. The evaluation shows our model and
method can achieve a good recommendation effect.

Keywords: Recommender Systems, Trust Model, Reputation, Trust Propagation,

Topic Maps.

1 Introduction

With rapid development of Internet, more and more people use online systems to buy
products and services (hereafter called items). However, with overwhelming amount
of information about items available on the Internet, it is extremely difficult for users
to easily find and determine what they would like to buy. Recommender systems aim
to recommend the target users with the items which are considered to have high pos-
sibilities of meeting their preferences.
Given a huge number of users making commercial transactions online and an even
larger amount of items available for sale online, recommender systems have to face
two major challenges: data sparsity the average number of ratings given by users is
often very small compared to the huge number of items, and cold start the dumb
users who review few items and provide little information. It causes the problem that
a recommender system cannot decide what should be recommended to the target users
since it can only directly access to the users opinions on a small proportion of items.
Therefore, the data sparsity is now taken into account by many recommendation me-
thods [9]. The cold start problem is a challenge to recommender systems due to its
lack of sufficient information to justify their interests. The traditional solutions to the

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 208219, 2013.
Springer-Verlag Berlin Heidelberg 2013
 A Recommender System Model Combining Trust with Topic Maps 209

problems are to use a combination of content-based matching and collaborative filter-

ing [21]. Recently, researchers have considered using trust to deal with them [10, 14].
Trust is an assumed reliance on some person or thing, namely, a confident depen-
dence on the characteristics, ability, strength or truth of someone or something [12].
As pointed by Massa [15], recommender systems that make use of trust information
are the most effective in term of accuracy while preserving a good coverage. In a
trust-based recommender system, trust propagation is computed based on the trust
network to derive indirect trust relationships between users, such as in the case of
FOAF (friend of a friend) [7] the framework for representing information about
people and their social connections.
In general, the key concepts considered in modeling recommender systems include
users, items, and the characteristics of them. Trust-based methods take only users
social links (Trust) into account but ignore the relations among items, which are help-
ful in predicting users interest. In order to achieve a good recommendation method,
we will combine trust relations among users with similarity relations among items
derived by topic maps in an integrated model of recommender systems. This model
maintains three types of relationships. The first type is the trust relationships among
users. The second is the rating scores given to items by users, expressing how much
users like items. The third one is called relatedness of an item to a topic, representing
how much the item belongs to this topic, which leads to the computation of similarity
relations among items. We use Topic Maps to represent the third type of relationships.
Topic Maps is defined as an abstract model for semantically structured, self-
describing link networks laid over a pool of addressable information objects [13]. A
topic map includes three key components: topics, associations, and occurrences [22].
Using these elements, topic maps can be built in many domains.
The rest of this paper is organized as follows. We introduce the related work in re-
commender systems, trust and Topic Maps in Section 2. Then in Section 3 we pro-
pose our model based on trust and Topic Maps, and we novelly propose a method to
propagate trust to determine users interests and to apply Topic Maps to consider the
relationship between items. Then we introduce our datasets for experiments and ana-
lyze the experiment results in Section 4. We conclude this paper and point out our
future research in the last Section.

2 Related Work!

Many efforts have been put in studying and developing recommender systems, aiming
to support users doing business online. The major ones include content-based me-
thods, collaborative filtering (CF) methods, hybrid methods - a combination of con-
tent-based methods, collaborative filtering methods and others. Recently, researchers
are developing methods using trust to analyze users or items thereby to recommend
users with items they might like.
The content-based methods analyze the items that are rated by users and use the
contents of items to infer users profiles, which is used in recommending items of
interest to these users [2] More specifically, the TF-IDF (term frequencyinverse
210 Z. Yu et al.

document frequency) method, a computation method reflecting how important a word

is to a document in a collection, is used to compute similarities of contents [20]. In
this method, if two users collect more items with the similar content, they would be
more likely to prefer the same items. These methods have been extended to considera-
tion of using the attributes of items (for similarity computation) and the ratings given
to items by users (for users profile computation) in construction of recommendations
[23]. However, the content-based methods suffer from some shortcomings. Such kind
of methods usually needs to collect user profiles. This is a problem of privacy. What
is more, the content-based methods may cause overspecialized recommendations that
only include items very similar to those of which the user already knows [3].
The CF method uses a database about users preferences to predict more items us-
ers might like [4]. Papagelis et al. [19] propose a recommendation method based on
incremental collaborative filtering of users similarities. This method expresses the
new similarity values between two users in relation to the old similarity values, so as
to maintain an incremental update of their associated similarity. Zhang et al. [26]
propose a recommendation algorithm based on an integrated diffusion method making
use of both the user-item relations and the collaborative tagging information. They
use a so called user-item-tag tripartite graph as the base of the diffusion process to
generate recommendations. This method uses both the user-item relations and the
collaborative tagging information to improve the algorithmic performance. The short-
coming of CF methods is that they do not explicitly incorporate feature information
and face the sparsity problem and cold-start problem.
The authors of [18] propose a method using a weighted combination to fusion rat-
ings obtained by content-based methods and CF-based methods separately. The
weights are adjusted based on the strength of both the content-based method and the
CF-based method. As the number of users and ratings given by them increase, the CF-
based method is usually weighted more heavily, to take advantage of the wisdom of
crowd via globally computing all the ratings. Melville et al. [17] propose an approach
using a content-based predictor to enhance existing user data thereby to provide per-
sonalized suggestions through collaborative filtering. The content-based predictor
accepts the item with a high rating score and rejects the item with a low rating score.
However, these methods ignore the important social information which can reflect
users interest.
Trust among users and reputation of users are becoming important and elementary
issues in social networking study. As pointed out by Guha and Kumar et al. [8], a user
trust network is a fundamental buildings block in many of todays most successful e-
commerce and recommendation systems. The authors propose a framework of trust
propagation schemes, which appears to be the first to incorporate distrust in a compu-
tational trust propagation setting. This paper shows that a small amount of expressed
trust or distrust information can be used to predict trust between any two people in the
system with high accuracy. Ziegler and Lausen [27] introduce a classification scheme
for trust metrics. They present some model constituents for semantic web trust infra-
structure in the case FOAF (friend of a friend). However this paper has a limitation
that it assumes all trust information is publicly accessible, which, in practice, is nearly
impossible. Vydiswaran et al. [25] propose a trust propagation framework to compute
 A Recommender System Model Combining Trust with Topic Maps 211

how freetext claims of internet and their sources can be trusted by using an iterative
algorithm to compute the scores of trust propagation. But this work utilizes a weak
supervision at the evidence level, which makes it difficult to be used in other do-
mains. Apart from trust, reputation - a global aggregation of the local trust scores by
all the users [11] - is also important because it represents users trustworthiness from
a systemic perspective. Kamvar et al. [11] describe an algorithm to decrease the
number of downloads of inauthentic files in a peer-to-peer file-sharing network. This
paper assigns each peer a unique global trust value, based on the peers history of
uploads. Adler et al. [1] propose a content-driven reputation system for Wikipedia
authors, which can be used to flag new contributions from low-reputation authors, or
to allow only high-reputation authors to edit critical pages. Trust-based methods use
trust relationships among users to build a social network to link users and use it to
derive users interest. But these methods have a shortcoming that they fail to analyze
the relationships among items.
It is important for recommendation to take into account the relationships of items.
We will apply Topic Maps technology to model relationships between items. Topic
Maps related technologies are used in different research work. Dichev et al. [5] try to
use Topic Maps to organize and retrieve online information in the context of e-
learning courseware. They think Topic Maps offers a standard-based approach for
experts knowledge. This allows further reusing, sharing and interoperability of
knowledge and teaching units between courseware authors and developers. Dong and
Li [6] propose a new set of hyper-graph operations on XTM (XML Topic Map),
called HyO-XTM, to manage the distributed knowledge resources. In the HyO-XTM,
the set of vertices is the union of the vertices and the hyper-edges sets of the hyper-
graph; the set of edges is defined by the relation of incidence between vertices and
hyper-edges of the hyper-graph. The hyper-graph model matches the Topic Maps
with Hyper-graph vertices mapping to topic nodes and edges mapping to association
nodes. Topic Maps is shown to be a new way to graphically manage the knowledge.
Based on the previous work, we will first time try to use a topic map to represent
relationships among items of recommender systems, so as to be aware of the relation-
ships of items.

3 Topic Maps Based Trusted Recommender Model

We propose a graphical conceptual model for recommender systems, in which we

describe three types of nodes, i.e., user nodes, item nodes, and topic nodesIn this
model we also consider three types of relationships: trust (from one user to another),
rating (from a user to an item), and belonging (from an item to a topic), see Fig. 1.

3.1 Model Description

We use a user graph G (V , E ) to model the user trust network, where V is a set of user
nodes, representing users, and E a set of directed edges, representing trust relation-
ships, see Fig. 1 (left). A trust relationship e E , is an edge in G, from a user node vi
212 Z. Yu et al.

to another user node vj. Each e in E is associated to a value in [-1, 1], indicating the
weight of e. A negative trust value between two users means that they distrust each
other and a positive trust value means that they trust each other. We use e(vi, vj) to
denote the trust value. We define a truster function of a user node v, yielding the set
of all the users who have a trust relationship (an edge) to v, as
truster(v) = {u V | (u, v) E} . We also define a trustee function of v, representing
the set of all the users who have a trust relationship from v, as
trustee(v) = {u V | (v, u) E} . For each user node v, we define a reputation func-
tion, denoted as (v), taking values from [0, 1].

(v2 )
(v2 , i1)

e(v1 , v2 ) e(v2 , v4 ) (t1, t2 )

(v1 ) (v4 )
(v2 , i4 )
(t1, t3 )
(v4 , i5 )
e(v1, v3 ) e(v3 , v4 )
(v3 ) (v4 , i6 )

Fig. 1. Example of Topic Maps based trust recommender model Users Trust Network (left)
and Topic Maps (right)

Now we consider this type of relationships: rating R, from the user node set V to
the item node set I. Each element (v, i ) R is an edge from a user who rates the item
to the item with a rating function (v, i ) . We define the set of items rated by the user v
by the function item(v) = {i I | (v, i) R} . Similarly, we use the function
reviewer(i) ={vV |(v,i)R} to represent the set of users who rate the item i.
According to the concept of Topic Maps, an item is an occurrence which belongs to
a topic. We define the type of relationship, belonging B, from the item set I to the topic
set T. Each element (i, t ) B is a direct edge from an item i to a topic t. An item
might belong to a number of topics. So we use the function topic(i) = {t T | (i, t ) B}
to define the set of all the topics the item i belongs to. We use the function
occurrence(t ) = {i I | (i, t ) B} to define the set of all the items belonging to the
topic t. To model the association between topics in a topic map, we use a function
(tm , tn ) to represent the association degree between two topics tm and tn. Its value is a
real number in [0, 1]. The higher the value is, the closer the two topics are.
 A Recommender System Model Combining Trust with Topic Maps 213

3.2 Topic Maps Based Trusted Recommendation Methods

Using Topic Maps, we conceptually describe users, items, and topics, as well as vari-
ous relationships between them. In this section we will introduce a computation me-
thod to quantitatively calculate these functions defined on the nodes and relation-
ships.

(v2 )

e(v1 , v2 ) e(v2 , v4 )
v2
(v1 ) (v4 )
e (1) (v1 , v4 )

v1 v4
(v3 )
e(v1 , v3 ) e(v3 , v4 )

v3
Fig. 2. Example of trust propagation

3.3 Trust Model

In the trust model, the trust relationships among users are the basic ones, from which
we derive users reputations and propagate new trust relationships among users. The
reputation of a user v is computed by averaging all the trust values of the trust rela-
tionships to v from other users, i.e.

(v ) = e(u , v ) | tr u s te r ( v ) |
(1)
u tru ste r ( v )

If a user has no trust relationship from a truster, its reputation is set to 0. However,
for any two users u, v if they do not have a direct edge (i.e. trust relationship) between
them but have indirect edges (i.e. via one more users), we use trust propagation to
determine the trust value between u, v. In other words, we aim to use existing values
in the trust network to gain more trust values between the users without direct trust
relationships. For simplicity, we only consider to propagate trust relationships but not
distrust relationships in this paper. We use a step-by-step trust transitivity to derive
indirect trust, in which a single step means to derive trust by the intermediate user
nodes which have direct edge to both the start user node and the target user node in
the user trust network, see Fig. 2. We use solid lines to represent existing trust rela-
tionships among users and dashed lines to represent the derived trust relationship. The
trust relationships between v1 to v4 can be derived in the first single transitivity of trust
propagation and that between v1 and v5 can be derived in the second single transitivi-
ty. In order to clearly explain the propagation process of trust, we use e ( n ) (vi , v j ) to
denote a newly derived trust value in the nth single step of trust transitivity. The origi-
nal trust value e(vi , v j ) is denoted as e (0) (vi , v j ) , which means the existed trust
edges in the original trust network.
214 Z. Yu et al.

For example, to derive a new trust value from the user node v1 to the user node v4,
there might be a number of paths, where a path means a chain of edges from v1 to v4
via an intermediate user node, e.g. v2. Here we define a function for the set of the
common user nodes as com(vi, vj) = trustee(vi)truster(vj). We consider all the paths
to derive the trust value:
(e({0,1,2,...n}) (vi , vk ) (vk ) + e({0,1,2,...n}) (vk , vj ) (vj ))
vk trustee(vi ) truster(vj )
e(n+1) (vi , vj ) = (2)
vk trustee(vi ) truster(vj ) (| (vk )| + | (vj )|)

Here, e({0,1,2,...n}) (vi , vk ) is a trust relationship in the set of all the trust relationships
including the original trust relationships and those generated in the 1st, 2nd,, and nth
step of the single step transitivity. The formula above gives a new trust value in [-1, 1]
by applying once the single step transitivity. For a trust network, we can propagate
trust relationships using a number of steps of the single step transitivity. We use a
parameter s to control the number of steps for trust propagation.

3.4 Deriving Users Opinion Based on the User Trust Network

In this section, we discuss how to generate a rating value between a user and an item
if there was none there. We use the trust values from the user trust network and the
rating values from the existing ratings given by the users to the items to derive users
new opinions on items together with new rating values. We denote the newly generat-
ed rating relationships through the user trust network as R. We indirectly compute
users opinion on items through rating on the items from the intermediate users if they
rated them directly. These intermediate users should have positive reputation values.
The users with negative reputation values are thought as malicious users and not al-
lowed to give advices to others. The newly generated rating relationships with rating
values are computed by the function given below.

(v,i) = e({0,1,2,...n}) (v, ) (, i) (u) (u) (3)

trustee(v) reviewer(i) trustee(v) reviewer(i)

3.5 Deriving Users Opinion Based Topic Maps

In section 3.4 we generated a new set of rating relationships using trust propaga-
tion on the user trust network in the section 3.4. However, if the user trust network
contains a lot of isolated clusters, the trust values and rating values from one cluster
would not be possible to be used for other clusters. We call this the isolated trust clus-
ter problem and will further explain it in section 4.1.
Here we consider using the Topic Maps to solve this problem. We use the topic map
to derive users opinion on topics. To do so, we first define a function g(v, t) to be an
 A Recommender System Model Combining Trust with Topic Maps 215

user vs opinion on a topic t, deriving from users opinion on the items and items
belonging to the topic t, as follows:

g (v , t ) = (v , i ) | item (v ) occurrences (t ) | (4)

iitem(v ) occurrences (t )
Let us consider how to construct a topic map for a recommender system. We adopt
association rule for building the association relationships between topics, because the
association rule is frequently used to investigate sales transactions in market basket
analysis and navigation paths within websites [24]. We define a function (tm , tn )
to measure associations between the topics t m and tn , in terms of the users who
have rated the topics, as follows:
(tm, tn ) = | user(tm) user(tn)| | user(tm)| (5)
In formula (5), the function user (t ) represents the set of users who rated the topic t.
We set a threshold, denoted as confidence , to be used in filtering out the associa-
tions with value less than the threshold. The associations after filtering can be used to
derive users opinion on a topic which they did not rate. The computation formula is
as follows:

g (v, tm ) (tm , t )
tm{tT |g (v,t )0, (tm ,t )0}
g (v, t ) = (6)
| {t T | g (v, t ) 0, (tm , t ) 0} |
Finally, we consider how to decide a users opinion on an item through both trust and
topic map. We use a combination of users opinion derived based on the topic map
and users opinion on items, to compute users opinion (v, i ) based on the user
trust network.

(v, i) = (1 ) (v, i) + ( ( g(v, t )) | topic(i) | ) (7)

ttopic(i)

Here, is a weight parameter in [0, 1] to help controlling the weight of implicit

opinion by themselves on an item, i.e., the weight of users own opinion.

4 Experiments

In this section, first we describe the dataset used in our experiments, and then we
discuss the experiments and their results.

4.1 Dataset

We use a data collection of the real review data from Epinions.com, provided at Mas-
sas website [16], as the input dataset. We use two datasets, the trust relationships
216 Z. Yu et al.

(and their values) between users and the users ratings on items (and their values). In
order to reduce the time and space complexities of algorithm, we use two subsets
obtained by the method described below:
Extract all the users in the first dataset and select 100 different users.
Extract all the trust relationships among the 100 users. We obtained a subset, called
subset 1. It contains 110 trust relationships, 82 of which have trust value 1 and 28
have trust value -1.
Extract all the ratings given by the 100 users in the second dataset. We first obtain
282418 ratings on 230126 items from the 64 users who rated the items. Through
sampling the ratings, we obtain the subset 2. It contains 10141 ratings, with 8858
items and 34 users. The subset 2 has 6 different levels of rating values and 27 top-
ics. As shown in Figure 3, most items are rated 5 and 4. Only a small proportion of
them are rated 1, 2, 3 and 6.

From the sample datasets we constructed for the experiments, we clearly observe the
problems of data sparsity and cold start users, which we discussed in Section 1.
There are 10141 ratings for 100 users and 8858 items in this recommender system. So
the subset of this recommender system is sparse because the ratio of the number of
ratings to the total number of the matrix (number of users times number of items) is
10141/(100*8858)= 1.14%. Only one out of 100 users rated some items, so there are
more than 98 cold start users. The maximum ratings for one user is 7507, on average
every user 101 ratings.
The trust relationships in the subset 1 form a user trust network, as shown in Fig. 4.
We use the trust propagation method discussed in section 3 to obtain more trust rela-
tionships. But this user trust network consists of many isolated trust clusters. They
do not connect to each other and contribute little to the trust propagation computation.

Fig. 3. The distribution of rating values Fig. 4. The structure of user trust network

4.2 Results
We use MAE (Mean Absolute Error) and MAUE (Mean Absolute User Error) [15] to
evaluate our recommendation method, because the MAE is the most commonly used
and the easiest to be understood, and based on MAE, the MAUE provides the aver-
ages of evaluation.
 A Recomm
mender System Model Combining Trust with Topic Maps 217

Based on the observation that the most derived trust values have been obtainedd in
the first three steps of trustt propagation on the user trust network, we consider too set
s=3. We evaluate the variaation of accuracy as the weight parameter changes witth a
confidence =0.95 in consttructing a topic map. As show in Fig. 5, both the MAE and
MAUE decline nearly in a straight
s line with the weight parameter. We can see that the
MAUE is always less than the MAE due to the many cold start users in our dataaset.
So we find that the two meetrics keep consistent in the relationship between accurracy
and weight parameter.

Fig. 5. The variation of accurracy with weight Fig. 6. The coverage changing as confideence
parameter of Topic Map

Furthermore, we considerr using coverage to measure the performance of our m me-

ght parameter to be 0.1 and 0.2 respectively. We select
thod. First we set the weig
the top 10 most-suited item
ms in the recommendation lists of items. In Fig. 6, the lline
named 0.1 represents the coverage
c with the weight parameter to be 0.1 and the
t coverage with the weight parameter to be 0.2. As
line named 0.2 represents the
the confidence decreasees from 1 to 0, the coverage rises. The coverage in the ttwo
lines rises sharply when thee confidence is 0.7 and then it does not vary considerabbly.
The reason for the changiing of coverage is that when the confidence becom mes
smaller, the topic map will give more association relationships between topics to ssup-
port deriving users opinionns on items. The result tells that the Topic Maps does hhelp
improve the coverage of reccommender systems.

5 Conclusion and Future Work

We have three contribution ns in this work. First we propose a trust model based on the
user trust network and a meethod to compute the trust propagation. We use the repuuta-
tion as the weight in trustt propagation, which mimics considering friends adviices
before people buying produ ucts or services in the real world. That is, advices from the
users with the greater reputation values are trusted by their friends with higher degrree.
Second we propose a reco ommendation method, applying Topic Maps in analyzzing
users opinions on topics, which
w better and further supports deriving users opiniions
on items. Third we evaluatee our method using the two metrics accuracy and coveraage.
218 Z. Yu et al.

The result tells that the recommender system based on our method provides better
accuracy and coverage in coping with the problems of the sparsity and cold start us-
ers.
For the next step of study, we plan to scale up our method for recommender sys-
tems to a reasonably large number of users in the user trust network as in the reality a
recommender system should be able to deal with millions of users and hundred mil-
lions of items. We also consider to include the computation of distrust relationships in
the trust propagation method as we believe it will greatly contribute to the accuracy
and coverage of recommender systems.
Acknowledgments. The authors would like to thank Dalarna University, Sweden for
providing a visiting position and Zhejiang University, China for providing financial
support of this visit.

References
1. Adler, B.T., Alfaro, L.D.: A Content-Driven Reputation System for the Wikipedia. Tech-
nical Report ucsc-crl-06-18. School of Engineering, University of California, Santa Cruz
(2006)
2. Basu, C., Hirsh, H., Cohen, W.: Recommendation as Classification: Using Social and Con-
tent-Based Information in Recommendation, pp. 714720. AAAI Press (1998)
3. Blanco-Fernndez, Y., Pazos-Arias, J.J., Gil-Solla, A., Ramos-Cabrer, M., Lpez-Nores,
M., Garca-Duque, J., Fernndez-Vilas, A., Daz-Redondo, R.P., Bermejo-Muoz, J.: A
flexible semantic inference methodology to reason about user preferences in knowledge-
based recommender systems. Knowl-Based Syst. 21, 305320 (2008)
4. Breese, J.S., Heckerman, D., Kadie, C.: Empirical analysis of predictive algorithms for
collaborative filtering. In: Proceedings of the Fourteenth Conference on Uncertainty in Ar-
tificial Intelligence, pp. 4352. Morgan Kaufmann Publishers Inc., Madison (1998)
5. Christo, D., Darina, D., Lora, A.: Using topic maps for e-learning. In: Proceedings of In-
ternational Conference on Computers and Advanced Technology in Education including
the IASTED International Symposium on Web-Based Education, Rhodes, Greece, pp. 26
31 (2003)
6. Dong, Y., Li, M.: HyO-XTM: a set of hyper-graph operations on XML Topic Map toward
knowledge management. Future Gener. Comp. Sy. 20, 81100 (2004)
7. Golbeck, J., Rothstein, M.: Linking social networks on the web with FOAF: a semantic
web case study. In: AAAI 2008, pp. 11381143. AAAI Press (2008)
8. Guha, R., Kumar, R., Raghavan, P., Tomkins, A.: Propagation of trust and distrust. In:
Proceedings of the 13th International Conference on World Wide Web, pp. 403412.
ACM, New York (2004)
9. Huang, Z., Chen, H., Zeng, D.: Applying associative retrieval techniques to alleviate the
sparsity problem in collaborative filtering. ACM Trans. Inf. Syst. 22, 116142 (2004)
10. Jamali, M., Ester, M.: TrustWalker: a random walk model for combining trust-based and
item-based recommendation. In: KDD 2009, pp. 397406. ACM, New York (2009)
11. Kamvar, S.D., Schlosser, M.T., Garcia-Molina, H.: The eigentrust algorithm for reputation
management in P2P networks. In: Proceedings of the 12th International Conference on
World Wide Web, pp. 640651. ACM, Budapest (2003)
 A Recommender System Model Combining Trust with Topic Maps 219

12. Kini, A., Choobineh, J.: Trust in electronic commerce: definition and theoretical considera-
tions. In: Proceedings of the Thirty-First Hawaii International Conference on System Sci-
ences, vol. 4, pp. 5161 (1998)
13. Knud, S.: Topic Maps-An Enabling Technology for Knowledge Management. In: Interna-
tional Workshop on Database and Expert Systems Applications, p. 472 (2001)
14. Massa, P., Bhattacharjee, B.: Using Trust in Recommender Systems: An Experimental
Analysis, pp. 221235 (2004)
15. Massa, P., Avesani, P.: Trust-aware recommender systems. In: RecSys 2007, pp. 1724.
ACM, New York (2007)
16. Massa, P.:
http://www.trustlet.org/wiki/Extended_Epinions_dataset.
17. Melville, P., Mooney, R.J., Nagarajan, R.: Content-boosted collaborative filtering for im-
proved recommendations. In: Eighteenth National Conference on Artificial Intelligence
(AAAI 2002)/Fourteenth Innovative Applications of Artificial Intelligence Conference, pp.
187192 (2002)
18. Miranda, T., Claypool, M., Gokhale, A., Murnikov, P., Netes, D., Sartin, M.: Combining
Content-Based and Collaborative Filters in an Online Newspaper (1999)
19. Papagelis, M., Rousidis, I., Plexousakis, D., Theoharopoulos, E.: Incremental Collabora-
tive Filtering for Highly-Scalable Recommendation Algorithms. In: Hacid, M.-S., Murray,
N.V., Ra , Z.W., Tsumoto, S. (eds.) ISMIS 2005. LNCS (LNAI), vol. 3488, pp. 553561.
Springer, Heidelberg (2005)
20. Salton, G., McGill, M.J.: Introduction to Modern Information Retrieval. McGraw-Hill,
Inc., New York (1986)
21. Schein, A.I., Popescul, A., Popescul, R., Ungar, L.H., Pennock, D.M.: Methods and Met-
rics for Cold-Start Recommendations, pp. 253260. ACM Press (2002)
22. G. M. O. The Topic Maps: XML Topic Maps (XTM) 1.0. (2001)
23. Tso, K., Schmidt-Thieme, L.: Empirical Analysis of Attribute-Aware Recommendation
Algorithms with Variable Synthetic Data. Data Science and Classification, 271278
(2006)
24. Vercellis, C.: Business Intelligence: Data Mining and Optimization for Decision Making,
p. 277. John Wiley & Sons, Ltd., Chichester (2009)
25. Vydiswaran, V.G.V., Zhai, C., Roth, D.: Content-driven trust propagation framework. In:
Proceedings of the 17th ACM SIGKDD International Conference on Knowledge Discov-
ery and Data Mining, pp. 974982. ACM, San Diego (2011)
26. Zhang, Z.K., Zhou, T., Zhang, Y.C.: Personalized recommendation via integrated diffusion
on user-item-tag tripartite graphs. Physica A 389, 179186 (2010)
27. Ziegler, C., Lausen, G.: Spreading activation models for trust propagation. In: Proceedings
of 2004 IEEE International Conference on e-Technology, e-Commerce and e-Service, pp.
8397 (2004)
 A Novel Approach
to Large-Scale Services Composition

Hongbing Wang and Xiaojun Wang

School of Computer Science and Engineering, Southeast University, Nanjing, China

{hbw,seuwxj}@seu.edu.cn

Abstract. We investigate a multi-agent reinforcement learning model

for the optimization of Web service composition in this paper. Based
on the model, a multi-agent Q-learning algorithm was proposed, where
agents in a team would benefit from one another. In contrast to single-
agent reinforcement-learning, our algorithm can speed up the conver-
gence to optimal policy. In addition, it allows composite service to
dynamically adjust itself to fit a varying environment, where the proper-
ties of the component services continue changing. A set of experiments
is given to prove the eciency of the analysis. The advantages and the
limitations of the proposed approach are also discussed.

Keywords: Web Service composition, multi-agent.

1 Introduction

As a common understanding, Web services are self-describing and open building

blocks for rapid, low-cost composition of distributed applications [8]. In prac-
tice, a single Web service may not be sucient at performing complex tasks. We
usually need to combine multiple existing services together to meet customers
complex requests. With todays SOC technology, such composition is usually per-
formed by human engineers. However, the Web environment is highly dynamic,
and most Web services are evolving at all time. A service engineer cannot al-
ways foresee all the changes that could happen in the future. A manual service
composition can also be too rigid to adapt to a dynamic environment. There-
fore, dynamic service composition is regarded as a crucial functionality for the
Web of the future. Dierent technologies of computational intelligence have been
investigated for solving the problem of dynamic service composition.
AI planning is a typical type of techniques used to automate Web services
composition [9] [2]. Doshi [3] and Gao [4] have studied the application of MDPs
(Markov Decision Processes) in Web service composition. MDPs assume a fully
observable world and require explicit reward functions and state transition func-
tions. Such requirements are too strict to a real world scenario. Wang et al. [11]
proposed to use reinforcement learning (RL) for service composition, so as to

This work is partially supported by NSFC (61232007) and JSNSF of China

(No.BK2010417).

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 220227, 2013.
c Springer-Verlag Berlin Heidelberg 2013

 A Novel Approach to Large-Scale Services Composition 221

avoid complex modeling of the real world. Although it has been proved to be
eective for small scale service compositions, it can be overly computationally
expensive when working on a large number of services.
In this paper, we present a novel mechanism based on multi-agent reinforce-
ment learning to enable adaptive service composition. The model proposed in
this paper extends the reinforcement learning model that we have previously
introduced in [11]. In order to reduce the time of convergence, we introduce
a sharing strategy to share the policies among the agents, through which one
agent can use the policies explored by the others. As the learning process con-
tinues throughout the life-cycle of a service composition, the composition can
automatically adapt to the change of the environment and the evolvement of its
component services. Experimental evaluation on large scale service compositions
has demonstrated that the proposed model can provide good results.

2 A MAMDP Model for Service Composition

In this section we present the reinforcement learning model that we have previ-
ously introduced in [11] for solving the dynamic web service composition prob-
lem. The RL model introduced in [11] will be extended in this section towards
a distributed architecture.
Reinforcement learning is the problem faced by an agent that must learn
behavior through trial-and-error interactions with a dynamic environment [7].
One key aspect of reinforcement learning is a trade-o between exploitation and
exploration. To accumulate a lot of reward, the learning system must prefer the
best experienced actions, however, it has to try new actions in order to discover
better action selections for the future. One such method is greedy, when
the agent chooses the action that it believes has the best long-term eect with
probability 1 , and it chooses an action uniformly at random, otherwise.
We formally define the key concepts used in the model.
Definition 1. (Web Service). A Web service is modeled as a triple W S =<
P r; E; QoS >,where
P r represents the precondition of W S, which specifies the states of the world
in which W S can be executed.
E represents the eect of W S, which describes how W S changes the state
of the world.
QoS is a n-tuple < att1 ; att2 ; :::; attn >, where each atti denotes a QoS
attribute of W S.
As mentioned earlier, we use Multi-agent Markov Decision Process (MAMDP)
to model service composition. A MAMDP involves multiple actions and paths
for each agent to choose. For agent m, we call our service composition model
WSC-MDP, which simply replaces the actions in a MDP with Web services.
Definition 2. (Web service composition MDP (WSC-MDP). A Web
service composition MDP is a 7-tuple WSC-MDP =< m, sm m m m m
0 , S , A (.), P , R ,
m
sr >, where
222 H. Wang and X. Wang

Query Find
Flight Find
Book Book Book Restaurant Map
s4 Hotel Find s1
Flight1 s4 Hotel
Query s6
s0(m) s2
Query s6 s7
Book Restaurant s12 sr(m)
Flight1 Weather
Book Hotel s7 Find Forecast Hotel
Book Map
Flight2
Query Flight1 Book s12
Flight s2 Book Workflow1
Flight2
Flight1 sr(m)
Query s8 Query
Hotel Query Book
s0(m) Flight Book Find Flight Flight2
Find Book Hotel s5 Map
Weather s1 s3 Map s11 Flight2 s0(m) s1 s3 s9 s10
Forecast Weather Find s11 sr(m)
Query Book Forecast Restaurant
Hotel Hotel s5
Find s9 s10 Query
Flight
Restaurant Workflow2
Find
Map

Fig. 1. The WSC-MDP of a Com- Fig. 2. Two Workflows Contained

posite Service for Travel Plan by the WSC-MDP in Fig. 1

m: denote the agent m.

S m : a finite set of state spaces of agent m.
sm0 Sm : is the initial state of agent m and also an execution of the service
compositions starts from this state.
smr Sm : is the set of terminal states for agent m. Upon arriving at one of
the states, an execution of the service composition terminates.
Am (.): represents the set of Web services that can be executed in state s S m
. A service ws belongs to Am , only if the precondition ws P is satisfied by s.
P m : [pm m m
iaj ] S A S
m
[0, 1]: The transitions function for the agent.
It expresses the probability that the agent m goes to state j if it executes web
service a in state i is pmiaj .
Rm : [riajm
] S m R: defines the rewards that agent m receives if it is in
state i and goes to state j with the execution of web service a.

A WSC- MDP can be visualized as a transition graph. As illustrated by Fig. 1,

the graph contains two kinds of nodes, i.e. state nodes and service nodes, which
are represented by open circles and solid circles respectively. s0(m) represents
the initial state node. The terminal states nodes are sr(m). A state node can
be followed by a number of service nodes, representing the possible services that
can be invoked in the state. There is at least one arrow pointing from a service
node to the next state node. Each arrow is labeled with a transition probability
pm m
iaj , and the expected reward for that transition riaj . (For simplicity, we omit
the labels in Fig. 1.) The transition probabilities on the arrows rooted at a single
action node always sum to one.

Definition 3. Service Workflow. Let wf be a subgraph of a WSC-MDP. wf is

a service workflow if and only if there is at most one service that can be invoked
at each state wf . In other words, a service workflow is actually equivalent to a
deterministic state machine. A tradition service composition usually builds on a
single such workflow.

Example 1. Fig. 2 shows two of multiple service workflows. Which workflow to

be executed is determined by the policy of the Markov decision process.

Definition 4. (Policy): A policy is a mapping from state s S to a service

ws A, which tells which service ws = (s) to execute when in state s.
 A Novel Approach to Large-Scale Services Composition 223

Each policy of a WSC-MDP can determine a single workflow. By executing a

workflow, the service customer is supposed to receive a certain amount of reward,
which is equivalent to the cumulative reward of all the executed services. Given a
WSC-MDP, the task of our service composition system is to identify the optimal
policy or workflow that oers the best cumulative reward. As the environment of
a service composition keeps changing, the transition function P and the reward
function R of a WSC-MDP change too. As a result, the optimal policy changes
with time. If our system is able to identify the optimal policy at any given time,
the service composition will be highly adaptive to the environment.

3 Algorithm for Service Composition

The proposed Distributed Approach

The Q-learning is the most popular and seems to be the eective model-free
algorithm about the RL problems. It does not, however, address any of the issues
involved in generalizing over large state and/ or action spaces [7]. That is why, in
order to speed up the training process, we extend the proposed approach towards
a distributed one, in which multiple cooperative agents learn to coordinate in
order to find the optimal policy in their environment.
Experience sharing can help agents with similar tasks to learn faster and
better. For instance, agents can exchange information using communication [10].
Furthermore, by design, most multiagent systems also allow the easy insertion
of new agents into the system, leading to a high degree of scalability [1].
From agent-ms standpoint, its control task could be thought of as an ordinary
reinforcement problem except that their action selecting strategy is dependent
on other agents optimal policies at the beginning of the learning. So at a cer-
tain state, one agents policy may be useful to other agents which can help them
to find optimal strategy quickly. But assume that each agent can simultane-
ously send its current policy to other agents, the communication information
is huge. For the purpose of reducing the communication information we dont
let the agent to communicate with each others, but introduce supervisor agent
which supervises the learning process and synchronizes the computations of the
individual agents. In our algorithm, each agent m use the global Q-values esti-
mations stored in the blackboard which stores the global Q-values estimations
and communicate to the supervisor agent their intention to update a Q-value
estimation.
So we have two types of agents in our architecture:

WSCA (Web Service Composition Agents). Each WSCA agent runs in a

separate process or thread and is trained using the Q-learning algorithm.
Each local agent performs local Q-values estimations updating from its own
point of view.
a WSCS(Web Service Composition Supervisor) agent which supervises the
learning process and synchronizes the computations of the individual WSCA
224 H. Wang and X. Wang

agents. It keeps a blackboard [5] which stores the global Q-values estima-
tions. The local WSCA agents use the global Q-values estimations stored
in the blackboard and communicate to the WSCS agent their intention to
update Q-value estimation. If a local agent tries to update a certain Q-value,
the WSCS agent will update the global Q-value estimation only if the new
estimation received from the local agent is greater than the Q-values esti-
mations existing in the blackboard.

In this study, each reinforcement-learning agent uses the one-step Q-learning

algorithm. Its learning decision policy is determined by the central Q-table and
the state/action value function which estimates long-term discounted rewards for
each state/action pair. In order to get more knowledge for the agent and jump
out of the sub-superior strategy trap, searching strategy was introduced into the
Q-learning. The agent is allowed to take the action which isnt the superior at the
current view, so greedy strategy was proposed. Thus the agent can explore
the state-action space by choose the viable action randomly at some degree, and
avoid arriving at the local superior solution via only choosing the action with
the maximal Q-value.
For agent m, given a current state s and available actions A(m), a Q-learning
agent selects action ai with a probability given by the rule below:
4
m (1 ) if ai = argM axa Q[s, a]
p (ai |s) =
others

In each time step, the agent m updates Q(s, a) by recursively discounting future
utilities and weighting them by a positive learning rate :

Q(s, a) = (1 ) Q(s, a) + [r(s, a) + maxQ(s, a)]. (1)

Here (0 < 1) is a discount parameter.
If Q(s, a) is the biggest among the actions in state s, than the WSCS will
update Q(s, a) in the blackboard.
The training process consists of three phases and will be briefly described in
the following.

Phase 1. Initial phase

The WSCS supervisor agent initializes the Q-values from the blackboard.

Phase 2. Training phase of each WSCA agent

During some training episodes, the individual WSCA agents will experiment
some paths from the initial to a final state, using the greedy mechanism and
updating the Q-values estimations according to the algorithm described below
(algorithm 1). We denote in following by Q(s, a) the Q-value estimate associated
to the state s and a, as stored by the blackboard of the WSCS agent.
 A Novel Approach to Large-Scale Services Composition 225

Algorithm 1. Multiagent Q-learning Algorithm for Agent m

Require:
The WSC-MDP for agent m;
The WSCS agent;
repeat
evaluate the starting state s
select action a from s using policy derived from Q ( Greddy)
repeat
Learning:(for each step of episode)
Take action a, observe the reward r(s, a) and the next state s.
WSCA agent asks WSCS agent for Q(s, a).
WSCS retrieves Q(s, a) from the blackboard.
WSCS sends the retrieved Q(s, a) to WSCA.
WSCA agent updates the table entry Q(s, a) ad follows

Q(s, a) = (1 ) Q(s, a) + [r(s, a) + maxQ(s, a)]. (2)

WSCA sends the new Q(s, a) to WSCS.

WSCS updates the Q(s, a) if it is greater than the old.
WSCA agent go to state s
until s is the terminal state
until the Q-values tiny changes

Phase 3. Executing phase

After the training of the multi-agent system has been completed, the solution
learned by the WSCS supervisor agent is constructed by starting from the initial
state and following the Greedy mechanism until a solution is reached. The system
applies the solution as a service workflow to execute. At the same time, the
execution is also treated as an episode of the learning process. The Q-functions
are updated afterwards, based on the newly observed reward.
By combining execution and learning, our framework achieves self-adaptively
automatically. As the environment changes, service composition will change its
policy accordingly, based on its new observation of reward. It does not require
prior knowledge about the QoS attributes of the component services, but is able
to achieve the optimal execution policy through learning.

4 Experimental Evaluation
In order to evaluate the methods, we conducted simulation to evaluate the prop-
erties of our service composition mechanism based on the methods discussed in
this paper. The PC configuration: Intel Xeon E7320 2.13GHZ with 8GB RAM,
Windows 2003, jdk1.6.0.
We considered two QoS attributes of services. They ware service fee and ex-
ecution time. We assigned each service node in a simulated WSC-MDP graph
with random QoS values. The value followed normal distribution. To simulate
226 H. Wang and X. Wang

the dynamic environment, we periodically varied the QoS values of existing ser-
vices based on a certain frequency. We applied the algorithms introduced in
Section 3 to execute the simulated service compositions. The reward function
used by the each learner was solely based on the two QoS attributes. After an
execution of a service , the learners get a reward , whose value is:
f eemax f eesi timemax timesi
R(s) = f eemax
i
f eemin
+ timemax
i
timemin
i i i i

The reward was always positive, however service consumers always prefer low
execution time and service fee.
We will show that such cooperative agents can speed up learning, measured
by the average cumulative values in training, even though they will eventually
reach the same asymptotic performance as independent agents.

50 50 50

45 45 45

40 40 40
Mean cumulative reward

Mean cumulative reward

Mean cumulative reward

35 35 35

30 30 30

25 25 25

20 20 20

15 Q-learning 15 Q-learning 15 Q-learning

20 agents 20 agents 20 agents
10 10 10
100 400 700 1000 1300 1600 1900 100 400 700 1000 1300 1600 1900 100 400 700 1000 1300 1600 1900
Episodes Episodes Episodes

(a) (b) (c)

50 50 50

45 45 45

40 40 40
Mean cumulative reward

Mean cumulative reward

Mean cumulative reward

35 35 35

30 30 30

25 25 25

20 20 20

15 Q-learning 15 Q-learning 15 Q-learning

20 agents 20 agents 20 agents
10 10 10
200 600 1000 1400 1800 2200 2600 3000 500 1000 1500 2000 2500 3000 3500 4000 500 1000 1500 2000 2500 3000 3500 4000
Episodes Episodes Episodes

(d) (e) (f )
Fig. 3. (a) Results of comparison with 20 services in each state; (b) Results of com-
parison with 30 services in each state; (c) Results of comparison with 40 services in
each state; (d) Results of comparison with 100 services in each state; (e) Results of
comparison with 150 services in each state; (f) Results of comparison with 200 services
in each state;

Scalability with respect to the number of services

In this stage of our evaluation, we studied the eect of distributed RL approach
with varied number of services in each state. We fix the state number on 500
and vary the services from 20 to 200. As the Fig. 3 shows, the distributed RL
learns more quickly than Q-learning, and when the number of services increases,
the reducing of convergence time is more considerable, because they may have
explored the dierent parts of a state space and share their knowledge. If agents
perform the similar task, two agents can complement each other by exchanging
their policies or use what the other agent had already learned for its own benefit.
Assume that each agent can simultaneously send its current policy at some state
to blackboard Q-table by WSCS agent, if some agent finds a better choice, it
may update blackboard Q-table through WSCS agent, then other agents can
adopt that policy with certain probability in that state.
 A Novel Approach to Large-Scale Services Composition 227

The results in all cases clearly indicate that distributed RL approach presented
in this paper learns more quickly and reduces the overall computational time
compared againest the Q-learning.

5 Conclusion
This paper studied a novel framework for large scale service composition. In or-
der to reduce the time of convergence we introduce a sharing strategy to share
the policies among agents in a team. The experimental results show that the
strategy of sharing state-action space improves the learning eciency signifi-
cantly. Additionally, the problem that has to be further investigated is how to
reduce the communication cost between the WSCA agents and WSCS agent and
explore other local search mechanisms. Next, we will concentrate on these issues
and improve our algorithm further.

References
1. Busoniu, L., Babuska, R., De Schutter, B.: A comprehensive survey of multiagent
reinforcement learning. IEEE Transactions on Systems, Man, and Cybernetics,
Part C: Applications and Reviews 38(2), 156172 (2008)
2. Carman, M., Serafini, L., Traverso, P.: Web service composition as planning. In:
ICAPS 2003 Workshop on Planning for Web Services, pp. 16361642 (2003)
3. Doshi, P., Goodwin, R., Akkiraju, R., Verma, K.: Dynamic workflow composition
using markov decision processes. In: IEEE International Conference on Web Ser-
vices, pp. 576582. IEEE (2004)
4. Gao, A., Yang, D., Tang, S., Zhang, M.: Web service composition using markov de-
cision processes. In: Fan, W., Wu, Z., Yang, J. (eds.) WAIM 2005. LNCS, vol. 3739,
pp. 308319. Springer, Heidelberg (2005)
5. Gonzaga, T., Bentes, C., Farias, R., de Castro, M., Garcia, A.: Using distributed-
shared memory mechanisms for agents communication in a distributed system. In:
Seventh International Conference on Intelligent Systems Design and Applications,
ISDA 2007, pp. 3946. IEEE (2007)
6. Hwang, S.Y., Lim, E.P., Lee, C.H., Chen, C.H.: Dynamic web service selection for
reliable web service composition. IEEE Transactions on Services Computing 1(2),
104116 (2008)
7. Kaelbling, L., Littman, M., Moore, A.: Reinforcement learning: A survey. Arxiv
preprint cs/9605103 (1996)
8. Papazoglou, M., Georgakopoulos, D.: Service-oriented computing. Communications
of the ACM 46(10), 2528 (2003)
9. Sirin, E., Parsia, B., Wu, D., Hendler, J., Nau, D.: Htn planning for web service
composition using shop2. Web Semantics: Science, Services and Agents on the
World Wide Web 1(4), 377396 (2004)
10. Sutton, R., Barto, A.: Reinforcement learning. Journal of Cognitive Neuro-
science 11(1), 126134 (1999)
11. Wang, H., Zhou, X., Zhou, X., Liu, W., Li, W., Bouguettaya, A.: Adaptive ser-
vice composition based on reinforcement learning. In: Maglio, P.P., Weske, M.,
Yang, J., Fantinato, M. (eds.) ICSOC 2010. LNCS, vol. 6470, pp. 92107. Springer,
Heidelberg (2010)
 The Consistency and Absolute Consistency
Problems of XML Schema Mappings
between Restricted DTDs

Hayato Kuwada1, Kenji Hashimoto2 , Yasunori Ishihara1, and Toru Fujiwara1

1
Osaka University, Japan
{h-kuwada,ishihara,fujiwara}@ist.osaka-u.ac.jp
2
Nara Institute of Science and Technology, Japan
[email protected]

Abstract. Consistency of XML schema mappings, which means that

some document conforming to the source schema can be mapped into a
document conforming to the target schema, is an essentially necessary
property. It is also important for XML schema mappings to be absolutely
consistent, that is, every document conforming to the source schema can
be mapped into a document conforming to the target schema. As a known
result, consistency of a mapping between general DTDs is EXPTIME-
complete even if the class of document patterns for defining mappings
is restricted to downward axes and qualifiers. In addition, the known
tractability result is only on a restricted class of document patterns un-
der restricted DTDs called nested-relational DTDs. Moreover, there are
few known results on the tractability of absolute consistency. In this
paper, we discuss the consistency and absolute consistency problems
under restricted DTDs called disjunction-capsuled DTDs, which were
proposed by Ishihara et al. We show that for many document pattern
classes, both problems are solvable in polynomial time under disjunction-
capsuled DTDs. Although disjunction-capsuled DTDs are an incompara-
ble class to nested-relational DTDs, a part of our results can be extended
to a proper superclass of nested-relational DTDs.

1 Introduction

XML (Extensible Markup Language) is a markup language for describing struc-

tured data and used for a variety of applications and database systems. Accord-
ingly, many schemas are defined to specify document structures, and schema
description languages such as DTD (Document Type Definition) are used for
specifying schemas. A schema mapping is the basis of data transformation from
one schema to another caused by data exchange or schema evolution. In partic-
ular, schema mappings for XML are called XML schema mappings. A mapping
is a triple of a source schema, a target schema, and a set of dependencies, which
are pairs of document patterns and specify correspondence of data. Consistency
of schema mappings, which means that some document conforming to the source

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 228239, 2013.
c Springer-Verlag Berlin Heidelberg 2013

 The Consistency Problems of XML Schema Mappings 229

schema can be mapped into a document conforming to the target schema ac-
cording to the given dependencies, is an essentially necessary property. It is
also important for schema mappings to be absolutely consistent, that is, every
document conforming to the source schema can be mapped into a document
conforming to the target schema according to the given dependencies.
Example 1. Define the source and target schemas DS and DT as follows:
DS : DT :
root -> student+ root -> student+
student -> name phone student -> name phone
phone -> home mobile other phone -> home | mobile | other
First, consider the following dependency set 1 :

1 = {student/phone[home][mobile] student/phone[home][mobile]}.

1 states that if there is a document of DS such that a phone node has both
home and mobile nodes as its children, then the document is mapped into a
document of DT with a phone node which has, again, both home and mobile
nodes as its children. However, 1 is not consistent because in any document
of DT , a phone node cannot have both home and mobile nodes as its children.
Next, consider the following dependency set 2 :

2 = {student/phone[home][other] student/phone[home][other]}.

2 is consistent because there is a document of DS which has no other nodes.

Such a document vacuously satisfies 2 . On the other hand, 2 is not absolutely
consistent because of the similar reason to the case of 1 . Lastly, consider the
following dependency set 3 :

3 = {student/phone[home][mobile][other]
student[phone/home][phone/mobile][phone/other]}.

3 is absolutely consistent because for every document of DS , if it has a phone

node with children home, mobile, and other, then it can be mapped into a
document of DT which has (at least three) phone nodes one of which has home,
another has mobile, and the other has other as their children.

Amano et al. investigated the computational complexity of the problems of deter-

mining consistency and absolute consistency of XML schema mappings [1]. The
main focus of [1] is XML schema mappings with data values, but they addressed
also mappings without data values. The results of the case without data values
are shown in Table 1. In this paper we use XPath expressions for specifying doc-
ument patterns, although in [1] tree patterns is used. Using XPath expressions
allows us to utilize the rich known results on the satisfiability problem for our
problem. XPath expressions consist of (child axis), (descendant-or-self axis),
(parent axis), (ancestor-or-self axis), (next-sibling axis), (following-
sibling-or-self axis), + (following-sibling axis), + (preceding-sibling axis),
230 H. Kuwada et al.

Table 1. Known results in [1]

XPath class General DTDs

+ + Absolute
[] Consistency
Consistency
EXPTIME-
+ + + + 2p -complete
complete
EXPTIME-
+ + + + + + 2p -complete
complete
XPath class Nested-relational DTDs
+ + Absolute
[] Consistency
Consistency
+ + PTIME PTIME
+ + + + PTIME 2p
+ + + + PSPACE-hard 2p

Table 2. Results of this paper

XPath class DC-DTDs

+ + Absolute
[] Consistency
Consistency
+ + + + + + + PTIME PTIME
+ + + + + + + + PTIME PTIME
+ + + + PTIME PTIME

(path union), [ ] (qualifier), and (wild card). Table 1 describes computa-

tional complexity of each class including only axes and operators with the +.
Nested relational DTDs are non-recursive DTDs such that each content model
is in the form of l1 lm , where all li s are distinct and each li is either li , li ,
+ ?
li , or li .
As can be seen from Table 1, constraints under which decision problems of con-
sistency and absolute consistency are solved in PTIME have been little known,
even if data values are not concerned. Thus, the purpose of this paper is to find
practically wide classes of DTDs and XPath expressions for which both problems
can be solved in PTIME. As a first step for achieving this, we focus on the case
without data values in this paper.
In this paper, we use a restricted class of DTDs called disjunction-capsuled
DTDs (DC-DTDs), which were proposed in [2]. DC-DTDs are an important class
from the theoretical point of view because many tractability results on XPath
satisfiability under DC-DTDs are known. Such results can be exploited for con-
sistency and absolute consistency problems. DC-DTDs are also a meaningful
class from the practical point of view, because 7 out of 22 real-world DTDs, 853
out of 950 real-world DTD rules are actually DC [3]. Although DC-DTDs are
incomparable to nested-relational DTDs, the tractability results on the satisfia-
bility under DC-DTDs can be extended to a wider class, called DC?+ -DTDs [3],
 The Consistency Problems of XML Schema Mappings 231

which is a proper superclass of nested-relational DTDs. 13 out of 22 DTDs, 928

out of 950 rules are DC?+ .
The results of this paper are shown in Table 2. + acrossing multiple cells
represents one of them. First, it is shown that consistency of mappings be-
tween DC-DTDs is solvable by deciding satisfiability and validity of document
patterns linearly many times. Precisely speaking, we show that under DC-DTDs,
satisfiability of a set of dependencies can be decomposed into satisfiability and
validity of each document pattern in the dependencies. Hence, if the document
patterns are in an XPath class for which satisfiability and validity are tractable,
then consistency is also tractable. Next, it is shown that absolute consistency of
mappings between DC-DTDs is solvable by deciding satisfiability of document
patterns linearly many times. Again, the technically important point is that sat-
isfiability of a set of dependencies can be done by independently checking the
satisfiability of each document pattern in the dependencies. Hence, if the doc-
ument patterns are in an XPath class for which satisfiability is tractable, then
absolute consistency is also tractable.
The rest of this paper is organized as follows. Several definitions are given in
Section 2. Sections 3 and 4 present tractability results of consistency and absolute
consistency of mappings between DC-DTDs, respectively. Section 5 summarizes
the paper.

2 Preliminaries
2.1 XML Documents and DTDs
In this paper, we regard elements of XML documents as labeled nodes, and
entire XML documents as labeled ordered trees. In what follows, we write trees
to mean labeled ordered trees. Let (v) denote the label of a node v. For a
sequence of nodes v1 v2 vn , define (v1 v2 vn ) as (v1 )(v2 ) (vn ).
A regular expression over consists of (empty sequence), the symbols in ,
and operators (concatenation, usually omitted in the notation), | (disjunction),
and (repetition). We exclude (empty set) because we are interested in only
nonempty regular expressions. Let L(e) denote the string language represented
by a regular expression e.

Definition 1. A DTD D is a triple (, r, P ), where

is a finite set of labels,
r is the root label, and
P is a mapping from to the set of regular expressions over . P (l) is called
the content model of label l.
A tree T conforms to a DTD D = (, r, P ) if the root label of T is r, and for each
node v of T and its children sequence v1 v2 vn , it holds that (v1 v2 vn )
L(P ((v))). Let TL(D) denote the set of all the trees conforming to D.

Next, we introduce DC-DTDs, a restricted class of DTDs.

232 H. Kuwada et al.

Definition 2. A regular expression e is disjunction-capsuled [2], or DC for

short, if e is in the form of e1 e2 en (n 1), where each ei is either a symbol
in or in the form of (ei ) for an arbitrary regular expression ei .
For example, (a|b ) ba is disjunction-capsuled, and is also disjunction-
capsuled because is . On the other hand, a |b is not disjunction-capsuled.
Definition 3. Let D = (, r, P ). D is a disjunction-capsuled DTD, or DC-
DTD for short, if P (l) is disjunction-capsuled for each l .

2.2 XPath Expressions

XPath is a query language for XML documents.
Definition 4. The syntax of an XPath expression p is defined as follows:
p ::= :: l | p/p | p p | p[q],
::= | | | | | + | + ,
q ::= p | q q | q q,
where l . Let X denote the set of all XPath expressions.
Note that (next-sibling axes) and (following-sibling-or-self axes) are ex-
cluded from X because they are not proper axes in W3C XPath. Moreover,
(wild card) is also excluded because it can be represented by using (path
union). On the other hand, X contains (self axis), which is not considered in
previous research. Self axis makes our discussion simpler without destroying the
known tractability results of XPath satisfiability.
Definition 5. The semantics of an XPath expression p over a tree T is defined
as follows, where p and q are regarded as binary and unary predicates over nodes
of T , respectively:
T |= ( :: l)(v, v) if (v) = l;
T |= (:: l)(v, v ) if v is a child of v and (v ) = l;
T |= (:: l)(v, v ) if v is a parent of v and (v ) = l;
T |= ( :: l)(v, v ) if v is v or a descendant of v, and (v ) = l;
T |= ( :: l)(v, v ) if v is v or an ancestor of v, and (v ) = l;
T |= (+ :: l)(v, v ) if v is a following sibling of v and (v ) = l;
T |= (+ :: l)(v, v ) if v is a preceding sibling of v and (v ) = l;
T |= (p/p )(v, v ) if there is v such that T |= p(v, v ) and T |= p (v , v );
T |= (p p )(v, v ) if T |= p(v, v ) or T |= p (v, v );
T |= (p[q])(v, v ) if T |= p(v, v ) and T |= q(v );
T |= p(v) if there is v such that T |= p(v, v );
T |= (q q )(v) if T |= q(v) and T |= q (v);
T |= (q q )(v) if T |= q(v) or T |= q (v).
A tree T satisfies an XPath expression p if there is a node v such that T |=
p(v0 , v), where v0 is the root node of T .
Definition 6. An XPath expression p is satisfiable under a DTD D if some
T TL(D) satisfies p. Moreover, an XPath expression p is valid under a DTD
D if every T TL(D) satisfies p.
 The Consistency Problems of XML Schema Mappings 233

2.3 XML Schema Mappings

In this section, we define XML schema mappings according to [1,4,5].
Definition 7. A dependency is an expression of the form , where and
are XPath expressions called document patterns. An XML schema mapping
M is a triple (DS , DT , ), where DS is a DTD representing a source schema,
DT is a DTD representing a target schema, and is a finite set of dependencies.
Hereafter, XML schema mappings are referred to as just mappings.
Originally, in [1,4,5], document patterns are defined by means of tree patterns.
However, we use XPath expressions in order to exploit many results on XPath
satisfiability and validity.
Definition 8. A pair of trees TS TL(DS ) and TT TL(DT ) satisfies a de-
pendency if the following condition is satisfied: If TS satisfies , then
TT satisfies .
The set of all pairs of trees which satisfy all dependencies of M is denoted as
!M". Then, consistency and absolute consistency are defined as follows:
Definition 9. Let M = (DS , DT , ). M is consistent if !M" = . On the
other hand, M is absolutely consistent if, for every tree TS conforming to DS ,
there exists a tree TT conforming to DT such that (TS , TT ) !M".

3 Deciding Consistency
3.1 Deciding Consistency of Mappings between General DTDs
To begin with, we show that the consistency decision problem can be solved
by deciding validity and satisfiability of combinations of document patterns in
dependencies. The following lemma is essentially equivalent to the statement
given in the proof of the EXPTIME-completeness of deciding consistency under
general DTDs [5].
Lemma 1. Let M = (DS , DT , ) and = {i i | i [n]} where [n] =
{1, 2, . . . , n}. Then, M is consistent if and only if there are two disjoint subsets
I = {i1 , , ik } and J = {j1 , , jk } of [n] such that I J = [n],
:: rS [i1 ik ] is not valid under DS , and
:: rT [j 1 j k ] is satisfiable under DT
where rS and rT are the root labels of DS and DT , respectively.
Proof. Suppose M is consistent. Then, a pair of trees (TS , TT ) belongs to !M".
Let J = {j1 , , jk } be a subset of [n] such that TT satisfies j l for each jl J.
Then, because of the semantics of qualifier [ ], TT satisfies :: rT [j 1 j k ].
On the other hand, let I = {i1 , , ik } = [n] J. Then, TT does not satisfy i
for any i I. So, because (TS , TT ) !M", TS must not satisfy i for any i I.
Therefore, because of the semantics of , TS does not satisfy :: rS [i1 ik ].
234 H. Kuwada et al.

Conversely, suppose that :: rS [i1 ik ] is not valid under DS and

:: rT [j 1 j k ] is satisfiable under DT for some I = {i1 , , ik } and
J = {j1 , , jk } such that I J = and I J = [n]. The non-validity of
:: rS [i1 ik ] under DS means that there is an instance TS TL(DS )
which does not satisfy i for any i I. On the other hand, if :: rT [j 1 j k ]
is satisfiable under DT , there is TT TL(DT ) which satisfies j for all j J =
[n] I. Therefore, the pair (TS , TT ) belongs to !M" because TS does not satisfy
l or TT satisfies l for each l l where l [n]. =

According to Lemma 1, the problem of deciding consistency of schema map-

pings can be solved by using the decision procedure of validity and satisfiability
of XPath expressions under DTDs. However, the complexity of this decision
procedure for consistency of schema mappings between general DTDs is high,
because deciding validity and satisfiability of XPath expressions under general
DTDs is known to be intractable [6] and combinatorial explosion occurs in the
choice of disjoint subsets I and J of [n]. In fact, consistency of mappings be-
tween general DTDs is known to be EXPTIME-complete [5] even if the class of
document patterns is restricted to downward axes and qualifiers.

3.2 Deciding Consistency of Mappings between DC-DTDs

Now we restrict source and target schemas to be DC-DTDs. We shall give
tractability results of consistency of mappings between DC-DTDs by showing
the following facts:
1. Satisfiability for broad XPath subclasses under DC-DTDs is decidable in
polynomial time [2,3].
2. Validity for XPath expressions in X (i.e., the whole XPath class of this
paper) under DC-DTDs is decidable in polynomial time.
3. Under DC-DTDs, we can decide consistency of a schema mapping by decid-
ing validity or satisfiability of each individual document pattern in a mapping
instead of combinations of document patterns.
First, we recall that satisfiability under DC-DTDs for several XPath subclasses
can be decided in polynomial time [2,3]. Table 3 shows the known tractability
results of XPath satisfiability under DC-DTDs.
Next, we show that validity of XPath expressions in X under DC-DTDs is
tractable. For a DC regular expression e = e1 e2 en , let e denote the regular
expression obtained by replacing with each subexpression ei such that ei =
(ei ) for some regular expression ei . For a DC-DTD D = (, r, P ), let D =

(, r, P ) such that P (l) = e when P (l) = e. Note that TL(D) is a singleton
because each content model e in D does not contain any disjunction operator
and thus it represents only one sequence over . The DTD D has the following
property on validity of XPath expressions under a DC-DTD D.
Lemma 2. For an XPath expression in X and a DC-DTD D, the unique tree
satisfies if and only if is valid under D.
t in TL(D)
 The Consistency Problems of XML Schema Mappings 235

Table 3. Computational complexity of the satisfiability decision under DC-DTDs

+ + [] Satisfiability
+ + + + + + PTIME
+ + + + + + + PTIME
+ + + + PTIME
+ + + + NP-complete
+ + + + NP-complete
+ + + NP-complete
+ + + + + NP-complete

Proof. For the if part, assume that is valid under D. Then, since t TL(D),
t satisfies . For the only if part, assume that t satisfies . From the definition
of D, for each l , the unique sequence in L(P (l)) is a subsequence of any
w L(P (l)). Let t be an arbitrary tree in TL(D). Then there is an injective
mapping from the set of nodes of t to that of t such that
(v) = ((v)) for each node v of t,
for any two nodes v1 and v2 of t, if v1 is a parent of v2 , then (v1 ) is a parent
of (v2 ), and
for any two nodes v1 and v2 of t, if v1 is a following sibling of v2 , then (v1 )
is a following sibling of (v2 ).
For this, we can see by induction on the structure of that t |= (v, v ) implies
t |= ((v), (v )). Thus, t also satisfies . =

Lemma 2 leads to tractability of XPath validity checking under DC-DTDs. That

is, validity of an XPath expression in X under a DC-DTD D is linear-time
reducible to evaluation of on the tree t TL(D). Since a polynomial-time
algorithm for evaluation is known [7], validity of XPath expressions in X under
DC-DTDs can be decided in polynomial time.
Lastly, we can avoid combinatorial explosion mentioned in Section 3.1 by the
following lemmas.
Lemma 3. Let 1 and 2 be XPath expressions in X and D be a DC-DTD.
Either or both of 1 and 2 is valid under D if and only if :: r[1 2 ] is valid
under D, where r is the root label of D.

Proof. The only if part is trivial. For the if part, assume that :: r[1 2 ] is
Because t TL(D), t satisfies :: r[1 2 ].
valid under D. Consider t TL(D).
By the semantics of , t satisfies at least one of 1 and 2 . Without loss of
generality, we assume that t satisfies 1 . From Lemma 2, 1 is valid under D.
=

Lemma 4. Let 1 and 2 be XPath expressions in X and D be a DC-DTD.

Both 1 and 2 are satisfiable under D if and only if :: r[1 2 ] is satisfiable
under D, where r is the root label of D.
236 H. Kuwada et al.

! node of

Fig. 1. Construction of t12 from t1 and t2 by inserting subtrees

Proof. The if part is trivial. For the only if part, assume that both 1 and 2
are satisfiable under D. Then, there are trees t1 and t2 in TL(D) such that
t1 and t2 satisfy 1 and 2 , respectively. Because X does not include negation
operator and next-sibling axis, if a tree t satisfies XPath expression X , any
tree obtained by inserting arbitrary subtrees to t also satisfies [3]. Thus, if we
can obtain the same tree t12 from t1 and t2 by inserting subtrees (see Fig. 1), t12
satisfies :: r[1 2 ]. To show that such t12 exists, we should prove that for any
regular DC expression e, if w1 , w2 L(e), there is w L(e) such that w1 and
w2 are subsequences of w. This property holds because DC regular expressions
cannot specify non-co-occurrence among labels. =

From the above, we can refine Lemma 1 for deciding consistency of mappings
between DC-DTDs as follows.
Lemma 5. Let M be a mapping (DS , DT , ) such that the schemas DS and
DT are DC-DTDs. Then, M is consistent if and only if there is no dependency
in such that is valid under DS and is not satisfiable under DT .
Proof. Let = {i i | i [n]} and rS and rT be the root labels of DS
and DT , respectively. From Lemma 1, we prove this lemma by showing the next
condition: there is no such that is valid under DS and is not
satisfiable under DT if and only if there are two disjoint subsets I = {i1 , . . . , ik }
and J = {j1 , . . . , jk } of [n] such that I J = [n], :: rS [i1 ik ] is not
valid under DS , and :: rT [j 1 j k ] is satisfiable under DT .
Assume that there is no such that is valid under DS and is
not satisfiable under DT . Then, for each l [n], l is not valid under DS or l
is satisfiable under DT . Let I = {i1 , . . . , ik } be the set of all the indexes il in [n]
such that il is not valid under DS . Let J = {j1 , . . . , jk } = [n] I. Then j l is
satisfiable under DT for each jl J. From Lemmas 3 and 4, :: rS [i1 ik ]
is not valid under DS and :: rT [j 1 j k ] is satisfiable under DT .
Conversely, assume that there are two disjoint subsets I = {i1 , . . . , ik } and
J = {j1 , . . . , jk } of [n] such that I J = [n], :: rS [i1 ik ] is not valid
under DS , and :: rT [j 1 j k ] is satisfiable under DT . From Lemmas 3
and 4, l is not valid under DS for any l I and l is satisfiable under DT for
any l J. Since I J = [n], there is no such that is valid under
DS and is not satisfiable under DT . =

 The Consistency Problems of XML Schema Mappings 237

Here, in order to decide consistency of mappings between DC-DTDs, we just have

to check validity of under DS and satisfiability of under DT for each
in . Therefore, if we choose as the class of document patterns the XPath
subclasses in which validity and satisfiability under DC-DTDs are tractable,
deciding consistency of schema mappings is tractable.

Theorem 1. Let M be a mapping (DS , DT , ), where DS , DT are DC-DTDs,

such that for each , X and XT , where XT is an XPath
subclass for which satisfiability is decidable in polynomial time under DC-DTD
(see Table 3). Then, consistency of M can be decided in polynomial time.

4 Deciding Absolute Consistency

4.1 Deciding Absolute Consistency of Mappings between General

DTDs

We begin with a lemma stating that absolute consistency of a mapping M =

(DS , DT , ) can be determined by checking satisfiability of all XPath expres-
sions induced by all subset of . Again, this lemma is essentially equivalent to
the statement in the explanation of 2p -completeness of absolute consistency of
mappings between general DTDs [1].

Lemma 6. Mapping M = (DS , DT , ) is absolutely consistent if and only if

the following condition holds: For every subset = {1 1 , . . . , k k }
of , if :: rS [1 k ] is satisfiable under DS , then :: rT [1 k ]
is satisfiable under DT , where rS and rT are the root labels of DS and DT ,
respectively.

Proof. Suppose that mapping M = (DS , DT , ) is absolutely consistent, and

consider an arbitrary subset = {1 1 , . . . , k k } of . Also suppose
that TS TL(DS ) satisfies :: rS [1 k ]. Since M is absolutely consistent,
there is TT TL(DT ) such that (TS , TT ) !M". Hence, TT satisfies :: rT [1
k ] and the only if part holds.
Conversely, suppose that for every subset = {1 1 , . . . , k k }
of , if :: rS [1 k ] is satisfiable under DS , then :: rT [1 k ] is
satisfiable under DT . Consider an arbitrary tree TS TL(DS ). Let = {1
1 , . . . , k k } be a subset of such that TS satisfies i for each i (1 i
k). Then, TS satisfies :: rS [1 k ], and therefore, :: rS [1 k ] is
satisfiable under DS . Then, by assumption, there must be TT TL(DT ) such
that TT satisfies :: rT [1 k ], and hence, (TS , TT ) !M". Consequently,
M is absolutely consistent. =

Similarly to the case of consistency, the above lemma tells us that the hardness
of deciding absolute consistency stems from the combinatorial explosion caused
by checking all the subset of dependencies as well as the hardness of deciding
XPath satisfiability.
238 H. Kuwada et al.

4.2 Deciding Absolute Consistency of Mappings between DC-DTDs

To obtain tractability results on absolute consistency, we focus on DC-DTDs
again. Similarly to the case of consistency, the combinatorial explosion of sub-
sets of dependencies can be avoided, that is, satisfiability can be independently
checked for each dependency.
Lemma 7. Let DS and DT be DC-DTDs. Mapping M = (DS , DT , ) is ab-
solutely consistent if and only if for every dependency , if is
satisfiable under DS , then is satisfiable under DT .
Proof. Suppose that mapping M = (DS , DT , ) is absolutely consistent. Then,
by Lemma 6, for every subset = {1 1 , . . . , k k } of , if ::
rS [1 k ] is satisfiable under DS , then :: rT [1 k ] is satisfiable
under DT . This immediately implies the only if part of the lemma.
Conversely, suppose that for every dependency , if is satisfi-
able under DS , then is satisfiable under DT . By the property of DC-DTDs
(Lemma 4), two XPath expressions 1 and 2 are satisfiable under DC-DTD D
if and only if :: r[1 2 ] is satisfiable under D, where r is the root label of
D. Hence, for any subset = {1 1 , . . . , k k } of , 1 ,. . . , k are
satisfiable under DS if and only if :: rS [1 k ] is satisfiable under DS . On
the other hand, by the assumption, if 1 ,. . . , k are satisfiable under DS , then
1 ,. . . , k are satisfiable under DT , and hence, :: rT [1 k ] is satisfiable
under DT . Therefore, the if part holds. =

Now, the next theorem immediately follows from the above lemma:
Theorem 2. Let M be a mapping (DS , DT , ), where DS , DT are DC-DTDs,
such that for each , , XT , where XT is an XPath subclass for
which satisfiability is decidable in polynomial time under DC-DTD (see Table 3).
Then, absolute consistency of M can be decided in polynomial time.

5 Conclusion
This paper has discussed the consistency and absolute consistency problems of
XML schema mappings between DC-DTDs. First, we have shown that consis-
tency of mappings between DC-DTDs can be solved by deciding validity and sat-
isfiability of XPath expressions linearly many times. Moreover, we have proved
that the XPath validity problem in the presence of DC-DTDs can be solved
in polynomial time. From these facts, we have proved that the consistency of
mappings between DC-DTDs can be solved in polynomial time if the document
patterns are in an XPath class for which satisfiability and validity are tractable
under DC-DTDs. Next, we have shown that absolute consistency of the map-
pings between DC-DTDs can also be solved by deciding satisfiability of XPath
expressions in the presence of DC-DTD linearly many times. So the absolute
consistency of mappings between DC-DTDs can also be solved in polynomial
time if the document patterns are in an XPath class for which satisfiability is
tractable under DC-DTDs.
 The Consistency Problems of XML Schema Mappings 239

As stated in Section 1, there is a superclass of DC-DTDs, called DC?+ -

DTDs [3]. They allow two operators ? (zero or one occurrence) and + (one or
more occurrences) of regular expressions in a restricted manner. Actually, DC?+ -
DTDs are also a proper superclass of nested-relational DTDs. The tractability re-
sults of absolute consistency of mappings between DC-DTDs can be extended to
mappings between DC?+ -DTDs because DC?+ -DTDs inherit all the tractability
of XPath satisfiability of DC-DTDs. However, as for consistency, the tractability
results do not seem to be extended to mappings between DC?+ -DTDs because
we are conjecturing that XPath validity under DC?+ -DTDs is coNP-hard. Now
we are trying to prove the coNP-hardness and then propose a DTD class which
is a superclass of both nested-relational DTDs and DC-DTDs but under which
XPath validity is tractable.
We are also planning to incorporate data values into mappings. As stated
in [1], consistency of mappings with data values can be reduced to consistency
of mappings without data values. However, such reduction is not possible for
absolute consistency. It is interesting to find XPath classes such that absolute
consistency of mappings between DC-DTDs with data values becomes tractable.

Acknowledgment. The authors thank Mr. Yohei Kusunoki of Osaka Univer-

sity for his insightful comments and suggestions on deciding XPath validity under
DC-DTDs. The authors are also thankful to the anonymous reviewers for their
helpful comments. This research is supported in part by Grant-in-Aid for Scien-
tific Research (C) 23500120 from Japan Society for the Promotion of Science.

References
1. Amano, S., Libkin, L., Murlak, F.: XML schema mappings. In: Proceedings of
the Twenty-Eigth ACM SIGMOD-SIGACT-SIGART Symposium on Principles of
Database Systems, pp. 3342 (2009)
2. Ishihara, Y., Morimoto, T., Shimizu, S., Hashimoto, K., Fujiwara, T.: A tractable
subclass of DTDs for XPath satisfiability with sibling axes. In: Gardner, P., Geerts,
F. (eds.) DBPL 2009. LNCS, vol. 5708, pp. 6883. Springer, Heidelberg (2009)
3. Ishihara, Y., Shimizu, S., Fujiwara, T.: Extending the tractability results on XPath
satisfiability with sibling axes. In: Lee, M.L., Yu, J.X., Bellahs`ene, Z., Unland, R.
(eds.) XSym 2010. LNCS, vol. 6309, pp. 3347. Springer, Heidelberg (2010)
4. Arenas, M., Libkin, L.: XML data exchange: Consistency and query answering.
Journal of the ACM 55(2) (2008)
5. Arenas, M., Barcelo, P., Libkin, L., Murlak, F.: Relational and XML Data Exchange.
Morgan & Claypool (2010)
6. Benedikt, M., Fan, W., Geerts, F.: XPath satisfiability in the presence of DTDs.
Journal of the ACM 55(2) (2008)
7. Bojanczyk, M., Parys, P.: XPath evaluation in linear time. Journal of the
ACM 58(4), 17 (2011)
 Linking Entities in Unstructured Texts
with RDF Knowledge Bases

Fang Du1,2 , Yueguo Chen1 , and Xiaoyong Du1

1
School of Information, Renmin University of China, Beijing, China
2
School of Mathematics and Computer Science, Ningxia University, China
{dfang,chenyueguo,duyong}@ruc.edu.cn

Abstract. Entity linking (entity annotation) is the task of linking named entity
mentions on Web pages with the entities of a knowledge base (KB). With the con-
tinued progress of information extraction and semantic search techniques, entity
linking has received much attention in both research and industrial communi-
ties. The challenge of the task is mainly on entity disambiguation. To our best
knowledge, the huge existing RDF KBs have not been fully exploited for entity
linking. In this paper, we study the entity linking problem via the usage of RDF
KBs. Besides the accuracy of entity linking, the scalability of handling huge Web
corpus and large RDF KBs are also studied. The experimental results show that
our solution on entity linking achieves not only very good accuracy but also good
scalability.

Keywords: RDF knowledge base, entity linking, named entity disambiguation.

1 Introduction
Nowadays, search engines have been widely used as the most convenient way of ac-
cessing information within the huge amount of unstructured texts on the Web. How-
ever, techniques adopted by most existing search engines are based on straightforward
matches of terms within queries and those within the unstructured texts. This often re-
sults in that, users are frustrated in frequently adjusting their query terms to retrieve
the desired results. To address this problem, semantic search has been proposed and
widely studied [1]. In semantic search, the contextual meaning of terms in the unstruc-
tured texts is very important to enrich the semantics of terms. In this paper, we study
the problem of enriching contexts of named entities within unstructured texts, by link-
ing the detected named entities with existing RDF knowledge bases. By doing this, we
are able to extract important concepts and topics out of the plain texts, and effectively
support the semantic search over the unstructured texts.
With the continued progress of Semantic Web and information extraction techniques,
more and more RDF data emerge on the web. They form many huge RDF knowledge
bases (KBs) such as Yago[2], Freebase[3], DBPedia[4] et al. Such RDF KBs contain
billions of RDF triple facts either extracted from Web pages or contributed manually by
users, describing information about hundreds of millions of entities. The entity infor-
mation in these KBs is so abundant and diverse that they are very suitable for enriching
the contexts of named entity mentions within the unstructured texts.

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 240251, 2013.
c Springer-Verlag Berlin Heidelberg 2013
 Linking Entities in Unstructured Texts with RDF Knowledge Bases 241

The entity linking (also called entity annotation in the paper) task is defined as to
map a named entity mention m in the free texts of a Web page to a corresponding entity
e in a KB. In our work, a KB refers to an RDF KB without further specification. We list
some applications that such a way of entity linking can be applied as follows:
Enhancing search results of semantic search. Through linking entities with RDF
KBs, the search results of unstructured texts can be expanded by utilizing relevant
semantic information derived from RDF KBs.
Cleaning the extracted named entities. The named entity mentions extracted by
automatic information extraction tools often have quality issues such as errors or
duplicates. Linking entities with proper entities in RDF KBs can help to clean the
extracted named entities.
Finding relations between entities. Given a pair of entities, their relations can be
learned from the contexts. The accuracy of entity relation discovery will highly
depend on the accuracy of entity resolution during the annotation process.

One major challenge of entity linking is the disambiguation problem. A named en-
tity mention has the disambiguation problem because a number of entities (of different
types) in the KB can have same name mention. When this happens, we need to accu-
rately link the entity mention to a proper KB entity. This is called entity resolution (or
entity disambiguation) problem. For example, a mention Michael Jordan may map to
at least two entities in a KB. One is a famous retired NBA basketball player, and the
other is a famous computer scientist in U.C. Berkeley. Many existing approaches use
a lot of features from contexts to address the entity disambiguation problem. However,
an RDF KB can be treated as a huge labeled graph, by considering the properties of the
graph, we observe that the relevant entities of mentions in the same context often are
close pairs in the KB graph. Based on the observation, we propose an effective approach
that accomplish the entity disambiguation problem with only two proposed features.
An open domain RDF KB has a large number of entities. However, many existing
approaches[8,9] are designed on small-scale data set, or can only process one document
a time[21], which are not scalable. For a huge RDF KB, entity linking will have to be
done in a parallel and distributed fashion for guaranteeing the efficiency and scalability.
Fortunately, the MapReduce[6] framework on cloud computing provides an easy-to-use
way of dispatching the expensive annotating and disambiguating tasks over clusters. We
therefore propose an efficient entity disambiguation algorithm that is able to conduct the
entity linking task in parallel based on the MapReduce framework.
The main contributions of the paper are as follows:
We propose a simple but effective algorithm to accurately link named entity men-
tions in the unstructured texts with proper entities in an RDF KB.
We propose an MapReduce-based approach to linking mentions of unstructured
texts with entities in the RDF KB. The annotation of multiple documents can be
conducted parallelly and in batch.
We test the performance of our approaches over two real-life RDF KBs. The re-
sults show that our solution can achieve high accuracy on both KBs. Meanwhile,
by designing algorithms running over MapReduce framework, the solution is very
efficient and scalable.
242 F. Du, Y. Chen, and X. Du

The rest of the paper is organized as follows. Section 2 reviews the related work. Section
3 describes the solution of entity disambiguation using RDF KBs. In Section 4, we
propose an efficient framework of conducting the entity linking task on MapReduce
framework. Section 5 presents the results of experimental study. Finally, the paper ends
with conclusion and future work in Section 6.

2 Related Work
Entity linking has received much attention recently[7,15,14], especially after the emerg-
ing of large knowledge bases such as Yago[2], Freebase[3], DBPedia[4]. Most of these
works[8,9] link free texts to real world knowledge bases. The most important and chal-
lenging issue in entity linking task is the named entity disambiguation problem.
Named entity disambiguation is also called co-reference resolution or word sense
disambiguation in some other studies. Most studies in this area tackle the challenges
based on rich contexts where named entity mentions occur. Bagga and Baldwin[11]
use vector space model to resolve the ambiguities in persons names. They represent
the context of the entity mentions with the bag of words model and compute similarity
between vectors. An early work [7] identifies the most proper meaning of ambiguous
words by measuring their context overlaps. In [10], the authors derive a set of features
from Wikipedia to compute similarity between context of mentions and the texts of
Wikipedia.
In word sense disambiguation, there are two main streams of solutions[17,18]. Some
researchers use knowledge extracted from dictionaries to identify the correct word in
a given context[8,16]. The others collect probabilities from large amounts of sense-
annotated data, and then use machine learning approaches to solve the problem[17,18].
The authors of [12] implement and combine these two approaches. They choose the
right and left word of the ambiguous word as its context. For each word, they extract
a training feature vector from Wikipedia links, and then integrate the features in Naive
Bayes classifier. These works in word sense disambiguation are similar to the entity
linking task. However, they are more likely to assign dictionary meanings to the poly-
semous words, which is not as difficult as the entity linking task.
Weikum and Theobald[19] propose their named entity disambiguation approach in
facts harvesting research works. In [20], they interconnect RDF data and Web contents
via LOD (Linked Open Data)[5]. They construct an entity mention graph and use a
coherence graph algorithm to solve the entity disambiguation problem. Through com-
puting the coherence between ambiguous entities, the mentions are connected to at most
one entity in a KB.
The most similar work to ours are Linden[21]. Linden first builds a dictionary based
on four kinds of Wikipedia pages: entity page, redirect page, disambiguation page and
hyper links. By using the dictionary, Linden generates a candidate linking list for each
mention. In order to conduct entity disambiguation, it creates feature vectors of four
dimensions to rank the entities in candidate list. However, Linden can process only
one document at the same time. For real world applications, there will be much more
mentions from more than one document to be processed simultaneously. Moreover, it
generates many features to rank the candidate linking entities. Some of them (e.g., the
global coherence feature) are relatively hard to be accurately computed.
 Linking Entities in Unstructured Texts with RDF Knowledge Bases 243

RDF (Resource Description Framework) is recommended by W3C Consortium to

describe information on the Web. The basic information unit of RDF data is a triple
t =< s, p, o > which stands for subject, predicate and object respectively. An RDF
KB contains a finite set of RDF triples. Assuming subjects and objects are entities,
predicates are relations between entities, an RDF KB then forms labeled entity relation
graphs. An example is shown in Fig. 1.

Fig. 1. An example of RDF KB graph. Circles Fig. 2. An example of semantic KB graph.

refer to entities in RDF KBs. Square nodes refer to the classes of entities.

To the best of our knowledge, there is no entity linking work based on RDF knowl-
edge base over MapReduce framework. Because RDF data can be represented as graphs,
we can use the properties of the graph to design a simple but effective approach.
The frequently used notations in this paper are summarized in Table1.

Table 1. Frequently Used Notations

D all documents need to be processed

D one document in D
M all mentions in D
M the set of mentions in D
C the set of contexts of mention m M
L the candidate linking entity list of m M
E the set of linking entities of each c C

3 Entity Disambiguation via RDF Knowledge Bases

The entity linking task basically links named entity mentions with entities in a KB. If
it is navely conducted based on string match, quite often, a mention will be mapped to
more than one entities in the KB because of the homonymy problem. In our work, we
use Jaccard to compute the similarity between two strings (entity mention and entity in
244 F. Du, Y. Chen, and X. Du

KB), an entity mention will be mapped to an entity in KB if the similarity between them
is higher than some threshold. For example, a mention Michael Jordan in unstructured
texts will be mapped to more than one entity in KB, for instance, Michael I. Jordan
and Michael J. Jordan. This requires a solution of entity disambiguation.
In this study, we take two measures into account for addressing the entity disam-
biguation problem. One is the relevance, which measures the relation between a men-
tion and an entity. The other one is the homogeneity, that is to find the proper linking
entity which has closer hyponyms with the mention. By studying the properties of exist-
ing RDF KBs, we introduce two features to measure the relevance and the homogeneity
of an entity in a KB to a given mention of certain context.

3.1 Construct Semantic RDF KB Graph

An RDF KB can be represented as a graph, which is called RDF KB graph. Based on

the RDF KB graph G, we construct a semantic KB graph SG. Through SG we propose
approaches to linking mentions with proper entities in an RDF KB.

Definition 1 (RDF KB Graph). Given an RDF KB, let be the set of all entities in
the KB, R be the set of relations between entities in the KB. An RDF KB Graph is
G = (V, E, L), where V is the set of vertexes, E is the set of edges and L is the set of
literals. V , E R, Lv is the set of labels on vertex, while Le is the set of labels on
edges. L = Lv Le .
Definition 2 (Semantic KB Graph). Given an RDF KB graph, a semantic KB graph
(SG) is SG = (V, E, L, C), where V is the set of vertexes, E is the set of edges, L is
the set of literals and C is the set of classes. v V, Cv C.
To construct the semantic KB Graph, we firstly recognize entities and relations in a
given RDF KB, where an entity is the triples with same subject(s), and a relation be-
tween entities is the predicate (p) which connects two entities. Based on entities and
relations, we can easily construct an RDF KB graph. Then, we pick predicates such as
rdf:type in the RDF KB graph as initial class nodes set. Duplicates and noises in the
set are cleaned. Some new class nodes are added based on WordNet. We finally built
the taxonomy of the RDF KB. For pages limitation, we will not discuss the details of
taxonomy construction. An RDF KB graph expanded by the taxonomy is called seman-
tic KB graph. Fig. 1 is an example of an RDF KB graph, while Fig. 2 is an example of
the corresponding semantic KB graph. For instance, after expanded as Semantic Graph,
node Michael Jordan and node David E. Rumelhart in Fig. 1 have class node as
Professor in Fig. 2.

3.2 Context Dependency

The mentions appearing in the same context often have a high probability to be under
the same topic. Meanwhile, in an RDF KB graph, the relevant entities (e.g., under the
same topic) are likely to be close with each other in terms of the graph distances. For
example, if Michael Jordan refers to the famous basketball player, mentions such
as NBA and Chicago Bulls are more likely to be found than mentions such as
 Linking Entities in Unstructured Texts with RDF Knowledge Bases 245

machine learning in the context. Moreover, in the KB graph, the shortest path distance
between the entitiesMichael Jordan (basketball player) and NBA is likely to be
shorter than the distance between Michael Jordan (computer scientist) and NBA.
Accordingly, we are able to measure the relevance of entities to a given mention.
We denote the mentions in same document with m as the contexts of m, that is C. An
entity e E is the linking entity of a named entity mention m M. Suppose ck C
is the kth context mention, ek E is the linking entity of ck (where we choose the cs
from C which have only one linking entity. Mentions have no linking entity are useless.
Mentions with more than one linking entities will lead to a recursive processing.), we
formally define the context dependency of an entity e, denoted as CD(e) as:
n
"
1
CD(e) = ck D Dist(ek , e), (1)
N
k=1,k=i

where Dist(ek , e) is a function measuring the distance between ek and e as

1
Dist(ek , e) = (2)
Dk
where Dk is the shortest path distance from e to ck . If the average distance between
ei L and all ek E is shorter than some other e L, the value of CD(e) for ei will
be the highest.

3.3 Semantic Similarity

Most entity nodes in an RDF KB have at least one corresponding class node (see square
nodes in Fig.2). A class node may have a super class node as its ancestors. Therefore,
when a mention m M is linked to an entity e, m will have the same class node
and super class nodes as e. For these mentions without direct connected class node,
we choose the class with the shortest path to m as its class node. As shown in Fig.2,
if a mention links to an entity Michael Jordan (the basketball player), it will have a
basketball player as its class node, and sportsman as a super class node.
Assuming that the taxonomy of the KB is a tree, for two entities ei and ej in KB, the
classes in taxonomy for them are Ci and Cj respectively, where C0 (i, j) is the lowest
common ancestor (LCA) class (super class) of Ci and Cj . Suppose ei is the linking
entity of mi , while cik is one of the context mention of mi , ej is the linking entity of
cik . The semantic similarity between ei and ej can be computed by equation 3.
r(i, j)
Sim(Ci , Cj ) = (3)
2
P
where r(i, j) is the reverse weight of class C. r(i, j) = LC 0 ij , Suppose C0 (i, j) is at
the level L in the taxonomy tree, P is the shortest path distance between C0 (i, j) and
Ci in the tree. Larger P leads to less r(i, j), which means less similarity of Cj and Ci .
Therefore, the semantic similarity for entity e can be defined as:
n
"
1
SS(e) = Sim(C, Cj ), (4)
N
j=1,j=i
246 F. Du, Y. Chen, and X. Du

SS(e) shows the type similarity between the context mention c and the mention m,
it measures the homogeneity of e. For example, e1 and e2 are two candidate linking
entities of m, if the class nodes of linking entities (e) for c are more closer to e1 than
e2 , e1 is more likely to be the proper linking entity of m. Higher value of SS(e) means
closer class structure to the candidate linking entity.

3.4 Ranking Candidate Entities

The score of each candidate entity in the candidate list of a mention m can be computed
by combining CD(e) and SS(e):

Score(e) = CD(e) + (1 ) SS(e) (5)

where are weights of CD(e) and SS(e). A max-margin technique introduced in
[23,21] is applied to learn the weights.
When there are more than one entities in the candidate linking list (CL) for a mention
m, the one with the highest score etop = maxeCL Score(e) is chosen as the proper
linking entity of m. Algorithm 1 provides the details of linking entity disambiguation.

Algorithm 1. Linking Entity Disambiguation (LED)

Input: CL1 , . . . , CLt : A set of candidate entity lists, each CL has a corresponding mention m
Input: SG: A Semantic KB graph
Input: , weight parameters
:
Output: M E1 , . . . , M En : a list of < m, e > pairs, where m is the mention and e is the proper
corresponding linking entity
1. ME = ;
2. for each mention m and corresponding CL do
3. map CL of m and CL of context mention C of m to same node;
4. for each key value pair (e,ek ) do
5. compute Dist(e!k , e);
6. reduce e as key, Dist(mathf rakek , e) as value;
7. compute CD(e);
8. for each class of ei and class of ej do
9. compute Sim(Ci , Cj ) !
10. for each key value pair (e, Sim(Ci , Cj )) do
11. compute SS(e);
12. for each key value pair (e,(list of CD(e), SS(e))) do
13. Score(e) = CD(e) + SS(e);
14. e = max Score(e);
15. add e into M E;
16. return a set of M E;

4 Fast Entity Linking

Nowadays, as numerous data emerge both on Web pages and RDF KBs, the entity link-
ing task faces with the big data problem. Therefore, efficiency and scalability become
 Linking Entities in Unstructured Texts with RDF Knowledge Bases 247

important issues of the entity linking solution. In this section, we propose a frame-
work for efficient entity linking based on the entity disambiguation algorithm proposed
above. We assume that the mentions need to be linked to entities have been extracted
and recognized from web pages. We also assume the mentions have potential linking
entities in the KB. In order to linking a mention to an RDF KB, we propose a framework
including the following modules:
Generate Candidate Entity Linking List. For each mention m M , obtain the
list of entities having the name m (in the KB) as the candidate linking list. For the
mentions that do not have any linking entity in the KB, NIL will be returned as the
result.
Linking Entities Disambiguation. If the candidate linking list has more than one
candidate, we need to find the proper linking entity of m. For each mention m, we
use a defined score function to compute the score of each entities e in the candi-
date linking list. Firstly, we compare the context dependency to see which candi-
date linking entity is more relevant to the context mentions. Secondly, we compute
semantic similarity to check the similarity of candidate entities in terms of their se-
mantic types. Finally, a score is calculated for each entity in candidate linking list,
the one with the maximal score will be the proper entity to mention m.
Each mention m in Web pages will be compared with the entities e in KB, if they
match, e will be add into the candidate linking list of m. The length of linking list, |L|,
must be in one of the following three cases. (1) |L| = 0, which means no matching
entity exists in the knowledge base. (2) |L| = 1, which means there is only one match
in KB for m. (3) |L| 1, which means two or more e in knowledge base are the
possible entity links of m. For case 1, NIL will be returned to m; for case 2, the only
e is returned to show there is a link between m and e; only for case 3 we need to use
the disambiguation algorithm described in Section3 to find the proper linking entity of
m. The algorithm is executed on MapReduce framework. In the map phrase, mentions
and entities with the same key are dispatched to the same reduce node; In the reduce
phrase, mentions(entities) and their neighbors in the semantic KB graph are gathered
to compute CD and SS. After several iterative map and reduce phases, the algorithm
can output the proper linking entity of m. With MapReduce framework, the algorithm
is conducted in a parallel fashion, which makes the approach more scalable.

5 Experimental Study
5.1 Experimental Setup
All the experiments are run on the Renda-Cloud platform, which is a Hadoop cluster.
Each node in the cluster has 30GB memory, 2TB disk space and 2.40 GHz core 24
processor, running Hadoop 0.20.2 under Ubuntu 10.04 Linux 2.6.32-24.
In order to evaluate our approach, we choose two real world RDF KBs, YAGO and
FreeBase, to conduct the experimental study. YAGO is a large knowledge base devel-
oped at Max-Planck-Institute Sarrbrucken. It contains information from Wikipedia and
linked to Wordnet. There are more than 10 million entities and 120 million facts about
the entities. YAGO contains not only entities and relations between entities but also
248 F. Du, Y. Chen, and X. Du

semantics of entities. The semantics relations include TYPE, subClassOf, domain et

al. Based on the structure, we can build semantic KB graph over YAGO. FreeBase is
a large collaborative knowledge base developed by the Metaweb company. It contains
data from Wikipedia, NNDB et al., with 125M tuples and more than 4000 types of enti-
ties. The facts in FreeBase are on different fields, such as sports, arts and entertainment.
In the experiments, we use Accuracy(Accu.) to evaluate the performance of our so-
lution. It is calculated as in Equation 6:

Nl
Accu. = (6)
Nm
where Nl is the number of mentions which are correctly linked to entities, Nm is the
total number of mentions.
Sections 5.2 and 5.3 give the experimental results on accuracy and scalability over
the two RDF KBs respectively.

5.2 Experiments on YAGO

Accuracy. In order to evaluate the performance, we firstly extract 1000 mentions from
50 Wiki pages as 50 sets of Mentions. mentions in each set are contexts for each other.
We test Accu. by using different features in score function. Table 2 shows the highest,
lowest and average Accu. with different features for the 50 groups of mentions. Then
we compare our solution with Linden, which is the most similar work with ours in the
existing solutions. we use the 1000 mentions in one group and test our solution on 2
Hadoop nodes, while test Linden on a PC with a 1.8Ghz Core 4 Duo processor, 10 GB
memory and running 64-bit Linux Redhat kernel. Table 3 shows the results.

Table 2. Experimental results on YAGO: Accu. Table 3. Experimental results on YAGO: com-
pared with Linden
Accu. our solution Linden
feature set
Max Min Avg. feature set Accu. feature set Accu.
CD 0.9406 0.9198 0.9325 CD 0.9395 LP 0.8665
SS 0.9264 0.8813 0.9054 SS 0.9240 LP+SA 0.9189
CD + SS 0.9487 0.9232 0.9411 CD + SS 0.9436 LP+SA+SS 0.9391

From Table 2 we can see that the accuracy of CD is higher than SS. By combining
the two features, our solution achieves the best accuracy. It indicates that in most cases,
relatedness is more useful in dealing with entity linking problem. Although Linden has
good performance, the comparison results in Table 3 shows that the accuracy of our
solution is better than that of Linden.
Scalability. In order to test the scalability of the algorithm, we varying the number
of mentions from 1000 to 100,000 and execute the algorithms on 2 Hadoop nodes, 8
Hadoop nodes, 20 Hadoop nodes respectively. The results are shown in Table 4. The
time consumption includes generating candidate linking list and scoring candidate link-
ing entities, for these are the main jobs in linking entity task.
 Linking Entities in Unstructured Texts with RDF Knowledge Bases 249

Table 4. Time Consumption in Seconds over different number of mentions

# of mentions
# of cluster nodes 1000 5000 10000 100000
2 23.39 26.27 38.61 108.05
8 20.45 23.14 29.38 42.06
20 16.07 18.27 22.75 35.11

Results in Table 4 show the scalability of our solution. With more nodes in Hadoop
cluster, the computing capability are efficiently enhanced. The minimum execution time
in our solution is above 10 seconds. Note that the initial time for MapReduce job is at
least 10 seconds.

5.3 Experiments on Freebase

In this part of experiment, we choose 30GB Freebase KB data and more than 250,000
mentions to evaluate our solution in processing a large amount of data. The mentions
are extracted from IMDB, which is a database about movies, directors, actors and rating
from users et al. We store the data from IMDB into a flat table, each row in table are
extracted from one page, therefore the cell contents in same row are treated as context
to each other. We also test the accuracy and the scalability to see the performance of
our solution.
Accuracy. In the experiment, we firstly process 250,000 mentions in 50 group, the
results in Table 5 show the maximum, the minimum and the average accuracy with
different features used in the score function. Then we compare our solution of feature
SS with baseline: LCA(Lowest Common Ancestor) with one group of 5,000 mentions
(results are shown in Table 6), for both SS in our approach and LCA are annotated
entities by using the information of their classes (types). Where LCA is simply assign
all cells in one column with their lowest common ancestor.

Table 5. Experimental results on Freebase: Table 6. Experimental results on Freebase:

Accu. Compare with Baseline

Accu. Our solution Baseline

feature set
Max Min Avg. Accu. Accu.
CD 0.9377 0.9104 0.9265 SS LCA
SS 0.9212 0.8795 0.9033 0.9033 0.6680
CD + SS 0.9407 0.9185 0.9366

From Table 5 we can see that the accuracy results are worse than that in Table 2.
We just store data extracted from IMDB without any preprocessing, because they have
little noisy than data form Wikipedia. However, the accuracy is still above 0.93. The
results in Table 6 show that with only the feature SS our solution still outperforms the
baseline.
250 F. Du, Y. Chen, and X. Du

Table 7. RDF KB preprocessing run-time in seconds over different nodes of Hadoop cluster.

# of nodes in Hadoop cluster.

Size of KB
2 nodes 8 nodes 20 nodes
5M 42 17 13
600M 286 134 58
30G > 30min 1305 392

Scalability. In this part of experiments, we test the time consumption in KB preprocess-

ing to see the scalability of our solution. The main work of KB preprocessing includes
construction of RDF KB graph and semantic KB graph. We treat it as a preprocessing
step, because once the RDF KB graph and semantic KB graph has been constructed,
we can do entity linking on top of them when needed without of reconstruction.
We choose part of connected data from Freebase, so that we can vary the KB size
from 5M to 30G. We execute KB preprocessing steps on 2 Hadoop nodes, 8 Hadoop
nodes, 20 Hadoop nodes respectively. The time consumption results are shown in Table
7. The results show that with more nodes in Hadoop cluster ,we can handle the linking
entity task over KB size above tens of Gigabytes. It also shows that the algorithm we
proposed in the paper has high scalability. Our solution gives an efficient way to handle
the linking entity problem on big data.

6 Conclusions

In this paper, we propose a simple but effective algorithm to accurately link entity men-
tions with proper entities in an RDF KB. The algorithm is designed on the MapReduce
framework, which can annotate multiple documents in parallel. The experimental study
over huge real-life RDF knowledge bases demonstrates the accuracy and scalability of
the approach.
Based on our solution we can study semantic search techniques using the annotated
unstructured texts. For example, by linking unstructured texts to RDF KB, we are able
to build a huge extended labeled graph, through which we are able to apply and extend
semantic keyword search over text graphs.
Acknowledgements. This work is supported by the National Science Foundation of
China under grant No. 61170010 and No. 61003085, and HGJ PROJECT 2010ZX01042-
002-002-03.

References

1. Guha, R.V., McCool, R., Miller, E.: Semantic search. In: WWW, pp. 700709 (2003)
2. Suchanek, F.M., Kasneci, G., Weikum, G.: Yago: A core of semantic knowledge unifying
wordnet and wikipedia. In: WWW, pp. 697706 (2007)
3. Bollacker, K.D., Evans, C., Paritosh, P., Sturge, T., Taylor, J.: Freebase: a collaboratively cre-
ated graph database for structuring human knowledge. In: SIGMOD, pp. 12471250 (2008)
 Linking Entities in Unstructured Texts with RDF Knowledge Bases 251

4. Auer, S., Bizer, C., Kobilarov, G., Lehmann, J., Cyganiak, R., Ives, Z.G.: DBpedia: A Nucleus
for a Web of Open Data. In: Aberer, K., Choi, K.-S., Noy, N., Allemang, D., Lee, K.-I., Nixon,
L.J.B., Golbeck, J., Mika, P., Maynard, D., Mizoguchi, R., Schreiber, G., Cudre-Mauroux,
P. (eds.) ISWC/ASWC 2007. LNCS, vol. 4825, pp. 722735. Springer, Heidelberg (2007)
5. Linking Open Data, http://www.w3.org/wiki/SweoIG/TaskForces/
CommunityProjects/LinkingOpenData
6. Dean, J., Ghemawat, S.: MapReduce: Simplified Data Processing on Large Clusters. In:
OSDI 2004 (2004)
7. Lesk, M.: Automatic sense disambiguation using machine readable dictionaries: How to tell
a pine cone from an ice cream cone. In: SIGDOC, Toronto (June 1986)
8. Mihalcea, R.: Large vocabulary unsupervised word sense disambiguation with graph-based
algorithms for sequence data labeling. In: Proceedings of the Human Language Technol-
ogy/Empirical Methods in Natural Language Processing Conference, Vancouver (2005)
9. Navigli, R., Velardi, P.: Structural semantic interconnections: a knowledge-based approach to
word sense disambiguation. IEEE Transactions on Pattern Analysis and Machine Intelligence
(PAMI) 27 (2005)
10. Bunescu, R., Pasca, M.: Using Encyclopedic Knowledge for Named Entity Disambiguation.
In: EACL, pp. 916 (2006)
11. Bagga, A., Baldwin, B.: Entity-based cross-document coreferencing using the vector space
model. In: COLING, pp. 7985 (1998)
12. Dredze, M., McNamee, P., Rao, D., Gerber, A., Finin, T.: Entity disambiguation for knowl-
edge base population. In: COLING, pp. 277285 (2010)
13. Hasegawa, T., Sekine, S., Grishman, R.: Discovering relations among named entities from
large corpora. In: ACL, pp. 415422 (2004)
14. Demartini, G., Difallah, D.E., Cudre-Mauroux, P.: ZenCrowd: leveraging probabilistic rea-
soning and crowdsourcing techniques for large-scale entity linking. In: WWW, pp. 469478
(2012)
15. Stoyanov, V., Mayfield, J., Xu, T., Oard, D.W., Lawrie, D., Oates, T., Finnin, T.: A context
aware approach to entity linking. In: The Joint Workshop on Automatic Knowledge Base
Construction and Web-scale Knowledge Extraction, NAACL-HLT 2012 (2012)
16. Navigli, R., Velardi, P.: Structural semantic interconnections: a knowledge-based approach
to word sense disambiguation. IEEE Transactions on Pattern Analysis and Machine Intelli-
gence 27, 10751086 (2005)
17. Gliozzo, A., Giuliano, C., Strapparava, C.: Domain kernels for word sense disambiguation.
In: ACL (2005)
18. Ng, H., Lee, H.: Integrating multiple knowledge sources to disambiguate word sense: An
examplar-based approach. CoRR, vol. 9606032 (1996)
19. Weikum, G., Theobald, M.: From information to knowledge: harvesting entities and relation-
ships from web sources. In: PODS, pp. 6576 (2010)
20. Hoffart, J., Yosef, M.A., Bordino, I., Furstenau, H., Pinkal, M., Spaniol, M., Taneva, B.,
Thater, S., Weikum, G.: Robust Disambiguation of Named Entities in Text. In: Proceedings
of EMNLP, pp. 782792 (2011)
21. Shen, W., Wang, J.Y., Luo, P., Wang, M.: LINDEN: linking named entities with knowledge
base via semantic knowledge. In: WWW 2012, pp. 449458 (2012)
22. Shen, W., Wang, J.Y., Luo, P., Wang, M.: LIEGE: Link Entities in Web Lists with Knowledge
Base. In: Proceedins of KDD 2012 (2012)
23. Carlos, B.T., Guestrin, C., Koller, D.: Max-margin markov networks. In: NIPS (2003)
24. Nadeau, D., Turney, P.D., Matwin, S.: Unsupervised Named-Entity Recognition: Generating
Gazetteers and Resolving Ambiguity. In: Lamontagne, L., Marchand, M. (eds.) Canadian AI
2006. LNCS (LNAI), vol. 4013, pp. 266277. Springer, Heidelberg (2006)
25. Stanford NER, http://nlp.stanford.edu/ner/index.shtml
26. Hadoop, http://hadoop.apache.org/
 An Approach to Retrieving Similar Source Codes
by Control Structure and Method Identifiers

Yoshihisa Udagawa

Faculty of Engineering, Tokyo Polytechnic University

1583 Iiyama, Atsugi-city, Kanagawa, 243-0297 Japan
[email protected]

Abstract. In this paper, we present an approach to improve source code retriev-

al using the structure of control statements. We develop a lexical parser that ex-
tracts structural information for each method including control statements and
method identifiers. We present a source code retrieval model, named derived
structure retrieval model, in which a retrieval conditions are defined as a se-
quence of control statements and/or method identifiers, and meaningful search
conditions are derived from the given sequence. Experiments on the source
code of Struts 2 Core show that the derived structure retrieval model outper-
forms the vector space retrieval model.

Keywords: Derived structure retrieval model, Control statement, Method iden-

tifier, Vector space retrieval model.

1 Introduction

Source code retrieval is an important task in order to maintain the quality of software
in the development cycle. Duplicated code complicates this activity and leads to high-
er maintenance cost because the same bugs will need to be fixed and consequently
more code will need to be tested. Many developers put significant effort into finding
software defects. However, due to the vast amounts of source codes, it is difficult to
find efficiently the code segments that we want.
Various techniques have been proposed to collect similar source codes. These
techniques can be classified into four categories:

Text-based comparison

This approach compares source codes in the same partition. The key idea of this ap-
proach is to identify source-code fragments using similar identifiers [3].

Token comparison

In this approach, before comparison, tokens of identifiers (data type names, variable
names, etc.) are replaced by special tokens, and then similar subsequences of tokens

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 252259, 2013.
Springer-Verlag Berlin Heidelberg 2013
 An Approach to Retrieving Similar Source Codes 253

are identified [2]. Because the encoding of tokens abstracts from their concrete val-
ues, code fragments that are different only in parameter naming can be detected.

Metrics comparison

This approach characterizes code fragments using some metrics, and compares these
metric vectors instead of directly comparing the code [4].

Structure-based comparison

This approach applies pattern matching and complex algorithms on abstract syntax
trees or dependency graphs. Baxter et al. [1] propose a method using abstract syntax
trees for detecting exact and near-miss program source fragments. Ouddan et al. [5]
propose a multi-language source code retrieval system using the structural content of
the source code and the semantic one. Their approach is focused on detecting plagiar-
ism in source code between programs written in different programming languages,
such as C, C++, Java, and C#.
Our approach is the structure-based comparison that takes a sequence of statements
as a retrieval condition. We developed a lexical parser and extracted structures of
source codes for control statements and method identifiers. The extracted structural
information is an input to the vector space model and a proposed source code retrieval
model, named derived structure retrieval model. Our retrieval model takes a se-
quence of statements as a retrieval condition and derives meaningful search condi-
tions from the given sequence. Because a program is composed of a sequence of
statements, our retrieval model practically improves the performance of source code
retrieval.
The rest of this paper is organized as follows. In Section 2, we present architecture
of the developed tools. In Section 3, we present overview of the Struts 2 framework
[7] and extracted structures. In Section 4, we discuss a similarity retrieval approach
based on the vector space model. In Section 5, we discuss the derived structure re-
trieval model and results of code retrieval. Section 6 concludes the paper.

2 Overview --- Developed Tools

Figure 1 illustrates a high-level architecture of the tools we developed. The structure

extraction tool is implemented in C-language. It extracts control statements and me-
thod identifiers from every method of a class in Java. Then, these extracted structures
are inputted to a statistic tool, a vector space retrieval tool, and a derived structure
retrieval tool, which are written in Visual Basic language. Finally, the outputs of those
modules are fed into a source code viewer.
The vector space retrieval tool retrieves a set of similar methods by assigning con-
trol statements and method identifiers to a document vector, and then computing the
cosine of the angle between a document vector and a query vector as similarity.
Details are discussed in Section 4.
The derived structure retrieval tool takes a sequence of control statements and/or
method identifiers as a retrieval condition. Then, it generates derived sequences by
254 Y. Udagawa

Fig. 1. Architecture of developed tools

replacing each element of the given sequence with a wildcard. In case, the sequence
n
consists of n elements, 2 - 2 sequences are derived to retrieve a set of similar me-
thods. Details are discussed in Section 5.

3 Struts 2 Core and Extracted Structures

3.1 Framework and Struts 2

A framework automates common tasks, and thereby providing a user platform that
simplifies web development. The Struts 2 framework [7] implements the model-view-
controller (MVC) design pattern.
We estimated the volume of the source codes using file metrics. Struts 2.3.1.1 Core
consists of 46,100 lines in source code including comment lines and blank lines. As
for the number of lines, Struts 2.3.1.1 Core is a middle scale application in industry.
We can estimate the volume of the source codes using file metrics. Typical file me-
trics used are as follows:
Number of Java Files ---- 368
Number of Classes ---- 414
Number of Methods ---- 2,667
Number of Code Lines ---- 21,543
Number of Comment Lines ---- 17,954
Number of Total Lines ---- 46,100
The number of Java files are 368, which differs from the number of declared classes
414 because some java files include definitions of inner classes and anonymous
classes.

3.2 Extracted Control Statements and Method Identifiers

Fig. 2 shows the extracted structures of the evaluateClientSideJsEnablement() method
in Form.java file in org.apache.struts2.components package. Three numbers preceded
by # symbol are the number of comment lines, the number of blank lines and the
number of code lines, respectively. Because the extracted structures include the
depth of nesting of the control statements, they supply enough information for retriev-
ing methods using a substructure of source codes.
 An Approach to Retrieving Similar Source Codes 255

Fig. 2. An example of extracted control statements and method identifiers

4 Code Retrieval Using the Vector Space Model

4.1 Structural Metrics as Vector Components

The vector space model [6] is an algebraic model for representing text documents as
vectors of identifiers or terms. Given a set of documents D, a document dj in D is
represented as a vector of term weights:

(1)

where N is the total number of terms in document dj and wi, j is the weight of the i-th
term. A user query can be similarly converted into a vector q:

(2)

The similarity between document dj and query q can be computed as the cosine of the
angle between the two vectors dj and q in the N-dimensional space:

(3)
256 Y. Udagawa

4.2 Results of Code Retrieval by the Vector Space Model

The left part of Table 1 shows the top 27 methods obtained by a query vector with the
components consisted by one if-statement and one addParameter method identifier.
The addParameter method maintains a pair of a parameter identifier and its value in
an internal list of Struts 2 Core. It is typically used after checking the existence of a
parameter identifier. Thus, one if-statement and one addParameter method identifi-
er are a reasonable retrieval condition in the vector space retrieval model.
392 methods are retrieved with the similarity between 0.997 and 0.009 because the
392 methods include at least one if-statement. However, only 21 methods whose No
columns are shaded in Table 1 are valid candidate methods in the sense that an add-
Parameter method is called after an if-statement. Details are discussed in Section 5.

Table 1. The top 27 retrieved methods

5 Code Retrieval Using Derived Structure Retrieval Model

5.1 Derived Structure Retrieval Model

In the vector space retrieval model, a document is represented as a vector of terms

that comprise the document. Similarity of a document and a query is calculated as the
cosine of the angle between the document vector and the query vector. This means
 An Approach to Retrieving Similar Source Codes 257

that the order in which the terms appear in the document is lost in the vector space
model. On the other hand, a program is essentially a sequence of statements. So, the
vector space model has limitations in performance when it is applied to retrieving
program source codes.
The derived structure retrieval model, a source code retrieval model proposed in
the paper, takes a sequence of statements as a retrieval condition, and derives mea-
ningful search conditions from the given sequence. Let S1 and S2 are statements ex-
tracted by the structure extraction tool. [S1 S2] denotes a sequence of S1 followed by
S2. In general, for a positive integer n, let Si (i ranges between 1 and n) be a state-
ment. [S1 S2 ... Sn] denotes a sequence of n statements, i.e., S1, S2, and Sn.
The essential idea of the derived structure retrieval model is matching on wildcard
characters. Given a sequence [S1 S2 ... Sn] and a method structure m, our model
first try to search structure of m that matches the sequence [S1 S2 ... Sn]. Next,
derive a sequence, such as [* S2 ... Sn], by replacing one of elements in
[S1 S2 ... Sn], and try to search structure of m that matches the derived sequence.
The algorithm continue to evaluate structure of m until all of derivable sequences,
except for the sequence [* * ... *], are evaluated.
The sequence [S1 S2 ... Sn] is more "restrictive" than [* S2 ... Sn] and
n
[S1 * ... Sn], etc. In general, 2 - 1 sequences are derived from an n-element se-
quence including the sequence whose all elements are replaced by wildcards. Because
the sequence [* * ... *] matches all substructures, it does not work as a retrieval
n
condition. A given sequence [S1 S2 ... Sn] and 2 - 2 derived sequences define a
domain of discourse for retrieval conditions.
Let Count (m, [S1 S2 ... Sn], r) is a function that counts the number of sub-
structures of m that match derived sequences of [S1 S2 ... Sn] whose r elements
are replaced by the wildcard. Fig.3 shows an algorithm to compute the similarity of
methods that near match a retrieval condition [S1 S2 ... Sn]. In Fig.3, it is as-
sumed that getMethodStructure(j) returns a structure of the j-th method extracted by
the structure extraction tool. Note that the similarity is 1.0 when a method includes
the sequence [S1 S2 ... Sn] and it does not include any of derived sequences from
[S1 S2 ... Sn].

5.2 Results of Code Retrieval by Derived Structure Retrieval Model

A sequence [if addParameter] is used as a retrieval condition in order to compare
the result obtained in Subsection 4.2. The right part of Table 1 shows retrieved me-
thods that match sequences [if addParameter], [* addParameter], and [if *].
FormButton::void evaluateExtraParams() method is ranked 21th with similarity
0.200 in the derived structure retrieval model, whereas it is ranked 10th with similarity
0.913 in the vector space model. The method includes five if-statements, three addPa-
rameter identifiers, and one [if addParameter] sequence. Because the similarity in
the vector space model is computed by the number of statements and/or method iden-
tifiers, FormButton::void evaluateExtraParams() method is ranked higher than it is
ranked in the derived structure retrieval model.
258 Y. Udagawa

Input: set_of_structure M;
Input: sequence [S1 S2 ... Sn];
Output: Sim[M.length];
// Declare array Sim[].
double Sim[M.length];
// Compute Similarity for each method in M.
int Nume; int Deno;
for ( int j=0; j < M.length; j++ ) {
Nume= Count(getMethodStructure(j),[S1 S2 ... Sn],0);
Deno= 0;
for ( int r=1; r < [S1 S2 ... Sn].length; r++ ){
Deno= Deno+ Count(getMethodStructure(j),[S1 S2 ... Sn],r);
}
Sim[j]= (double) Nume / (double)( Nume + Deno );
}

Fig. 3. Algorithm to compute the similarity of methods for a sequence [S1 S2 ... Sn]

The similarity of ActionError::void evaluateExtraParams() method is 0.000 in the

derived structure retrieval model, whereas it is 0.894 and is ranked 11th in the vector
space model. The method includes two if-statements, two addParameter identifiers
and zero [if addParameter] sequences.
A program is essentially represented by a sequence of statements. Because the vec-
tor space model computes similarity of two document vectors based on a set of terms
that comprise the document vectors, it theoretically has a limitation on performance in
source code retrieval. On the other hand, because the derived structure model com-
putes similarity based on a sequence of statements, it achieves higher performance
than the vector space model.
As shown in Table 1, 6 methods do not satisfy the [if addParameter] sequence
condition, the derived structure retrieval model achieves (27-21) / 27= 22.2% im-
provement. Table 2 shows a summary of performance improvement of other expe-
riments on source code retrieval.

Table 2. Summary of performance improvement on source code retrieval

 An Approach to Retrieving Similar Source Codes 259

6 Conclusions

In this paper, we presented an approach that improves source code retrieval using the
structural information of control statements. Our key contribution in this research is
the development of an algorithm that derives meaningful search conditions from a
given sequence, and then performs retrieval using all of the derived conditions. Thus
our source code retrieval model retrieves all of source codes that partially match the
given sequence.
The results are promising enough to warrant future research. In future work, we
will work on improving our methods by combining information, such as an inherit-
ance of a class and implementation of an abstract class, etc. into structural informa-
tion. We will also conduct experiments on various types of open source programs
available on the Internet.

References
1. Baxter, I.D., Yahin, A., Moura, L., SantAnna, M., Bier, L.: Clone detection using abstract
syntax trees. In: Proc. of the 14th International Conference on Software Maintenance, pp.
368377 (November 1998)
2. Kamiya, T., Kusumoto, S., Inoue, K.: CCFinder: A multi-linguistic tokenbased code clone
detection system for large scale source code. IEEE Transactions on Software Engineer-
ing 28(7), 654670 (2002)
3. Marcus, A., Maletic, J.: Identification of high-level concept clones in source code. In: Proc.
of the 16th International Conference on Automated Software Engineering, pp. 107114
(November 2001)
4. Mayland, J., Leblanc, C., Merlo, E.M.: Experiment on the automatic detection of function
clones in a software system using metrics. In: Proc. of the 12th International Conference on
Software Maintenance, pp. 244253 (November 1996)
5. Ouddan, A.M., Essafi, H.: A Multilanguage Source Code Retrieval System Using Structur-
al-Semantic Fingerprints. International Journal of Computer and Information Science and
Engineering, 96101 (Spring 2007)
6. Salton, G., Buckley, C.: Term-Weighting approaches in automatic text retrieval. Informa-
tion Processing and Management 24(5), 513523 (1988)
7. Struts - The Apache Software Foundation (2012), http://struts.apache.org/
 Complementary Information for Wikipedia
by Comparing Multilingual Articles

Yuya Fujiwara1 , Yu Suzuki2 , Yukio Konishi1 , and Akiyo Nadamoto1

1
Konan University, 8-9-1 Okamoto, Higashi-Nada, Kobe, Hyogo 6588501, Japan
2
Nagoya University, Furo, Chikusa, Nagoya, Aichi 4648601, Japan

Abstract. Information of many articles is lacking in Wikipedia because

users can create and edit the information freely. We specifically exam-
ined the multilinguality of Wikipedia and proposed a method to comple-
ment information of articles which lack information based on comparing
dierent language articles that have similar contents. However, much
non-complementary information is unrelated to a users browsing article
in the results. Herein, we propose improvement of the comparison area
based on the classified complementary target.

1 Introduction

Wikipedia has become an extremely popular website, with over 284 language
versions in 2012. Nevertheless, information of many articles is lacking because
users can create and edit the information freely. Furthermore, Wikipedia has
dierent levels of value of its information depending on the language version
of the site because dierent users create and edit articles of the respective lan-
guage versions. We specifically examine the multilinguality of Wikipedia and
propose a method that complements information of articles that lacks infor-
mation based on comparing dierent language articles that have the similar
contents[3].
In our proposed method, a user first browses an article in Wikipedia of the
users language. Then the system extracts complementary information from user-
specified articles that have been composed in other languages. As described in
our research, we designate the lacking information as complementary infor-
mation, an article that a user is browsing and which lacks information as a
browsing article, its Wikipedia as a browsing Wikipedia, and its language
as a browsing language. We also designate a dierent language article as a
comparison article, its Wikipedia as a comparison Wikipedia, and its lan-
guage as a comparison language.
When we compare a browsing article with a comparison article, the infor-
mation granularity is shown to dier. The browsing article is only one page,
but the comparison article might have multiple pages because these articles are
written by intercultural authors. Our proposed method has been used to extract
lacking information from all content in target articles in comparison Wikipedia.

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 260267, 2013.
c Springer-Verlag Berlin Heidelberg 2013

 Complementary Information for Wikipedia 261

However, much un-complementary information is unrelated to a users browsing

article in the results.
In this paper, we propose improvement of the method, which is to extract
complementary information from dierent language articles in Wikipedia. The
improvement points are the following.
Classifying complementary target articles.
Improving the comparison area based on a classified complementary target.
Experimenting to ascertain the usefulness of our proposed method.
The process used for our proposed system is (1)A user who browses the Wikipedia
article clicks the complement button on the web page, and inputs the compari-
son language. (2)The system extracts comparison articles from the comparison
Wikipedia using a link graph of Wikipedia. (3)It classifies the comparison articles
and extracts a comparison area from each comparison article. (4)It compares the
browsing article with the comparison articles extracted in step (3), and extracts
complementary information. (5)It browses the browsing article and the comple-
ment information.
In our prototype system, we use the English version of Wikipedia as the
browsing Wikipedia, and the Japanese version of Wikipedia as the compari-
son Wikipedia. Nevertheless, our system is language-independent. When we use
other versions of Wikipedia, we can develop a version for any language. In ad-
dition, our target users are people who browse Wikipedia articles.

2 Related Work
Our proposed approach uses link structure analysis techniques. Many studies
use the Wikipedia link structure. Milne,[4] and Strub et al.[7] extract measures
of semantic relatedness of terms using the Wikipedia link structure.
Some studies have been conducted to examine complementary information.
Ching et al. [1] propose a framework to assist Wikipedia editors to enrich
Wikipedia articles. They extract complementary information by comparing mul-
tilingual Wikipedia. However, they are not interested in dierences of informa-
tion granularity between the languages. In contrast, we consider the coverage
of comparison articles and create a Wikipedia article to extract complementary
information from comparison articles. Eklou et al.[2] proposed a method of in-
formation complementation of Wikipedia by the web. They extract complemen-
tary information from web pages using LDA. Their goal is similar to ours, but
the target complement information is dierent. Ma et al.[6] proposed integrat-
ing cross-media news contents such as TV programs and Web pages to provide
users with complementary information. They extract complementary informa-
tion from web pages using their proposed topic structure. Their research issue,
which is complement information, is similar to our issue. However, the target of
the content is dierent.
262 Y. Fujiwara et al.

3 Extraction of Complementary Information from

Articles
3.1 Extraction of Comparison Articles from Comparison Wikipedia
We extract comparison articles based on the Wikipedia link graph and our pro-
posed relevance degree.

Link Structure Analysis

We create a link graph for a comparison Wikipedia based on the users browsing
title. First, the system extracts an article having the same title as the users
input from the comparison Wikipedia. We designate the article as the basic
article. We regard the basic article as the root node of the link graph. Next,
we regard the interactive linked articles which are the subjects of link-out and
link-in connections with the basic article as important and relevant articles to
the basic article. The system extracts all interactive linked articles from the
comparison Wikipedia. It includes the articles as nodes and thereby creates the
link graph.
Calculating Relevance Degree
The article of an interactive link does not always have a high relation to a basic
node. We calculate the relevance degree between a root node (basic article) and
other nodes (interactive linked articles) in the link graph. An important anchor
text appears many times in a basic article, and appears also in the summary area
in Wikipedia. We use that structural feature and calculate the relevance degree
between the root node and other node in the link graph as described below.
1. The system divides the basic article according to the section which is the
structure of the table of contents of the basic article. We designate the di-
vided parts as segments.
2. From the link graph, the system extracts a title from a interactive linked
article. The title becomes the keyword when we extract the anchor text
from the basic article. We designate the title as the target title.
3. The system counts the anchor text of the target title in the summary area
of the basic article. The number of the anchor text is T F sumi . Therein, i is
the identification number of the interactive link article.
4. It also counts of the anchor text in each segment of the basic article. The
number of the anchor text is T Fik . Therein, k demotes the segment number.
5. It calculates the relevance degree between each segment and the interactive
linked article for which the title is the target title. We use the following
expression to calculate the degree of relevance.
n
"
Ri = ( T F sumi Ssumi + (T Fik Sik ))/ max(Rim ) (1)
k=1

In the equation presented above, Ri is the relevance degree of each interactive

link article i. Ssumi represents the cosine similarity between the summary
area of the basic article and an interactive link article i. n is the number
 Complementary Information for Wikipedia 263

of segment in the basic article. Sik represents the cosine similarity between
segment k of the basic article and an interactive link article i. max(Rim )
signifies the maximum value in all Ri .
The system calculates from 2. to 5. for all interactive linked articles in the link
graph. When relevance degree Ri is less than the threshold value, we delete
the article from the link graph. When the system extracts complementary in-
formation from the comparison articles, nodes that consist of a root node and
other nodes become comparison articles.

3.2 Comparison Area in Comparison Articles

Then some comparison articles have information that is unrelated to the title
of browsing article. For example, we extract My Neighbor Totoro, which is
famous Japanese animated movie as a basic article, then we extract Sayama
Hill as a comparison article. (Sayama Hill is the setting of the movie. The hill
actually exists in Japan.) However, the Sayama Hill article invariably mentions
the geographic information and the animals that live on the hill. The information
about My Neighbor Totoro is only a small amount of the information included in
the Sayama Hill article. In this case, our past proposed method extracted infor-
mation such the geographically information and the animals in Sayama Hill as
complementary information for My Neighbor Totoro. However this information
is not related to My Neighbor Totoro. In this paper, we propose a new method
that incorporates consideration of the comparison area of comparison articles.
In our new method, we first cluster three kinds of articles. Next we extract com-
parison areas from articles of each kind and extract complementary information
from them.

3.2.1 Clustering of Articles

We divide the comparison articles into three types.
Basic article
A basic article has the same title as that of a browsing article. It is a root
node in the link graph.
Inclusive relation article
When a comparison article has an inclusive relation to a basic article, it
becomes an inclusive relation article. That is, an inclusive relation article is
a kind of is-a relation to a basic article. For example, if Sushi is a basic
article, then Inari-sushi and Maki-sushi are inclusive relation articles because
they are kinds of sushi. When we extract the inclusive relation article, we
use Lead Sentence Parsing (LSP) method[5]. LSP method is that the lead
part is often defined is-a relation to other articles in Wikipedia. That is,
when there is the anchor text that links to the basic article in a lead part of
comparison article, the article becomes an inclusive relation article.
Partial match article
When a partial content in a comparison article is related to the basic article,
it becomes a partial match article. We define a partial match article as a
264 Y. Fujiwara et al.

comparison article that has no anchor text which links to the basic article
in the lead part, but it has anchor text in other parts of comparison articles.
That is, the partial match article excludes inclusively related articles from
comparison articles.
3.2.2 Comparison Area
We extract complementary information from the comparison area, which is based
on articles of three types described above.
Basic Article
A comparison area of the basic article is all contents of a page (article) because
a basic article has the same title as the browsing article. It is strongly related to
the browsing article. (Fig. 1(a)).
Inclusive Relation Article
An inclusive relation article is a subset of a basic article. We regard a comparison
area of the inclusive relation article as all contents of a article. That is, the
comparison area is the same as its basic article. (Fig. 1(a)).
Partial Match Article
We specifically examine the position of the anchor text which links to the basic
article, in a partial match article. We extract a comparison area as follows:

An anchor text in a title area of section or sub-section.

Anchor text in a title of section
We consider that the section and its sub-section are related to a basic
article. A comparison area is both section and sub-section (Fig. 1(b)).
Anchor text in a title of sub-section
We consider that the sub-section and its parent section are related to
the basic article. A comparison area is both its parent section and the
sub-section of itself. In this case, we regard sibling sub-sections as not
related to the sub-section. They are not a comparison area (Fig. 1(c)).
Anchor text only in the context area of a partial match article.
We consider a paragraph that includes an anchor text is related to a basic
article. We regard only the paragraph as a comparison area (Fig. 1(d)).

3.3 Extracting Complementary Information from Comparison

Articles
After the system extracts comparison articles, it calculates and extracts comple-
mentary information by comparing the browsing article with each comparison
article. Almost all Wikipedia articles are divisible into segments based on the
table of contents, which means that the segments are divided semantically. When
comparing the similarity of multilingual Wikipedia, we specifically examine the
segments of the table of contents of Wikipedia specifically. Particularly, we com-
pare each article based on each segment.
 Complementary Information for Wikipedia 265

Fig. 1. Comparing browsing article with comparison article

1. We extract nouns from respective articles.

2. We translate comparison articles to browsing language using dictionaries. At
this time, the translation is insucient because it is done solely using the
dictionary. We use a translation API such as Google API. Furthermore, we
translate proper nouns using Wikipedias language link. As described in this
paper, we are unconcerned about word sense disambiguation. That is left as
a subject for future work.
3. We extract complementary information from each type of comparison article.
Basic Article and Inclusive Relation Article
The comparison area of these articles includes all contents of the article.
However, it is dicult to extract complementary information based on com-
paring whole contents of the browsing article with whole contents of the
comparison area. We specifically examine the structure of an article. Almost
all Wikipedia articles are divided into segments based on the table of con-
tents, meaning that the segments are divided semantically. We calculate the
similarity between the browsing article and each comparison article using co-
sine similarity. When comparing the similarity of multilingual Wikipedia, we
examine the segment of the table of contents of Wikipedia specifically(show
Fig. 1(a)). Particularly, we compare each article based on each segment.
#
xi yi
cos(xk , yl ) = < # 2 # 2 (2)
xi yi
In that equation, xk signifies a segment of the browsing article k; yl denotes a
segment of comparison article l. In addition, xi represents the term frequency
of noun i in a browsing articles segment and yi stands for the term frequency
of noun i in a comparison target articles segment.
If the cosine similarity is less than the threshold value, then it is regarded
as dierent information.
Partial Match Article
We extract complementary information by comparing the browsing article
with the comparison area in section 3.2.2. When we compare articles, we
266 Y. Fujiwara et al.

Table 1. Dataset and results of experiment

Dataset Our proposed method Baseline

Query
A1 B2 C3 Precision Recall F-measure Precision Recall F-measure
My Neighbor Totoro 9 2 25 0.82 0.86 0.84 0.75 0.85 0.8
Doraemon 9 5 38 0.65 0.77 0.71 0.76 0.7 0.74
Iaido 10 4 24 1 0.82 0.9 0.45 0.82 0.58
Manzai 3 4 22 1 0.48 0.65 1 0.85 0.92
Yukata 1 3 7 0.75 0.5 0.6 0.8 0.67 0.72
Urashima Taro 5 5 39 1 0.48 0.64 0.48 0.57 0.52
Pikachu 6 2 25 1 0.55 0.71 0.46 0.55 0.5
Kinkaku-ji 24 6 28 0.78 0.88 0.82 0.33 0.88 0.48
Hello Kitty 15 2 60 0.71 0.48 0.57 0.6 0.48 0.53
Kyudo 16 3 47 0.92 0.45 0.69 0.37 0.45 0.41
Average - - - 0.86 0.62 0.71 0.6 0.68 0.62

use cosine similarity, as shown in equation (2). At this time, xk signifies

a segment of the browsing article k; yl denotes a paragraph of comparison
article l which is comparison area in section 3.2.2. In addition, xi represents
the term frequency of noun i in a browsing articles segment and yi stands
for the term frequency of noun i in a comparison area in the partial match
article.

4 Experimental Evaluation
We assessed the availability of extracting complementary information based on
comparing our proposed method which is a calculated comparison area with a
baseline which is not a calculated comparison area. In both of our experiments,
we set an appropriate weight =3.0 and threshold =0.2 of the relevance degree
described in equation (1) in section 3, and threshold =0.2 of cosine similar-
ity in equation (2) in our earlier experiment[3]. Table 1 presents results of the
experiment. The precision is 0.26 higher than the baseline, which means that
our proposed method is useful for extracting complementary information. An
example of a good result is the browsing article is written for the plot and cast
of My Neighbor Totoro. One comparison article is written about Sayama Hill,
the geographic information related to Sayama Hill diers, but it is unrelated to
the movie. This time, in our proposed method, the information is not a com-
parison area. We do not extract the information as complementary information.
However, the exemplary bad results are those for Manzai, which is a type of
Japanese comedy. We can not extract the information written about the associ-
ation of Manzai, Roukyoku Manzai which is a kind of manzai, and so on because
the anchor text in each case is in an itemization area. Our proposed method
1
The total number of section in browsing article.
2
The total number of comparison articles.
3
The total number of sections in comparison articles.
 Complementary Information for Wikipedia 267

processes an itemization area is a segment. Few contents are included in seg-

ments. Therefore, we can not extract such information. This is left as a problem
for future study.

5 Conclusion

As described in this paper, we proposed a method for extracting complementary

information from Wikipedia by comparing multilingual Wikipedia. Specifically,
we improve our proposed method to examine the comparison area based on a
classified complementary target. This paper presented three important technical
points, (1)Classifying complementary target articles, (2)Improving the compar-
ison area based on a classified complementary target, and (3)Experiment to
measure the usefulness of our new proposed method.
Future work is that we should consider word sense disambiguation, because
many instances of word sense disambiguation exist, but we have ignored such
cases in the analyses presented in this paper.

Acknowledgments. This work was partially supported by Research Institute

of Konan University.

References
1. Ching-Man, A.Y., Duh, K., Nagata, M.: Assisting wikipedia users in cross-lingual
editing. In: WebDB Forum 2010, pp. 1112 (2010)
2. Eklou, D., Asano, Y., Yoshikawa, M.: How the web can help wikipedia: a study on
information complementation of wikipedia by the web. In: Proceedings of the 6th
International Conference on Ubiquitous Information Management and Communica-
tion, ICUIMC 2012 (2012)
3. Fujiwara, Y., Suzuki, Y., Konishi, Y., Nadamoto, A.: Extracting Dierence Infor-
mation from Multilingual Wikipedia. In: Sheng, Q.Z., Wang, G., Jensen, C.S., Xu,
G. (eds.) APWeb 2012. LNCS, vol. 7235, pp. 496503. Springer, Heidelberg (2012)
4. Milne, D.: Computing semantic relatedness using wikipedia link structure. In: Proc.
of New Zealand Computer Science Research Student Conference, NZCSRSC 2007.
CDROM (2007)
5. Nakayama, K.: Wikipedia mining for triple extraction enhanced by co-reference
resolution. In: Proceedings of the 1st International Workshop on Social Data on the
Web, SDoW 2008 (2008)
6. Qiang, M., Nadamoto, A., Tanaka, K.: Complementary information retrieval for
cross-media news content. Elsevier ARTICLE Information Systems 31(7), 659678
(2006)
7. Strube, M., Ponzetto, S.P.: WikiRelate! computing semantic relatedness using
wikipedia. In: Proceedings of the 21st International Conference on Artificial In-
telligence, AAAI 2006, pp. 14191424 (2006)
Identification of Sybil Communities Generating
Context-Aware Spam on Online Social Networks

Faraz Ahmed and Muhammad Abulaish

Center of Excellence in Information Assurance

King Saud University, Riyadh, Saudi Arabia
{fahmed.c,mabulaish}@ksu.edu.sa

Abstract. This paper presents a hybrid approach to identify coordi-

nated spam or malware attacks carried out using sybil accounts on online
social networks. It also presents an online social network data collection
methodology, with a special focus on Facebook social network. The pages
crawled from Facebook network are grouped according to users interests
and analyzed to retrieve users profiles from each of them. As a result,
based on the users page-likes behavior, a total number of six groups has
been identified. Each group is treated separately and modeled using a
graph structure, termed as profile graph, in which a node represents a
profile and a weighted edge connecting a pair of profiles represents the
degree of their behavior similarity. Behavior similarity is calculated as a
function of common shared links, common page-likes, and cosine similar-
ity of the posts, and used to determine weights of the edges of the profile
graph. Louvains community detection algorithm is applied on the profile
graphs to identify various communities. Finally, a set of statistical fea-
tures identified in one of our previous works is used classify the obtained
communities either as malicious or benign. The experimental results on a
real dataset show that profiles belonging to a malicious community have
high closeness-centrality representing high behavioral similarity, whereas
those of a benign community have low closeness-centrality.

Keywords: Social network analysis, social network security, sybil com-

munity detection.

1 Introduction
Online social networking sites have attracted a large number of internet users.
Among many existing Online Social Networks (OSNs), Facebook and Twitter
are the most popular social networking sites with over 800 million and 100 mil-
lion active users, respectively. However, due to this popularity and existence of
a rich set of potential users, malicious third parties have also diverted their at-
tention towards exploiting various features of these social networking platforms.
Though, the exploitation methodologies vary according to the features provided
by the social networking platforms, malware infections, spam, and phishing are
the most common security concerns for all of these platforms. In addition, a
number of social botnets have emerged that utilize social networking features to

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 268279, 2013.
c Springer-Verlag Berlin Heidelberg 2013

 Identification of Sybil Communities Generating Context-Aware Spam 269

spreading infections as command and control channels [1], [2], [3]. The root cause
of all these security concerns is the social network sybils or fake accounts created
by malicious users to increase the ecacy of their attacks that are commonly
known as sybil attacks [4]. Generally, an attacker uses multiple fake identities to
unfairly increase its ranking or influence in a target community. Moreover, sev-
eral underground communities exist, which trade sybil accounts with users and
organizations looking for online publicity [5], [6]. Recent studies have shown that
with the increase in the popularity of social media, sybil attacks are becoming
more widespread [7]. Several sybil communities have been reported so far that
forward spam and malwares on Facebook [8] network. In online social networks,
third-party nodes are most vulnerable to sybil attacks, where the third-party
nodes are communities and groups on OSN platforms which bring together users
from dierent real-world communities on the basis of their interests. In case of
Facebook, a third-party node can be defined as a Facebook community page
which is used to connect two users from entirely dierent regions. Sybil accounts
hired for carrying out spam campaigns target such vulnerable nodes. Recently,
the rapid increase in the number of spam on popular online social networking
sites has attracted the attention of researchers from security and related fields.
Though a significant amount of research works has been reported for the
detection and characterization of spam on Facebook and Twitter networks [9],
[10], [11], [12], [13], the existing techniques do not focus on the detection of
coordinated spam campaigns carried out by the communities of sybil accounts.
Similarly, several techniques have been presented for the identification of sybil
communities [4], [14], [15],[16], [17], but all of them focus on the decentralized
detection of sybil accounts. Moreover, the existing techniques are based on two
common assumptions about the behavior of sybil nodes. Firstly, sybil nodes can
form edges between them in a social graph and secondly, the number of edges
connecting sybil and normal nodes is less as compared to the number of edges
connecting either only normal nodes or only sybil nodes. These assumptions
were based on the intuition that normal users do not readily accept friendship
requests from seemingly unknown users. Although empirical studies from [17]
showed existence of such sybil communities in the Tuenti social network, another
study of Renren social network [7] showed that sybil nodes rarely created edges
between themselves. This implies that the community behavior of sybil nodes in a
social graph is mercurial and the assumption that sybil nodes form communities
cannot be generalized [18].
In this work, the authors utilize the rich corpus of prior research works on
spam detection and sybil community identification as a basis and present a
hybrid approach to identify coordinated spam or malware attacks carried out
using sybil accounts. The proposed approach is independent of the assumptions
discussed above by the previous researchers. Although the proposed approach is
generic in nature, this paper focuses on the sybil accounts present on Facebook
social network for experiment and evaluation purposes. The contributions of this
paper can be summarized as follows:
270 F. Ahmed and M. Abulaish

An online social network data collection methodology is introduced which

is based on the intuition that sybil accounts under the control of a single
user tend to attack dierent nodes of the same community; they may not be
connected to each other, but may have a common target.
A new social graph generation mechanism is presented, in which a node
represents a profile and an edge represents an association between a pair of
connecting profiles. The weight of an edge is determined as a function of the
features extracted from the content of linked profiles. In this way, the weight
of an edges is independent of the actual friendship link between the profiles,
and consequently profiles with similar behavior are interlinked together to
form a single group.
Each group of related profiles is modeled as a social graph and analyzed
independently using a community detection algorithm.
A statistical approach is applied on the obtained communities from each
profile group to identify sybil communities.

The rest of the paper is organized as follows. After a brief review of the existing
state-of-the-art techniques for spam identification in online social networks in
Section 2, Section 3 presents a data collection methodology from Facebook social
network. Section 4 presents the profile grouping methodology to generate various
groups of similar profiles in the original social network dataset. This Section also
presents the experimental results obtained from a real dataset and their analyses.
Finally, Section 5 presents conclusions.

2 Related Work

A significant number of research works has been reported in last few years for
spam detection on online social networks. In [19], the authors proposed a real
time URL-spam detection scheme for Twitter. They proposed a browser moni-
toring approach, which takes into account a number of details including HTTP
redirects, web domains contacted while constructing a page, HTML content be-
ing loaded, HTTP headers, and invocation of JavaScript plug-ins. In [11], the
authors created honey-profiles representing dierent age, nationality, and so on.
Their study is based on a dataset collected from the profiles of several regions,
including USA, Middle-East, and Europe. They logged all types of requests,
wall posts, status updates, and private messages on Facebook. Based on the
users activities over social networking sites, they distinguished spam and nor-
mal profiles. The authors in [12] utilized the concept of social honeypot to lure
content polluters on Twitter. The harvested users are analyzed to identify a set
of features for classification purpose. The technique is evaluated on a dataset of
Twitter spammers collected using the @spam mention to flag spammers. In [8],
the authors analyzed a large dataset of wall posts on Facebook user profiles to
detect spam accounts. They built wall posts similarity graph for the detection of
malicious wall posts. Similarly, in [13] the authors presented a thorough analysis
 Identification of Sybil Communities Generating Context-Aware Spam 271

of profile-based and content-based evasion tactics employed by Twitter spam-

mers. The authors proposed a set of 24 features consisting of graph-, neighbor-,
automation-, and timing-based features that are evaluated using dierent ma-
chine learning techniques. In [20] and [10], the authors proposed combination of
content-based and user-based features for the detection of spam profiles on Twit-
ter. In order to evaluate the importance of these features, the collected dataset
is fed into traditional classifiers.
Similar to spam detection on online social networks that has received a lot
of attention from researchers, a significant eort have been diverted towards
the detection of sybil accounts. Initial studies [4], [14], [15], [16], [7] focus on
detecting sybil users. However, individual users do not pose a great threat to
normal users of OSNs. The situation becomes alarming when a large number
of sybil accounts generate a coordinated attack. Several techniques have been
presented to detect groups of accounts coordinating with each other [4], [14],
[15],[16], [17]. All these techniques focus on the decentralized detection of sybil
accounts. Moreover, they are based on two common assumptions: i) sybil nodes
can form edges between them in a social graph, and ii) the number of edges
connecting sybil and normal nodes is less as compared to the number of edges
connecting either only normal nodes or only sybil nodes. However, later studies
have shown that these assumptions cannot be applied in general [7]. Despite the
presence of rich amount of works for spam and sybil detection, there has been
little attention towards the identification of sybil accounts that are particularly
responsible for spam proliferation. Therefore, this paper focuses on the detection
of coordinated spam campaigns that are carried out by sybil accounts under the
control of a single user.

3 Dataset
Based on the analyses reported in [21], it is found that a significant amount of
spam posts on Facebook are directed towards those Facebook pages that are
publicly accessible and any user in the network can post on them. Spammers
generally utilize such openly accessible public pages to spread spam in the net-
work. This type of spam spreading mechanism not only relieves the spammers
from their dependence on friendship requests, but also increases the number of
target users. Once a spam post is visible on a pages wall, it can be visible to
every user who likes that page. In addition, users page-like information can help
spammers to spread context-aware spam through Facebook pages in normal user
communities. Recently, there has been an increasing number of evidence about
the existence of underground communities trading groups of accounts that carry
spam campaigns [18]. Therefore, a group of accounts bought by a party could
be used for a single purpose, resulting in a high correlation in their behavior.
This work exploits the intuition that a spam targeting a community is most
likely the spam generated by a community. A dataset [21] containing Facebook
spam profiles is analyzed to identify Facebook pages that have been mostly
272 F. Ahmed and M. Abulaish

I
634
22
H 6
621 D

24
682
25

15
5
12
20
J C B N K
21 9 14 11
858 655 691 917 395
44

8 8 8 5

L
791 G F E M
4 4 5
688 883 326 431
X X
Z

24
Y Y 7
X = Page ID
Y = No. of Users
A
Z = No. of Common
Users 358

Fig. 1. Graphical illustration of Facebook pages and users

targeted by spammers. As a result, a total number of 14 Facebook pages is found

that are heavily spammed by the spam profiles identified in [21]. All these pages
are accessed to identify active users and to group profiles based on the number of
their common pages-likes. Figure 1 shows a graph in which each node represents
a Facebook page and the weight associated to an edge represents the number of
users commonly shared by the connected pair of nodes. The weights of the edges
can be used to divide the users into various groups based on their interests in
the network. In Figure 1, there are six groups of pages that are close to each
other in terms of their common interests, and each group is treated separately for
detection of profiles that are under the control of a single spammer and generate
context-aware spam towards a community of normal users. Table 1 shows the
various groups and the number of users belonging to each of them. The names
of the groups in Table 1 has been derived from the node levels used in Figure 1.
The next Section explains the methodology used to identify the groups of sybil
accounts.

Table 1. Various profile groups along with the number of users

Groups FEM BGFA BNDIC AFGNK JHLCB DBINKF

n 1 2 3 4 5 6
No. of users 1631 2575 3465 3166 3482 4059
 Identification of Sybil Communities Generating Context-Aware Spam 273

4 Methodology
To detect communities of sybil accounts generating context-aware spam, the rich
amount of textual information contained in each profile is used to generate an
undirected-weighted social graph, in which a node represents a profile and an
edge connecting a pair of nodes represents a link between them. The connections
initiated through a friendship request are independent of the links created in the
actual social graph. A total number of three important features has been used
to determine the weight of an edge in the social graph. For each group of profiles
identified in Section 3, a social graph is generated as G = (Vn , E, W ), where n
represent the group id, Vn is the set of profiles in group n, E V V is the
set of edges, and W is the set of weights assigned to edges. For each node
v Vn , a 3-dimensional feature vector comprising profile similarity, page likes,
and URLs shared is generated, which is then used to calculate the weight of
an edge eij = (vi , vj ). Further details about the features and weight calculation
process are presented in the following Subsections.

4.1 Social Graph Generation

To generate social graph, a set of features has been identified to determine the
weight of an edge highlighting the degree of similarity of the connected profiles.
The following paragraphs present a detailed discussion on the identified features
and edges weight calculation mechanism.
Profile Similarity: The profile similarity of a pair of connected users represents
the degree of match in their posts. This is calculated as a similarity index, Is ,
for each edge eij = (vi , vj ) that connects a pair of nodes. The similarity index
uses vector-space model to represent users posts and applies cosine function to
measure their degree of similarity. The first criteria for two profiles to be similar
is the number of times a profile has posted on its own wall. For example, a profile
vi with a large number of posts as compared to a profile vj with a small number
of posts on their own walls is clearly dissimilar. In this elimination process, the
posts from other profiles on the subject profiles wall are not considered. For two
profiles vi and vj containing x and y number of posts, respectively, a squareness
measure, as shown in Equation 1, is used to determine the eligibility of the
two profiles to be considered for further comparison. In Equation 1, Sij is the
squareness measure of nodes vi and vj , which must be greater than or equal to
4 before considering them for similarity index calculation.

Sij = x/y|x > y (1)

For the nodes vi and vj that fulfil the squareness measure criterion, the similarity
index is calculated as follows. Considering x and y as the number of posts of vi
and vj on their own walls, respectively, a cosine similarity matrix C of dimensions
x y is created in such a way that each post of vi is compared with all the posts
of vj . For cosine similarity, each post is converted into a tf -idf feature vector,
where tf -idf of a term t is calculated using Equations 2 and 3. In these equations,
274 F. Ahmed and M. Abulaish

d is the post under consideration, D is the set of posts present in nodes vi and
vj , and tf (t, d) is calculated as the number of times t appears in d.
tf -idf (t, d, D) = tf (t, d) idf (t, D) (2)

|D|
idf (t, D) = log (3)
|{d D : t d}|
For any two posts a and b with their corresponding tf -idf feature vectors A and
B, the value of an element cij of the matrix C is calculated as a cosine similarity
using Equation 4, where l is the length of feature vectors.
#l
k=1 Ak Bk
cij = 5# 5
#l (4)
l 2 2
k=1 (Ak ) k=1 (B k )

Finally, after smoothing the values of the matrix C using Equation 5, the simi-
larity index for the edge eij = (vi , vj ) is calculated as the normalized cardinality
of the set of non-zero elements in C, as shown in Equation 6.
4
1 if cij > 0.1
cij = (5)
0 if cij < 0.1

|{cij C|cij = 1}|

Is = (6)
xy
Page-Likes: This feature is similar to the feature considered in [21]. However, in
this work, the value of this feature is normalized along the lines of the similarity
index normalization process. This feature captures the page-likes behavior of the
users in a social network. For an edge eij = (vi , vj ), the common page-likes of
vi and vj , Pij , is calculated as a fraction of the page-likes commonly shared by
them, as given in Equation 7. In this equation, Pi and Pj represent the set of
page-likes by nodes vi and vj , respectively.
|Pi Pj |
Pij = (7)
|Pi Pj |

URL Sharing: Like page-likes feature, the value of URL sharing feature of
nodes vi and vj is calculated as the fraction of the URLs commonly shared by
them, as shown in Equation 8. In this equation, Ui and Uj represent the set of
URLs shared by nodes vi and vj , respectively.
|Ui Uj |
Uij = (8)
|Ui Uj |
Based on the values of the features discussed above, the final weight of edge
eij = (vi , vj ), (eij ), is calculated using Equation 9, where 1 , 2 , and 3 are
constants such that each i > 0 and 1 + 2 + 3 = 1.
(eij ) = 1 Is + 2 Pij + 3 Uij (9)
 Identification of Sybil Communities Generating Context-Aware Spam 275

4.2 Community Detection

To identify the communities in a social graph, the proposed approach uses the
Louvain algorithm, which has been implemented as a part of an open source
social network analysis tool Gephi 0.8.1 [22]. It has been widely used for social
network analysis [23]. The algorithm supports community detection in various
types of graphs and provides the flexibility to identify communities at dierent
levels of granularity. It implements a greedy approach for optimizing modular-
ity of a network divisions. The modularity measures the strength/ability of a
network to be divided into groups or communities. Initially, the algorithm op-
timizes the modularity of smaller individual communities, then nodes from the
same communities are added to form a new network in which each node repre-
sents a community. This process is repeated until maximum possible modularity
is obtained. The result is a hierarchy of communities present in the underlying
social graph.

Fig. 2. Community structures in FEM group of profiles

Figure 2 shows a subgraph of the FEM group present in the dataset. The
graph shows 4 major communities out of total 14 communities obtained through
Louvain algorithm. In the experiment, the default resolution values of Louvains
implementation in Gephi has been used. In Figure 2, the legend describes the
percentage of nodes in each community. It can be observed in Figure 2 that
nodes with modularity class 2 are dispersed, whereas nodes of classes 1, 3 and 4
are more closely related. In Section 4.3, the analysis has been further extended
276 F. Ahmed and M. Abulaish

Table 2. Modularity percentages of communities identified for each group of the

dataset

Groups FEM BGFA BNDIC AFGNK JHLCB DBINKF

Class-1 30.09 18.25 17 16.33 10.60 20.92
Class-2 27.59 10.72 12.35 11.37 10.43 11.06
Class-3 24.45 9.28 7.22 11.24 7.96 6.48
Class-4 14.42 6.8 4.7 9.79 6, 15 3.03
Class-5 0.18 0.08 3.17 0.25 5.14 2.81

Table 3. Communities with closeness-centrality values

Groups FEM BGFA BNDIC AFGNK JHLCB DBINKF

Class-1 0.651 0.573 0.421 0.546 0.592 0.589
Class-2 0.549 0.600 0.596 0.548 0.609 0.611
Class-3 0.562 0.620 0.582 0.545 0.644 0.590
Class-4 0.628 0.568 0.592 0.548 0.610 0.570
Class-5 0.621 0.448 0.574 0.451 0.601 0.575

to classify the identified communities as sybils or normal. Table 2 shows details

about the percentage of nodes in communities along with the highest modularity
in each group of the dataset.

4.3 Sybil Community Identification

Once the communities are identified, profiles of each community with the highest
closeness-centrality have been processed separately to classify them either as ma-
licious or benign. Table 3 provides the details about the nodes with highest values
of closeness-centrality. A set of features and JRip rules identified from a locally
crawled dataset have been used to classify the nodes with highest closeness-
centrality as malicious or benign. Table 4 shows the final results obtained after
identifying communities as normal or malicious on the basis of the nodes close-
ness centrality values. After having a close look at the closeness centrality values
and the final results, it can be found that, in most of the cases, the nodes of a
normal user communities have low closeness centrality values. This mainly hap-
pens because the weights assigned to the edges are according to the degree of
similarity among the nodes. A higher similarity between a pair of nodes produces
a higher weight for the edge connecting them in the social graph. Therefore, in
the generated social graph, nodes with high closeness centrality values are similar
to the majority of the nodes in the set, and as a result, a higher weight is as-
signed to all the edges connecting the similar nodes. Moreover, because the sybil
accounts are controlled by a single spammer, they have high similarity among
them as compared to normal users. Hence, nodes belonging to sybil communities
have higher closeness centrality values in comparison to normal users.
 Identification of Sybil Communities Generating Context-Aware Spam 277

Table 4. Communities identified as malicious (M) or benign (B)

Groups FEM BGFA BNDIC AFGNK JHLCB DBINKF

Class-1 M B B B B B
Class-2 B M M M M M
Class-3 B M B B M M
Class-4 M M M M M B
Class-5 M B M B B B

5 Conclusions

Along the lines of the previous research works, this paper has presented a hybrid
approach to detect communities of sybil accounts that are under the control of
spammers and generate context-aware spam towards normal user communities.
The proposed approach is independent of the assumptions made by the previous
eorts and identifies six dierent profiles groups in the dataset based on the
users interests on Facebook network. The users with most common page-likes
have been grouped together for further analysis. Three dierent types of features
have been identified and used to model each group as a social graph in which
profiles are represented as nodes and their links as edges. The weight of a link is
calculated as a function of the degree of similarity of the nodes. Louvain commu-
nity detection algorithm is applied on the social graphs to identify communities
embedded within them. Thereafter, based on the class (malicious or benign)
of the nodes with high closeness-centrality values, the underlying community is
marked either as malicious or benign. The obtained results highlight that gener-
ally nodes with high closeness-centrality values are malicious and belong to sybil
communities, whereas nodes with low closeness-centrality values are benign and
constitute normal user communities.

Acknowledgment. The authors would like to thank King Abdulaziz City

for Science and Technology (KACST) and King Saud University for their sup-
port. This work has been funded by KACST under the NPST project number
11-INF1594-02.

References

1. Boshmaf, Y., Muslukhov, I., Beznosov, K., Ripeanu, M.: Key challenges in defend-
ing against malicious socialbots. In: Proceedings of the 5th USENIX Conference
on Large-scale Exploits and Emergent Threats, LEET, vol. 12 (2012)
2. Nagaraja, S., Houmansadr, A., Piyawongwisal, P., Singh, V., Agarwal, P., Borisov,
N.: Stegobot: A covert social network botnet. In: Filler, T., Pevn
y, T., Craver, S.,
Ker, A. (eds.) IH 2011. LNCS, vol. 6958, pp. 299313. Springer, Heidelberg (2011)
3. Thomas, K., Nicol, D.: The koobface botnet and the rise of social malware. In:
IEEE 2010 5th International Conference on Malicious and Unwanted Software,
MALWARE, pp. 6370 (2010)
278 F. Ahmed and M. Abulaish

4. Yu, H., Kaminsky, M., Gibbons, P., Flaxman, A.: Sybilguard: defending against
sybil attacks via social networks. ACM SIGCOMM Computer Communication Re-
view 36, 267278 (2006)
5. Boyd, S., Ghosh, A., Prabhakar, B., Shah, D.: Gossip algorithms: Design, analysis
and applications. In: Proceedings IEEE 24th Annual Joint Conference of the IEEE
Computer and Communications Societies, INFOCOM 2005, vol. 3, pp. 16531664.
IEEE (2005)
6. Danezis, G., Lesniewski-Laas, C., Frans Kaashoek, M., Anderson, R.: Sybil-
resistant DHT routing. In: De Capitani di Vimercati, S., Syverson, P.F., Goll-
mann, D. (eds.) ESORICS 2005. LNCS, vol. 3679, pp. 305318. Springer, Heidel-
berg (2005)
7. Yang, Z., Wilson, C., Wang, X., Gao, T., Zhao, B., Dai, Y.: Uncovering social
network sybils in the wild. In: Conference on Internet Measurement (2011)
8. Gao, H., Hu, J., Wilson, C., Li, Z., Chen, Y., Zhao, B.: Detecting and characterizing
social spam campaigns. In: Proceedings of the 10th Annual Conference on Internet
Measurement, pp. 3547. ACM (2010)
9. Lee, K., Caverlee, J., Cheng, Z., Sui, D.: Content-driven detection of campaigns in
social media (2011)
10. Jin, X., Lin, C., Luo, J., Han, J.: A data mining-based spam detection system for
social media networks. Proceedings of the VLDB Endowment 4(12) (2011)
11. Stringhini, G., Kruegel, C., Vigna, G.: Detecting spammers on social networks. In:
Proceedings of the 26th Annual Computer Security Applications Conference, pp.
19. ACM (2010)
12. Lee, K., Eo, B., Caverlee, J.: Seven months with the devils: A long-term study
of content polluters on twitter. In: Intl AAAI Conference on Weblogs and Social
Media, ICWSM (2011)
13. Yang, C., Harkreader, R.C., Gu, G.: Die free or live hard? Empirical evaluation and
new design for fighting evolving twitter spammers. In: Sommer, R., Balzarotti, D.,
Maier, G. (eds.) RAID 2011. LNCS, vol. 6961, pp. 318337. Springer, Heidelberg
(2011)
14. Yu, H., Gibbons, P., Kaminsky, M., Xiao, F.: Sybillimit: A near-optimal social net-
work defense against sybil attacks. In: IEEE Symposium on Security and Privacy,
SP 2008, pp. 317. IEEE (2008)
15. Danezis, G., Mittal, P.: Sybilinfer: Detecting sybil nodes using social networks.
NDSS (2009)
16. Tran, N., Min, B., Li, J., Subramanian, L.: Sybil-resilient online content voting.
In: Proceedings of the 6th USENIX Symposium on Networked Systems Design and
Implementation, pp. 1528. USENIX Association (2009)
17. Cao, Q., Sirivianos, M., Yang, X., Pregueiro, T.: Aiding the detection of fake
accounts in large scale social online services. Technical Report (2011),
http://www.cs.duke.edu/~ qiangcao/publications/sybilrank_tr.pdf
18. Wang, G., Mohanlal, M., Wilson, C., Wang, X., Metzger, M., Zheng, H., Zhao, B.:
Social turing tests: Crowdsourcing sybil detection. Arxiv preprint arXiv:1205.3856
(2012)
19. Thomas, K., Grier, C., Ma, J., Paxson, V., Song, D.: Design and evaluation of a
real-time url spam filtering service. In: IEEE Symposium on Security and Privacy
(2011)
 Identification of Sybil Communities Generating Context-Aware Spam 279

20. McCord, M., Chuah, M.: Spam detection on twitter using traditional classifiers. In:
Calero, J.M.A., Yang, L.T., M armol, F.G., Garca Villalba, L.J., Li, A.X., Wang,
Y. (eds.) ATC 2011. LNCS, vol. 6906, pp. 175186. Springer, Heidelberg (2011)
21. Ahmed, F., Abulaish, M.: An mcl-based approach for spam profile detection in on-
line social networks. In: The 11th IEEE International Conference on Trust, Security
and Privacy in Computing and Communications, TrustCom 2012, IEEE (2012)
22. Blondel, V., Guillaume, J., Lambiotte, R., Lefebvre, E.: Fast unfolding of commu-
nities in large networks. Journal of Statistical Mechanics: Theory and Experiment
2008, 10008 (2008)
23. Blondel, V.: The Louvain method for community detection in large networks
(2011), http://perso.uclouvain.be/vincent.blondel/research/louvain.html
(accessed July 11, 2012)
 Location-Based Emerging Event Detection
in Social Networks

Sayan Unankard, Xue Li, and Mohamed A. Sharaf

School of Information Technology and Electrical Engineering,

The University of Queensland, Australia
{uqsunank,m.sharaf}@uq.edu.au, [email protected]

Abstract. This paper proposes a system for the early detection of emerg-
ing events by grouping micro-blog messages into events and using the
message-mentioned locations to identify the locations of events. In our re-
search we correlate user locations with event locations in order to identify
the strong correlations between locations and events that are emerging.
We have evaluated our approach on a real-world Twitter dataset with dif-
ferent granularity of location levels. Our experiments show that the pro-
posed approach can eectively detect the top-k ranked emerging events
with respect to the locations of the users in the dierent granularity of
location scales.

Keywords: Emerging event detection, Location-based social networks,

Micro-blogs.

1 Introduction
The detection of emerging events with respect to the locations and participants
of events provides a better understanding of ongoing events. Emerging events like
natural disasters may need to be reported in real time when they are observed
by many people. Emerging events such as infectious diseases and cyberspace-
initiated/plotted attacks/unrest need to be detected at their early stages.
Currently, a User Generated Content (UGC) system, such as micro-blog ser-
vices, provides a wealth of online discussions about real-world events. Micro-blog,
e.g. Twitter is being considered as a powerful means of communication for people
who are looking for sharing and exchanging information over the social media.
The social events range from popular, widely known ones such as events con-
cerning well-known people or political aairs to local events such as accidents,
protests or natural disasters. Moreover social events could be malicious, danger-
ous, or disastrous. For instance, in August 2011, rioters used instant messaging
and social network services to have arranged meetings and agitated people across
England. Our research interest is to understand where, when, and what is hap-
pening (emerging), so to predict its occurrence via the real-time monitoring of
the social networks. More specifically, our research interest is focused on the
emerging hotspot events. We define that a hotspot event is a tuple (location,
time, topic) that a social network user is associated with through the posting of
a micro-blog.

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 280291, 2013.
c Springer-Verlag Berlin Heidelberg 2013

 Location-Based Emerging Event Detection in Social Networks 281

Table 1. An example of Twitter data

tweet created user user geo-
content
id time id location tagged
2012-08-21 Brisbane, -27.480835, STOP THE CUTS: march on Qld
1 08:26:45 22855027 Australia 153.030392 Parliament. Meet 5pm at King George
Square, Brisbane this Thursday 23rd
2012-08-21 - Stop the Cuts Rally this afternoon
2 08:32:55 19573441 Queensland 5pm King George Square. Ill be there.
http://t.co/UqH3EMOl

In this paper, we developed a Location-based Emerging Event Detection

(LEED) system in social networks. We used this system to detect hotspot events
from micro-blog messages that can help government or organizations prepare for
and response to unexpected events. With the large range of events discussed on
social networks, we do not know how many emerging hotspot events are likely
to happen. Moreover, dealing with a very large number of messages and noisy
data is dicult. The research challenges of our study are: (1) how to eectively
detect events in terms of keywords in micro-blogs? (2) how to detect hotspot
events (i.e. associate the message-mentioned location(s) to an event)? (3) how
do we know a hotspot event is emerging?
Our approach has five stages. Firstly, the pre-processing is performed to im-
prove quality of the dataset. Secondly, we propose a clustering approach to au-
tomatically group the messages into events. Thirdly, we propose a hotspot event
detection method. Fourthly, emerging hotspot event detection is performed. Fi-
nally, we develop a visualization model for representing emerging events. The
experiments are conducted to demonstrate the eectiveness of our approach.
This paper is organised as follows, Section 2 is about micro-blog data and
events. The architecture of proposed system is presented in Section 3. In Section
4, we present the experimental setup and results, Section 5 describes the related
work. The conclusions are given in Section 6.

2 Micro-blog Data and Events

A micro-blog message is a short text message such as a tweet that is restricted to
140 characters and therefore is much concise than a blog post. The twitter data
consists of tweet ID, creating time, user ID, user location, tweet location, text
content and the message-mentioned locations. A tweet location is known when
a user posts the message using a smart phone. An example of Twitter data is
shown in Table 1.
An event is something that occurs in a certain place during a particular in-
terval of time1 . An event location is a place where the event will happen or is
happening while a topic location is a place that is included in the topic content.
In real-world, an event location can also be a topic location, or they can be dif-
ferent to each other. In this research we only consider the event locations. The
topic locations are not considered exclusively, e.g., an earthquake location is
1
http://dictionary.reference.com/browse/event?s=t
282 S. Unankard, X. Li, and M.A. Sharaf

Table 2. Statistics of message locations

List Dataset1 Dataset2

#msgs % #msgs %
Messages with 3,121 1.42 5,055 2.57
geo-tagged location
Messages with user 170,334 77.45 132,661 67.40
profile location
Messages with
IP address based 46,478 21.13 59,118 30.03
location
Messages that
contain geographical 16,875 7.67 19,529 9.92
location in contents
Messages that
contain geographical 744 4.41 1,337 6.85
Fig. 1. A bar chart of micro-blog location in contents
messages location more than 1 location

both the topic as well as the event location. An emerging event is an event
that has significantly increased in the number of messages but rarely posted in
the past. A hotspot event is an event where there is a strong association between
event location and user location. User location is the location where the message
is sent from. Message-referred location is the location mentioned in the message.
It could be event location which is the location where the event occurs or other
locations referred to within the messages.
In order to understand where the locations of messages are found in social
media, we conducted statistical studies. Firstly, we try to understand the avail-
ability of the locations within the micro-blogs. We used Twitter API2 to have
crawled the micro-blog datasets. Dataset1 is crawled from the messages sent by
users around Brisbane Australia, from the dates 17 March 2012 to 25 March
2012 with 219,933 messages while Dataset2 is crawled from the messages sent
by users around the USA, from the dates 21 June 2012 and 27 June 2012 with
196,834 messages. The statistical information of locations from the two datasets
is shown in Figure 1 and Table 2.
As we can see from Datasets 1 and 2 that, the user locations are divided into
three types i.e. the geo-tagged locations, from the user profiles, and implied by IP
addresses (first three rows in the Table 2). The majority of user locations come
from user profiles, approximately 77% of Dataset1 and 67% of Dataset2. Also
from Datasets 1 and 2 we can see that the message-mentioned locations are only
available within a small proportion of the micro-blogs and appeared in one or two
times in average in the micro-blog messages. Our most important observation on
the locations of micro-blogs is that most of the micro-blog messages contain only
one geographical location while messages that contain geographical locations of
more than one constituted approximately 4% of Dataset1 and 7% of Dataset2.
When a location is mentioned in the micro-blog message, it can be either an
event location or a topic location.
In Table 3, we also try to understand what the locations used for in the
messages contents. Dataset3 is downloaded from G. R. Boynton, University of
2
https://dev.twitter.com/docs/api/1.1
 Location-Based Emerging Event Detection in Social Networks 283

Table 3. Message-mentioned locations

Total Messages contain #of times #of times Confidence of
Dataset Event No. mentioned that event that other location mentioned
of msgs location location locations in the content
#of msgs % occur occur as event location
3 H1N1 958 79 8.25 73 6 92.41%
Earthquake 801 798 99.66 846 37 100%
QLD
4 Election 629 87 13.83 86 1 98.85%
2012

Iowa which consists of two events: USA H1N1 and the Indonesia Earthquake3.
Dataset4 is crawled by using hashtag #qldvote for Queensland Election 2012.
As we can see from Datasets 3 and 4 that the message-mentioned locations are
mostly the event locations. The message-mentioned location can appear more
than one time in the message. Our most important observation on the message-
mentioned locations is that the more frequently a message-mentioned location,
the more likely it is the event location. Therefore, the use of message-mentioned
locations to identify event location is the key goal behind this study.

3 Proposed System Architecture

In order to provide a complete coverage of emerging hotspot event detection we
proposed our system which has five stages presented in Figure 2. The following
information provides detail of each stage.

3.1 Data Pre-processing for Events Detection

In particular dealing with micro-blog messages, the message is short and often
noisy. In order to improve the quality of our dataset and the performance of
the subsequent steps, the pre-processing was designed to ignore common words
that carry less important meaning than keywords and remove irrelevant data
e.g re-tweet keyword, web address and message-mentioned username. A micro-
blog loader is developed to collect the Twitter data from public users via the
Twitter API service. The messages are converted into lower case and are removed
by web addresses and a single character word. We also remove the keyword
RT(ReTweet) and the message-mentioned username such as @Sayan. The stop
words are removed and all words are converted into a seed word (stemming word)
by using Lucene 3.1.0 Java API4 . All messages after pre-processing are stored
in the database.

3.2 Text Stream Clustering

The problem that we address in this paper is how to identify events from a given
set of micro-blog messages. We consider a set of messages where each message
is associate with an event. With the large range of events discussed on social
3
http://ir.uiowa.edu/polisci_nmp/2/ and http://ir.uiowa.edu/polisci_nmp/5/
4
http://lucene.apache.org
284 S. Unankard, X. Li, and M.A. Sharaf

Fig. 2. Architecture of LEED system

networks, we do not know the number of events in advance. On the other hand,
our approach requires no prior knowledge of the number of events. Therefore,
the hierarchical clustering is used. In this stage, we aim to automatically group
messages into the same event. We also need a fast and ecient message clustering
system to overcome the problem of the high arrival rate of messages. In order
to dealing with the high incoming rate of messages, we use a sliding window
manager to keep track of messages arriving in the system. The size of the sliding
window can be defined as the number of messages or time interval. In our case,
we use time intervals such as one hour, two hours or one day depending on the
user given preference. Additionally, the number of previous time slots needs to
be specific because it is not necessary to consider the complete usage history of
data to compute an emerging of the event.
In order to find the best representation of term weight for tweet messages, we
compare four dierent term weight formulas (i.e. Term frequency, Augmented
Normalized Term Frequency[13], TFIDF[13], and Smooth-TFIDF[10]). We also
compare four dierent similarity functions (i.e. Jaccard index, Euclidean dis-
tance, Manhattan distance and Cosine similarity) for finding the best similarity
function. To evaluate an eective clustering method, we manually label 16,144
tweets into 17 topics. We evaluate our algorithm by using pair-wise precision,
recall and F1-score. Due to space limit, we do not show our preliminary exper-
iments. The clustering method performs well when using the augmented nor-
malized term frequency and cosine similarity function. Therefore, we calculate
the weight wi,t of term t in message i by using augmented normalized term
frequency:
 Location-Based Emerging Event Detection in Social Networks 285

tfi,t
wi,t = 0.5 + 0.5 (1)
tfimax
where tfi,t is the term frequency value of the term t in message i, tfimax is the
highest term frequency value of the message i. The cosine similarity function is
used to calculate the similarity between the existing cluster and the new message:
#
(wm,ti wc,ti )
cos(m, c) = 5# i 5# (2)
w 2 w 2
j m,tj j c,tj

where m is a message, c is a cluster centroid, and wm,ti is the weight of term ti

in message m.
We use a clustering method called leader-follower clustering [5]. Message clus-
tering is executed when the timestamp of the coming message is greater than the
sliding window size. The system keeps history clusters within previous time slots
to decrease time of computation. Every new message is compared with previous
clusters. The algorithm creates a new cluster for the message if there is no cluster
whose similarity to the message is greater than threshold. We use a centroid repre-
sentation of the cluster because centroid is agglomerative, using the mean, which
trades memory use for speed of clustering. In the final step, we find the most simi-
lar pair of clusters and merge them into the same cluster. In order to find the most
similar cluster, we calculate cosine similarity of two clusters. If the cosine similar-
ity value exceeds the margin threshold, we merge them. The detailed algorithm is
not presented here due to the limitation of the space.

3.3 Hotspot Event Identification

From the previous stage results, all clusters can not be assigned as event clusters
because it can be private conversations, advertisements or others. In this step,
we focus on how to identify the hotspot event clusters. A cluster will be assigned
as the hotspot event if there is strong correlation between the event location and
user location. The tasks are divided into two parts; to find a correlation between
the event location and user location in the cluster and to extract the event topic.

Find a Correlation between the Event Location and User Location:

we calculate the correlation scores between user locations and event locations
for identifying the hotspot event. In order to calculate a correlation score, we
need to extract the user locations and event location first.
Firstly, for user location, we extract from the geo-tag and from the user profile.
The geo-tagged information is generated from smart phone applications while
the location from user profile is a free format text where the user fills in the
user profile. For those users who can post messages from dierent locations, for
a given message we use geo-tagged information to locate the message location
firstly, because it can provide precise location of the user. In order to convert
a latitude/longitude pair into an address, we use Google Maps API5 . If geo-
tagged information is not available we use user profile to query the gazetteer
5
https://developers.google.com/maps/documentation/geocoding/
286 S. Unankard, X. Li, and M.A. Sharaf

database, the database of geographic locations, for getting the location. Finally,
if neither of them is not available we set message location equal to World.
For the gazetteer database, we downloaded a list of geographic locations from
GeoNames6 and stored a local copy of the gazetteer in a database.
Secondly, for event location, we find all terms or phrases reference to geo-
graphic location (e.g. country, state and city) from tweet contents. Since the
location extraction from text is one of the challenging problems of this research
area, this task will be more focused in future work. In this research, we simply
extract the message-mentioned locations via Named Entity Recognition (NER).
We use the Stanford Name Entity Recognizer [7] to identify locations within
the messages. We also use the Part-of-Speech Tagging for Twitter which is in-
troduced in [6] to extract proper nouns. We use extracted terms query into the
gazetteer database to obtain candidate locations of the event. We find the most
possible location of the event using the frequency of each location in the cluster.
The location which has the highest frequency is assigned as event location.
Finally, for finding the correlation between event location and user location,
a correlation score is computed by comparing the level of location granularity.
The granularity level is defined as Country>State>City>PlaceName. We assign
scores for each level if both of them have the same value. The equation is shown
below:
CorrelateScore = 1(F (uCountry, eCountry)) + 2(F (uState, eState))
+3(F (uCity, eCity)) + 4(F (uP lace, eP lace)) (3)
where 1 - 4 are the weight of granularity levels, 1 = 2 = 3 = 4 = 0.25,
uCountry, uState, uCity and uP lace are user location, eCountry, eState, eCity
and eP lace are event location, and F (x, y) = 1; if x = y and the higher granu-
larity level has the same value otherwise F (x, y) = 0.
To identify which cluster is a hotspot event, the LocScore is computed. The
range of LocScore is 0 to 1. It will be used to compute emerging score in the
next section and the top k ranking emerging hotspot events will be selected.
The LocScore of cluster c is defined as:
#
CorrelateScoreu
LocScorec = uU (4)
Np
where N p is the number of users who post messages in cluster c and U represents
users who post messages in cluster c.

Event Topic Extraction: In order to understand what the event cluster is

about, we need to find the set of keywords to represent the event topic. Our in-
tuition is that keywords co-occur when there is a meaningful topical relationship
between them. To extract the set of co-occurring keywords, firstly, we create
a directed, edge-weighted graph. The edge is created if the correlation weight
between the two terms exceeds the threshold. The threshold is defined as the
average of correlation weights in the cluster. We adopt the smoothed correlation
6
http://www.geonames.org
 Location-Based Emerging Event Detection in Social Networks 287

weight function which is introduced in [12], to calculate the semantic correlation

weight between terms. The function is shown below:
(nk,z + nk /N )/(nz nk,z + 1)
ck,z = log( ) (5)
(nk nk,z + nk /N )/(N nk nz + nk,z + 1)
where nk is number of posts containing term k, nz is number of posts containing
term z, nk,z is number of posts containing the terms k and z, and N is the
total number of posts. We identify the event topic by extracting the Strongly
Connected Components (SCCs) from the graph. In the case of SCCs extraction,
if the number of SCCs is more than one sub-graph we calculate the sum of edge
weight on each SCC sub-graph. The SCC which has the highest score is defined
as an event topic.

3.4 Emerging Hotspot Event Detection

As our research interest is to detect emerging hotspot events, all events from pre-
vious stage are not emerging hotspot events. We need to extract only emerging
hotspot events. We find the event that has significantly increased in the amount
of messages but rarely posted in the past. A burst in data is one technique for de-
tecting emerging events. It is calculated by comparing the number of messages in
the current time slot with the mean and the standard deviation of the number of
messages in the previous time slots. Any data point which is higher than the sum
of mean and two standard deviations can be considered as an emerging point.
In order to compute an emerging score for a given events cluster we calculate
the mean and the standard deviation of the number of messages in the previous
time slots. If the number of messages in the current time slot is greater than the
sum of mean and two standard deviations, we calculate the emerging score of
this event. The emerging score of the event e in the current slot is computed by
the following equation:
Ne
EmergScoree = (1 + LocScoree ) (6)
(M eanprev + 2SDprev )
where LocScoree is the location score of event e, Ne is the number of messages
of event e in the current time slot, and M eanprev and SDprev are the mean and
standard deviation of the number of messages in the previous time slots of the
given event, respectively.
To detect emerging hotspot event in dierent location granularity e.g. state
and city. We firstly segment event clusters into users location groups according
to location granularity and follow all of the steps above.

3.5 Visualization
For usability and understanding issues of visualizing model, we use a motion
chart7 to represent emerging hotspot events. Google Map8 is used to represent
7
http://code.google.com/apis/chart/interactive/docs/
gallery/motionchart.html
8
https://developers.google.com/maps/documentation/javascript/examples
288 S. Unankard, X. Li, and M.A. Sharaf

Fig. 3. Example of a motion chart (left) and a geo-map (right)

Table 4. Detection results of KeyGraph and LEED in Country level

Method No. of detected No. of detected Precision Recall
events real-world events
KeyGraph 54 23 0.426 15
Hashtags 949 23 0.024 17
LEED 151 121 0.801 95

the top k emerging hotspot events for a specific location. Examples of a motion
chart and a geo-map are shown in Figure 3. Figure 3 (left) shows a given emerging
hotspot event within a time period (x-axis is the number of users who post
massages; y-axis is the number of messages; colour represents location and bubble
size is emerging score). Figure 3 (right) shows the top five events in a selected
State.

4 Experiments and Evaluation

In order to evaluate our approach, we use search API from Twitter to collect
the messages sent by users around the USA, from the dates 21 June 2012 and 27
June 2012 with 196,834 messages. Since no ground-truth labels are available for
us on realistic events within the data collection period, we manually search local
news from Google to check the events detected by our system. It is impractical
to manually label the overly large number of tweets in the dataset. We follow
the definition of Precision used in [17], which is defined as follows:
detected realworld events
P recision = (7)
total detected events
However, Recall was not defined in [17] because it is not feasible to enumerate
all the real-life events which happened in the dataset. Therefore, we follow the
definition of Recall used in [8], which is defined as the number of distinct realistic
events detected from the dataset. Note that, more than one detected event can
relate to the same real-world event, then they are considered correct in terms of
precision but only one event is considered in counting recall.
Recall = no distinct realworld events (8)
 Location-Based Emerging Event Detection in Social Networks 289

Table 5. Sample of 5 events detected by the KeyGraph approach on 25/06/2012

Detected events Description
1) thunder,spawns,stalls,tropical,storm, Tropical Storm Debby stalls in Gulf of Mexico
tornadoes,threatened
2) park,police,oce Police ocer shot dead at jazz concert in Denver park
3) truck,sell,right Citys red fire trucks to be transformed into billboards
4) township,vehicle,south,unknown Vehicle accident, Hempfield and Manheim Township
5) alarm,calories,brush,burn,acres Brush fire burns 50 acres on Kent Island

Table 6. Sample of 5 events detected by the Hashtags approach on 25/06/2012

Detected events Description
1) #Debby,#debbie,#dc Tropical Storm Debby
2) #nascar,#teamlaurenscartomney,#winning Car racing
3) #phillies,#debby,#philly Tropical Storm Debby
4) #ripmj,#legend,#timeflys Michael Jackson has been dead 3 years today
5) #sussexde,#pasco,#florida Strong thunderstorm heading into SW, Sussex

Table 7. Sample of top 5 events detected by the LEED approach on 25/06/2012. Data
is represented as event topic (event location).
Detected events Description
1) behold, dalla, debbi, editor, storm, tropic A tropical storm named Debbie,
(US>Texas>Dallas) headed for Dallas
2) debbi, florida, gulf, move,slowli, soak, storm, Moving Slowly in Gulf,Tropical Storm
tropic (US>Florida) Debby Soaks Florida (Related to event 1)
3) attack, coast, debbi, flood, florida, Tropical Storm Debby attacks
lead, neighborhood, storm, tornado, touch, tropic, Floridas coast, flooding neighborhoods
weather, latest, listen, newscast (US>Florida) (Related to event 1)
4) concert, denver, oc, park, polic, shot Police ocer shot dead at jazz concert
(US>Colorado>Denver) in Denver
5) accid, buse, circl, injuri, involv, metro, Accident Washington Circle 2 Metro buses
report, washington (US>Washington, D.C.) involved.

We compare the performance of our approach with the KeyGraph approach9

described in [15] and the Semantic Expansion of Hashtags approach described
in [11]. Our proposed approach can eectively detect emerging hotspot events
with a precision of 0.801 which is significantly larger than the baselines. On the
other hand, the correlation between user location and event location can filter
non real-world event clusters out from our system. Our approach can also detect
the larger number of real-world events than the baselines. However, problems
with using a gazetteer are the granularity of locations required and speed of
processing required. For illustration purposes, we present the precision and recall
in Table 4. Examples of detected events are shown in Table 5, 6 and 7.

5 Related Work
5.1 Event Detection in Social Streams
Event detection on micro-blogs as a challenging research topic has been increas-
ingly reported recently [1,3,8,11,15,17]. Sayyadi et al. in [15] developed a new
event detection algorithm by using a keyword graph and community detection
algorithm to discover events from blog posts. Keywords with low document
9
http://keygraph.codeplex.com
290 S. Unankard, X. Li, and M.A. Sharaf

frequency are filtered. An edge is removed if it does not satisfy the conditions.
Communities of keywords are detected by removing the edges with a high be-
tweenness centrality score. However, the number of detected events depends on
the threshold parameters and there was no evaluation conducted. Ozdikis et al.
in [11] proposed an event detection method in Twitter base on clustering of hash-
tags, the # symbol is used to mark keywords or topics in Twitter, and applied a
semantic expansion to message vectors. For each hashtag the most similar three
hashtags are extracted by using cosine similarity. A tweet vector with a single
hashtag is expanded with three similar hashtags and then used in the clustering
process. However, using only messages with a single hashtag can lead to ignore
some important events. Also they did not implement any credibility filter in
order to decide whether a tweet is about an event or not.

5.2 Emerging Topic Detection in Social Networks

A significant amount of research has previously been conducted on emerging topic
detection [2,4,9]. Cataldi et al. presented the emerging topic detection on Twit-
ter [4]. Their work use aging theory to model the life cycle of each term. The im-
portance of Twitter users is studied, which represents an important weighting of
contents. Emerging terms are identified within a given time interval. Terms that
frequently correlated with emerging terms are extracted and are reported together
as emerging topics. However, the system need user-authority for weighting term
which is dicult to collect complete user-network. More recent work is presented
by Alvanaki et al. [2]. Correlation of tag pairs which contains at least one seed
tag are tracked. The seed tags are selected from current sliding window based on
term frequency. Emerging events are identified as top k highest shift scores of tag
pairs. However, the performance is highly sensitive to the selected seed tags and
it is focused on detecting emerging topics rather than emerging events.

5.3 Location-Based Social Network

The geographical scope of social networks content has been studied in the last
decade. TwitterStand has been proposed by Sankaranarayanan et al. in [14]. The
2,000 handpicked users of Twitter are used as seeders who are known to publish
news. The online clustering method is used to group the messages into the news
topic. User location and content location are used to locate geographic content
from each news topic. However, the results relied on handpicked users and there
was no evaluation conducted. Watanabe et al. proposed a real-time local-event
detection system [16]. Local events are detected using geo-tagged from Twitter
data. The place name are extracted from check-in messages such as Im at Time
Square. Watanabe et al. searched for non geo-tagged messages which contained
a distinct place name and allocated a geo-tagged to the location. However, the
results relied on the number of geo-tagged data they had.

6 Conclusions
In this paper, we develop a system namely LEED, to automatically detect
emerging hotspot events over micro-blogs. Our contributions can be summarized
 Location-Based Emerging Event Detection in Social Networks 291

as: (1) We proposed a solution to detect the emerging hotspot events which can
help governments or organizations prepare for and respond to unexpected events
such as natural disasters and protests. (2) We correlate user location with event
location to establish a strong correlation between them. (3) We provide an eval-
uation for the eective event detection on a real-world Twitter dataset with
dierent granularity of location levels. Our experiments show that the proposed
approach can eectively detect emerging events over the baselines. However,
some clusters that we found are not belong to the real-world events. It can be
chat topic. In future work, the dierentiation between event location and topic
location which are mentioned in the text will be studied. Moreover, each com-
ponent may aect the final detection results so a break-down staged evaluation
will be assessed.

References
1. Aggarwal, C., Subbian, K.: Event detection in social streams. In: SDM (2012)
2. Alvanaki, F., Michel, S., Ramamritham, K., Weikum, G.: See whats enblogue:
real-time emergent topic identification in social media. In: EDBT (2012)
3. Becker, H., Naaman, M., Gravano, L.: Beyond trending topics: Real-world event
identification on twitter. In: ICWSM (2011)
4. Cataldi, M., Caro, L.D., Schifanella, C.: Emerging topic detection on twitter based
on temporal and social terms evaluation. In: MDMKDD (2010)
5. Duda, R.O., Hart, P.E.: Pattern Classification and Scene Analysis, 1st edn. John
Wiley & Sons Inc. (February 1973)
6. Gimpel, K., Schneider, N., OConnor, B., Das, D., Mills, D., Eisenstein, J., Heil-
man, M., Yogatama, D., Flanigan, J., Smith, N.A.: Part-of-speech tagging for twit-
ter: Annotation, features, and experiments. In: ACL (2011)
7. Jurafsky, D., Martin, J.H.: Speech and Language Processing: An Introduction to
Natural Language Processing, Computational Linguistics, and Speech Recognition,
vol. 163. Prentice Hall (2000)
8. Li, C., Sun, A., Datta, A.: Twevent: segment-based event detection from tweets.
In: CIKM (2012)
9. Mathioudakis, M., Koudas, N.: Twittermonitor: trend detection over the twitter
stream. In: SIGMOD (2010)
10. Ni, X., Quan, X., Lu, Z., Wenyin, L., Hua, B.: Short text clustering by finding core
terms. Knowl. Inf. Syst. 27(3), 345365 (2011)
11. Ozdikis, O., Senkul, P., Oguztuzun, H.: Semantic expansion of hashtags for en-
hanced event detection in twitter. In: VLDB-WOSS (2012)
12. Ruthven, I., Lalmas, M.: A survey on the use of relevance feedback for information
access systems. Knowl. Eng. Rev. 18(2), 95145 (2003)
13. Salton, G., Buckley, C.: Term-weighting approaches in automatic text retrieval.
Inf. Process. Manage. 24(5), 513523 (1988)
14. Sankaranarayanan, J., Samet, H., Teitler, B.E., Lieberman, M.D., Sperling, J.:
Twitterstand: news in tweets. In: GIS (2009)
15. Sayyadi, H., Hurst, M., Maykov, A.: Event detection and tracking in social streams.
In: ICWSM (2009)
16. Watanabe, K., Ochi, M., Okabe, M., Onai, R.: Jasmine: a real-time local-event
detection system based on geolocation information propagated to microblogs. In:
CIKM (2011)
17. Weng, J., Lee, B.-S.: Event detection in twitter. In: ICWSM (2011)
 Measuring Strength of Ties in Social Network

Dakui Sheng, Tao Sun, Sheng Wang, Ziqi Wang, and Ming Zhang

School of Electronic Engineering and Computer Science

Peking University
Beijing 100871, China
{shengdakui,suntao,wangshengpkucn,wangziqi,mzhang_cs}@pku.edu.cn

Abstract. Measuring the strength of ties has been a fundamental task

for a long time. However, most of previous work treated it as classifying
a tie as strong or weak and were not able to quantitatively estimate the
strength, which limits their scope of contribution. To tackle the problem,
through leveraging user similarities and social interactions, we propose
a latent variable model to calculate a continuous value that measures
the strength. By bringing real users as judge, we demonstrate that the
proposed method can outperform previous methods. Further, we utilize
it to measure the strength of ties among a large set of microblogging
users, and conduct statistical analysis on triads. We find that comparing
to other types of triads, the one with three significantly strong ties are
more likely to be created, which verifies the theory of sociology.

1 Introduction
Granovetter [4] introduced the concept of strength of ties in his landmark paper
The Strength of Weak Ties. Strong ties connect people you really trust, and
those whose social circles tightly overlap with yours. Often, they are also the
people who are most like you while weak ties are merely acquaintances.
Weak ties often provide people with access to information and resources out of
their own circles [3] while strong ties have greater motivation to be of assistance
and are typically more easily available [3]. So the study of social ties can bring
lots of benefits to daily life [10, 5] which is well worth working on. However, in
most studies of social network, social connections are often presented as being
friends or not, neglecting the dierence between strong and weak ties. And the
problem hasnt been well explored yet.
Previous studies have investigated the strength of ties, e.g., by exploiting
human annotation as training data [2, 6]. The problem is that performance
largely depends on the quality of annotation. At the same time, they treat the
problem as a binary classification task, i.e., to classify a relationship as strong or

This work is partially supported by the the National Natural Science Foundation of
China (NSFC Grant No. 61272343), as well as The Specialized Research Fund for
the Doctoral Program of Higher Education of China under Grant (FSSP Grant
No. 20120001110112).

Corresponding author.

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 292300, 2013.
c Springer-Verlag Berlin Heidelberg 2013

 Measuring Strength of Ties in Social Network 293

weak and were not able to quantitatively estimate the strength, which limits the
scope of their contribution in real world. In this paper, we harness the homophily
principles and empirical analysis to develop a latent variable model for estimating
tie strength based on user similarities and social interactions. Comparing to the
existing methods, the proposed method requires no human annotation, and the
outputs are real numbers (instead of 0 or 1) representing the strength of ties.
To evaluate the proposed method, we conduct experiments on data collected
from Sina Weibo1 , and experimental results demonstrate that our model out-
perform previous methods. Further, with the method to calculate tie strength,
we investigate the phenomenon of social triads. In sociology a triad is a group of
three people which is mostly investigated by microsociology. We examine social
triads and our findings are consistent with Granovetters conclusion [3].
The rest part of this paper is organized as follows. We review the related
works in Section 2. In Section 3, we introduce the framework of measuring the
strength of ties in social network. In Section 4, we show the experimental results
of our work. Finally, in Section 5, we conclude and talk about future work.

2 Related Work
Some recent eorts have been made on the task of measuring the strength of
ties. Most of them focus on supervised model which involves eorts on human
annotation [2, 6]. However, measuring the strength of ties is only treated as a
binary classification task by these works, i.e., to classify a relationship as strong
or weak and they were unable to quantitatively estimate the strength between
dierent users. Zhao et al. [13] proposed a novel framework for measuring the
relationship strength between dierent users on various activity fields in online
social networks. However, it only represented the strength of ties on various
activity fields rather than the overall relationship strength.
The most relevant work was proposed by Xiang et al. [12] which is a latent
variable model that measures the strength of ties by outputting continuous values
without human annotation. User similarity and social interactions are exploited
while constructing the model. Our model is the same with it in the aspects
mentioned above. However, there are several dierences: 1)we consider both user
similarities and social interactions as decisive factor of tie strength while they
only considered user similarities. 2)The way in which we utilize social interactions
is rather dierent. In our work, we believe that social interactions directly impact
on the strength of ties and it in turn impacts on the interaction probability of
the user pairs in the future. In the previous work, researchers believe that the
strength of ties impacts on the interaction probability of user pairs all the time.
Experimental results confirm that our model outperforms the previous methods.
At present, there has been work concerning about triads. However these works
either investigated the network structure using social triad theory [11, 1] or

1
http://weibo.com. It is one of the most popular microblog systems in China, which
is similar to twitter.
294 D. Sheng et al.

studied the triads composed of positive and negative edges(friends and ene-
mies) [8] in social media . To the best of our knowledge, our work is the first
that investigates the triads whose edges are measured by strength of ties.

3 Model
One key assumption is the theory of homophily [9]. It postulates that people
tend to form ties with other people who have similar characteristics. Moreover,
it is likely that the stronger the tie, the higher the similarity. Therefore, we can
model the strength of ties as a hidden eect of user similarities.
Another assumption is that the frequency of social interactions between user
pairs directly impact on their strength of ties. Since each user has a finite amount
of resources (e.g., time) to form and maintain relationships, it is likely that they
direct the sources(interacts) towards the ties that they deem more important [12].
As a result, user pairs with strong ties are more likely to interact in the future. On
the other hand, it can be inferred that user pairs with more social interactions in
the past are more likely to form strong ties between them. In this way, we model
the strength of ties as the hidden cause of users future interaction probability
and the hidden eect of the users historical social interactions. Graphical model
representation of the tie strength model is shown in Fig. 1.

s(i,j)

z(i,j)

y1(i,j) y2(i,j) ....... yt(i,j)

Fig. 1. Graphical model representation of the tie strength model

3.1 Model Specification

Formally, let s(i,j) be the attribute vectors of two individuals i and j which rep-
resents the user similarity and social interactions between them in the past. Let
z (i,j) be the hidden strength of ties between i and j. Interactions between them
are divided into t categories(e.g., in weibo system, interactions can be divided
(i,j)
into mention, retweet, comment, reply and so on), and let ym (1 m t)
to be the probability of whether the m th kind of interaction would happen
in the last month. The model can be viewed as a hybrid of discriminative
and generative models: the upper part is discriminative (p(Z|S)), while p(Y |Z)
 Measuring Strength of Ties in Social Network 295

is generative. Our model represents the possible casual relationships among these
variables by modeling the conditional dependencies, so it decomposes as follows:

t
=
P (z (i,j) , y (i,j) |s(i,j) ) = P (z (i,j) |s(i,j) ) (i,j) (i,j)
P (ym |s ) (1)
m=1

The general latent variable model of tie strength can be instantiated in dif-
ferent ways and we adopt the Gaussian distribution to model the conditional
probability of the tie strength given the vectors of the similarity and historical
interactions information. Then the dependency between z (i,j) and s(i,j) is :

P (z (i,j) |s(i,j) ) = N (wT s(i,j) , v) (2)

where s(i,j) is n-dimensional, w is a n-dimensional weight vector to be estimated,

and v is the variance in Gaussian model which is configured to be 0.5.
(i,j)
The probability distribution of each ym given z (i,j) is considered to be con-
(i,j)
ditionally independent. ym (1 m t) is the probability of whether the
(i,j)
m th kind of interaction would happen in the last month. For example, ym
may denote whether i has commented on js statuses in the last month. We use
(i,j)
logistic function to model the conditional probability of ym given z (i,j) :

(i,j) 1
P (ym = 1|z (i,j) ) = (3)
1+ e(m z(i,j) +b)

where m (1 m t) and b are the parameters estimated to adjust the proba-

(i,j)
bility of ym given z (i,j) . b is defined to be zero for simplicity.
The data are represented as N user pair samples, denoted by D = {(i1 , j1 ), (i2 ,
(i,j)
j2 )....(iN , jN )}. During training, ym and s(i,j) can be observed. Given all the
observed variables, based on Eq.(1), the joint probability is as follows:

P (D|w, )P (w, ) = P (y, z, w, |s)

& = t
= '
= (P (z (i,j) |s(i,j) , w) (i,j) (i,j)
P (ym |z , m ) P (w)P ()
(i,j)D m=1
& = "t T (i,j) (i,j)
1 T
s(i,j) z (i,j) )2 e(m z +b) (1 ym ) w wT w (4)
(e 2v (w e 2
(i,j)D m=1
1 + e(m z(i,j) +b) )
t
= '
T
m m
e 2

m=1

Finally, to avoid over-fitting, we put L2 regularizers(i.e., and w ) on the

parameters w and , which can be regarded as Gaussian priors.
z (i,j) , w and are unvisible, so the task can be modified as to find point
estimates w, > z>, > that maximize the likehood P (y, z>, w, >
> |s)
296 D. Sheng et al.

3.2 Inference

Taking the logarithm of Eq.(4), we get the log-likehood:

L = log(P (y, z, w, |s))

" 7 "t
1 T (i,j) ; w T T
= (w s z (i,j) )2 w w m +
2v 2 m=1
2 m (5)
(i,j)D
t
" 7 (i,j) ;
(i,j)
(1 ym )(m z (i,j) + b) log(1 + e(m z +b)
) +C
m=1

Note that both the quadratic terms and the logarithm of logistic function are
concave. Since the sum of concave functions is concave, the function L is concave.
Therefore, a gradient based method will allow us to optimize over the parameters
w, m (m = 1, 2, . . . , t) and the latent variables z (i,j) , (i, j) D, to find the
maximum of L. We derive a coordinate ascent method for the optimization.
The coordinate-wise gradients are:

"t
L 1 T (i,j) (i,j) (i,j) 1
= (w s z ) + (ym )m (6)
z (i,j) v m=1
1 + e m z(i,j) +b)
(

L " 1
(i,j)
= (ym ( (i,j) +b)
)z (i,j) t (7)
m 1 + e mz
(i,j)D

L 1 "
= (z (i,j) wT s(i,j) )s(i,j) w w (8)
m v
(i,j)D

A coordinate ascent optimization scheme will update w, z (i,j) , and m iteratively

until convergence. For z (i,j) and m , we use Newton-Raphson update in each
iteration, where the 2nd order derivatives are given by:
t
= 2 (m T (i,j)
2L 1 m e z +b)

(i,j) 2
= ( (i,j) +b) 2
(9)
(z ) v m=1 (1 + e m z )

2L = T (i,j)
(z (i,j) )2 e(m z +b)
T
= I (10)
t t
(i,j)D
(1 + e(m z(i,j) +b) )2

For w, the root of Eq.(8) can be found analytically as in usual ridge regression:

W = (w I + S T S)1 S T Z (11)

where S = (s(i1 ,j1 ) , s(i2 ,j2 ) ....s(iN ,jN ) )T , and Z = (z (i1 ,j1 ) , z (i2 ,j2 ) ....z (iN ,jN ) )T .
Implementing the model , given y (i,j) and s(i,j) , we can gradually obtain the
parameter w, and z (i,j) ( continuous). So the problem can be solved.
 Measuring Strength of Ties in Social Network 297

Algorithm 1. The process of the Newton-Raphson update

1: while not converged: do
2: for m = 1, 2...t: do
3: update m based on Eq.(7) and Eq.(10).
4: end for
5: for (i, j) D : do
6: update z (i,j) based on Eq.(6) and Eq.(9).
7: end for
8: update W based on Eq.(11).
9: end while
10: return Z, W, .

4 Experiment
4.1 Model Evaluation
The experiments are conducted on a real-world dataset crawled from Sina Weibo
using API2 . Since two-way communication are mostly be observed between bi-
followers3, we focus on ties between bi-followers in the experiments. We recruits
several people to label the category of ties(strong or weak) between them and
their bi-followers. The criterion of human annotation is based on the definition
of strength of ties proposed by Granovetter [4] talked about in Section 1. Then
we crawl data, extract features, and apply our model to the dataset. Meanwhile,
the human annotation results are not used while calculating the strength of ties.
To evaluate the model, we use the continuous value obtained to identify whether
a tie is strong or weak. The accuracy and AUC of the classification task are used
to evaluate the model since there are no other criterion suitable for it. Several
baselines are compared with our model. One is the model proposed by Xiang et
al. [12] referred as UBL model here. One uses social interaction features to classify
while the other exploits both social interaction and user similarity features. They
are referred as UC model and UIC model respectively while our model is referred
as UIBL model. The experiment results are shown in Fig. 2(a).

Models Accuracy AUC

Sports Technology Entertainment
UIBL 0.754 0.638
Nodes 9,798 23,358 27,728
UBL 0.695 0.545
Edges 176,276 811,410 991,759
UIC 0.714 0.609
Triads 688,224 4,906,228 4,759,689
UC 0.697 0.571
(b) Dataset statistics
(a) Evaluation results

Fig. 2. Experiment information

From the result, it can be seen that our model outperforms the others on both
criterions. Despite of the fact that our model and UBL model are quite similar4
2
http://open.weibo.com/.
3
Your bi-followers follow you and you follow them as well.
4
We have discussed about the dierences and common point in Section 2.
298 D. Sheng et al.

to some extent, our model increases the accuracy of it by 6 percent and achieves
much better in AUC mainly for the eect of historical social interactions on tie
strength that has been into account. Although UIC model has exploited all the
features of our model as well as the results of human annotation, it is still beat
by our model on both criterions. We explain it for that theory of homophily
and essence of tie strength are taken into account while proposing our model. In
other words, we exploited the inner relation between features and tie strength
rather than just training based on the features and training set.

4.2 Statistically Analysis on Social Triads

Relations between social media sites often reflect a mixture of friendly and an-
tagonistic interactions [7] while in social network, it often reflects a mixture of
strong and weak ties. Dierent from triads signed in social media [8], we classify
edges of triads based on the strength of ties. The edge(tie) in social triads can
be classified to three categories, strong,weak and atypical5 . Triads with atypical
ties are not considered in experiments. In this way, social triads composed of
strong and weak ties can be classified into four categories shown in Fig. 3.

(a) 3 strong (b) 2 strong & 1 weak (c) 1 strong & 2 weak (d) 3 weak

Fig. 3. Four types of social triads

We use snowball sampling to crawl the data from Sina Weibo. We randomly
pick up several celebrities as seed , utilize snowball sampling to crawl all their bi-
followers and repeat it several times. Since users are more likely to be friends with
people of common interests, the datasets can be seen as communities composed
of celebrities in one field to some extent. By choosing celebrities from Sports,
Technology and Entertainment field as seed respectively, we obtain three datasets
of celebrities in dierent fields which are referred to as dataset Sports, Technology
and Entertainment. Statistics of them are shown in Fig. 2(b).
We conduct several experiments to reveal the phenomena behind social tri-
ads. Significance level() are defined to classify between strong , weak ties and
atypical ties. It is denoted by k%(0 k 100). We sort ties by their strength
in ascending order, and considere that the first k% ties are strong ties while the
last k% ones are weak ties. The others are atypical ties. In this way, the strong
5
Considering that it is rather hard to define the boundary between strong ties and
weak ties, we add another type of tie which is not significant enough to be a strong
tie or weak tie and refer it as atypical tie. They are not considered to be statistically
meaningful here.
 Measuring Strength of Ties in Social Network 299

(a) Sports (b) Technology (c) Entertainment

Fig. 4. The proportions of four triads change with the level of significance, in data set
Sports, Technology and Entertainment, respectively. The horizontal ordinate here is
the significance level.

ties and weak ties we defined here are significantly strong or weak ties when
is small. In the experiments, we gradually take k as 5, 10, 15, 20, 25, 30 for study.
Fig. 4 shows changes of triads in proportion when varies. They appear to be
similar for dierent datasets. When is less than 0.15, there are only triads with
three strong ties since few weak ties make up triads then. When increases from
0.15 to 0.30, triads with three strong ties remains the most. So we can infer that
triads with three strong ties are more likely to form. Meanwhile, the number of
triads with two strong and one weak tie is much smaller which matches with the
theory of sociology that your friends friend are more likely to be your friend.

5 Conclusions and Future Work

In this paper, through leveraging user similarities and social interactions, we

propose a latent variable model to calculate a continuous value that measures
the strength of ties. Our experiments show that using this model to identify
whether a tie is strong or weak gives rise to better performance than baselines.
Further, we utilize our method to measure the strength of ties among a large
set of microblogging users, and conduct statistical analysis on triads. We find
that comparing to other types of triads, the one with three significantly strong
ties are more likely to be created. It is consistent with Granovetters conclusion,
which confirms the eectiveness of our model from an alternative point of view.
In the future, we will consider about applying it to dierent social net-
works,constructing bigger data set and proposing a more eective model.

References

[1] Cartwright, D., Harary, F.: Structural balance: a generalization of heiders theory.
Psychological Review 63(5), 277 (1956)
[2] Gilbert, E., Karahalios, K.: Predicting tie strength with social media. In: Pro-
ceedings of the 27th International Conference on Human Factors in Computing
Systems, pp. 211220. ACM (2009)
[3] Granovetter, M.: The strength of weak ties: A network theory revisited. Sociolog-
ical Theory 1(1), 201233 (1983)
300 D. Sheng et al.

[4] Granovetter, M.: The strength of weak ties. American Journal of Sociology, 1360
1380 (1973)
[5] Harrison, F., Sciberras, J., James, R.: Strength of social tie predicts cooperative
investment in a human social network. PloS One 6(3), e18338 (2011)
[6] Kahanda, I., Neville, J.: Using transactional information to predict link strength
in online social networks. In: Proceedings of the Third International Conference
on Weblogs and Social Media, ICWSM (2009)
[7] Leskovec, J., Huttenlocher, D., Kleinberg, J.: Predicting positive and negative links
in online social networks. In: Proceedings of the 19th International Conference on
World Wide Web, pp. 641650. ACM (2010a)
[8] Leskovec, J., Huttenlocher, D., Kleinberg, J.: Signed networks in social media. In:
Proceedings of the 28th International Conference on Human Factors in Computing
Systems, pp. 13611370. ACM (2010b)
[9] McPherson, M., Smith-Lovin, L., Cook, J.: Birds of a feather: Homophily in social
networks. Annual Review of Sociology, 415444 (2001)
[10] Montgomery, J.: Job search and network composition: Implications of the strength-
of-weak-ties hypothesis. American Sociological Review, 586596 (1992)
[11] Weber, C., Weber, B.: Exploring the antecedents of social liabilities in cvc triadsa
dynamic social network perspective. Journal of Business Venturing 26(2), 255272
(2011)
[12] Xiang, R., Neville, J., Rogati, M.: Modeling relationship strength in online social
networks. In: Proceedings of the 19th International Conference on World Wide
Web, pp. 981990. ACM (2010)
[13] Zhao, X., Li, G., Yuan, J., et al.: Relationship strength estimation for online social
networks with the study on facebook. Neurocomputing (2012)
 Finding Diverse Friends in Social Networks

Syed Khairuzzaman Tanbeer and Carson Kai-Sang Leung

Department of Computer Science, University of Manitoba, Canada

[email protected]

Abstract. Social networks are usually made of users linked by friend-

ship, which can be dependent on (or influenced by) user characteristics
(e.g., connectivity, centrality, weight, importance, activity in the net-
works). Among many friends of these social network users, some friends
are more diverse (e.g., more influential, prominent, and/or active in a
wide range of domains) than other friends in the networks. Recognizing
these diverse friends can provide valuable information for various real-life
applications when analyzing and mining huge volumes of social network
data. In this paper, we propose a tree-based mining algorithm that finds
diverse friends, who are highly influential across multiple domains, in
social networks.

1 Introduction and Related Works

Social networks [9,11] are made of social entities (e.g., users) who are linked by
some specific types of relationships (e.g., friendship, common interest, kinship).
Facebook, Google+, LinkedIn, Twitter and Weibo [13,16] are some examples of
social networks. Within these networks, a user fi usually can create a personal
profile, add other users as friends, endorse their skills/expertise, exchange mes-
sages among friends. These social networks may consist of thousands or millions
of users, and each user fi can have dierent number of friends. Among them,
some are more important or influential than others [2,6,7,15,17].
Over the past few years, several data mining techniques [5,8,12,14] have been
developed to help users extract implicit, previously unknown, and potentially
useful information about the important friends. Recent works on social network
mining include the discovery of strong friends [3] and significant friends [10]
based on the degree of one-to-one interactions (e.g., based on the number of
postings to a friends wall).
However, in some situations, it is also important to discover users who (i) are
influential in the social networks, (ii) have high level of expertise in some do-
mains, and/or (iii) have diverse interest in multiple domains. In other words,
users may want to find important friends based on their influence, prominence,
and/or diversity. For instance, some users may be narrowly interested in one
specific domain (e.g., computers). Other users may be interested in a wide range
of domains (e.g., computers, music, sports), but their expertise level may vary
from one domain to another (e.g., a user fi may be a computer expert but only
a beginner in music).

Corresponding author.

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 301309, 2013.
c Springer-Verlag Berlin Heidelberg 2013

302 S.K. Tanbeer and C.K.-S. Leung

Table 1. Prominence values & lists of interest groups

(b) Lists of interest groups in FSN

(a) Prominence of friends Domain Interest-group list Lj
Friend Prominence Prom(fi ) L1 = {Amy, Bob, Don}
(fi ) Domain D1 Domain D2 Domain D3 D1 L2 = {Cathy, Don}
Amy 0.45 0.60 0.50 L3 = {Amy, Bob}
Bob 0.90 0.70 0.30 L4 = {Bob, Greg}
Cathy 0.20 0.60 0.70 D2 L5 = {Bob, Cathy, Don}
Don 0.30 0.50 0.40 L6 = {Cathy, Ed}
Ed 0.50 0.40 0.45 L7 = {Bob, Cathy, Ed}
F red 0.42 0.24 0.70 L8 = {Amy, Cathy, Ed}
Greg 0.57 0.10 0.20 D3 L9 = {Amy, F red}
L10 = {Amy, Cathy}

In this paper, one of our key contributions is an ecient tree-based algorithm

called Div-growth for mining diverse friends from social networks. Div-growth
takes into account multiple properties (e.g., influence, prominence, and/or diver-
sity) of friends in the networks. Another key contribution is a prefix-tree based
structure called Div-tree for capturing the social network data in a memory-
ecient manner. Once the Div-tree is constructed, Div-growth computes the
diversity of users based on both their influence and prominence to mine diverse
groups of friends.
The remainder of this paper is organized as follows. We introduce the notion
of diverse friends in the next section. Section 3 introduces our Div-growth algo-
rithm, which mines diverse friends from our Div-tree. Experimental results are
reported in Section 4; conclusions are presented in Section 5.

2 Notion of Groups of Diverse Friends

Consider a social network on three dierent domains (domains D1 , D2 , D3 ) and
seven individuals (Amy, Bob, Cathy, Don, Ed, Fred & Greg) with prominence
values in each domain, as shown in Table 1(a). Each domain represents a sub-
category (e.g., sports, arts, education) of interest. The prominence value of
an individual reveals his level of expertise (e.g., importance, weight, value, rep-
utation, belief, position, status, or significance) in a domain. In other words, the
prominence value indicates how important, valued, significant, or well-positioned
the individual is in each domain. The prominence value can be measured by us-
ing a common scale, which could be (i) specified by users or (ii) automatically
calculated based on some user-centric parameters (e.g., connectivity, centrality,
expertise in the domain, years of membership in the domain, degree of involve-
ment in activities in the domain, numbers of involved activities in the domain). In
this paper, the prominence value is normalized into the range (0, 1]. As the same
individual may have dierent levels of expertise in dierent domains, his corre-
sponding prominence value may vary from one domain to another. For example,
prominence value Prom D1 (Amy) of Amy in domain D1 is 0.45, which (i) is dif-
ferent from Prom D2 (Amy)=0.60 and (ii) is higher than Prom D1 (Cathy)=0.20
(implying that Amy is more influential than Cathy in D1 ).
 Finding Diverse Friends in Social Networks 303

Consistent with existing settings of a social network [3,5,10], let F = {f1 , f2 ,

. . . , fm } be a set of individuals/friends in a social network. An interest-group list
L F is a list of individuals who are connected as friends due to some common
interests. Let G = {f1 , f2 , . . . , fk } F be a group of friends (i.e., friend group)
with k friends. Then, Size(G) = k, which represents the number of individuals
in G. A friend network FSN = {L1 , L2 , . . . , Ln } is the set of all n interest-group
lists in the entire social network. These lists belong to some domains, and each
domain contains at least one list. The set of lists in a particular domain D is
called a domain database (denoted as FD ). Here, we assume that there exists an
G
interest-group list in every domain. The projected list FD of G in FD is the set of
lists in FD that contains group G. The frequency Freq D (G) of G in FD indicates
G
the number of lists Lj s in FD , and the frequencies of G in multiple domains are
represented as Freq D1,2,...,d (G) = Freq D1 (G), Freq D2 (G), . . . , Freq Dd (G).

Example 1. Consider FSN shown in Table 1(b), which consists of n=10 interest-group
lists L1 , ..., L10 for m=7 friends in Table 1(a). Each row in the table represents the list
of an interest group. These 10 interest groups are distributed into d=3 domains D1 , D2
and D3 . For instance, FD1 = {L1 , L2 , L3 }. For group G = {Cathy, Ed}, its Size(G)=2.
G G G
As its projected lists on the 3 domains are FD 1
=, FD 2
={L6 , L7 } and FD 3
={L8 }, its
frequencies Freq D1,2,3 (G) = 0, 2, 1. 1

Definition 1. The prominence value Prom D (G) of a friend group G in a single

domain D is defined as the average of all prominence values for all the friends in G:
"Size(G)
Prom (f )
Prom D (G) = i=1 Size(G) D i . Prominence values of G in multiple domains are
represented as Prom D1,2,...,d (G) = Prom D1 (G), Prom D2 (G), . . . , Prom Dd (G). 1

Definition 2. The influence Inf D (G) of a friend group G in a domain D in FD is

defined as the product of the prominence value of G in the domain D and its frequency
in the domain database FD , i.e., Inf D (G) = Prom D (G) Freq D (G). For multiple
domains, Inf D1,2,...,d (G) = Inf D1 (G), Inf D2 (G), . . . , Inf Dd (G). 1

Definition 3. The diversity Div(G) of a friend group G among all d domains in

FSN is defined as the average
"
of all the influence values of G in all domains in the social
d
j=1 Inf D (G)
j
network: Div(G) = d
. 1

Example 2. Revisit FSN in Table 1(b). The prominence value of friend group G =
Prom (Cathy)+Prom (Ed)
{Cathy, Ed} in D1 = D1
Size(G)
D1
= 0.20+0.50
2
= 0.35. We apply similar
0.60+0.40 0.70+0.45
computation and get Prom D1,2,3 (G) = 0.35, 2
, 2
= 0.35, 0.5, 0.575.
Recall from Example 1 that Freq D1,2,3 (G) = 0, 2, 1. So, the overall influence of G in all
3 domains can be calculated as Inf D1,2,3 (G) = 0.350, 0.52, 0.5751 = 0, 1, 0.575.
Thus, the diversity of G in these d=3 domains in FSN is Div(G)= 0+1+0.575 3
=0.525. 1

A group G of friends in a social network FSN is considered diverse if its diversity

value Div (G) the user-specified minimum threshold minDiv, which can be
expressed as an absolute (non-negative real) number or a relative percentage
(with respect to the size of FSN ). Given FSN and minDiv, the research problem
of mining diverse friends from social networks is to find every group G
of friends having Div (G) minDiv.
304 S.K. Tanbeer and C.K.-S. Leung

Example 3. Recall from Example 2 that Div({Cathy, Ed})=0.525. Given (i) FSN in
Table 1(b) and (ii) the user-specified minDiv=0.5, group G={Cathy, Ed} is diverse
because Div(G)=0.525 0.5=minDiv. But, group G ={Ed} is not diverse because
Div(G ) = (0.50)+(0.42)+(0.451)
3
= 0+0.8+0.45
3
= 0.417 < minDiv. 1

3 Our Div-growth Algorithm for Mining Diverse Friends

When mining frequent patterns, the frequency/support measure [1,4] satisfies the
downward closure property (i.e., all supersets of an infrequent patterns are infre-
quent). This helps reduce the search/solution space by pruning infrequent pat-
terns, which in turn speeds up the mining process. However, when mining diverse
friends, diversity does not satisfy the downward closure property. Recall from Ex-
ample 3, group G ={Ed}) is not diverse but its super-group G={Cathy, Ed})
is diverse. As we cannot prune those groups that are not diverse, the mining of
diverse friends can be challenging.
To handle this challenge, for each domain D, we identify the (global) max-
imum prominence value GMProm D among all friends. Then, for each
friend fi , we calculate an upper bound of the influence value Inf UD (fi ) by mul-
tiplying GMProm D (instead of the actual Prom D (fi )) with the corresponding
frequency Freq D (fi ). The upper bound of diversity value Div U (fi ) can then be
computed by using Inf U D (fi ).
Lemma 1. Let G be a group of friends in FSN such that a friend fi G. If Div U (fi ) <
minDiv, then Div(G) must also be less than minDiv. 1

Example 4. Revisit FSN in Table 1(b). Note that GMProm D1 =0.90, GMProm D2 =
0.70, and GMProm D3 =0.70. Recall from Example 3 that Freq D1,2,3 ({Ed})=0, 2, 1.
Then, we can compute Div U ({Ed})= (0.900)+(0.702)+(0.701)
3
=0.7 minDiv. So, we
do not prune {Ed} to avoid missing its super-group {Cathy, Ed}, which is diverse.
Similarly, we can compute Div U ({F red})= (0.900)+(0.700)+(0.701)
3
=0.23 < minDiv.
Due to Lemma 1, we prune Fred as none of its super-groups can be diverse. 1

3.1 Phase 1: Constructing a Div-tree Structure

Given FSN and minDiv, our proposed Div-growth algorithm constructs a Div-
tree as follows. It first scans FSN to calculate Freq Dj (fi ) for each friend fi in
each domain Dj . For each fi , Div-growth then uses GMProm D to compute the
upper bound of the diversity value Div U (fi ), which is used to prune groups of
friends who are not potentially diverse. Every potentially diverse friend fi , along
with its Freq D1,...,d (fi ), is stored in the header table.
Afterwards, Div-growth scans FSN the second time to capture the important
information about potentially diverse friends in a user-defined order in the Div-
tree. Each tree node consists of (i) a friend name and (ii) its frequency counters
for all d domains in the respective path. The basic construction process of a
Div-tree is similar to that of the FP-tree [4]. A key dierence is that, rather
than using only a single frequency counter capturing either the maximum or
average frequency for all domains (which may lead to loss of information), we use
d frequency counters capturing the frequency for all d domains. See Example 5.
 Finding Diverse Friends in Social Networks 305

Fig. 1. Construction of a Div-tree

Example 5. To construct a Div-tree for FSN shown in Table 1(b) when minDiv=0.5,
Div-growth scans FSN to compute (i) GMProm D1,2,3 = 0.9, 0.7, 0.7 for all d=3 do-
mains, (ii) frequencies of all 7 friends in d=3 domains (e.g., Freq D1,2,3 ({Amy}) =
2, 0, 3), (iii) upper bound of diversity values of all 7 friends (e.g., Div U ({Amy}) =
(0.92)+(0.70)+(0.73)
3
= 1.3 using Inf UD1,2,3 ({Amy})). Based on Lemma 1, we safely
remove F red and Greg having Div U ({F red})=0.23 and Div U ({Greg})=0.23 both be-
low minDiv as their super-groups cannot be diverse. So, the header table includes only
the remaining 5 friendssorted in some order (e.g., lexicographical order of friend
names)with their Freq D1,2,3 ({fi }). To facilitate a fast tree traversal, like the FP-
tree, the Div-tree also maintains horizontal node traversal pointers from the header
table to nodes of the same fi .
Div-growth then scans each Lj FSN , removes any friend fi Lj having Div U (fi )
<minDiv, sorts the remaining friends according to the order in the header table, and
inserts the sorted list into the Div-tree. Each tree node captures (i) fi representing the
group G consisting of all friends from the root to fi and (ii) its frequencies in each
domain Freq D1,2,3 (G). For example, the rightmost node Ed:0,1,0 of the Div-tree in
Fig. 1(b) captures G={Cathy, Ed} and Freq D1,2,3 (G)=0, 1, 0. Tree paths of common
prefix (i.e., same friends) are shared, and their corresponding frequencies are added.
See Figs. 1(a) and 1(b) for Div-trees after reading all interest-group lists in domain D1
and the entire FSN , respectively. 1

With this tree construction process, the size of#the Div-tree for FSN with a given
minDiv is observed to be bounded above by Lj FSN |Lj |.

3.2 Phase 2: Mining Diverse Friend Groups

Once the Div-tree is constructed, Div-growth recursively mines diverse friend
groups by building projected and conditional trees in a fashion similar to that
of FP-growth [4].
Recall that Div (G) computed based on Prom D (G) does not satisfy the down-
ward closure property. To facilitate pruning, we use GMProm D (fi ) to com-
pute Div U (fi ), which then satisfies the downward closure property. However,
if Div U (G) was computed as an upper bound to super-group G of fi , then it
may overestimate diversity of G and may lead to false positives. To reduce the
number of false positives, Div-growth uses the local maximum prominence
value LMProm D (G) = maxfi FDG {Prom D (G)} for the projected and condi-
tional trees for G. See Lemma 2 and Example 6.
306 S.K. Tanbeer and C.K.-S. Leung

Fig. 2. Tree-based mining of diverse friend groups

Lemma 2. The diversity value of a friend group G computed based on LMProm D (G)
is a tighter upper bound than Div U (G) computed based on GMProm D . 1

Example 6. To mine potentially diverse friend groups from the Div-tree in Fig. 1(b)
using minDiv =0.5, Div-growth first builds the {Ed}-projected treeas shown in
Fig. 2(a)by extracting the paths Amy, Cathy, Ed:0,0,1, Bob, Cathy, Ed:0,1,0 and
Ed
Cathy, Ed:0,1,0 from the Div-tree in Fig. 1(b). For FD 1,2,3
={Amy, Bob, Cathy, Ed},
Ed
Div-growth also uses LMProm D1,2,3 (FD1,2,3 ) = 0.9, 0.7, 0.7 to compute the tightened
Div U (G) such as tightened Div U ({Amy, Ed})= (0.90)+(0.70)+(0.71)
3
=0.23 < minsup.
As Div U ({Amy, Ed}) and Div U ({Bob, Ed}) are both below minsup, Div-growth
prunes Amy and Bob from the {Ed}-projected tree to get the {Ed}-conditional tree
Ed
as shown in Fig. 2(b). Due to pruning, Div-growth recomputes LMProm D1,2,3 (FD 1,2,3
)
=0.5, 0.6, 0.7 and the tightened Div U ({Cathy, Ed})= (0.50)+(0.62)+(0.71)
3
=0.63 for
Ed
the updated FD 1,2,3
={Cathy, Ed}. This completes the mining for {Ed}.
Next, Div-growth builds {Don}-, {Cathy}- & {Bob}-projected and conditional
trees, from which potentially diverse friend groups can be mined. Finally, Div-growth
computes the true diversity value Div(G) for each of these mined groups to check if it
is truly diverse (i.e., to remove all false positives). 1

4 Experimental Results
To evaluate the eectiveness of our proposed Div-growth algorithm and its as-
sociated Div-tree structure, we compared them with a closely related weighted
frequent pattern mining algorithm called Weight [18] (but it does not use dier-
ent weights for individual items). As Weight was designed for frequent pattern
mining (instead of social network mining), we apply those datasets commonly
used in frequent pattern mining for a fair comparison: (i) IBM synthetic datasets
(e.g., T10I4D100K) and (ii) real datasets (e.g., mushroom, kosarak) from the Fre-
quent Itemset Mining Dataset Repository fimi.cs.helsinki.fi/data. See Ta-
ble 2 for more detail. Items in transactions in these datasets are mapped into
friends in interest-group lists. To reflect the concept of domains, we subdivided
the datasets into several batches. Moreover, a random number in the range (0, 1]
is generated as a prominence value for each friend in every domain.
All programs were written in C++ and run on the Windows XP operating
system with a 2.13 GHz CPU and 1 GB main memory. The runtime specified
indicates the total execution time (i.e., CPU and I/Os). The reported results
 Finding Diverse Friends in Social Networks 307

Table 2. Dataset characteristics

Dataset #transactions #items maxL avgTL Density
mushroom 8,124 119 23 23.0 Dense
T10I4D100K 100,000 870 29 10.1 Sparse
kosarak 990,002 41,270 2498 8.1 Sparse

Fig. 3. Experimental results

are based on the average of multiple runs for each case. We obtained consistent
results for all of these datasets.
Runtime. First, we compared the runtime of Div-growth (which includes the
construction of the Div-tree, the mining of potentially diverse friend groups from
the Div-tree, and the removal of false positives) with that of Weight. Fig. 3(a)
shows the results for a dense dataset (mushroom), which were consistent with
those for sparse datasets (e.g., T10I4D100K). Due to page limitation, we omit the
results for sparse datasets. Runtimes of both algorithms increased when mining
larger datasets (social networks), more batches (domains), and/or with lower
minDiv thresholds. Between the two algorithms, our tree-based Div-growth al-
gorithm outperformed the Apriori-based Weight algorithm. Note that, although
FP-growth [4] is also a tree-based algorithm, it was not design to capture weights.
To avoid distraction, we omit experimental results on FP-growth and only show
those on Weight (which captures weights).
Compactness of the Div-tree. Next, we evaluated the memory consumption.
Fig. 3(b) shows the amount of memory required by our Div-tree for capturing
the content of social networks with the lowest minDiv threshold (i.e., without
removing any friends who were not diverse). Although this simulated the worst-
case scenario for our Div-tree, Div-tree was observed (i) to consume a reasonable
amount of memory and (ii) to require less memory than Weight (because our
Div-tree is compact due to the prefix sharing).
Scalability. Then, we tested the scalability of our Div-growth algorithm by
varying the number of transactions (interest-group lists). We used the kosarak
dataset as it is a huge sparse dataset with a large number of distinct items
(individual users). We divided this dataset into five portions, and each portion is
subdivided into multiple batches (domains). We set minDiv=5% of each portion.
Fig. 3(c) shows that, when the size of the dataset increased, the runtime also
increased proportionally implying that Div-growth is scalable.
308 S.K. Tanbeer and C.K.-S. Leung

Additional Evaluation. So far, we have evaluated the eciency (e.g.,

runtime, compactness or memory consumption, as well as scalability) of our
Div-growth algorithm. Experimental results show that Div-growth is time- and
space-ecient as well as scalable. As ongoing work, we plan to evaluate the
quality (e.g., precision) of Div-growth in finding diverse friend groups. Moreover,
for a fair comparison with Weight, we have used those datasets that are com-
monly used in frequent pattern mining. As ongoing work, we plan to evaluate
Div-growth using real-life social network datasets.

5 Conclusions
In this paper, we (i) introduced a new notion of diverse friends for social net-
works, (ii) proposed a compact tree structure called Div-tree to capture impor-
tant information from social networks, and (iii) designed a tree-based mining
algorithm called Div-growth to find diverse (groups of) friends from social net-
works. Diversity of friends is measured based on their prominence, frequency
and influence in dierent domains on the networks. Although diversity does not
satisfy the downward closure property, we managed to address this issue by us-
ing the global and local maximum prominence values of users as upper bounds.
Experimental results showed that (i) our Div-tree is compact and space-eective
and (ii) our Div-growth algorithm is fast and scalable for both sparse and dense
datasets. As ongoing work, we conduct more extensive experimental evaluation
to measure other aspects (e.g., precision) of our Div-growth algorithm in finding
diverse friends. We also plan to (i) design a more sophisticated way to mea-
sure influence and (ii) incorporate other computational metrics (e.g., popularity,
significance, strength) with prominence into our discovery of useful information
from social networks.

Acknowledgements. This project is partially supported by NSERC (Canada)

and University of Manitoba.

References
1. Agrawal, R., Srikant, R.: Fast algorithms for mining association rules in large
databases. In: VLDB 1994, pp. 487499 (1994)
2. Anagnostopoulos, A., Kumar, R., Mahdian, M.: Influence and correlation in social
networks. In: ACM KDD 2008, pp. 715 (2008)
3. Cameron, J.J., Leung, C.K.-S., Tanbeer, S.K.: Finding strong groups of friends
among friends in social networks. In: IEEE DASC/SCA 2011, pp. 824831 (2011)
4. Han, J., Pei, J., Yin, Y.: Mining frequent patterns without candidate generation.
In: ACM SIGMOD 2000, pp. 112 (2000)
5. Jiang, F., Leung, C.K.-S., Tanbeer, S.K.: Finding popular friends in social net-
works. In: CGC/SCA 2012, pp. 501508. IEEE (2012)
6. Kamath, K.Y., Caverlee, J., Cheng, Z., Sui, D.Z.: Spatial influence vs. community
influence: modeling the global spread of social media. In: ACM CIKM 2012, pp.
962971 (2012)
 Finding Diverse Friends in Social Networks 309

7. Lee, W., Leung, C.K.-S., Song, J.J., Eom, C.S.-H.: A network-flow based influence
propagation model for social networks. In: CGC/SCA 2012, pp. 601608. IEEE
(2012)
8. Lee, W., Song, J.J., Leung, C.K.-S.: Categorical data skyline using classification
tree. In: Du, X., Fan, W., Wang, J., Peng, Z., Sharaf, M.A. (eds.) APWeb 2011.
LNCS, vol. 6612, pp. 181187. Springer, Heidelberg (2011)
9. Leung, C.K.-S., Carmichael, C.L.: Exploring social networks: a frequent pattern
visualization approach. In: IEEE SocialCom 2010, pp. 419424 (2010)
10. Leung, C.K.-S., Tanbeer, S.K.: Mining social networks for significant friend groups.
In: Yu, H., Yu, G., Hsu, W., Moon, Y.-S., Unland, R., Yoo, J. (eds.) DASFAA
Workshops 2012. LNCS, vol. 7240, pp. 180192. Springer, Heidelberg (2012)
11. Peng, Z., Wang, C., Han, L., Hao, J., Ou, X.: Discovering the most potential stars in
social networks with infra-skyline queries. In: Sheng, Q.Z., Wang, G., Jensen, C.S.,
Xu, G. (eds.) APWeb 2012. LNCS, vol. 7235, pp. 134145. Springer, Heidelberg
(2012)
12. Sachan, M., Contractor, D., Faruquie, T.A., Subramaniam, L.V.: Using content
and interactions for discovering communities in social networks. In: ACM WWW
2012, pp. 331340 (2012)
13. Schaal, M., ODonovan, J., Smyth, B.: An analysis of topical proximity in the
twitter social graph. In: Aberer, K., Flache, A., Jager, W., Liu, L., Tang, J., Gueret,
C. (eds.) SocInfo 2012. LNCS, vol. 7710, pp. 232245. Springer, Heidelberg (2012)
14. Sun, Y., Barber, R., Gupta, M., Aggarwal, C.C., Han, J.: Co-author relationship
prediction in heterogeneous bibliographic networks. In: ASONAM 2011, pp. 121
128. IEEE (2011)
15. Tanbeer, S.K., Leung, C.K.-S., Cameron, J.J.: DIFSoN: discovering influential
friends from social networks. In: CASoN 2012, pp. 120125. IEEE (2012)
16. Yang, X., Ghoting, A., Ruan, Y., Parthasarathy, S.: A framework for summarizing
and analyzing twitter feeds. In: ACM KDD 2012, pp. 370378 (2012)
17. Zhang, C., Shou, L., Chen, K., Chen, G., Bei, Y.: Evaluating geo-social influence
in location-based social networks. In: ACM CIKM 2012, pp. 14421451 (2012)
18. Zhang, S., Zhang, C., Yan, X.: Post-mining: maintenance of association rules by
weighting. Information Systems 28(7), 691707 (2003)
 Social Network User Influence Dynamics
Prediction

Jingxuan Li1 , Wei Peng2 , Tao Li1 , and Tong Sun2

1
School of Computer Science, Florida International University
{jli003,taoli}@cs.fiu.edu
2
Xerox Innovation Group, Xerox Corporation
{Wei.Peng,Tong.Sun}@xerox.com

Abstract. Identifying influential users and predicting their network

impact on social networks have attracted tremendous interest from both
academia and industry. Most of the developed algorithms and tools are
mainly dependent on the static network structure instead of the dynamic
diusion process over the network, and are thus essentially based on de-
scriptive models instead of predictive models. In this paper, we propose
a dynamic information propagation model based on Continuous-Time
Markov Process to predict the influence dynamics of social network users,
where the nodes in the propagation sequences are the users, and the
edges connect users who refer to the same topic contiguously on time.
Our proposed model is compared with two baselines, including a well-
known time-series forecasting model Autoregressive Integrated Moving
Average and a widely accepted information diusion model Indepen-
dent cascade. Experimental results validate our ideas and demonstrate
the prediction performance of our proposed algorithms.

Keywords: Social media, Influential users, Dynamic information diu-

sion, Continuous-Time Markov Process.

1 Introduction
1.1 Identifying Influential Users
One of most popular topics of the social network analysis is identifying influen-
tial users and their network impact. More interestingly, knowing the influence of
users and being able to predict it can be leveraged for many applications. The most
famous application to researchers and marketers is the viral marketing [5], which
aims at targeting a group of influential users to maximize the marketing campaign
ROI (Return of Investment). Other interesting applications include search [1], ex-
pertise/tweets recommendation [22], trust/information propagation [9], and cus-
tomer handling prioritization in social customer relationship management.

1.2 Limitations of Current Research Eorts

There is one main limitation of current research eorts on identifying influential
users in social network analysis: lacking of an eective approach for modeling,
predicting, and measuring the influence.

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 310322, 2013.
c Springer-Verlag Berlin Heidelberg 2013

 Social Network User Influence Dynamics Prediction 311

Influence Models and Measures. Currently most applications and tools com-
pute user influence based on their static network properties, such as, the num-
ber of friends/followers in the social graph, the number of posted tweets/received
retweets/mentions/replies in the activity graph, or users centrality (e.g. PageR-
ank, Betweeness-centrality, etc.). All of them make implicit assumptions about
the underlying dynamic process that social network users can only influence their
followers/friends, with equal impact strength, equal infected probabilities, and
at the same propagation rate; or the influence is assumed to take a random walk
on the corresponding static network.
A few works investigate adoption behaviors of social network users as the
dynamic influence propagation 1 or diusion process [18]. The adoption behaviors
refer to some activities or topics (tweets, products, Hashtags, URLs, etc.) shared
among users implicitly and explicitly such as users forwarding a message to their
friends, recommending a product to others, joining some groups with the similar
musical favor, and posting messages about the same topics, etc. According to
the diusion theory, the information cascades from social leaders to followers. In
most diusion models, propagators have certain probabilities to influence their
receivers, and the receivers also have certain thresholds to be influenced. Finding
the social leaders or the users who can maximize the influence coverage in the
network is the major goal of most diusion models. Targeting these influential
users is the idea behind viral marketing, aiming to activate a chain-reaction of
influence, with a small marketing cost.
Some drawbacks of existing social network influence models based on either
static networks or the influence maximization diusion process are: (1) Most
existing models are descriptive models rather than predictive models. For exam-
ple, the number of friends or the centrality score of a given user describes his/her
underlying network connectivity. The number of tweets that a user posted or
get retweeted indicates the trust/insterest that his/her followers have on his/her
tweets. All these measures/models are descriptive and very few models are able
to predict users future influence. (2) Existing influence maximization diusion
process is often modeled by discrete-time models such as Independent Cascade
Model or Linear Threshold Model. Using discrete-time models to model diusion
process in continuous time is very computationally expensive. The time step t
needs to be defined as a very small number. Thus parameter estimation (e.g. ac-
tivation probability) tends to come out very small numbers as well, and diusion
process simulation could take forever to run.

1.3 Content of the Paper

The aforesaid limitation motivates our study on social network user influence
and dynamics prediction in this paper. In particular,
To address the limitation, we propose a dynamic information diusion model
based on Continuous-Time Markov Process (CTMP) to predict the influence dy-
namics of social network users. CTMP assumes that the number of activations
1
We use information/influence propagation, information/influence diusion, and
information cascade, interchangeably to represent the same concept.
312 J. Li et al.

NumberofAdoptions
4

3
Hashtag
URL
2

0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
Day

Fig. 1. The average number of topic adoptions over the time on our Twitter dataset.
The x axis indicates the time, while the y axis denotes the average number of the
summation of the adoption frequency of all topics at the specific time point.

from a given node is following an exponential distribution over the time, which can
be observed in the real-world data [12]. Figure 1 shows that the average number
of topic adoptions decreases exponentially over the time. Hashtags receive more
adoptions compared with URLs, and the number of Hashtag adoptions decreases
more slowly. Furthermore, transition rates q are calculated and treated as the tran-
sition probabilities (or activation probability) of the embedded Markov chain in
CTMP. Then the transition probability P (t) can be computed from q, given any
time t. In this paper, the nodes in the propagation sequences are the users, and the
edges connect users who refer to the same topic contiguously on time. Topics here
particularly refer to Hashtags (expressed as # followed by a word) and short URLs
(e.g. bit.ly) on twitter, which is one of the most popular microblog services, was
launched since July 13, 2006. Hashtags and URLs are both unique identifiers tag-
ging distinct tweets with certain topic labels. We regard the temporal sequences
of Hashtags and URLs as the diusion paths, where the topics are reposted subse-
quently. Although retweeting is not included in our paper as a diusion approach, it
is implicitly considered because the retweets would usually contain the same Hash-
tags and URLs as in the original tweets. Essentially marketers hope by targeting
a small set of influential users not only can their marketing campaign messages
receive many retweets, but also can the influential users create topics about their
products and stimulate a lot of tweets about them continuously. Our experimental
results on a large-scale twitter dataset demonstrates a promising prediction per-
formance on estimating the number of influenced users within a given time.
The contributions of this paper can be summarized as: (1) A dynamic informa-
tion diusion model based on the Continuous-Time Markov Process is proposed
to model and predict the influence coverage of social network users (See Sec-
tion 4). (2) The proposed model is compared with two other baseline models. In
addition, the experiment is conducted over a large-scale twitter dataset, which
validates the capabilities of our model in the real world.

2 Related Work
Some recent research aims at studying the content in the social network, e.g.,
Petrovic et al [16] studied the problem of deciding whether a tweet can be
 Social Network User Influence Dynamics Prediction 313

retweeted, Tsur and Rappoport [23] did content based prediction using hash-
tags provided by microblogging websites. Meanwhile, a number of works have
addressed the matter of user influence on social network. Many of them regard
user influence as their network metrics. Kwak et al. [12] found the dierence
between three influence measures: number of followers, page-rank, and number
of retweets. Cha et al. [4] also compared these three measures, and discovered
that the number of retweets and the number of mentions are correlated well
with each other while the number of friends does not correlated well with the
other two measures. Their hypothesis is that the number of followers of user may
not be a good influence measure. Weng et al. [24] regarded the central users of
each topic-sensitive subnetwork of the follower-and-followee graph as influential
users. Other work such as [7], mined users influence from their static network
properties derived from either their social graphs or activity graphs.
Various dynamic diusion models have also been proposed to discover the influ-
ential users. They are shown to outperform influence models based on static net-
work metrics [17,7]. A lot of work in this direction are devoted to viral marketing.
Domingos and Richardson [5,17] were the first to mine customer network values for
influence maximization for viral marketing in data mining domain. The proposed
approach is a probability optimization method with the hill-climbing heuristics.
Kemper et al. [11] further showed that a natural greedy strategy can achieve 63%
of optimal for two fundamental discrete-time propagation models - Independent
Cascade Model (IC) and Linear Threshold Model (LT) . Many diusion models
assume the influence probabilities on the edges or the probability of acceptance on
the nodes are given or randomly simulated. Goyal et al. [8] proposed to mine these
probabilities by analyzing the past behavior of users. Saito et al. [19,20] extend
IC model and LT model to incorporate asynchronous time delay. Model parame-
ters including activation probabilities and continuous time delay are estimated by
Maximum Likelihood. Our proposed diusion model is dierent from the above
discussed models: (1) We model the dynamic probabilities of edge diusion and
node threshold changing over the time, rather than computing the static probabil-
ities. (2) Our model is a Continuous-Time diusion model instead of a discrete-time
one. Although Saito et al. also proposed Continuous-Time models, the fundamen-
tal diusion process of their models are following LT and IC models. For example,
in asynchronous IC, an active node can only infect one of its neighbors in one itera-
tion, while our proposed models does not assume iterations so that nodes can be ac-
tivated at any time without resetting the clock in the new iteration. Moreover, the
models proposed by Saito et al. supposed only one initial active user and focused on
model parameter estimation, not much on prediction. The experiments are eval-
uated on simulated data from some real network topology. Our proposed model
estimates the model parameters from the real large-scale social network data, al-
lows many initial active users asynchronously or simultaneously to influence other
users, and predicts the real diusion sizes in the future.
In addition, most of influence models are basically descriptive models instead
of predictive models. Bakshy et al. [3] studied the diusion tree of URLs on
twitter, and train a regression tree model on a set of user network attributes,
314 J. Li et al.

user past influence, and URL content to predict users future influence. Our work
is quite dierent from the work of Bakshy et al. in the following aspects : (1) They
predict users average spreading size in the next month based on the data from
the previous month. However, the dynamic nature of word-of-mouth marketing
determines that the influence coverage vary over the time. Thus our work aims
at predicting the spreading size of each individual user within a specific given
date, so we can answer what is the spreading size of user A within 2 hours,
1 day, or 1 month, etc.. (2) Their work is built on top of a regression model,
while our work proposes a real-time stochastic model. The input and output
from these two models are quite dierent. (3) Besides URLs diusion, we also
study the diusion of Hashtags on twitter, which usually have longer lifetime.
Continuous-Time Markov Process (CTMP) has showed its applicability for
various research topics, including simulation, economics, optimal control, genet-
ics, queues, and etc.. The detailed description of CTMP can be found in the
book by Norris [15]. Huang et al [10] adopted it to model the web user visiting
patterns. Liu et al. [13] also utilized CTMP to model user web browsing patterns
for ranking web pages. Song et al. [21] employed CTMP to mine document and
movie browsing patterns for recommendation. To the best of our knowledge,
our work is the first to construct influence diusion model based on CTMP for
spreading coverage prediction and user influence on social networks.

3 Influence Network and Influence Models

3.1 Influence Network
A social graph can be denoted as G(V, E), where V represents social network
users, and E is the set of edges/relations between users. The follower-followee
graph is one type of social graphs, where the edges indicate following relations.
Activity graphs are another type of social graphs, which are extracted from users
tweeting behaviors, e.g., the tweet-retweet graph.
The above networks can be viewed as static networks, which do not demon-
strate the dynamic propagation process over the time. In order to analyze how
topics are passing on social networks progressively, we construct a temporal
influence network by considering the continuous time. Given a Hashtag/URL
(topic), a group of users can be ordered based on the time when they post this
topic. As shown in Figure 2, user i is linked to user j if they post the same topic
contiguously and user j follows/friend with user i. The number on the top of
each arrow is the time taken to transfer a topic from a user to another user.
Definition 1 (temporal influence network). The temporal influence net-
work is G(V, E, T (E)), where V = {V0 , V1 , , Vn } contains all users who posted
at least one Hashtag or URL, E = {Vi Vj |Vi posted a topic earlier than
Vj }, where edges can be constrained to only exist between followers and fol-
lowees or between friends. So the propagation is along the paths from followees
to followers over continuous time. The function T (Vi Vj ) = {t0ij , t1ij , , tlij }.
tm 0 1 l
ij {tij , tij , , tij } is the time dierence between user i posting a topic and
user j posting the same topic.
 Social Network User Influence Dynamics Prediction 315

A B
E H

C D

F
10 70 150 100 700
iPad E A B H D F
100 60 30
Barack Obama G B C A
20 80 100 70 75 40
Earthquake E A D F G B H

Fig. 2. The example of Temporal Influence Network construction

There can be multiple entries in T (Vi Vj ) since user i and user j can post
the same set of topics or one topic at multiple times. Note that we aggregate all
topics together to form this temporal influencer network in this paper.

4 Information Diusion Model Based on

Continuous-Time Markov Process
The descriptive influence models answer questions such as How many followers
that user A has? and How many followers post the topic ipad after user A?,
etc. In this section, we introduce our proposed predictive Information Diusion
Model based on Continuous-Time Markov Process, abbreviated as IDM-CTMP
for convenience. IDM-CTMP is able to answer the following question, In the
next month, how many users would post the topic ipad estimably if user A
posts it now., or even a harder question In order to make a maximal number
of people to talk about our product in the next week, who are the seed users we
should target?. Note the influential users discovered by IDM-CTMP maximize
not only the information coverage, but also the rate of information cascade given
a certain period of time.

4.1 Model Formulation

A trending topic (a Hashtag/URL) is propagated by social network users within
the temporal influence network defined in Definition 1. Suppose X(t) denotes
the user who posts a specific topic at time point t, X = X(t), t 0 forms a
Continuous-Time Markov Process (CTMP) [2], in which the user who will discuss
this topic next only depends on the current user given the whole history of the
topic propagation. Formally, this markov property can be defined by:
Pij (t) = P {X(t + ) = j|X() = i, X() = x(), 0 < } = P {X(t + ) = j|X() = i},
(1)
where Pij (t) is the transition probability from i to j within time t, i is the
current user who discusses the trending topic, j is the next user who posts the
topic following i, and x() denotes the history of the topic propagation before
316 J. Li et al.

the time point . We assume that the transition probability Pij (t) does not
depend on the actual starting time of the propagation process, thus the CTMP
is time-homogeneous:
Pij (t) = P {X(t + ) = j|X() = i} = P {X(t) = j|X(0) = i}. (2)
In order to estimate the diusion size of user i given a pre-defined time window
t, we need to compute the transition probability from user i to all the others,
then determine the number of users being aected by i at the end of t. The
diusion size of user i over t based on CTMP can be defined as
#
DSi,t = Pij (t) ni , (3)
j

where ni is the number of times that user i occurs at time t. It can be estimated
by supposing that it linearly increments on t. However, it is impractical to esti-
mate the transition probability matrix P (t) with infinite possible t. Thus instead
of estimating P (t) directly, we calculate the transition rate matrix Q, and then
P (t) can be estimated from Q.
Before jumping into the details of Q matrix and P (t) matrix, a simple example
is presented in Figure 3 to illustrate the procedure of predicting the diusion
size of a particular user.































































 



















































Fig. 3. An example to show the application of our model over a simple dataset which
composed of only three users and two topics. Based on the two topic transitions in the
dataset, we can calculate the Q matrix. Then, the P(t) matrix can be obtained via the
assistance of Q matrix, assuming the predicting time is t. Once P(t) matrix is ready,
the diusion size can be predicted.

4.2 Estimation of Transition Rate Matrix

The transition rate matrix Q is also called the infinitesimal generator of the
Continuous-Time Markov Process. It is defined as the derivative of P (t) when t
goes to 0. The entry qij is the transition rate to propagate
# a topic from user i
to user j. The sum of the rows in Q is zero, with j,j=i qij = qii .
Pij (t)
qij = lim = Pij (0) (i = j). (4)
t0 t
 Social Network User Influence Dynamics Prediction 317

Note that qij reflects a change in the transition probability from user i to user
j. qi , namely out-user transition rate in this paper, is equal to qii . It indicates
the rate of user i propagating topics to any other users. As shown in Figure 1,
the average number of topic adoptions decreases exponentially over the time.
Thus, in order to compute qi , we assume that the time for user i to propagate a
topic to all the other users is following an exponential distribution as observed
for many users in our data, where the rate parameter is qi . It is known that the
expected value of an exponentially distributed random variable Ti (in this case,
the topic propagation time for user i) with rate parameter qi is given by:
1
E[Ti ] = . (5)
qi

Thus qi is one divided by the mean of j (T (Vi Vj )), which is defined in the
temporal influence network.
According to the theory of Continuous-Time Markov Process, if a propagation
occurs on user i, the probability that the other user j would post the #
topic forms
an embedded Markov chain. The transition probability is Sij , and j Sij = 1
(i = j) and Sii = 0. One important property is that qij = qi Sij . Then, the
transition rate from user i to j can be estimated by:
#
qij = qi2 exp(qi tm
ij ), (6)
m

where m is the number of topics diused from use i to j, and tm

ij denotes the
transition time from user i to j on the m-th topic.

4.3 Estimation of Transition Probability Matrix

Now we obtain all the entries of the transition rate matrix Q. Next, we will
specify how to derive the transition probability matrix P (t). The well accepted
Kolmogorovs Backward Equations [6] in the Continuous-Time Markov Process
can be utilized:
#
Pij (t) = qi Pik (t) Pkj (t) qi Pij (t). (7)
i=k

By performing some algebra operations, the above equation can be written as

the following matrix form:
P (t) = QP (t). (8)
The general solution for this equation is given by:
P (t) = eQt . (9)

P (t) is a stochastic, irreducible matrix for any time t. We approximate it using

Taylor expansion, so that P (t) can be estimated by [10]:
P (t) = eQt = lim (I + Qt/n)n . (10)
n

We raise the power of (I + Qt/n) to a suciently large n.

318 J. Li et al.

5 Experiment

5.1 The Dataset Description

We continuously collected 22-day twitter data, ranging from March 2 to March

24, 2011, using Twitter Gardenhose streaming API. The first 12-day data is
used for our training purpose, and the remaining 10-day data is for testing
and validation. We removed all non-English tweets and those tweets posted by
users with less than 20 followers (authors of these tweets are less likely to be
influential users.) Notice that the data is represented by tweet records, one of
which may include the tweet and the corresponding retweets, thus more than one
tweet. Then there are 48,113,490 tweet records in the 12-day training data and
27,237,631 tweet records in the 10-day testing data (tweets without Hashtags or
URLs are also removed from the testing data). After, those tweets with more
than three Hashtags or URLs are filtered, since they tend to be spam ones as
introduced in [12]. Finally, we obtained totally 78,858,046 tweets, in which there
are 9,431,404 unique users, 3,209,330 Hashtags, and 21,107,164 URLs.

5.2 Predicting Spreading Size Using IDM-CTMP

This paper aims at modeling, predicting, and measuring users influence in the
social network. Hence, to validate the dynamic metric and the predictive model
IDM-CTMP, instead of focusing on identifying influential users like the previous
work, we mainly evaluate its capability of predicting how many users would
adopt one topic given a certain period of time after a user posts it. Specifically,
in our experiment, we first train IDM-CTMP model on the first 12-day training
data, then calculate the spreading coverage of each user for each day from day
1 to day 22, including both training and testing periods.

The Ground Truth. In order to evaluate the prediction performance of IDM-

CTMP and present its feasibility in real world applications, we need to provide
the ground truth. Suppose user u posts a topic at time t1 , and subsequently n
users post the same topic till time t2 , then the ground-truth spreading size of
t1 t2
u from t1 to t2 with regards to topic , denoted as DSu, , is n. For example,
we know that user B first posts a topic #ipad in day 2, afterwards there are 10
users posting #ipad in day 2, and 20 users posting #ipad in day 3. Then the
ground-truth spreading size of B is 0 for day 1, 10 for day 2, and 20 for day 3.
After a users spreading sizes over dierent topics in a particular period of
time are computed, we can obtain the average spreading size over all topics in
that time period by dividing the number of involved topics:
! t1 t2
DSu,
DSut1 t2 =
, (11)
#( )

where DS is the spreading size, u denotes a user, t1 t2 indicates a time window,

and is a topic.
 Social Network User Influence Dynamics Prediction 319

Baselines. To our best knowledge, this is the first attempt to predict the
continuous-time spreading coverage of social network users. Therefore, we em-
ploy the Autoregressive Integrated Moving Average (ARIMA) model [14], which
is widely used for fitting and forecasting the time series data in the area of statis-
tics and econometrics, as one baseline. This model can first fit to time series data
(in our case, a users spreading sizes of dierent days in the history), then pre-
dict this users spreading size in the future. Thus, the spreading sizes of first 12
days are used to build the ARIMA model. Then, it predicts the entire 22 days.
Note that the optimal ARIMA is always selected based on Akaike information
criterion (AIC) and Bayesian information criterion (BIC) for comparison.
In addition to ARIMA, one of the basic information diusion models Inde-
pendent Cascade (IC) [11] is used as the second baseline. In the IC model, a user
u who mentions a topic in his/her tweets at the current time step t is treated
as a new activated user. An activated user has one chance to activate each of
his/her neighbors (i.e., make them adopt this topic) with a certain probability.
If a neighbor v posts the same topic after time step t, it is said that v becomes
active at time step t + 1. If v becomes active at time step t + 1, u cannot activate
v in the subsequent rounds.
In order to apply the IC model to calculate users spreading sizes, the activa-
tion probability for every pair of users needs to be estimated. Specifically, for a
user u, we first obtain the spreading size of each of his/her topics during the first
12 days, thus we can get average spreading size over all of his/her topics. Then,
the daily average spreading size (DDS) is computed from dividing the average
spreading size by 12 days. Finally, 1/DDS is taken as the activation probability
of u and each of his/her neighbors.

Prediction. To compare the performance of IDM-CTMP, we choose 10, 000

top users computed by IDM-CTMP given the entire 22 days and 10, 000 random
users. Three well-known metrics for measuring prediction accuracy are utilized
in our experiment for evaluation: MAE (Mean Absolute Error), RMSE (Root
Mean Square Error), and MASE (Mean Absolute Scaled Error). The average
values of three metrics for IDM-CTMP, ARIMA, and IC are listed in Table 1
and Table 2. It can be seen that our proposed method IDM-CTMP performs
better than baseline methods ARIMA and IC, because ARIMA fits the overall
trend of the time series data and does not consider the underlying network
cascading causing the change of the spreading sizes. The basic IC model needs
predefined time step, which is set to be 1 day. It might be too large to capture
the real-time topic propagation. However, if setting it to be small, it would take
long time to run. The parameter estimation of the basic IC model assumes the
constant activation probability for all neighbors, which could be another reason
of poor performance.
In Figure 4, we plot the ground truth spreading sizes and the predicted spread-
ing sizes, using IDM-CTMP, ARIMA, and IC for both top 5 users and 5 random
users. We can observe that even though the predicted results by IDM-CTMP are
320 J. Li et al.

Table 1. The comparison over the 10,000 Table 2. The comparison over the 10,000
top users random users
Methods MAE RMSE MASE Methods MAE RMSE MASE
IDM-CTMP 3.290 4.231 0.714 IDM-CTMP 1.686 2.055 0.702
ARIMA 4.369 5.470 1.294 ARIMA 2.026 2.855 0.764
IC 5.858 7.209 2.355 IC 3.928 4.834 2.091

Top user 1 Random user 1

100 50
80 40
60 30
Size

Size
40 20
20 10
0 0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
Day Day
Top user 2 Random user 2
100 50
80 40
60 30
Size

40 Size 20
20 10
0 0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
Day Day
Top user 3 Random user 3
100 50
80 40
60 30
Size

Size

40 20
20 10
0 0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
Day Day
Top user 4 Random user 4
100 50
80 40
60 30
Size

Size

40 20
20 10
0 0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
Day Day Ground Truth
Top user 5 Random user 5 IDMCTMP
100 50 ARIMA
80 40
IC
60 30
Size

Size

40 20
20 10
0 0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
Day Day

Fig. 4. The comparison between the predicted spreading size of top ranked 5 users (left
side) and randomly picked 5 users (right side) by IDM-CTMP and baseline against the
ground truth

not exactly as same as the ground truth, most of the predicted curves fit very
close to the true curves. In particular, most of the peaks and valleys can
be well captured by our proposed method. However, ARIMA and IC does not
perform well, missing many peaks and valleys and having wrong predictions.
In addition to the prediction capability, we also conduct experiments to com-
pare the running time of three models. It turns out that IDM-CTMP spends
similar running time as IC model for prediction tasks, while ARIMA requires
the least running time. However, to achieve the best performance of ARIMA,
a lot of parameter tuning work needs to be done beforehand. Considering this,
ARIMA cannot be the best choice in terms of the running time.
 Social Network User Influence Dynamics Prediction 321

6 Conclusion

In this paper, we propose IDM-CTMP, an information diusion model based

on Continuous-Time Markov Process. IDM-CTMP is able to predict the influ-
ence dynamics of social network users, i.e., it can predict the spreading cover-
age of a user within a given period of time. Our experiment results show that
IDM-CTMP achieves very promising performance as its predicted spreading size
demonstrated can fit closely to the ground truth.

Acknowledgements. The work is partially supported by a Xerox Universit

Aair Committee (UAC) Award and by US National Science Foundation under
grants IIS-0546280, CCF-0830659, HRD-0833093, and IIS-1213026.

References
1. Adamic, L.A., Adar, E.: How to search a social network. Social Networks 27, 2005
(2005)
2. Anderson, W.J., James, W.: Continuous-time Markov chains: An applications-
oriented approach, vol. 7. Springer, New York (1991)
3. Bakshy, E., Hofman, J.M., Mason, W.A., Watts, D.J.: Everyones an influencer:
quantifying influence on twitter. In: WSDM, pp. 6574 (2011)
4. Cha, M., Haddadi, H., Benevenuto, F., Gummadi, K.P.: Measuring user influence
in twitter: The million follower fallacy. In: AAAI, ICWSM (2010)
5. Domingos, P., Richardson, M.: Mining the network value of customers. In:
SIGKDD, pp. 5766. ACM (2001)
6. Gardiner, C.W.: Handbook of stochastic methods. Springer, Berlin (1985)
7. Ghosh, R., Lerman, K.: Predicting influential users in online social networks.
CoRR, abs/1005.4882 (2010)
8. Goyal, A., Bonchi, F., Lakshmanan, L.V.S.: Learning influence probabilities in
social networks. In: WSDM, pp. 241250 (2010)
9. Gruhl, D., Guha, R., Liben-Nowell, D., Tomkins, A.: Information diusion through
blogspace. In: WWW, pp. 491501. ACM (2004)
10. Huang, Q., Yang, Q., Huang, J.Z., Ng, M.K.: Mining of web-page visiting patterns
with continuous-time markov models. In: Dai, H., Srikant, R., Zhang, C. (eds.)
PAKDD 2004. LNCS (LNAI), vol. 3056, pp. 549558. Springer, Heidelberg (2004)
11. Kempe, D., Kleinberg, J., Tardos, E.: Maximizing the spread of influence through
a social network. In: SIGKDD, pp. 137146 (2003)
12. Kwak, H., Lee, C., Park, H., Moon, S.: What is twitter, a social network or a news
media? In: WWW, pp. 591600 (2010)
13. Liu, Y., Gao, B., Liu, T.Y., Zhang, Y., Ma, Z., He, S., Li, H.: Browserank: letting
web users vote for page importance. In: SIGIR, pp. 451458 (2008)
14. Mills, T.C.: Time series techniques for economists. Cambridge Univ. Pr. (1991)
15. Norris, J.R.: Markov chains. Number 2008. Cambridge University Press (1998)
16. Petrovic, S., Osborne, M., Lavrenko, V.: Rt to win! predicting message propagation
in twitter. In: 5th ICWSM (2011)
17. Richardson, M., Domingos, P.: Mining knowledge-sharing sites for viral marketing.
In: SIGKDD, pp. 6170 (2002)
18. Rogers, E.M.: Diusion of Innovations, vol. 27. Free Press (2003)
322 J. Li et al.

19. Saito, K., Kimura, M., Ohara, K., Motoda, H.: Ecient estimation of cumulative
influence for multiple activation information diusion model with continuous time
delay. In: Zhang, B.-T., Orgun, M.A. (eds.) PRICAI 2010. LNCS, vol. 6230, pp.
244255. Springer, Heidelberg (2010)
20. Saito, K., Kimura, M., Ohara, K., Motoda, H.: Generative models of information
diusion with asynchronous timedelay. JMLR - Proceedings Track 13, 193208
(2010)
21. Song, X., Chi, Y., Hino, K., Tseng, B.L.: Information flow modeling based on
diusion rate for prediction and ranking. In: WWW, pp. 191200 (2007)
22. Song, X., Tseng, B.L., Lin, C.-Y., Sun, M.-T.: Personalized recommendation driven
by information flow. In: SIGIR, pp. 509516 (2006)
23. Tsur, O., Rappoport, A.: Whats in a hashtag?: content based prediction of the
spread of ideas in microblogging communities. In: Proceedings of the Fifth ACM
International Conference on Web Search and Data Mining. ACM (2012)
24. Weng, J., Lim, E.P., Jiang, J., He, Q.: Twitterrank: finding topic-sensitive influen-
tial twitterers. In: WSDM, pp. 261270 (2010)
 Credibility-Based Twitter Social Network Analysis

Jebrin Al-Sharawneh1, Suku Sinnappan2, and Mary-Anne Williams1

1
Centre for Quantum Computation & Intelligent Systems (QCIS),
University of Technology Sydney, Sydney, Australia
[email protected], [email protected]
2
Swinburne University, Australia
[email protected]

Abstract. Social Network (SN) members in Twitter communicate in varied

contexts such as crisis. Formations within social networks are unique as some
members have more influence over other members; members with more
influence are known as leaders or pioneers. Finding leaders in a social network is
the central factor in information dissemination; therefore, the main objective of
this paper is to utilize social network information to identify leaders in Twitter
during crisis situations. We propose a new approach to identify social network
leaders based on their credibility which is drawn from their trustworthiness and
expertise utilizing the Follow the Leader model based on the content quality
generated by SN members. Our approach is highly innovative and the empirical
results demonstrate that our approach can identify SN community leaders in a
specific context (crisis) and has significant practical benefits.

Keywords: social network, Twitter, crisis, leaders, credibility, follow the

leader, content quality.

1 Introduction
The immediacy of social networks play an excellent host to various social networks
formed at times of crisis. Examining complex Twitter communication can be
challenging; however, it is known that in Social Networks (SNs), some members have
more influence over other members; they are known as leaders or pioneers [6, 13].
Finding a credible member within an SN is a challenging issue, therefore, the main
objective of this paper is to utilize social network information to identify members
roles as Leader or Follower in Twitter Social Network Service (SNS), based on their
credibility and to assess the impact of Leaders in crisis situations.
Credibility refers to the objective and subjective components of the believability of
an agent. A credible SN member is defined as one who has performed consistently,
accurately, and its generated content is useful over a period of time in a specific
context. The credibility of an agent refers to its quality of being believable or
trustworthy and can be measured by its trustworthiness, expertise, and dynamism [8].
In SNSs members interact in varied contexts. We consider Victorian bushfires [11]
as the context of the interaction between Twitter members, where the worst day
known as Black Saturday when 173 people died.

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 323331, 2013.
Springer-Verlag Berlin Heidelberg 2013
324 J. Al-Sharawneh, S. Sinnappan, and M.-A. Williams

1.1 Motivations and Contributions

Identifying natural leaders in large SNSs such as Twitter is a challenging task.
Making assumptions from the number of followers and the number of tweets on
Twitter profiles can be deceiving and may not be accurate as both are cumulative
figures. Communication in Twitter is best analysed and understood in a context [11].
The Goldbaums Follow the Leader model [6] provides us with insights to identify
members roles in the SNS whether they are leaders or followers. Enriching the
Follow the Leader model with trust provides us with the potential to analyse SNSs
based on members credibility. Member credibility level is a measure that reflects
their trustworthiness and expertise, and provides the means to classify members in a
specific context; some members can be classified either leaders or followers.
Our key contributions in this paper as follows:
A credibility based approach that identifies SNS members roles.
An answer to what extent is leaders prominent and influential members in SNS.
An answer to what extent new tweets spread during crisis based on their quality.
The rest of this paper is organized as follows: Section 2 provides a brief overview of
some related research work followed by the proposed SN members credibility based
model in Section 3. Sections 4 and 5 report, analyse and evaluate the empirical study
of the proposed Follow the Leader model.

2 Related Works

2.1 Social Networks and Trust

In SNSs, members can express how much they trust other members through the
adoption of their postings [14]. Since most SN participants do not know each other
in real life, the trust inference mechanism is used when participants want to
establish a new trust relationship or measure trust values between connected
members. The connection between member similarity and trust was established in
[14]. They demonstrated that there is a significant correlation between the trust
expressed by the members and their similarity. The more similar two people are, the
greater the trust between them. Similarity can be interpreted as similarity in
interests or opinions.

2.2 Follow the Leader

Social psychology theory [10] points out that a persons role in a specific context has a
significant influence on trust evaluation if that person refers a person or object. The
Follow the Leader phenomenon in dynamic social networks [6] can be used as a model
of opinion formation with dynamic confidence in social networks; in which the
profiling of agents as leaders or followers is possible. Goldbaum [6] defines the leader
as somebody who possesses better knowledge and information than the general public
in a specific context.
 Credibility-Based Twitter Social Network Analysis 325

3 SNS Members Credibility Model

In SNS, the number and types of links that members have are keys to determining
their importance in the network, which constrains the range of opportunities,
influence and power that they have [7]. A Twitter member credibility reflects their
quality of being believable or trustworthy and can be measured by their
trustworthiness, expertise, and dynamism [8]. We adopt the credibility computation
from our previous work [2] with variation in the expertise computation.

3.1 SNS Members Trustworthiness

Trustworthiness is an attribute of individual exchange partners; a trustworthy member
is a member in which others place their trust. SNS members gain trustworthiness from
their direct followers and indirect followers using the trust transitivity feature of trust;
friends of friends who trust their nearest friend and also trust their friends next
friend, yield a trust score which is the product of the two scores.
Direct and indirect followers (trusters) of a member provide implicit trust
indicating their trust in that member. Members similarity in their opinions implies
their implicit trust. Trust in Twitter is considered as binary construct (i.e., Trust score
= 1, if there is a trust statement including @ tag referring a member, and 0 otherwise).
The aggregation of trust scores is a measure of the members trustworthiness;
consequently it is a measure of member credibility in the SNS. Formally, member m
credibility from followers trust type x (direct or indirect) is denoted by , and
defined as follows:

(1)

where value is a real number in [0, 1] representing member m credibility

gained from direct\indirect followers at time t, refers to number of
direct\indirect followers of the member with maximum direct\indirect followers in the
context, considered as a reference point. Notably, credibility from followers is
normalized, hence if the most trustworthy member received the highest trust score
from all friends, then . Notably, the credibility gained from lower level
followers in the hierarchy is ignored, since their contribution to the members
credibility is small compared to first and second level followers.

3.2 Expertise Component

Expertise, a key dimension of credibility, is defined as the degree of a members
competency to provide valuable quality tweets and exhibit high activity [9]. The
expertise dimension of the member credibility captures their perceived knowledge
and skills in a given context. SNS members vary in their knowledge and consequently
their level of expertise; the more knowledge a member has, the more power they
possess, to the extent that "knowledge is power" [7]. The credibility of a member
coincides with their generated tweets quality which consequently coincides with their
326 J. Al-Sharawneh, S. Sinnappan, and M.-A. Williams

level of expertise in a particular context. In Twitter, members who are seen to be

disseminating information laden have more followers and are more trusted.
Member m expertise in SNS Twitter is drawn from the following three types of
messages a member can post:
A new message originated by member m.
A retweet message adopting other member opinion and contains (RT @, or
Retweet, via, follow or similar expressions).
A mention message [3] or a reply message to another member with (@ sign).
Each type of message has its influence based on its content quality. We believe that
new messages are the most important messages, followed by retweets and replies.
Thus, member m expertise in posting type y messages, (new, retweet,
mention), with accumulated quality is given by the following relationship:

(2)

Each type y tweets accumulated quality posted by member m defined as

where refers to type y tweet i quality and represents the member with the
maximum type y tweets accumulated quality, i.e., the member who has the highest
quality of type y tweets issued in the context, considered as a reference point.
Quality of a tweet message in the range [0, 1] is drawn from its content that
reflects its importance and influence in the context based on the following assessment
components: (1) number of keywords KN, (2) number of hash tags #N, (3) number of
other Twitter references indicated with @ signs @N, and (4) number of URLs
included in the tweet UN; each with a corresponding weight factor W in that
component. For tweet i, Quality , is given by:

(3)

where W = (0.4, 0.1, 0.1, 0.1) represent each component weight impact in tweet
quality respectively, and finally is normalized to the max in the context.
Notably, the @ parameter does not appear in new messages, so we need to consider
this impact when computing new tweet quality by scaling new tweets quality by 7/6.

3.3 Tweet Message Quality Computation

Quality of a message in Twitter should reflect its influence and importance in the
context. Most of SNS contributors in crisis are from general public who want to
provide information that help to minimize the impact and consequences of the crisis.
Thus, our concern is to assess message quality based on the content of the message
which is reflected in the above four components. Although other researchers such as
[1] use textual features and semantic features as basis to assess content quality; we
believe that semantic metrics can be considered as added-value to the content quality.
Our approach as follows: (1) For each assessment component, we build a list of
keywords and items from tweets after removing stop words from each tweet message.
(2) Ignore each keyword with a frequency less than four. (3) Assign a weight for each
 Credibility-Based Twitter Social Network Analysis 327

item based on its frequency in the context as the standard deviation of the keywords
set. (4) Compute quality of a tweet message as shown in equation 3.

3.4 Computing Member Credibility

By aggregating the credibility components, i.e. trustworthiness component credibility
from (direct followers and indirect followers ), and expertise
component (credibility gained from new tweets, retweets and replied tweets), thus
member m credibility at time t is given by:
(4)
where , represent trustworthiness credibility weighted importance factors gained
from direct followers and indirect followers respectively, and represent
expertise credibility weighted importance factors gained from new tweets, retweets
and mention/replied tweets respectively. Where are model tuning
parameters, and hese parameters can be normally set or
using automated machine learning techniques. In our experiment, we set the values to
(0.3, 0.2, 0.25, 0.15, 0.10) for model tuning parameters respectively, where
trustworthiness and expertise components each assigned equal weights 0.5.

3.5 Classifying Members Based on Credibility

To find the most trustworthy and expert SN members who can act as leaders, we
define a credibility threshold from which leaders and followers are identified as:
then member is a leader, otherwise member is a follower (6)
is a system parameter and is used to promote an adequate number of
Credibility Threshold
leaders from the crisis dataset based on the power-law degree distribution.

4 Experiments and Results

To demonstrate the feasibility and effectiveness of the proposed approach to identify
SNS leaders, in this section, we discuss the dataset and the results which demonstrate
that the notion of Follow the Leader in assessing credibility can be used as a means
to identify leaders and followers in the context.

4.1 SNS Crisis Dataset

Our Victorian bushfire crisis dataset was borrowed from [11]. The dataset contains
1684 tweets representing members communication in one week; as the event
unfolded from February 6-14, 2009. 697 unique users are identified in the dataset.
In Victorian bushfire crisis, 71% of the tweets are new tweets to disseminate new
information, followed by mentions/replies (21%) and retweets (8%). Although the
retweets ratio is less than the number of mentions/replies, retweets have more impact
on information diffusion than replies to spread critical crisis information among SNS
328 J. Al-Sharawneh, S. Sinnappan, and M.-A. Williams

members. This is because retweets reinforce a message and as such are considered
important, and hence they are passed on down the network.

4.2 Leaders and Followers Analysis

In the crisis dataset, about 78% of the population are inactive members who listen to
active members. They passively contribute to the conversation by posting limited
number of new messages, and they do not retweet messages of active members.
Similar findings were reported by Starbird et al. [12] in their study covering the Red
River flooding in America. Table 1 below provides a comparison between leaders
community and followers community for the active subset, i.e., after removal of
inactive members who do not have from/to links.

Table 1. Active Members: Leaders vs. Followers community

(NW : New, RP: Mention/Replies, RT: Retweets)

From Table 1, for active members; we note the following observations: (1) 59.4% of
the tweets are new, of which 15.5% of the new tweets are retweeted again;
furthermore, 25.1% of the tweets are mentions or replies. This indicates new tweets
are vital source of information distribution. (2) Leaders who represent 20% of the
population in the active set generate 59.3% of all tweets, they retweet 62% of all
retweet posts, and this for certain extent reflects power law distribution. (3) As shown
in (RT 2 Me) column, 84.2% of the Retweets in the communication are generated by
members addressing leaders content, i.e., originally leaders generated content as
new messages. On the other hand 15.8% of the Retweets are generated by members
addressing followers content. (4) 67% of followers follow leaders while 33% of
followers follow other members in the SNS. (5) Leaders community possess 72.8%
of indirect followers, on other hand, 27.2% of indirect followers follow other
followers' members, and those appear in the fourth level of the Follow the Leader
hierarchy. (6) Followers are more tweet repliers than leaders; they generate 53.1% of
reply messages. Leaders usually create new messages and pass (retweet) messages
more than followers. (7) Leaders community possesses 63.7% of SNS credibility,
while followers community possess 37.3% of SNS credibility; this tends to align
with the standard power law distribution.

4.3 SNS Leaders Influence Evaluation

To determine the efficiency of the proposed Credibility approach in identifying
leaders properly: first, we adopted centrality measure [4] which have been used as a
 Credibility-Based Twitter Social Network Analysis 329

proxy for power and influence; second, we conducted quality and spread analysis on
the new tweets and compare between leaders and followers influence.

4.3.1 Centrality Measure

Centrality reveals how influential and powerful a node is, which reflects the roles of
individuals in the network. Degree centrality is measured as the number of links of a
given node [5]. According to [7] if an agent receives many links, they are said to be
prominent; this justifies why many other agents seek to direct links to them and this
indicate their importance. Agents who have a significant high out-degree are able to
exchange opinions with many other members, or make many others aware of their
views and opinions. Agents who show high out-degree centrality are often said to be
influential agents [7]. So, to what extent leaders are prominent and influential agents?

Table 2. Centrality Measure and New Content Analysis

a. Centrality degree b. New content per quality category c. New content spread time

Table 2.a presents the centrality measures (i.e., in-degree and out-degree) for
leaders, followers and all members in the dataset. The in-degree centrality of a
particular node is defined as the number of in-links (trust statements this node
received); it represents the number of retweets and replied tweets addressing that
member. The out-degree centrality of a node is defined as the number of out-links
(trust statements issued from this node), it represents the number of retweets and
replied tweets issued by that member to other members in SNS. Table 2.a shows that
leaders in-degree 3.87 which is about 8 times followers in-degree (0.48).
Furthermore, leaders out-degree 5.39 is about 4 times followers out-degree (1.21).
Since leaders possess high average in-degree and high average out-degree
centrality, which in both cases is more than the followers and more than the average
member in the SNS, leaders are the most prominent and influential members in the
SNS. This outcome is drawn from their trustworthiness and expertise; leaders possess
the highest trustworthiness and expertise level among all members in the context.

4.3.2 New Content Quality and Spread Analysis

We conducted quality analysis on the new tweets for all population taking 10% as
leaders using new content quality computation described in section 3.2.1. We define
new content quality based on the quality distribution considering Quality Mean
(QMean) and Quality Standard Deviation (QSdev) from the mean of the new content
categorised as follows: (1) Low Quality: below QMean - QSdev (2) BelowAvg: equal
or greater than QMean - QSdev and less than QMean, (3) AboveAvg: equal or greater
than QMean and less than QMean + QSdev, and finally (4) High Quality: equal or
greater than QMean + QSdev.
330 J. Al-Sharawneh, S. Sinnappan, and M.-A. Williams

1. Quality Analysis Results: from Table 2.b, we note that: (1) High quality content
represents 16.75% of all new tweets content, of which 41.6% of this content is
generated by leaders. (2) As shown in column 3 (LdrsContent%), the percentage of
new postings content generated by leaders for each category is increasing by the
increase of the new content quality. This indicates that leaders are capable of
generating the most credible content. (3) As shown in column 4 (Retweeted%),
retweeting is increasing with the increase of the content quality; moreover, multiple
retweets correspond to quality above the average.
2. Spread / Diffusion Time Analysis Based on Quality Category Results: from
Table 2.c, we note that the elapsed time between new content posting time and first
time retweeted as shown in (1st Retweet Avg Time (Min)) column is decreasing with
the increase of new content quality. As shown for the High quality category, the time
elapse from new content posting to the first retweet time is 8.5 minutes, with overall
elapse time for all retweeting of 195.0 minutes. One notable phenomenon is that
tweets which carry bad news spread faster than other tweets.

5 Results Summary
In a Twitter SNS, members behaviour is the determinant of their credibility; different
members undertook varied activities and possess different interests. In a SN, the
credibility of a member acts as the predictor of their role. Leaders are the most
prominent and influential members in Twitter SNS. They generate the highest quality
content to be adopted and retweeted by other members.
SN leaders are aware and become involved in the crisis early; they produce
information-laden content and maintain high activity all times. Thus, their postings
are an important resource for information management and decision making.
In Twitter SNS, context leaders differ from Twitter leaders. In each context, there
are a set of leaders who are considered experts in the context, thus, it is difficult to
find same leaders across all contexts, simply because each Twitter member has varied
interests and expertise in varied contexts.

6 Conclusions and Future Work

Finding leaders who have more influence in a SNS context is a challenging issue. In
this paper, we proposed a new credibility based approach to identify leaders in SNSs
specific context utilizing the Follow the leader strategy. In SNS members
behaviour is the key determinant of their credibility; members credibility is drawn
from their trustworthiness and expertise. Members trustworthiness is assessed based
on other members trust in target member postings and their corresponding followers.
Members expertise is assessed using their activities and the type and quality of the
content they generate.
The experimental results demonstrated that the proposed credibility based
approach is a significantly innovative approach to identifying Twitter SNS
community leaders in a specific context such as crisis or terrorism. New content
quality plays an important role in information dissemination and diffusion.
 Credibility-Based Twitter Social Network Analysis 331

References
1. Agichtein, E., Castillo, C., Donato, D., Gionis, A., Mishne, G.: Finding high-quality
content in social media. In: WDSM 2008, pp. 183194. ACM (2008)
2. Al-Sharawneh, J., Williams, M.-A.: Credibility-based Social Network Recommendation:
Follow the Leader. In: ACIS 2010 Proceedings, Paper 24, pp. 110 (2010),
http://aisel.aisnet.org/acis2010/24
3. Bruns, A., Burgess, J.E., Crawford, K., Shaw, F.: # qldfloods and@ QPSMedia: Crisis
Communication on Twitter in the 2011 South East Queensland Floods (2012)
4. Burt, R.S.: Brokerage and closure: An introduction to social capital. Oxford University
Press, USA (2005)
5. Freeman, L.: Centrality in social networks conceptual clarification. Social Networks 1(3),
215239 (1979)
6. Goldbaum, D.: Follow the Leader: Simulations on a Dynamic Social Network, UTS
Finance and Economics Working Paper No 155 (2008),
http://www.business.uts.edu.au/finance/
research/wpapers/wp155.pdf
7. Hanneman, R., Riddle, M.: Introduction to social network methods. University of
California, Riverside (2005)
8. Kouzes, J.M., Posner, B.Z.: Credibility: How Leaders Gain and Lose It, Why People
Demand It, Revised Edition. Jossey-Bass, San Francisco (2003)
9. Kwon, K., Cho, J., Park, Y.: Multidimensional credibility model for neighbor selection in
collaborative recommendation. Expert Systems with Applications 36(3), 71147122
(2009)
10. Liu, G., Wang, Y., Orgun, M.: Trust Inference in Complex Trust-Oriented Social
Networks. In: Proceedings of the International Conference on Computational Science and
Engineering, Vancouver, Canada, vol. 4, pp. 9961001. IEEE (2009)
11. Sinnappan, S., Farrell, C., Stewart, E.: Priceless tweets! A study on Twitter messages
posted during crisis: Black Saturday. In: ACIS 2010 Proceedings, Paper 39 (2010)
12. Starbird, K., Palen, L., Hughes, A.L., Vieweg, S.: Chatter on the red: what hazards threat
reveals about the social life of microblogged information. In: Proceedings of the ACM
2010 Conference on Computer Supported Cooperative Work (CSCW 2010), pp. 241250
(2010)
13. Xu, Y., Ma, J., Sun, Y., Hao, J., Zhao, Y.: Using Social Network Analysis as a Strategy for
e-Commerce Recommendation. In: PACIS 2009 Proceedings, p. 106 (2009)
14. Ziegler, C.N., Golbeck, J.: Investigating interactions of trust and interest similarity.
Decision Support Systems 43(2), 460475 (2007)
 Design and Evaluation of Access Control Model
Based on Classification of Users Network Behaviors

Peipeng Liu1,2,3,4, Jinqiao Shi3,4, Fei Xu3,4, Lihong Wang, and Li Guo3,4
1
Institute of Computing Technology, CAS, China
2
Graduate University, CAS, China
3
Institute of Information Engineering, CAS, China
4
National Engineering Laboratory for Information Security Technologies, IIE, China
[email protected]

Abstract. Nowadays, the rapid development of the Internet brings great con-
venience to peoples daily life and work, but the existence of pornography
information seriously endangers the health of youngsters. So researches on net-
work access control model, aiming at preventing youngsters from accessing
pornography information, have become a hotspot. However, existing models
usually dont distinguish between adults and youngsters, and they block all the
accesses to pornography information, which does not meet the actual demand.
So preventing youngsters from accessing pornography information, while not
affecting the freedom of expressions in adults, has become a reasonable yet
challenging alternative. In this paper, an access control model based on classifi-
cation of users network behaviors is proposed, and this model can effectively
make a decision whether to allow ones access to pornography information ac-
cording to its age. Besides, an evaluation framework is also proposed, through
which the models controllability, cost-effectiveness and users accessibility are
analyzed respectively.

Keywords: Access Control, Evaluation, Controllability, Cost-effectiveness.

1 Introduction

With the rapid development of the Internet, more and more network services are pro-
vided for peoples daily life and work, but the Internet is also becoming flooded with
services involving pornography, which seriously endangers youngsters physiological
and mental health [8]. Therefore, researches on access control model aiming at pre-
venting youngsters from accessing pornography information have attracted more and
more attentions.

This work is supported by National Natural Science Foundation of China (Grant No.
61100174 and No.61272500), National High Technology Research and Development Pro-
gram of China, 863 Program (Grant No.2011AA010701 and No.2011AA01A103), National
Key Technology R&D Program (Grant No.2012BAH37B04).

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 332339, 2013.
Springer-Verlag Berlin Heidelberg 2013
 Design and Evaluation of Access Control Model Based on Classification 333

There have been raised many technical models for network access control [1, 7].
However, it has been shown that most of these control models work in an assertive
manner by which once the content or resource is identified as pornography, the model
refuses all accesses to it. Obviously, this does not meet the actual demand. So it is
necessary to develop a systematic model that can effectively prevent youngsters por-
nography access, while not affect the freedom of expression in adults.
In this paper, an access control model based on classification of users network be-
haviors is proposed, and this model first judges whether the user is a youngster or not,
and then based on the judgment, determines whether to allow the access. This paper
also proposed an evaluation framework, through which the models controllability,
users accessibility and the cost-effectiveness are analyzed respectively.
The rest of the paper is organized as follows. Section 2 discusses the existing
works related to access control model and its evaluation. In section 3, the proposed
access control model is described in detail. And section 4 is about the evaluation
framework. An example is given in section 5. At last, we summarize the paper and
give a conclusion in section 6.

2 Related Works

In [1], the authors proposed the control model (ICCON) for the ICS (Internet content
security) on the Internet. Although it involves the communicators identities, it mainly
takes the coarse-grained properties, such as IP address and URL, into consideration,
excluding the users fine-grained properties, such as the ages. As a result, it cannot
distinguish between adults and youngsters.
Another kind of popular access control model is based on the resource, that is once
a resource on the Internet is identified as pornography, and then all access to it is re-
fused. Otherwise, the resource is open to everyone.
As for evaluation, previous researches, such as [2, 6, 9, 10], usually focus on the
controllability from a technical point, which are lack of assessment of the users expe-
rience and the models cost-effectiveness.
All these works have played important roles in preventing the spread of the porno-
graphy information on the Internet and evaluating the existing models, but they are
sometimes not comprehensive enough. So its necessary to develop a more compre-
hensive access control model.

3 Access Control Model Based on Classification of Users

Network Behaviors

3.1 General Framework for Information Access Control

As described in [5], Internet information access control mechanism controls access
actions with static or dynamic control rules by RM (Reference Monitor). RM is a
core component of access control mechanism, which decides whether to allow the
334 P. Liu et al.

continuance of an access. In our model, RM is deployed on the transmission channel

(e.g., on the router connected to the Internet in primary/middle schools), and we call
RM of this deployment NRM (network-side reference monitor). To describe the gen-
eral framework for information access control, two definitions are given below.

Definition 1: Network Reference Monitor.

A network reference monitor is a 4-tuple:
NRM = < M , P, S , O >
Where M is set of hardware and software, P is set of access control strategies, S is
set of access subjects, and O is set of network resource objects. NRM controls actions
of subjects S accessing objects O based on the access control strategies P .

Definition 2: Internet Information Access Control.

Whether an access action a is permitted or not is determined by satisfaction between
control rules P ( x) and network resource O which subject S wants to access. If
resource O matches P ( x) , then the action a cannot be permitted, else action a is
permitted.

3.2 Access Control Model Based on Classification of Users Network

Behaviors
According to the general framework, an access control model for pornography infor-
mation, based on classification of users network behaviors, is proposed. This model
consists of three parts and each of them will be described in detail below.

Framework
At first, we give the framework for the access control model based on classification of
users network behaviors as in Fig. 1.

Fig. 1. Framework for access control model based on classification of users network behaviors
 Design and Evaluation of Access Control Model Based on Classification 335

The framework works as follows. At first, the model learns users surfing habits
and predicts the users ages depending on the learned knowledge, and secondly, when
an access action is captured, the model analyzes it together with the prediction results
and the control strategies. At last, based on the analysis, the model makes a decision
whether to allow the access action.

Learning and Prediction Module

There have been a lot of research on demographic prediction, and many of them have
achieved satisfactory results [3, 4]. In our model, a simple age-prediction method is
adopted. Firstly, based on users profiles and their browsing history, the users age
information is propagated to the browsed pages, and then, a supervised regression
model is trained to predict the probability distribution of the ages of a given Web-
pages readers. Secondly, within Bayesian framework, an internet users age is pre-
dicted based on the age tendency of the WebPages that he/she has browsed.

Fig. 2. Learning Module and Prediction Module

In Fig. 2, the learning module is in a training process, through which we can get
the probability distribution of ages in WebPages. And based on the results of learning
module, the prediction module takes the users ip and historical browsing behaviors as
inputs, and gives the prediction results in the form of ( IP, Age ) .

Action Discovering and Analysis

The action discovering and analysis module of our model is shown in Fig.3.

Fig. 3. Action Discovering and Analysis

In Fig. 3, extraction is applied to the network stream to get the ( IP,Webpage) pairs,
and then together with prediction results and control rules, namely, ( IP, Age) and
(Age,Webpage) are supplied to the analyze module as inputs, then the analyze mod-
ule analyzes all the inputs comprehensively, and at last gives the control rules satis-
faction, i.e., whether the access action should be allowed or not.
336 P. Liu et al.

The analysis process can be denoted by a tuple, that is < Lu , Lp , P,T , , D, Auth > ,
where Lu is the label of a requester; L p is the label of the webpage requested; P is the
prediction results according to Lu , existing in the form likes ( Li , Agei ) , where
Agei is the predicted age for Li ; T is a blacklisted control strategies, existing in the
form likes {( Age1 ,Webpage1 ), ( Age2 , Webpage2 )...( Agen ,Webpagen )} , where each
pair (Agei ,Webpagei ) means users at Agei should not be allowed to access
Webpagei ; is a set of non-negative real numbers; D : ( Lu , Lp ) P T , is a dis-
tance function, and the detailed calculation method for distance varies with different
applications. The smaller the distance is, the more likely that the user Lu should not
be allowed to access the service L p ; Auth : [0,1] , is an authentication function,
which is employed to compute the probability of the response. The input of the func-
tion is the distance calculated above. The greater Auth is, the greater the probability
with which a response should be made is.

Response
At last, in the stage of response, the active response is chosen for the proposed model.
Active response refers to taking some approaches such as blocking and cutting to
refuse the access request.

4 Evaluation
An evaluation framework is also proposed to quantitatively evaluate the proposed
control model from several aspects, including controllability, accessibility and cost-
effectiveness.

4.1 Controllability
We firstly give the formal definition of controllability. According to the Internet in-
formation access control model, we suppose extract, analyze, match and control as
atomic operation, which denote the process of extracting information from Internet,
analyzing information, matching information with control rules, and controlling
specific information or access action correspondingly.

Fig. 4. Evaluation Model for Controllability

 Design and Evaluation of Access Control Model Based on Classification 337

To quantitatively analyze the controllability of network access actions, we model

the network information access control by finite state machine as shown in Fig.4.
In Fig. 4, Pe is the probability that information is extracted correctly from net-
work streams; Ptn is the probability that information which actually mismatches the
rules is judged to mismatch; Pfn is the probability that information which actually
matches the rules is judged to mismatch; Ptp is the probability that information which
actually matches the rules is judged to match; Pfp is the probability that that is, infor-
mation which actually mismatches the rules is judged to match; Pres is the probability
that response is applied to the access actions correctly.
The evaluation model works as follows. State 0 is an original state, and state 6 de-
notes actions not matching control rules and state 7 are actions which match control
rules. State 8 is without response that actions are not extracted correctly or mismatch-
ing with control rules or being unsuccessful controlled. State 9 denotes the situation
that actions are processed with a response. In this model, the transitions failures from
state 1 to state 3 and state 5 correspond to two possibilities, one is NRM determines
actions satisfy rules P( x) but they do not in fact, and the other is NRM determines
actions do not satisfy rules P( x) but they actually do.
Then controllability can be defined as follow.

Pcon = (((extract match) ((extract match) > control )) (1)

That is, controllability is defined to contain two cases, that is, information extracted
from Internet is mismatching control rules or information must be controlled if it
matches control rules.
So controllability can be expressed by Pi in Fig. 4 as Pcon = Pe Ptn + Pe Ptp Pres .

4.2 Accessibility

Accessibility is proposed from a users point, namely, the impact brought to users by
the control model. The accessibility is divided into two types according to users dif-
ferent access requests. For the pornography request, the accessibility is expressed as
follow.

Ppornography = (1- Pe ) + Pe [ Pfn + Ptp (1- Pres )] (2)

Formula ( ) expresses the probability that a users request which actually is porno-
graphy is allowed.
While for the non-pornography requests, the accessibility is expressed as follow.

Pnon - pornography = (1- Pe ) + Pe [ Ptn + Pfp (1- Pres )] (3)

338 P. Liu et al.

Formula ( ) expresses the probability that a users request which actually is not por-
nography is allowed.
For accessibility, absolutely, the aim of the control model is to increase
Pnon - pornography while decrease Ppornography .

4.3 Cost-Effectiveness
In order to achieve the best performance while decrease the costs as much as possible,
an evaluation method based on linear programming is proposed. Here, we suppose it
will cost Wi to increase one percentage point of probability Pi in Fig. 4. And for
each component in the system, its ceiling of cost is limited, and we suppose for each
Pi , its costs can take up to Ci . Then the problem can be described as follows.

Minimize : Ppornography = (1- Pe ) + Pe [ Pfn + Ptp (1- Pres )]

Subject to :
100 Pi Wi Ci , Pi > 0

Here, according to different requirements, the aim can be set to maximize

Pnon - pornography or minimize Ppornography . With the solution to this linear problem, the con-
trol model not only achieves the optimization targets, but also meets the requirements
in cost.

5 Example

Considering a network management system for a community or a school, and here we

mainly discuss the actions analysis process. Suppose blacklist is used by the control
model. Now we give the corresponding model.
T is a blacklist in the form of (Age,Webpage) pairs, and each (Age,Webpage)
pair, means that users at such ages should not be allowed to access such WebPages.
P , in the form of ( IP, Age) , is the prediction result when an access action is cap-
tured, and A , in the form of ( IP,Webpage) , is the access action captured.
D is a distance function defined as follows:

= D (( P , A ), T ) ={ 1,0,otherwise
if ( P, A) is inT ,
.
(4)

Auth is an authentication function:

Auth ( ) = {1,undefined
if =0,
, =1.
(5)

The authentication function means that if the distance between the access actions and
the blacklist is 0, then the access should be responded with probability 1, otherwise,
the authentication is undefined.
 Design and Evaluation of Access Control Model Based on Classification 339

As for evaluation, in order to quantitatively evaluate the network management sys-

tem, we need to discuss controllability, accessibility and cost-effectiveness. Based on
practical statistic or experiments, the values of these indicators can be analyzed by the
way shown in section 4.

6 Conclusions and Future Work

In this paper, a fine-grained network control model based on classification of users

network behaviors is proposed. And this control model decides whether an access
action is allowed or not by analyzing the ages of users. Besides, a framework to eva-
luate the model is also proposed, through which the models controllability, users
accessibility and the cost-effectiveness are analyzed respectively.
However, several problems still need to be focused on in the future work. Firstly,
our learning and prediction module needs a period of time to achieve certain accura-
cy, so for situations where users ip changes frequently, the model is not very suitable.
Besides, the widely used encryption and anonymity also make the prediction process
difficult, because it becomes harder to realize which webpage the user want to access
and what the contents of the webpage are.

References
1. Fang, B., Guo, Y., Zhou, Y.: Information content security on the Internet: the control mod-
el and its evaluation. Science China Information Sciences (2010)
2. Hongli, Z.: A Survey on Internet Measurement and Analysis. Journal of Software (2003)
3. Fortuna, B., Mladenic, D., Grobelnik, M.: User Modeling Combining Access Logs, Page Con-
tent and Semantics. In: 1st International Workshop on Usage Analysis and the Web of Data
(USEWOD 2011) in the 20th International World Wide Web Conference (WWW 2011),
Hyderabad, India, March 28 (2011)
4. Hu, J., Zeng, H.-J., Li, H., Niu, C., Chen, Z.: Demographic prediction based on users
browsing behavior. In: Proceedings of the 16th International Conference on World Wide
Web, Banff, Alberta, Canada, pp. 151160. ACM (2007)
5. Lihua, Y., Binxing, F.: Research on controllability of internet information content security.
Networked Digital Technologies (2009)
6. Liu, Y.-Y.: Controllability of complex networks. Nature 473, 167173 (2011)
7. Park, J., Sandhu, R.: The UCONabc usage control model. ACM Transactions on Informa-
tion and System Security 7(1), 128174 (2004)
8. Peter, R.W.: There is a need to regular indecency on the internet. 6 CORNELL J.L. &
PUB.POLy 363 (1997)
9. Wiedenbeck, S., Waters, J., et al.: PassPoints: Design and longitudinal evaluation of a
graphical password system. International Journal of Human-Computer Studies 63(1-2),
102127 (2005)
10. Wei, J.: Evaluating Network Security and Optimal Active Defense Based on Attack-
Defense Game Model. Chinese Journal of Computers (2008)
 Two Phase Extraction Method
for Extracting Real Life Tweets Using LDA

Shuhei Yamamoto and Tetsuji Satoh

Graduate School of Library, Information and Media Studies,

University of Tsukuba
{yamahei,satoh}@ce.slis.tsukuba.ac.jp
http://ce.slis.tsukuba.ac.jp

Abstract. Nowadays, many twitter users tweet their personal aairs.

Some of these posts can be quite beneficial for real life, for example,
Eating, Appearance, Living, Disasters, and so on. In this paper, we pro-
pose a two phase extracting method for selecting beneficial tweets. In
the first phase, many topics are extracted from a sea of tweets using
Latent Dirichlet Allocation (LDA). In the second phase, associations be-
tween many topics and fewer aspects is built using a small set of labeled
tweets. To enhance accuracy, the weight of feature words is calculated by
information gain. Our prototype system demonstrates that the proposed
method can extract the aspects of each unknown tweet.

Keywords: Twitter, Real Life, LDA, Two Phase Extraction.

1 Introduction
Many information sharing services currently exist, such as community knowledge
sharing sites, blogs, and microblogs. Twitter is one of the most widely spread
microblogs. Articles in Twitter are posted very easily within 140 characters.
Users post their experiences, opinions, and other events that arise in their daily
life.
Most articles in Twitter are both useful and timely because they are written
based on current events. For example, commuter train delay information posted
by a passenger is timely and useful for waiting passengers. Supermarket sales and
bargain information is also useful for neighborhood consumers. Such information
has highly regionality and freshness. Thus, we call such posts Real Life Tweets.
Extracting real life tweets from a sea of tweets is quite an important research
issue for supporting life activities.
The Great East Japan Earthquake Disaster, which occurred in March of 2011,
is a perfect example of the benefits of real life tweets. There was a great amount
of confusion in the stricken area immediately following the earthquake. There
was a lack of food, suspension of water supply, and train service cancellations.
At that time, useful tweets reported the location of water supplies and food
distribution, as well as the service status of trains, demonstrating that such real
life tweets helped the users in the devastated region[1].

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 340347, 2013.
c Springer-Verlag Berlin Heidelberg 2013

 Two Phase Extraction Method for Extracting Real Life Tweets Using LDA 341

Table 1. Aspects of Real Life

Aspect typical terms

Appearance clothes, dress, wearing, decoration, makeup, barber ...
Contact appointment, set a meeting, invitation, family, friend ...
Disaster flood, tornado, earthquake, hazard, secondary disaster ...
Eating cooking food, dining out, eating, restaurant, junk food ...
Event festival, ceremonial functions, project, event schedules ...
Expense shopping, order, advertisement, discount, bargain sale ...
Health cold, physical condition, hurt, hospital, prophylaxis ...
Hobby leisure-time, pastime, entertainment, hobby, interest ...
Living home, lodgings, furniture, cleaning, washing, living ...
Locality sightseeing, regionally specific, local information ...
School study, class, examination, education, research, homework ...
Trac train, bus, airplane, timetables, trac information ...
Weather weather forecast, temperature, humidity, wind, pollen ...
Working job hunting, part-timer, coursework, open a store, sales ...

As mentioned above, highly useful real life tweets are increasingly posted on
Twitter. However, users posts various types of tweets. Nod and sympathetic
phrases frequently appear on Twitter. For example, Thank you and I see
often appear in posts. These posts do not directly support the real life situations
of other users. We believe that users want a method of locating beneficial tweets
on Twitter. These types of nods and sympathies simply impede the discovery of
substantive tweets.
Information is used in various aspects of life. Real life tweets can accommo-
date such aspects. For example, tweets such as The train is not coming! are
categorized in the Trac aspect and will support users who want to ride the
train. Posts such as Today, bargain sale items are 50% o! are categorized
as Expense aspects and will support users who are going shopping. In our
previous research, we detected 14 aspects of real life[2]. These 14 aspects, listed
in Table 1, are obtained from Local Community1 and Life2 in the Japanese
version of Wikipedia.
In this paper, we propose a two phase extraction method for extracting real life
tweets. In the first phase, many topics are extracted from a sea of tweets using
Latent Dirichlet Allocation (LDA). In the second phase, associations between
many topics and fewer aspects, shown in Table 1, are constructed using the
weight of feature terms calculated with information gain.

2 Related Works
Real life tweets consist of both the experiences and knowledge of users and re-
gional information. Several studies with experience mining have been conducted
1
http://ja.wikipedia.org/wiki/nR~ jeB
2
http://ja.wikipedia.org/wiki/
342 S. Yamamoto and T. Satoh

to extract experiences from documents. Kurashima et al. [3] reported that human
experience can be divided into five areas: Time, space, action, object, and feeling.
Inui et al. [4] described a method of indexing personal experience information
from the viewpoint of time, polarity, and speaker modality. This information is
indexed as Topic object, Experiencer, Event expression, Event type, and Factu-
ality. These mining methods are eective for relatively long documents such as
blogs. Hence, these methods are not appropriate for Twitter posts, which consist
of many short sentences. In addition, experience mining would be much more
dicult because subjects and objects in sentences are often omitted in Twitter.
The study of Twitter is flourishing. Ramage et al. [5] used large scale topic
models to represent Twitter feeds and users, showing improved performance
on tasks, such as post and user recommendations. Bollen et al. [6] analysed
sentiment on Twitter according to a six-dimensional mood (tension, depression,
anger, vigor, fatigue, and confusion) representation, determining that sentiment
on Twitter correlates with real-world values such as stock prices and coincides
with cultural events. Diakopoulous and Shamma [7] conducted inspirational work
in this vein, demonstrating the use of timeline analytics to explore the 2008
Presidential debates through Twitter sentiment. Sakaki et al. [8] assumed that
Twitter users act as sensors, discovering an event occurring in real time in the
real world. Zhao et al. [9] suggested a model, called Twitter-LDA, based on
the hypothesis that one tweet expresses one content of a topic. They classified
tweets into appropriate topics and extracted keywords to express the contents of
the topic. Mathioudakis et al. [10] extracted burst keywords in tweets collected
automatically. They found a trend fluctuating in real time by creating groups
using the co-occurrence of keywords.
Our paper contends that the information is not only based on user experience,
but also user knowledge, which we believe to be useful in real life.

3 Two Phase Extraction Method

Real life tweets contain the various aspects mentioned in Section 1. Therefore,
it is dicult to enumerate keywords related into whole aspects. Moreover, the
rule-based parsing approach using experience mining does not work well because
twitter posts are comprised of very short sentences.
We propose a two phase extraction method, shown in Figure 1. In the first
phase, a large number of topics are extracted from a sea of tweets using LDA.
LDA is an unsupervised learning model for clustering large amounts of docu-
ments [11]. In the second phase, we construct an association between the topics
and aspects.

3.1 Association Building Using Labeled Tweets

For building associations, we prepare a small set of labeled tweets. A set of

extraction terms from tweets labeled aspect ap is Wap . To calculate the weight
of terms that represent each aspect, we use information gain. Information gain
 Two Phase Extraction Method for Extracting Real Life Tweets Using LDA 343

Fig. 1. Two Phase Extraction Method

provides feature choice and represents a good feature. The information gain of
term w is calculated as follows:

IG(w) = H(AP ) (P (w)H(AP |w) + P (w)H(AP

|w)),
(1)
where AP denotes all aspects. P (w) is the probability of term ws appearance
in all tweets. P (w)
is the probability of term w not appearing in all tweets.
H(AP |w) is the conditional entropy of all aspects AP that appear in term w.
H(AP |w) is the conditional entropy of all aspects AP that do not appear in
term w. When IG(w) has a high value, w is a good feature.
Relevance R(ap, k) between topics k and aspects ap is then calculated as
follows:

1 "
R(ap, k) = IG(w) pw,k , (2)
length(Wap )
wWap

where pw,k denotes term w of occurrence probability in topic k. length(Wap ) is

the size of the terms set. Note that this equation calculates the relevance using
the occurrence probability and information gain of terms.
To fold in value 0 to 1, we normalize R(ap, k) in each aspect. Normalized

relevance R(ap, k) is shown as follows:

R(ap, k)
R(ap, k) = " , (3)
R(ap, k)
kT

where T denotes all topics extracted using LDA.

We make an association between topics and aspects when R(ap, k) exceeds a
threshold (ap). The threshold value in each aspect is calculated as follows:

(ap) = max R(ap, k) std(ap) d, (4)
kT
344 S. Yamamoto and T. Satoh

where std(ap) denotes the standard variation in R(ap, k) for all topics. Parameter
d varies for maximum extraction precision in Section 4.

3.2 Real Life Tweet Extraction

To extract real life tweets, we use the associations between topics and aspects.
The score between tweets and aspects is calculated as follows:

R(ap, k) "
Score(ap) = pw,k , (5)
length(W )
wW

where W denotes a set of terms extracted from an unknown tweet and pw,k ,
denotes the term w of occurrence probability in topic k.
The selected aspect in each unknown tweet is as follows:

Aspect = arg max Score(ap) (6)

apAP

4 Experimental Evaluations
We evaluate the precision of the aspect selected by equation (6). After that,
we should clarify why we drew a dierence between high precision and lower
precision. Because of this, we carefully observe the association between topics
and aspects of varying thresholds d.

4.1 Datasets and Parameters Setting

Datasets: We used datasets of tweets posted from April 15, 2012 to August 14,
2012 in the Japanese location information of Tsukuba or . The number of
tweets is 2,900,819. This large dataset is classified into topics using LDA.
We also prepared a small set of tweets labeled with each aspect. The collection
term and condition are the same as above. The number of tweets in each aspect
is 200. We also prepared 200 un-real life tweets to confirm classification. These
sets of tweets are used for building the association.

Parameters Setting: LDA requires some hyper parameters. According to re-

lated works[12], we set at 50/T and at 0.1. The iterative calculation count
in LDA is 100 times in every case. We set the number of topics at 500.

Extraction Precisions: We evaluate the precisions by using 10-fold cross-

validation. Thus, 180 tweets in each aspect are used for building association and
the remaining 20 tweets are used for testing.
 Two Phase Extraction Method for Extracting Real Life Tweets Using LDA 345

4.2 Experimental Results

Extraction Precisions: The extraction precisions of every aspect are shown

in Figure 2. The horizontal axis is threshold d that decides the association be-
tween topics and aspects. In some aspects, i.e., Appearance, Eating, and Trac,
maximum extraction precision is achieved at the minimum value of d(d = 1). In
some aspects, i.e., Contact, Event, and Health, maximum extraction precision is
achieved at the maximum value of d(d = 20).
To analyze the association between topics and aspects, we evaluate the number
of connections from topics to each aspect. The number of topics against threshold
d is shown in Figure 3. In every topic, the number of topics is increased according
to d. The five aspects of Eating, Living, Expense, Disaster, and Working mostly
connect one topic until d 12. On the other hand, the Contact and Event
aspects are rapidly reached to connect whole topics.

(a)aspect: Appearance Health (b)aspect: Hobby Working

Fig. 2. Extraction Precisions

Fig. 3. Number of Topics in each Aspect Fig. 4. Top 10 Topics

346 S. Yamamoto and T. Satoh

5 Discussion

According to Figure 2, Living, Eating, Disaster, and Expense achieved the high-
est precisions. These four aspects are associated with fewer topics, as shown in
Figure 3, even if parameter d was increased. Association details between each
aspect and topics are shown in Figure 4. These ten topics in the figure have
a larger sum of relevances R from topics to aspects. In this figure, we confirm
that these four aspects have strong relevance to the specific topics, Living to
Topic418, Disaster to Topic144, and so on, respectively.
The worst cases, Event and Contact, have a lower precision value as shown in
Figure 2. These aspects were associated with many topics even if d was small. In
particular, Contact is associated with the same topics, Topic297, Topic383, and
Topic479, as Eating. Moreover, the values of R are smaller than Eating. Event
is associated with the same topics, Topi465 and Topic320, as Locality.

6 Conclusion

In this paper, we propose a two phase extraction method for extracting real life
tweets. In the first phase, many topics are extracted from a sea of tweets using
LDA. In the second phase, associations between many topics and fewer aspects
are constructed using a small set of labeled tweets. To enhance accuracy, the
weight of the feature terms is calculated with information gain.
Based on the experimental evaluation results, our prototype system demon-
strates that our proposed method can extract aspects of each unknown tweet.
We confirm that high precision aspects are associated with fewer topics that are
similar to the aspects. However, low precision aspects are associated with many
topics. In this case, many topics are associated with many aspects. In the future,
we should consider evaluating the KL Divergence from aspect to aspect in order
to enhance separation among the aspects.

Acknowledgement. This work was partly supported by Research Projects of

Faculty of Library and Information and Media Science, University of Tsukuba.

References
1. Yamamoto, M., Ogasawara, H., Suzuki, I., Furukawa, M.: Tourism informatics:9.
information propagation network for 2012 tohoku earthquake and tsunami on twit-
ter. IPSJ Magazine 53(11), 11841191 (2012) (in Japanese)
2. Yamamoto, S., Satoh, T.: Real life information extraction method from twitter. In:
The 4th Forum on Data Engineering and Information Management (DEIM 2012)
F3-4 (2012) (in Japanese)
3. Kurashima, T., Tezuka, T., Tanaka, K.: Blog map of experiences: Extracting and
geographically mapping visitor experiences from urban blogs. In: Ngu, A.H.H.,
Kitsuregawa, M., Neuhold, E.J., Chung, J.-Y., Sheng, Q.Z. (eds.) WISE 2005.
LNCS, vol. 3806, pp. 496503. Springer, Heidelberg (2005)
 Two Phase Extraction Method for Extracting Real Life Tweets Using LDA 347

4. Inui, K., Abe, S., Morita, H., Eguchi, M., Sumida, A., Sao, C., Hara, K., Mu-
rakami, K., Matsuyoshi, S.: Experience mining: Building a large-scale database
of personal experiences and opinions from web documents. In: Proceedings of the
2008 IEEE/WIC/ACM International Conference on Web Intelligence, pp. 314321
(2008)
5. Ramage, D., Dumais, S., Liebling, D.: Characterizing microblogs with topic models.
In: Proceedings of ICWSM 2010, pp. 130137 (2010)
6. Bollen, J., Pepe, A., Mao, H.: Modeling public mood and emotion: Twitter senti-
ment and socio-economic phenomena. In: Proceedings of WWW 2010, pp. 450453
(2010)
7. Diakopoulous, N.A., Shamma, D.A.: Characterizing debate performance via aggre-
gated twitter sentiment. In: Proceedings of CHI 2010, pp. 11951198 (2010)
8. Sakaki, T., Okazaki, M., Matsuo, Y.: Earthquake shakes twitter users: Real-time
event detection by social sensors. In: Proceedings of 18th International World Wide
Web Conference, WWW 2010, pp. 851860 (2010)
9. Zhao, X., Jiang, J., He, J., Song, Y., Achananuparp, P., Lim, E.P., Li, X.: Topical
key phrase extraction from twitter. In: The 49th Annual Meeting of the Association
for Computational Linguistics, pp. 379388 (2011)
10. Mathioudakis, M., Koudas, N.: Twittermonitor: trend detection over the twitter
stream. In: Proceedings of the 2010 International Conference on Management of
Data, pp. 11551158 (2010)
11. Blei, D.M., Ng, A.Y., Jordan, M.I.: Latent dirichlet allocation. The Journal of
Machine Learning Research 3, 9931022 (2003)
12. Griths, T.L., Steyvers, M.: Finding scientific topics. Proceedings of the National
Academy of Science 101, 52285235 (2004)
 A Probabilistic Model
for Diversifying Recommendation Lists

Yutaka Kabutoya1, Tomoharu Iwata2 ,

Hiroyuki Toda1 , and Hiroyuki Kitagawa3
1
NTT Service Evolution Laboratories, NTT Corporation, Japan
2
NTT Communication Science Laboratories, NTT Corporation, Japan
3
Faculty of Engineering, Information and Systems, University of Tsukuba, Japan
{kabutoya.yutaka,iwata.tomoharu,toda.hiroyuki}@lab.ntt.co.jp,
[email protected]

Abstract. We propose a probabilistic method to diversify the results

of collaborative filtering. Recommendation diversity is being studied by
many researchers as a critical factor that significantly influences user
satisfaction. Unlike conventional approaches to recommendation diver-
sification, we theoretically derived a diversification method. Specifically,
our method naturally diversifies a recommendation list by maximizing
the probability that a user selects at most one item from the list. For
enhanced practicality, we formulate a model for the proposed method
on three policies robust estimation, the use of only purchase history,
and the elimination of any hyperparameters controlling the diversity. In
this paper, we formally demonstrate that our method is practically su-
perior to conventional diversification methods, and experimentally show
that our method is competitive with conventional methods in terms of
accuracy and diversity.

1 Introduction
Many online services, such as Amazon1 , Netflix2 , and YouTube3 , employ rec-
ommender systems because they are expected to improve user convenience and
service provider profit [6, 10, 12, 18]. This is promoting a lot of research on
recommendation methods [1]. Many conventional recommendation methods are
accuracy-oriented, i.e., their goal is to accurately predict items that the user will
purchase in the future.
These accuracy-oriented algorithms, however, sometimes degrade user satis-
faction. The recommendation results are often presented to users in the form of
a multiple item list, not one single item. The recommended items in a list gen-
erated by an accuracy-oriented method often have similar content. According to
[21], since such a list reflect only partial aspects of the users interests, it cannot
necessarily satisfy the user.
1
Amazon - Online shopping, http://www.amazon.com
2
Netflix - Online movie rentals, http://www.netflix.com
3
YouTube - Online video service, http://www.youtube.com

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 348359, 2013.
c Springer-Verlag Berlin Heidelberg 2013

 A Probabilistic Model for Diversifying Recommendation Lists 349

Researchers have recently addressed this problem by focusing on the diversity

of recommendation lists [9, 11, 17, 19, 21]. This novel evaluation metric indicates
balance of recommendation list entries, i.e. how well they satisfy the active users
full range of interests.
This paper proposes a theoretical method to diversify recommendation lists.
Our method make a recommendation list so as to maximize the probability that
a user purchases at most one item from the list. While exactly optimizing our
objective function is dicult, we derive a greedy heuristic for approximately
optimizing it. Consequently, the diversity of the recommendation lists oered by
our method is naturally enhanced. Specifically, estimating the probability that a
user will purchase the lth item from the list requires us to consider, in addition
to the users preferences, the additional condition that the first l 1th items are
not purchased. This additional condition suggests that the lth item should be
dissimilar to the first l 1 items.
We formulate a model and its parameter estimation for our method by setting
the following three policies as: First, we reduce the number of parameters for the
robust estimation. Second, parameters are estimated only from purchase history,
which can be automatically collected. Collecting the other kinds of information
used by conventional methods, such as item taxonomy [21] and ratings [11, 17, 19]
requires some manual tasks. Finally, we avoid the use of hyperparameters in con-
trolling the trade-o between accuracy and diversity, unlike [9, 17, 21]. Determin-
ing such hyperparameters is dicult, since we do not know the optimal balance
between them that will maximize end-user satisfaction in advance.
The rest of this paper is organized as follows. In 2, we explain the detail
of our method. In 3, we formulate a model and its parameter estimation for
our method. In 4, we briefly review related work. In 5, we formally and
experimentally evaluate our method, by comparing it with conventional methods.
Finally, we oer concluding remarks and a discussion of future work in 6.

2 Proposed Method

Our method oers a recommendation list consisting of L items maximizing the

probability that a user u purchases an item from the list. Our objective function
for determining the recommendation list is as follows:

P (y1 = 1 . . . yL = 1|u, i1 , . . . , iL ) , (1)

where il represents the lth item in the list, yl is a binary variable indicating that
the lth item is purchased or not (yl = 1 if purchased, yl = 0 if not). On the
other hand, the objective function of a conventional accuracy-oriented method
is individually set for each item in the list without considering a purchase from
the whole list [2].
Here, we assume that the target user purchases at most one item from the
list. This assumption is based on the behavior of many actual recommendation
services, such as on Amazon.com. On these services, a user cannot purchase
350 Y. Kabutoya et al.

multiple items from a list, since the screen is transited when the user selects
(click a link) an item from a recommendation list.
We use the greedy algorithm because exactly maximization of the objective
function expressed by Eq. (1) falls into a combinatorial optimization problem,
which is dicult in general. Since many users decide whether to purchase each
item in the list in order of rank, applying the greedy approach seems reasonable.
The first item can be easily obtained by maximizing P (y1 = 1|u, i1 ) as is
done in conventional methods. Having chosen the first item i1 , we determine the
second item i2 so as to maximize P (y1 = 1 y2 = 1|u, i1, i2 ). For simplicity, we
expand P (y1 = 1 y2 = 1|u, i1 , i2 ) as:

P (y1 = 1 y2 = 1|u, i1 , i2 )
= P (y1 = 1|u, i1 , i2 ) + P (y1 = 0 y2 = 1|u, i1, i2 )
P (y1 = 1|u, i1 ) + P (y2 = 1|u, i1 , i2 , y1 = 0) P (y1 = 0|u, i1 ) ,

where the partitioning of the event of y1 = 1 y2 = 1 into the disjoint events

y1 = 1 and y1 = 0 y2 = 1 in the second line, and the simplification in the last
line is based on the independence of y1 and i2 . Since only the second factor in
this equation depends on i2 , we can formulate the probability that the user u
selects the second item i2 from the list as follows:

P (y2 = 1|u, i2 , y1 = 0, i1 ) .

Recursively, we can express the probability that the user u selects the lth item
il from the list as follows:

P (yl = 1|u, il , y1 = 0, . . . , yl1 = 0, i1 , . . . , il1 ) . (2)

This probability means that for choosing il according to our two assumptions,
we should take into account not only the target user u but also the condition
that items i1 , . . . , il1 are not purchased.

3 Model and Parameter Estimation

We formulate a model and its parameter estimation for our method based on
the following three policies:

1. We reduce the number of parameters for the robust estimation. The direct
robust estimation of 7this model
; is dicult because the number of parameters
is of the order of O U N l , where U is the number of unique users and N
is the number of unique items, and so becomes excessive compared with the
size of purchase history for training.
2. We use only purchase history, which can be automatically collected. This
avoids the expensive manual tasks needed to acquire the other kinds of in-
formation needed for recommendations, such as item taxonomy and ratings.
 A Probabilistic Model for Diversifying Recommendation Lists 351

3. We avoid the use of hyperparameters in controlling the balance between

accuracy and diversity. According to [21], there is a trade-o between the two,
and end-user satisfaction demands that they be well balanced. Parameter
tuning is, however, dicult.
We expand Eq. (2) in accordance with Bayes rules as follows:
P (yl = 1|u, il , y1 = 0, . . . , yl1 = 0, i1 , . . . , il1 )
P (yl = 1 y1 = 0 . . . yl1 = 0|u, yl = 1, i1 , . . . , il )
= #1
y =0 P (yl y1 = 0 . . . yl1 = 0|u, yl , i1 , . . . , il )
l

P (yl = 1|u, i1, . . . , il ) P (y1 = 0 . . . yl1 = 0|u, yl = 1, i1 , . . . , il )

= #1
. (3)
y =0 P (yl |u, i1 , . . . , il ) P (y1 = 0 . . . yl1 = 0|u, yl , i1 , . . . , il )
l

We approximate the numerator of Eq. (3) so as to satisfy the policies. First,

to satisfy the first policy, we assume that yl (the event that il is purchased) is
independent of the items shown before, and the first factor becomes as follows:
P (yl = 1|u, i1 , . . . , il ) P (yl = 1|u, il ) .
The probability expressed as the right hand side can be equivalent to the rec-
ommendation score of conventional accuracy-oriented methods, and thus can be
regarded as a personalization factor focused on accuracy. To satisfy the first
and the third policies, we approximate the second factor by assuming the inde-
pendence of y1 = 0, . . . , yl1 = 0 (the event that the first l 1 items are not
purchased) and u (the target user) as:
P (y1 = 0 . . . yl1 = 0|u, yl = 1, i1 , . . . , il )
P (y1 = 0 . . . yl1 = 0|yl = 1, i1 , . . . , il ) .
The probability expressed as the second line is high when il is dissimilar to
the past items, i1 , . . . , il1 , and thus can be regarded as a diversification factor
that ensures diversity. This probability does not depend on any hyperparame-
ters controlling diversity. Additionally, to satisfy the first policy, we assume the
conditional independence of the event that the l th item is not purchased and
the event that the l th item is not purchased given the event that lth item is
purchased as:
l1
=
P (y1 = 0 . . . yl1 = 0|yl = 1, i1 , . . . , il ) P (yl = 0|yl = 1, il , il ) .
l =1

We also approximate the denominator of Eq. (3) in the same manner as the
numerator.
In summary, the probability that u will purchase the lth item from the list
can be expressed as follows:
P (yl = 1|u, il , y1 = 0, . . . , yl1 = 0, i1 , . . . , il1 )
?
P (yl = 1|u, il ) l1 l =1 P (yl = 0|yl = 1, il , il )
#1 ?l1 . (4)

y =0 P (yl |u, il )
l
l =1 P (yl = 0|yl , il , il )
352 Y. Kabutoya et al.
7 ;
The number
7 of
; parameters in this model is O U N + N 2 , which is far smaller
than O U N l , the number in the model expressed by Eq. (2).
We approximate the personalization factor as follows without considering po-
sition l:

P (yl = 1|u, il = i) P (y = 1|u, i) , (5)

where y is a binary variable representing the event that item i is purchased or

not (y = 1 if purchased, y = 0 if not). For the second policy, we estimate the per-
sonalization factor by analyzing purchase history via Latent Dirichlet Allocation
(LDA) [2], since it can oer high recommendation accuracy with only purchase
history. With LDA, P (i|u), which is the probability that the user will choose i
from items 1, . . . , N , is estimated instead of P (y = 1|u, i) for recommendations.
For estimating the probability expressed as Eq. (4), we formulate the estimation
of P (y = 1|u, i) using P (i|u). Since the order of the recommendation list for
each user must remain after the conversion, we assume that the ratio between
P (y = 1|u, i) and P (i|u) depends on only the user and is independent of each
item as follows:

P (y = 1|u, i) = u P (i|u) , (6)

In accordance with Bayes rule, u becomes as follows:

P (y = 1|u)
u = #I 2
, (7)
i=1 P (i|u)

where, for the readability, we skip the detailed derivation of u . Once P (y =

1|u, i) is estimated, we can easily estimate P (y = 0|u, i) included in the denomi-
nator of the Eq. (4) as 1 P (y = 1|u, i). With LDA, first we assign latent topics
to the past purchases using collapsed Gibbs sampling [7], and then P (i|u) is
calculated as:

"Z #Mu #U #Mu

m=1 I [zum = z] + m=1 I [zu m = z iu m = i] +
P (i|u) = u =1
#U #Mu ,
z=1
Mu + Z u =1 m=1 I [zu m = z] + N
(8)

where Z is the given number of topics, Mu is the number of us purchases, zum

is the assigned latent topic for the mth purchase of user u, ium is the item in the
mth purchase of user u, I[] is the indicator function, and and are Dirichlet
parameters. Z can be easily determined by cross validation as minimizing the
prediction error metric, which is called perplexity, and Dirichlet parameters
and can be inferred via Minkas fixed point iterations [13]. An expression for
the MAP estimation of P (y = 1|u) using purchase history can be written as:
#N
i=1 I [i iu ] +
P (y = 1|u) = , (9)
N + 2
 A Probabilistic Model for Diversifying Recommendation Lists 353

where is a hyperparameter for smoothing. The numerator of this expression is

the number of items which u purchased.
We approximate the diversification factor without considering positions l and
l as follows:
P (yl = 0|yl = 1, il = i , il = i) P (y = 0|y = 1, i , i) , (10)
where y represents the event that item i is purchased or not (y = 1 if purchased,
y = 0 if not).
For the second policy, we formulate the MAP estimation of this approximated
diversification factor from the purchase history as follows:
#U
I [i iu i / iu ] +
P (y = 0|y = 1, i , i) = u=1
#U , (11)
u=1 I [i iu ] + 2

where is a hyperparameter for smoothing. The denominator of this expression

is the number of users who have purchased item i, and the numerator is the
number of users who have purchased item i but not item i .

4 Related Work
Recently, diversification of recommendation lists has targeted by many researchers
[9, 11, 17, 19, 21] since it is known to be an important factor in improving user
satisfaction through recommendations. [21] proposed a re-ranking algorithm that
uses item taxonomy to realize topic diversification. Their method can control the
trade-o between accuracy and diversity by using a hyperparameter determined in
an actual user experiment. [9] formalized the intra-list topic diversification problem
by addressing a multi-objective optimization problem on diversity and preference
similarity. [11] proposed a method to make diversified recommendations over time,
and discovered the negative correlation between user profile length and the degree
of recommendation diversity. [19] raised the issue of evaluating the novelty and
diversity of recommendations. [17] studied the adaptive diversification problem for
recommendations by connecting latent factor models with portfolio retrieval.
Additionally, in the area of information retrieval and Web search, many re-
searchers have studied the diversity of search results [35, 8, 1416]. In particu-
lar, [4] is strongly related to our work since they are pioneers in the use of the
probabilistic approach to list diversification. Our method can be positioned as
an extension of [4] to recommendations. We expand Eq. (1) to Eq. (2) in the
same manner as the greedy approach to maximize the probability that least one
document in the search results is relevant.

5 Evaluation
5.1 Formal Evaluation
First, we evaluated our method formally in terms of practicality, by comparing it
to conventional methods for diversifying recommendation lists. We pick up three
conventional approaches, which we label Ziegler [21], Hurley [9], and Shi [17].
354 Y. Kabutoya et al.

First, our method and Hurley do not require any manual tasks for generating
information sources, while Ziegler and Shi do. Ziegler uses item taxonomy, which
is manually provided by service developer. Shi uses ratings, which are manually
given by users. On the other hand, our method and Hurley work even if only the
purchase history is available, and thus allow us to skip such manual tasks.
Second, our method and Shi do not use any hyperparameters for controlling
the balance between accuracy and diversity, while Ziegler and Hurley do. Accord-
ing to [21], there is a trade-o between accuracy and diversity, and balancing
the two maximizes end-user satisfaction. This balancing operation varies with
the service and/or the dataset, so a parameter tuning is needed to determine
the optimal balance. In general, such tuning is dicult since it requires end-user
experiments.
To summarize, we formally demonstrated that our method is superior to con-
ventional methods in terms of practicality.

5.2 Experimental Evaluation

Dataset. We experimentally evaluated our method using the MovieLens-1M
dataset4 . This publicly available dataset contains 1M ratings from about 6K
users on about 3.7K items. The data sparseness is 95.5%. Each user in this
dataset has at least 20 ratings. Additionally, genre labels of the items are pro-
vided. There are 18 genres, and each item can be associated with multiple genres.
The average number of genres per items is 1.62. Note that we interpreted the
MovieLens data as purchase data and discarded the ratings.
We evaluated our method by 5-fold cross validation. That is, we randomly
partitioned the data into 5 subsets. Of the 5 subsets, we held-out a single subset
as the test data (for evaluation), and we used the remaining 4 subsets as training
data (for estimating parameters), where we removed movies not appearing in the
training data from the test data and user-item pairs contained in the training
data from the test data. Then we repeated this cross-validation process 5 times,
with each of the 5 subsets used exactly once as the test data. Finally, we averaged
the 5 results to produce a single estimation.

Compared Methods. We compared the following recommendation methods

in terms of recommendation accuracy and diversity:

1. Proposed : This is the method we propose. In this paper, we used hyperpa-

rameters Z = 50, = 0.1, and = 0.1.
2. Non-diversified : Straight collaborative filtering using LDA [2]. The list of-
fered by this method is the same as the list oered by the personalization
factor of our method. In this method, we used the same number of topics
Z = 50.
3. Ziegler : The diversification method proposed by Ziegler et al. [21]. We used
the lists generated by LDA [2] where Z = 50 as the original lists, the metric
4
MovieLens dataset, http://www.grouplens.org/node/73
 A Probabilistic Model for Diversifying Recommendation Lists 355

proposed in [22] to measure similarity between two items, and genre informa-
tion for item taxonomy in calculating the similarity. In this paper, we used
the best settings yielded by the empirical analysis in [21, 22]. Specifically,
propagation factor = 0.75, and diversification factor F = 0.3.

Result. In our experiments, we evaluated the accuracy and the diversity of the
recommendation lists produced by the three methods.
We used precision, micro-averaged recall, and macro-averaged recall curves to
measure accuracy. It is important to use both micro- and macro-averaged recall
[20], as micro-averaged metrics stress performance on prolific users, while macro-
averaged ones target those with few transactions. Both are used in this paper for
fairness and for detail. The formulae of precision@K, micro-averaged recall@K,
and macro-averaged recall@K are given below:
#U #N K
I[i itest
u i iu ]
precision@K =100 u=1 i=1 ,
UK
#U #N K
I[i itest i iu ]
micro-averaged recall@K =100 u=1#Ui=1#N u ,
i=1 I[i iu ]
test
u=1
U #N K
1 " i=1 I[i itest
u i iu ]
macro-averaged recall@K =100 #N ,
U u=1 i=1 I[i iu ]
test

K
where itest
u is the data for u appearing in the test set, and iu is top-K recom-
mended items to u. Figure 1 show the recommendation accuracy of compared
methods via precision, micro-averaged recall, and macro-averaged recall curves.
These results reveal that our diversification also degrades the accuracy as does
Ziegler. The precision curves show that our method is of practical use because
its loss in accuracy at high rank is less pronounced than is true for Ziegler.
Specifically, our method degrades the accuracy less than Ziegler at the high
rank (K 5), but the accuracy curves for our method drastically drop, and
our method is then outperformed by Ziegler at the low rank (K > 8). Compar-
ing micro-averaged and macro-averaged recall, the drop of our method is less
for macro-averaged recall than for micro-averaged recall. This result means that
our method can oer more accurate recommendations to the users who make
few purchases.
To measure diversity, we employed intra-list similarity and original list overlap
as used in [21]. Intra-list similarity is calculated as follows:
U K K
1 "" "
ILS@K = sim(iuk , iuk ),
U u=1
k=1 k =k+1

where iuk is the kth recommendation to u, and sim() is the item similarity
measured by the metric proposed in [22]. The original list overlap can be calcu-
lated by counting the number of recommended items that stay the same after
356 Y. Kabutoya et al.

46
Proposed
44 Non-diversified
42 Ziegler
40

Precision
38
36
34
32
30
28
26
1 2 3 4 5 6 7 8 9 10
K: number of recommendations

(a) Precision@K
10 16
Proposed Proposed
9 Non-diversified 14 Non-diversified

Macro-averaged recall
Micro-averaged recall

8 Ziegler Ziegler
12
7
6 10
5 8
4
6
3
2 4
1 2
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
K: number of recommendations K: number of recommendations

(b) Micro-averaged recall@K (c) Macro-averaged recall@K

Fig. 1. Recommendation accuracy of compared methods

16 10
Proposed Proposed
14 Non-diversified 9 Non-diversified
12 Ziegler 8 Ziegler
Original list overlap
Intra-list similarity

10 7
8 6
6 5
4 4
2 3
0 2
-2 1
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
K: number of recommendations K: number of recommendations

(a) Intra-list similarity@K (b) Original list overlap@K

Fig. 2. Recommendation diversity of compared methods

the diversification. Higher values of these metrics denote lower diversity. Figure
2 shows recommendation diversity of the compared methods based on intra-
list similarity and original list overlap. These figures reveal that our method
can increase diversity even though it uses only purchase history. From Figure
2 (a), we found that our method can significantly lower the taxonomy-driven
pairwise similarity in lists without recourse to item taxonomy. Specifically, in
Figure 2 (a), the curve of our method is closer to the Ziegler curve than to the
 A Probabilistic Model for Diversifying Recommendation Lists 357

3
Proposed
2.8 Non-diversified
2.6 Ziegler
2.4
2.2

DCG
2
1.8
1.6
1.4
1.2
1
1 2 3 4 5 6 7 8 9 10
K: number of recommendations

Fig. 3. Combination of accuracy and diversity based on DCG

Non-diversified curve. Figure 2 (b) demonstrates that both diversification meth-

ods change recommended items drastically. For the top-10 recommended items,
only 3 items remain after diversification.
Finally, we investigate the combination of accuracy and diversity via DCG
[5] using the genres in MovieLens dataset:
#G u
U K
1 "" g=1 I[iuk itest
u g giuk ](1 )qg,k1
DCG@K = ,
U u=1 log2 (1 + k)
k=1

where G is the number of unique genres, iuk is the kth recommendation to u,

u
gi are the genres assigned to item i, qg,k1 denotes the number of items ranked
up to position k 1 that contain genre g in the recommended items to u. is
a constant controlling the magnitude of the penalty placed on recommendation
redundancy. Higher values of indicate larger penalties. In our experiments, we
used the setting discussed based on the ocial TREC judgments for QA tasks in
[5]: = 0.36. Although DCG is originally normalized via IDCG as NDCG,
we skipped normalization because our focus in this experiment was to compare
diversification methods. Figure 3 shows the DCG scores of the compared meth-
ods. This figure shows that our method is competitive with Ziegler at the high
rank (K 5).
Table 1 illustrates three sampled items from this user profile and an example
of the recommendation lists made by the three compared methods. From the
profile, we can see that this user watches various kinds of action movies. Intu-
itively, this example shows that our method produced a moderately diversified
recommendation list: The list includes three action movies, two comedy movies.
Ziegler yielded a similar recommendation list. On the other hand, the list oered
by Non-diversified is overly focused. Specifically, it provides only similar action
movies.
358 Y. Kabutoya et al.

Table 1. Example recommendation results for five items using a user profile. We refer
Ac to Action, Ad to Adventure, An to Animation, Ch to Childrens, Co
to Comedy, Cr to Crime, H to Horror, D to Drama, Rto Romance, SF to
Sci-Fi, T to Thriller, and W to War. ILS means intra-list similarity.

Three samples from this user profile

Aliens (Ac,SF,T,W)
Speed (Ac,R,T)
Jurassic Park (Ac,Ad,SF)

Rank Proposed Non-diversified Ziegler

1 Raiders (Ac,Ad) Raiders (Ac,Ad) Raiders (Ac,Ad)
2 Star Wars V (Ac,Ad,D,SF,W) Star Wars V (Ac,Ad,D,SF,W) Toy Story (An,Ch,Co)
3 Braveheart (Ac,D,W) Indiana Jones (Ac,Ad) Ghostbusters (Co,H)
4 American Beauty (Co,D) Lethal Weapon (Ac,Co,Cr,D) Star Wars V (Ac,Ad,D,SF,W)
5 Ghostbusters (Co,H) Hunt for Red Oct. (Ac,T) Hunt for Red Oct. (Ac,T)
ILS@5 1.998 4.239 0.549

To summarize, we experimentally demonstrated that end-user satisfaction for

the recommendations via our method may be competitive with Ziegler in terms
of accuracy and diversity.

6 Conclusion and Future Work

We developed a new method to diversify recommendation lists. Our key advances
were to naturally derive a diversification method by determining a recommenda-
tion list so as to maximize the probability that a user purchases at most one item
from a list, and and formulate a model and its parameter estimation based on
several policies. In this paper, we formally showed that our method can, unlike
conventional methods, avoid manual tasks such as the collection of information
and the control of the balance between accuracy and diversity. Additionally, ex-
periments on a real-world data set demonstrated that our method is competitive
with the conventional methods in terms of accuracy and diversity. These results
suggest that our method is practically promising.
Future work involves three issues. First, we would like to extend our method
to use recommendation information other than purchase history such as item
metadata automatically drawn from external knowledge bases. Second, we will
examine the eectiveness of diversification parameters, which are not used in
the current model. We would like to formulate a new model that uses a param-
eter to control the trade-o between accuracy and diversity, and automatically
estimate this parameter from user behavior logs. Finally, we would like to apply
our method to an online store and examine how well it can enhance end-user
satisfaction.

Acknowledgements. We would like to thank the GroupLens Research Group

for the use of the MovieLens 1M dataset.
 A Probabilistic Model for Diversifying Recommendation Lists 359

References
1. Adomavicious, G., Tuzhilin, A.: Toward the next generation of recommender sys-
tems: a survey of the state-of-the-art and possible extensions. IEEE TKDE 17(6),
734749 (2005)
2. Blei, D., Ng, A., Jordan, M.: Latent Dirichlet allocation. J. Machine Learning
Research 3, 9931022 (2003)
3. Carbonell, J., Goldstein, J.: The use of MMR, diversity-based reranking for reorder-
ing documents and producing summaries. In: SIGIR 1998, pp. 335336 (1998)
4. Chen, H., Karger, D.: Less is more: probabilistic models for retrieving fewer relevant
documents. In: SIGIR 2006, pp. 429436 (2006)
5. Clarke, C., Kolla, M., Cormack, G., Vechtomova, O., Ashkan, A., B uttcher, S.,
MacKinnon, I.: Novelty and diversity in information retrieval evaluation. In: SIGIR
2008, pp. 659666 (2008)
6. Davidson, J., Liebald, B., Liu, J., Nandy, P., Vleet, T.V.: The YouTube video
recommendation system. In: RecSys 2010, pp. 293296 (2010)
7. Griths, T., Steyvers, M.: Finding scientific topics. PNAS 101, 52285235 (2004)
8. Guo, S., Sanner, S.: Probabilistic latent maximal marginal relevance. In: SIGIR
2010, pp. 833834 (2010)
9. Hurley, N., Zhang, M.: Novelty and diversity in top-N recommendation analysis
and evaluation. ACM Trans. Internet Technol. 10(4), 14:114:30 (2011)
10. Koren, Y.: Collaborative filtering with temporal dynamics. In: KDD 2009, pp.
447456 (2009)
11. Lathia, N., Hailes, S., Capra, L., Amatriain, X.: Temporal diversity in recommender
systems. In: SIGIR 2010, pp. 210217 (2010)
12. Linden, G., Smith, B., York, J.: Amazon.com recommendations: item-to-item col-
laborative filtering. IEEE Internet Computing 7(1), 7680 (2003)
13. Minka, T.: Estimating a Dirichlet distribution. Tech. rep., M. I. T. (2000)
14. Radlinski, F., Dumais, S.: Improving personalized web search using result diversi-
fication. In: SIGIR 2006, pp. 691692 (2006)
15. Santos, R., Macdonald, C., Ounis, I.: Selectively diversifying web search results.
In: CIKM 2010, pp. 11791188 (2010)
16. Santos, R., Macdonald, C., Ounis, I.: Intent-aware search result diversification. In:
SIGIR 2011, pp. 595604 (2011)
17. Shi, Y., Zhao, X., Wang, J., Larson, M., Hanjalic, A.: Adaptive diversification of
recommendation results via latent factor portfolio. In: SIGIR 2012, pp. 175184
(2012)
18. Tosher, A., Jahrer, M., Bell, R.: The BigChaos solution to the Netflix
grand prize (2009), http://www.netflixprize.com/assets/GrandPrize2009 BPC
BigChaos.pdf
19. Vargas, S., Castells, P.: Rank and relevance in novelty and diversity metrics for
recommender systems. In: RecSys 2011, pp. 109116 (2011)
20. Yang, Y., Liu, X.: A re-examination of text categorization methods. In: SIGIR
1999, pp. 4249 (1999)
21. Ziegler, C.N., McNee, S.M., Konstan, J.A., Lausen, G.: Improving recommendation
lists through topic diversification. In: WWW 2005, pp. 2232 (2005)
22. Ziegler, C., Lausen, G., Schmidt-Thieme, L.: Taxonomy-driven computation of
product recommendations. In: CIKM 2004, pp. 406415 (2004)
 A Probabilistic Data Replacement Strategy
for Flash-Based Hybrid Storage System

Yanfei Lv, Xuexuan Chen, Guangyu Sun, and Bin Cui

School of Electronics Engineering and Computer Science, Peking University

Key Lab of High Confidence Software Technologies (Ministry of Education),
Peking University
{lvyf,bin.cui,gsun,xuexuan}@pku.edu.cn

Abstract. Currently, the popularization of flash memory is still limited

by its high price and low capacity. Thus, the magnetic disk and flash
memory will coexist over a long period of time. How to design an ef-
fective flash-hard disk hybrid storage system emerges as a critical issue.
Most of the existing works are designed based on traditional cache man-
agement approaches by taking the characteristics of flash into considera-
tion. In this paper, we revisit the existing hybrid storage approaches and
propose a novel probabilistic data replacement strategy for flash-based
hybrid storage system, named HyPro. Dierent from traditional deter-
ministic approaches, our approach moves the data probabilistically based
on the data access pattern. Such a method can statistically achieve a good
performance over massive memory operations of modern workloads. We
also present the detailed data replacement algorithm and discuss how
to determine the probability of data migration in the storage hierarchy
consisting of main memory, flash, and hard disk. Extensive experimen-
tal results on various hybrid storage systems show that our method can
yield better performance and achieve up to 50% improvements against
the competitors.

1 Introduction
Although most of the people believe that the magnetic disk will be replaced by
flash-based solid state drives (SSD) in the future, currently the popularization of
flash memory is still limited by its high price and low capacity. Hence magnetic
disk and flash memory will coexist over quite a long period of time. From the
comparison in Table 1, flash displays a moderate I/O performance and price per
GB between DRAM and hard disk. Consequently, it is straightforward to adopt
flash memory as a level of memory between the HDD and main memory because
of its advantages in performance [18,19]. The flash-hard disk hybrid storage is
more and more adopted. Seagate provides a mixed storage hard disk with 4GB
flash chip to improve the overall performance [2]. Windows operating system
support Quick Boost from vista to accelerate the booting [1]. In addition, some
companies start to replace some of the hard disk to SSD to build a hybrid storage
system. In this case, how to design an eective flash-hard disk hybrid storage
system emerges as a critical issue.

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 360371, 2013.
Springer-Verlag Berlin Heidelberg 2013
 A Probabilistic Data Replacement Strategy 361

Table 1. Comparison on dierent storage media

Price( ) Capacity(GB) Price( )/GB Read(s) Write(s)

DRAM(DDR3) 23.99 4GB 6.00 1 1
SSD 114.00 64GB 1.78 271 2012
Disk(7.2K) 119.99 1TB 0.12 12700 13700

The data migration among main memory, flash and disk is the most important
issue in hybrid storage design. Many works has been done on this problem.
TAC [7](Temperature-Aware Caching) adopts temperature to determine page
placement. In TAC, a global temperature table is maintained for each page.
The temperature of a page is decided by its access numbers and patterns in a
period. The pages with higher temperature are placed in main memory and flash
memory while a cold page evicted from main memory will be replaced to disk.
In contrast, the LC [11](Lazy Cleaning) and FaCE [13](Flash as Cache Exten-
sion) always cache a page exit from main memory to flash memory. LC handles
flash memory as a write-back cache. The dirty pages are kept on flash first and if
the percentage of dirty pages exceeds a threshold, these pages will be flushed to
disk. Whereas FaCE proposes FIFO replacement for flash memory management
which is an ideal pattern for flash write. In this way, FaCE can improve the
throughput and shortens the recovery time of database.
All the above methods are all in deterministic way. In some cases, such deter-
ministic migration policy is really inecient. For example, in some cases, pages
are only accessed once, and thus it is suboptimal to keep these cold pages on the
flash memory as designed in LC and FaCE. The temperature method in TAC
works well on hot page detection with stable pattern. However, on a workload
changing, TAC takes a rather long time to forget the history and learn the new
pattern. Another disadvantage of TAC is its high time and space consumption.
In order to overcome the problems in prior approaches, we propose a probability-
based policy named HyPro to manage data storage and migration in the storage
hierarchy, which is composed of main memory, SSD and HDD. In our approach,
the priority of data in each level of the hierarchy is maintained separately. The key
dierence from prior work is that the data migration among dierent levels are no
longer deterministic but based on probabilities. Compared to prior deterministic
approaches, HyPro has several advances:

Better management eciency. The probabilistic data migration can be con-

sidered as a statistical frequency-based implementation. Since existing cache
policies can also be applied to SSD in our approach, HyPro is a seamless
combination of the cache management and frequency-based migration poli-
cies, which can achieve better management eciency so that the total I/O
performance can be improved.
Less unnecessary data replacement. For deterministic approaches, there may
be some unnecessary data movements for cold pages. In our stochastic based
362 Y. Lv et al.

scheme, such unnecessary data replacement can be eectively eliminated

with the control of probabilities.
Lower overhead. To achieve the hot page detection, prior work needs to
maintain a global list to identify the hottest and coldest pages. The space
complexity is O(n) and the time complexity of the maintainable no less than
O(log(n)), where n is the number of all pages accessed including those on
the disk. In our approach, the space complexity is negligible and the time
complexity is O(1).

The remainder of the paper is organized as follows. Related work is described

in Section 2. Section 3 describes our framework. Parameter tuning method is
presented in Section 4. Experimental results are shown in Section 5, and we
make a conclusion in Section 6.

2 Related Work

Nowadays, flash-based hybrid storage has been gradually recognized as an eco-

nomical way for a practical system by more and more people. Some hard disks
leverage a small flash memory to improve I/O performance [5]. With the incre-
ment on the capacity, SSD is more and more widely deployed in storage systems.
The early SSD is skilled in reading but uncompetitive in writing. Thus migra-
tion methods [19,14] are proposed to dynamically transfer read intensive pages to
flash and write intensive ones to disk. Recently, SSD thoroughly surpasses disk on
both read and write speed, and hence the popular method is to adopt flash as a
middle-level cache between disk and main memory. Existing work includes static
deployment and dynamic loading. An object placement method [6] is developed
to give a proper deployment for the components of DBMS. By comparing the
object performance on SSD and disk beforehand, those with higher benefit per
size are chosen to place on SSD. Other methods suggest putting certain part of
the system to flash. FlashLogging [8] illustrates that storing the log of DMBS to
flash can largely improve the overall performance. Debnath et al. propose [9,10]
FlashStore and SkimpyStash to discuss the proper way to put the key-value
pair to SSD. The static methods need to know the specific information about
the application and is not self-adaptive to various environments. Some methods
based on dynamically page transferring are proposed. TAC(Temperature-Aware
Caching) [7] is the dynamic version of object placement strategy. It allocates
temperature to the extents according to access pattern and I/O cost and keeps
the data with higher temperature to higher level of the storage structure. Re-
searchers from Microsoft [11] discuss several possible designs for hybrid storage
methods. According to testing the LC(Lazy-Cleaning) is the best design. LC
method shows better performance than TAC on write intensive traces and sim-
ilar on read-intensive traces. hStorageDB [16] adopts semantic information to
exploit the capability of hybrid storage system, which is from another aspect to
solve the hybrid storage problem. FaCE [13] proposes to use the flash in FIFO
manner to improve throughput and provide faster recovery.
 A Probabilistic Data Replacement Strategy 363

3 Probabilistic Framework
3.1 The HyPro Approach
The typical structure of a hybrid storage system is illustrated in Figure 1. All the
data is stored on hard disk and organized as data pages. A page need to be loaded
into main memory before being accessed. Since flash has better performance
compared to disk, it works as the level between main memory and disk. When
a page miss happens in main memory, the flash will be checked first. The disk
is only accessed when the page is not found in the flash.

Main Memory

48 57 69 35 51
sh
Fla
m h
fro l as
ad to
F
Re
rite
W

Read from Disk

Write to Disk
32 23 21 45 36 39 99 81 76 18

Flash

Disk

48 5 69 10 18 32 57 2 21 76 1 45 77 23 35 36 39 99 81 51

Fig. 1. Illustration of hybrid storage system

In this system, the data placement and migration is a critical issue to achieve
better performance. In this part, we introduce our probabilistic approach for
hybrid storage management, named HyPro. The overall structure of HyPro is
shown in Fig 2. In our framework, the pages in main memory and flash memory
are exclusive from each other. In other words, we do not keep a page in flash
memory if it is already in main memory.
In the HyPro, we adopt two probabilities to control the data migration. If
some of the pages on flash are frequently accessed, its better to be elevated
to main memory. We call this process elevation. Once the elevation happens
we have to evict a page from main memory. Obviously, the elevation should be
managed carefully so that the benefits of accessing hot pages can oset the I/O
cost overhead caused by data movement. In our probabilistic data management,
we use a probability named pelevate to control the elevation frequency. As shown
in Fig 3, when a page is accessed, it has the chance of pelevate to be kept in
main memory; otherwise, this page will be evicted on the next data access. It is
obvious that the page has more chance to be elevated if it is more frequently used.
Hence, real hot pages are detected and promoted into main memory statistically
during a long runtime. In each elevation, a cold page in main memory need to
be evicted to flash and placed in the original space of the elevated page.
In the HyPro, some page may be evicted from the main memory, which we
name sinking operation in this paper. At first glance, this page is likely to be
364 Y. Lv et al.

Main Memory
pelevate
1.0
1.0 1.0- psink

Flash

psink

Disk

Fig. 2. Overall of Probabilistic Framework

Main Memory Main Memory

48 57 69 35 51 21 48 57 69 35

pelevate pelevate

32 23 21 45 36 39 99 81 76 18 32 23 51 45 36 39 99 81 76 18
Flash Flash

(a) Before elevation (b) After elevation

Fig. 3. An elevation example

hotter than pages on flash memory, and should replace one flash page. However,
sinking to flash incurs a flash write. Whether the future benefit is worth this
write depends on the cost ratio between flash read and write as well as the
hotness of the page being sunk. An example of sinking is illustrated in Fig 4. We
take a probability psink to control the ratio of sinking to flash. A main memory
evicted page has chance of psink to replace a flash page, otherwise, it will be
discarded directly or written back to disk if it is dirty. Lets see why this works.
We assume the main memory evicted page is M (page 51 in Fig 4 (a)), and the
replaced page from flash is F(page 18 in Fig 4 (a)), respectively. The larger psink
is, the more evicted pages are sunk to flash, and the closer the hotness of M
and F are. Consequently, the benefit of sinking to flash will be small for close
hotness of M and F. By setting proper value of pelevate and psink , we can achieve
a better trade-o between main memory and flash accesses and the overall I/O
cost diminishes. We will talk about the parameter tuning in the Section 4.
The pseudocode of HyPro are listed in Algorithm 1. A structure named frame
is used to store the position of each page, and the frames are organized in a
hash table to facilitate searching. The Algorithm 1 illustrates the routine of
page access. First, the position of the page is determined, and then dierent
operations are conducted according to the page position. The Algorithm 1(line
13) may invoke Algorithm 2. Algorithm 2 loads a page from disk and puts this
page to the right position according to the psink . Although LRU is adopted in
our experiments for main memory and flash memory management, HyPro can
support other strategies such as LIRS [12] and ARC [17]. The HyPro is easy to
 A Probabilistic Data Replacement Strategy 365

Main Memory
48 57 69 35 51

1.0
32 23 21 45 36 39 99 81 76 18
Flash

48 5 69 10 18 32 57 2 21 76 1 45 77 23 35 36 39 99 81 51
Disk

(a) Before sinking

Main Memory Main Memory

1 48 57 69 35 1 48 57 69 35

1.0 - psink psink

1.0 1.0
32 23 21 45 36 39 99 81 76 18 51 32 23 21 45 36 39 99 81 76
Flash Flash 1.0

48 5 69 10 18 ...... 1 45 77 ...... 81 51 48 5 69 10 18 ...... 1 45 77 ...... 81 51

Disk Disk

(b) Sink page to disk (c) Sink page to flash

Fig. 4. A sink example

implement and quick enough for online processing. The time complexity is O(1)
for each page access.
In this paper, we focus on the data migration design. Nevertheless, some
optimizations can be supported in HyPro applied to further improve the per-
formance e.g. by considering the asymmetric I/O and the access pattern (ran-
dom/sequential). For example, if the asymmetric I/O is considered, dierent
probabilities can be allocated to read and write operations respectively, which
can make one write operation equivalent to the eect of n read operations. Fur-
thermore, we could also manage flash in FIFO manner as FaCE to transform
flash space allocation into sequential pattern.

4 Parameter Tuning
The probabilities of transforming data among dierent memory levels are the
crucial part in our stochastic page management policy. In this section, we provide
study on how to automatically tune these probabilities based on the cost
analysis. Table 2 facilitates fast check on the notations used in this section.
Definition 1. For a cache management algorithm, we denote the place of the
page to be evicted as the evict position. (For example, the end of the LRU
queue). Nevict is defined as the total hit number on the evict position.
To begin with, a definition is introduced. Note that we consider the hit number
on a position instead of on a specific page in this definition. For example, if
LRU algorithm is adopted in main memory, the Nevict stands for the number of
accesses on the LRU end. If a read operation is hit on the least recently used
366 Y. Lv et al.

Algorithm 1. The HyPro algorithm

Input: an operation request on page p
1 if p exists in main memory then
2 Perform read/write operation in main memory;
3 else if p exists in Flash memory then
4 if rand() < pelevate then
5 AcquireFreeMemPage();
6 write the evicted page to flash memory if any;
7 read from Flash memory, or just write to main memory;
8 else
9 Perform read/write operation towards Flash memory;
10 end
11 else
12 call OnMemFull algorithm when main memory is full;
13 AcquireFreeMemPage();
14 read from disk, or just write to main memory;
15 end

Algorithm 2. The OnMemFull algorithm

Output: a free page in memory
1 if memory is full then
2 get a victim page from the buer manager;
3 if rand() < psink then
4 AcquireFreeFlashPage();
5 flush the page to Flash memory;
6 end
7 flush the page to disk if dirty;
8 return the page;
9 end
10 return a free page;

Table 2. Parameters used in this section

Parameters Description
Cdr ,Cf r The read time cost of one page on disk, flash re-
spectively
Cdw ,Cf w The write time cost of one page on disk, flash re-
spectively
pelevate The probability to elevate
psink The probability to sink
Rf evict ,Rmevict The number of read hit on the evict position of
flash, main memory respectively
Wf evict ,Wmevict The number of write hit on the evict position of
flash, main memory respectively
 A Probabilistic Data Replacement Strategy 367

page P , Nevict increases. However, at this time page P is moved to LRU head
and a new page named P is moved to LRU end. Then next time we increase
Nevict when the P is accessed rather than P .
Assume that in a certain time period, a page Pf on flash is read Rf times
and written Wf times. Rmevict and Wmevict stand for the read and write times
on the evict position in the main memory during the same time period. The
read and write costs of flash are denoted as Cf r and Cf w .
In HyPro, psink is adopted to balance the eviction to flash and disk. Hence,
we discuss the tuning of psink by comparing the cost of the two cases: 1)evict
the page to flash (Figure 4 (c)) and 2)evict the page to disk (Figure 4 (b)). The
I/O costs of two cases are calculated as follows respectively.
Case 1. Evict page Pmevict to flash, and if flash is full evict the Pf evict to disk
(when dirty). Consequently, Pmevict will be accessed from flash and Pf evict from
disk, and thereby the corresponding I/O cost is:

Csinkf =Rmevict Cf r + Wmevict Cf w + Rf evict Cdr + Wf evict Cdw + Cf w (1)

Case 2. Evict page Pmevict to disk. In this case Pmevict will be accessed from
disk while the Pmevict mentioned in Case 1 is still accessed from flash. Thus the
I/O cost is:

Csinkd =Rmevict Cdr + Wmevict Cdw + Rf evict Cf r + Wf evict Cf w (2)

In the case of Csinkf < Csinkd , it is more I/O ecient to evict a page to flash,
and hence, psink should be increased and vice versa. The above analysis only
takes the I/O cost of page transferred, that is, page evicted and page elevated
into consideration. Actually, the I/O costs of other pages will also be influenced
which are not the primary cost and experiments show that the obtained Csinkf
and Csinkd can deal with parameter adjusting eectively.
The Csinkf and Csinkd are very small and unstable for a single page on a short
period of time. In practice, we accumulate Csinkf and Csinkd on all the evicted
page in a certain window on the trace, so that the psink can be adjusted based
on the comparison of Csinkf and Csinkd on each accumulation. Note that the
calculation and parameter tuning described above need only O(1) time for each
access. The tuning will not significantly increase the overhead of whole strategy.
The tuning of pelevate can be performed in a similar way with psink , and thus
is omitted here due to space limitation.

5 Performance Evaluation
In this section, we conduct a trace-driven simulation to evaluate the eective-
ness of the proposed framework. The traces used include TPC-B, TATP [3] and
making Linux kernel (MLK for short) to further evaluate the performance on
various workloads. TAC and FaCE are chosen as the competitors. The simulation
368 Y. Lv et al.

is developed in Visual Studio 2010 using C#. All experiments are run on a
Windows 7 PC with a 2.4 GHz Intel Quad CPU and 2 GB of physical memory.

5.1 Experimental Setup

We use three traces mentioned above for performance evaluation. The bench-
marks namely TPC-B [4], and TATP [3] are run on PostgreSQL 9.0.0 with
default settings, e.g., the page size is 8KB. Dataset size of both TPC-B and
TATP are 2GB. The MLK is a record of the page accesses of making Linux
kernel 2.6.27.39. We use a tool named strace to monitor theses processes and
obtain the disk access history. Specification on the traces was given in Table 3.

Table 3. Specification on the Traces

Filename Page Number (103 ) Reference Number (106 ) Write Ratio

MLK 97.2 27.2 0.43%
TATP 135.1 2.5 4.59%
TPC-B 35.1 10.7 19.46%

In our experiment, the total I/O time is used as the primary metric to eval-
uate the performance. We employ The samsung SSD (64GB, 470 series) in our
experiment. We obtain its access latency by testing. Access latency of hard disk
is obtained from paper [15]. The parameters used in our experiments are listed in
Table 4, in which F lashsize/P ages denotes the ratio between flash and dataset
size. We fixed the main memory size to the 1% of the dataset size.

Table 4. Experimental Parameters

Parameter Value
Cr , Cw (s) 271, 803 (for SSD)
12700, 13700 (for hard disk)
F lashsize/P ages 1.25%, 2.5%, 5%, 10%, 20%
pelevate 0.001,0.01, 0.015,0.02,...,0.1
psink 0.01, 0.1,0.2,...,0.9

5.2 Parameter Tuning

To begin with, we inspect the eect of parameters on performance of algorithms.

We illustrate how the performance of HyPro varies with psink in Figure 5 (a),
along with the values of Csinkf and Csinkd . The HyPro achieves the best perfor-
mance at around psink = 0.15. psink reflects the chance for a page evicted from
main memory to be stored into flash. A low psink will result in a poor utility
of flash memory. On the contrary, an excessive high psink would incurs great
exchange cost between main memory and flash. Thus, our approach achieves the
best performance with proper psink .
 A Probabilistic Data Replacement Strategy 369

(a) psink tuning on HyPro (b) pelevate tuning on HyPro

Fig. 5. Parameters Tuning

The performance of our approach also significantly varies with the pelevate , as
shown in Figure 5 (b) , and the minimum I/O time is reached when it locates
between 0.01 and 0.02. As the pelevate controls whether a page on flash should be
elevated, large pelvevate will cause more exchanges between main memory and
flash, which may deteriorates the whole performance, while no exchange will
cause pages on flash has no opportunity to get into main memory (corresponding
to the case pelevate = 0). Other testings also show that the best parameter is
often around 0.02 and 0.2, and thus, we adopt these values as the initial value
and use dynamic parameter tuning in the following experiments.

5.3 Comparison with Other Approaches

In this part, we compare HyPro with FaCE and TAC. We test extensive configu-
rations, but only shows some results here in Figure 6 due to the space limitation.
In the following results, the main memory is 1% of the total workload size and
the flash size varies from 1.25% to 20%. Our approach shows similar or better
performance compared with FaCE and TAC approach. Our approach can re-
duce upto 50% of the total I/O time against other competitors. In MLK trace,
FaCE performs the worst, since many pages in MLK are accessed only once, but
the FaCE still cache these pages in flash memory. On TATP, HyPro and Face
has similar performance and better than TAC. Because the TATP is not a very
stable access pattern. FaCE and HyPro can adapt themselves to the workload
changes quickly, but TAC needs a rather long time to learn this change. On
TPC-B trace TAC has the best performance when the flash size is low, this is
because the temperature-based hot detection can accurately discover the hottest
pages, which has superiority when the cache size is low. However, When the flash
size is large, the performance of TAC degrades. This phenomenon is partially
because precisely hot page detection is not necessary for a larger cache size, and
partially because of the write through cache design, as the write ratio is very
high in TPC-B according to Table 3. The HyPro can adapt itself to the flash
size enlargement and shows a good performance in all the cases.
370 Y. Lv et al.

(a) Low-end SSD performance on MLK trace

(b) High-end SSD performance on TATP (c) Low-end SSD performance on TPC-B
trace trace

Fig. 6. I/O performance comparison on benchmark traces

6 Conclusion
In this paper we propose a novel stochastic approach for flash based hybrid stor-
age system management named HyPro. Dierent from the existing deterministic
models, HyPro controls the data migration between devices using two probabili-
ties. One probability describes the chance of which one page will be kept in main
memory after accessed from flash. Another probability is adopted to determine
the place where a page should be put after evicted from main memory. By doing
this, Hypro can achieves lower hard disk access with a little exchange overhead
increment on flash writes. We also developed an approach to determine the prob-
abilities based on cost analysis. The experiments show that HyPro outperforms
other competitors.

Acknowledgments. This research was supported by the grants of Natural Sci-

ence Foundation of China (No. 61272155, 61073019) and MIIT grant 2010ZX01042-
001-001-04.
 A Probabilistic Data Replacement Strategy 371

References
1. http://windows.microsoft.com/en-US/windows-vista/products/
features/performance
2. Momentus XT Solid State Hybrid Drives,
http://www.seagate.com/www/en-us/products/laptops/laptop-hdd/
3. Telecom Application Transaction Processing Benchmark,
http://tatpbenchmark.sourceforge.net/index.html
4. TPC Benchmark B (TPC-B), http://www.tpc.org/tpcb/
5. Bisson, T., Brandt, S.A., Long, D.D.E.: A hybrid disk-aware spin-down algorithm
with I/O subsystem support. In: IPCCC, pp. 236245 (2007)
6. Canim, M., Bhattacharjee, B., Mihaila, G.A., Lang, C.A., Ross, K.A.: An ob-
ject placement advisor for DB2 using solid state storage. PVLDB 2(2), 13181329
(2009)
7. Canim, M., Mihaila, G.A., Bhattacharjee, B., Ross, K.A., Lang, C.A.: SSD buer-
pool extensions for database systems. PVLDB 3(2), 14351446 (2010)
8. Chen, S.: Flashlogging: exploiting flash devices for synchronous logging perfor-
mance. In: SIGMOD Conference, pp. 7386 (2009)
9. Debnath, B.K., Sengupta, S., Li, J.: Flashstore: High throughput persistent key-
value store. PVLDB 3(2), 14141425 (2010)
10. Debnath, B.K., Sengupta, S., Li, J.: Skimpystash: RAM space skimpy key-value
store on flash-based storage. In: SIGMOD Conference, pp. 2536 (2011)
11. Do, J., Zhang, D., Patel, J.M., DeWitt, D.J., Naughton, J.F., Halverson, A.: Tur-
bocharging DBMS buer pool using SSDs. In: SIGMOD Conference, pp. 11131124
(2011)
12. Jiang, S., Zhang, X.: LIRS: an ecient low inter-reference recency set replacement
policy to improve buer cache performance. In: SIGMETRICS, pp. 3142 (2002)
13. Kang, W.-H., Lee, S.-W., Moon, B.: Flash-based extended cache for higher through-
put and faster recovery. Proc. VLDB Endow. 5(11), 16151626 (2012)
14. Koltsidas, I., Viglas, S.: Flashing up the storage layer. PVLDB 1(1), 514525 (2008)
15. Lee, S.-W., Moon, B.: Design of flash-based DBMS: an in-page logging approach.
In: SIGMOD Conference, pp. 5566 (2007)
16. Luo, T., Lee, R., Mesnier, M.P., Chen, F., Zhang, X.: hStorage-DB: Heterogeneity-
aware data management to exploit the full capability of hybrid storage systems.
CoRR abs/1207.0147 (2012)
17. Megiddo, N., Modha, D.S.: ARC: A self-tuning, low overhead replacement cache.
In: FAST (2003)
18. Ou, Y., H arder, T.: Trading memory for performance and energy. In: Xu, J., Yu,
G., Zhou, S., Unland, R. (eds.) DASFAA Workshops 2011. LNCS, vol. 6637, pp.
241253. Springer, Heidelberg (2011)
19. Wu, X., Reddy, A.L.N.: Managing storage space in a flash and disk hybrid storage
system. In: MASCOTS, pp. 14 (2009)
 An Influence Strength Measurement via Time-Aware
Probabilistic Generative Model for Microblogs

Zhaoyun Ding1, , Yan Jia1 , Bin Zhou1 , Jianfeng Zhang1, Yi Han1 , and Chunfeng Yu2
1
School of Computer, National University of Defense Technology,
Changsha, 410073, China
{zyding,yanjia,binzhou,jfzhang,yihan}@nudt.edu.cn
2
Naval Aeronautical Engineering Institute Qingdao Branch, Qingdao, Shandong, China
[email protected]

Abstract. Micro-blogging such as Twitter provides users a platform to partici-

pate in the discussion of topics and find more interesting friends. While this large
number can generate notable diversity and not all influence strengths between
users are the same, it also makes measure the influence strength more accurately,
that not only rated as binary friendship relations, challenging and interesting. In
this work, we develop a time-aware probabilistic generative model to estimate the
influence strength by taking the time interval, relationship of following, and the
post content into consideration. In particular, the Gibbs sampling is employed to
perform approximate inference, and the interval of time and the multi-path influ-
ence propagation is incorporated to estimate the indirect influence strength more
microscopically according to the propagation of words. Comprehensive experi-
ments has been conducted on a real data set from Twitter, which contains about
0.26 million users and 2.7 million tweets, to evaluate the performance of our pro-
posed approach. As indicated, the experimental results validate the eectiveness
of our approach. Furthermore, we also observe that the influence strength rank-
ing by our model is less correlative with the method which ranks the influence
strength according to the number of common friends.

Keywords: influence strength, probabilistic generative model, microblogs, so-

cial media, social network.

1 Introduction

Online social networking sites are becoming more popular each day, such as Facebook,
Twitter, and LinkedIn. Among all these sites, Twitter is the fastest growing one than any
other social network. Microblogs, such as Twitter, is a service that has emerged as a new
medium for communication recently. Unlike other social network services, the relation-
ship of following between users can be unidirectional; a user is allowed to choose who
she wants to follow without seeking any permission. Twitter employs a social network
called following relationship; the user whose updates are being followed is called the

This work was supported by NSFC (No.60933005, 91124002);863 (No. 012505,

2011AA010702, 2012AA01A401, 2012AA01A402);242 (No.2011A010);and NSTM
(No.2012BAH38B04, 2012BAH38B06).

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 372383, 2013.
c Springer-Verlag Berlin Heidelberg 2013

 Influence Strength Measurement for Microblogs 373

friend, while the one who is following is called the follower. In case the author is
not protecting his tweets, they appear in the so-called public timeline and his followers
will receive all messages from him.
It is well recognized that networks often contain both strong and weak ties, also
influence strength between dierent users is usually dierent and it is unidirectional
in microblogs, so treating all relationships as equal will increase the level of noise in
the learned models and likely lead to degradation in performance. Recently, influence
strength analysis has attracted considerable research interests. However, most of their
works only considered the network structure or the user behaviors in social network,
and the micro-level mechanisms of influence strength of a user on her followers at a
specific topic, have been largely ignored.
Studies by Kwak et al [1] have shown that Twitter is more likely to be a news media.
Users can not only find friends by microblogs, but also can publish messages by them-
selves, and microblogs transform people from content consumers into content produc-
ers. Usually, there are two aims for users to make use of microblog services, including
finding friends in real life and finding friends according to the homophily. So, it is in-
sucient to measure the influence strength of two users only by the number of common
friends.
+RZDUG
%U\DQW
UHODWLRQVKLS
-DPHV
VWUHQJWK 1DVK
%RVK 
:DGH
'XUDQW
1%$ IROORZLQJ $IWHU
%U\DQW +RZDUG )ULHQG% KRXU
/RV$QJHOHV
/DNHUV
1DVK
 
*DVRO IROORZLQJ
IRRWEDOO
%DUFHORQD $IWHU $IWHU
8VHU$
1%$ 5RQDOGR KRXU
UHODWLRQVKLS 5HDO KRXUV
%U\DQW
VWUHQJWK 0DGULG
0HVVL
 
)ULHQG&

Fig. 1. The correlation of the influence strength Fig. 2. An example of multi-level hierarchial
and the words structure

Moreover, if a user is influenced by her friends, she will retweet (by making use of
an informal convention such as RT @user) or reintroduce (tweets are similar with the
tweets previously posted by other users, but without acknowledgement of a source) the
messages of her friends, resulting that some words in tweets are the same. The more
the same words which are copied by a user are, the higher the influence strength on the
user is. Figure 1 gives an example to illustrate the correlation of the influence strength
and the number of the same words. In the topic of basketball, we can find the user A has
more similar words with her friend B and we infer the user A copes more ideas from her
friend B. In fact, figure 1 indicates two assumptions: 1) Users with similar interests have
a stronger influence on each other, 2) Users whose actions frequently correlate have a
stronger influence on each other, which have been validated in [2] [3].
Also, the interval of time is an important factor to measure the influence strength
and it stands for the diused rate of each word. If a user copes the ideas from her friend
much faster for a topic, it will indicate that the friend has stronger influence on this user.
The interval of time has been ignored in recent studies for influence strength analysis.
374 Z. Ding et al.

In addition, the network in microblogs is a multi-level hierarchial structure, imply-

ing that the influence is transitive. A small network of multi-level hierarchial structure
is illustrated in the figure 2. Two key factors need be considered to measure the indirect
influence strength. 1) The interval of time. Though the user 3 and the user 4 are influ-
enced indirectly by the user 1, the user 3 receives the idea indirectly from the user 1
faster than the user 4. We can infer the user 1 has stronger indirect influence strength
on the user 3. 2) Multi-path influence propagation. In the figure 2, we can find that the
user 4 is influenced indirectly by the user 1 through two users. How to combine the
multi-path indirect influence strength more microscopically is unknow.
In order to measure the influence strength in microblogs, we propose a novel time-
aware generative probabilistic model of a corpus to draw each word in tweet from the
corresponding probability distribution, considering the time interval, the relationship of
following, and the post content. The model is estimated by Gibbs sampling. Moreover,
we take advantage of the interval of time and the multi-path influence propagation to
measure the indirect influence strength more microscopically in microblogs.
Experiments are conducted on a real dataset from Twitter containing about 0.26 mil-
lion users and 2.7 million tweets, and the results show that the model in this paper has
better prediction performance than other methods. Furthermore, we find that the influ-
ence strength ranking of our model is less correlative with the method which ranks the
influence strength according to the number of common friends. Results indicate that
some users with higher influence strength by the number of common friends may be
friends in real life and they form relationships because they know each other in real life,
but they have little interaction in virtual life.

2 Related Work
Liu et al [2] [3] proposed a generative graphical model which leveraged both hetero-
geneous link information and textual content associated with each user in the network
to mine topic-level influence strength. Based on the learned direct influence, they fur-
ther studied the influence propagation and aggregation mechanisms. Guo et al [4] pro-
posed a generative model for modeling the documents linked by the citations, called
the Bernoulli Process Topic (BPT) model, to measure the influence strength in Citation
Networks. Dietz et al [5] devised a probabilistic topic model that explained the genera-
tion of documents; the model incorporated the aspects of topical innovation and topical
inheritance via citations to predict the citation influences.
Our work is similar to their works and we also propose a novel generative prob-
abilistic model of a corpus which leverages both the network of following and post
content associated with each user in the network to measure the influence strength in
microblogs. However, the time is considered in our work. Usually, the influence strength
between users is evolved with the variety of time. Moreover, the interval of time is an
important factor to decide the direct and indirect strength of influence between users.
Xiang et al [6] developed an unsupervised model to estimate relationship strength
from interaction activity (e.g., communication, tagging) and user similarity. More
specifically, they formulated a link-based latent variable model, along with a coordinate
ascent optimization procedure for the inference. However, influence strength analysis
microscopically is ignored in their work.
 Influence Strength Measurement for Microblogs 375

Also, interaction data has been used to predict relationship strength [7] [8] but this
work only considered two levels of relationship strength, namely weak and strong rela-
tionships.

3 Preliminary

Definition 1. (Word) A word is the basic unit of discrete data, defined to be an item
from a vocabulary indexed by {1, ..., v} . Words are represented with unit-basis vectors
that have a single component equal to one and all other components equal to zero.
Thus, the vth word in the vocabulary is represented by a V-vector w such that wv = 1
and wu = 0 for u ! v.

Definition 2. (Post/Tweet/Document) A post w is comprised of a bag of words from a

vocabulary V = {1, ..., v}, and is represented with vector w = (w1 , w2 , ..., wn ), where wi
denotes the occurrence number of word vi in document w.

Definition 3. (Topic) A semantic topic observed in a particular time period is defined

!
as a multinomial distribution of words {p(w|)}wV with the constrain p(w|) = 1.
wV

Definition 4. (Network) The network in microblogs is a directed graph G = (N, E). The
N stands for the set of users in microblogs. And the directed edge E N N stands
for the relationships of following in microblogs. For euv = (u, v) E, if there exists a
directed edge between u and v, euv = 1; otherwise euv = 0.

Definition 5. (Direct and Indirect Influence) Given two user nodes u, v in the network
G = (N, E), we denote Iu (v) R as the influence strength of user v on user u. Fur-
thermore, if euv = 1, we call Iu (v) the direct influence of user v on u; if euv = 0, we
call Iu (v) the indirect influence of user v on u. Because the relationship of following be-
tween users in microblogs can be unidirectional, the influence strength is asymmetric,
i.e., Iv (u) ! Iu (v).

4 Probabilistic Generative Model

Similar to the existing topic models, each document is represented as a mixture over
latent topics. In our model, the content of each document is a mixture of two sources:
1) the content of the given document, and 2) the content of other documents related to
the given document through the network structure. This perspective actually reflects the
process of writing a post (tweet) in microblogs: either depending on users interests or
influenced by one of their friends. when a user writes a post, she may create the idea
innovationally or copy it from one of her friends.
In the model, we use a parameter s to control the influence situation. The s is gen-
erated from a Bernoulli distribution whose parameter is . When s = 1, the behavior is
generated based on users own interests. When s = 0, it means the behavior of the user
is influenced by one of her friends. The parameter of the coin is learned by the model,
given an asymmetric beta prior = ( , ) which prefers the topic mixture of a
376 Z. Ding et al.

'

LII I

I V

LIV  LIV 
(DFKSDLU
WLPHVWDPS

t
LIV 

7RSLFV7 7RNHQVM LIV  7RNHQVL 7RSLFV7

)ULHQGVSRVWVI 8VHUSRVWV'

Fig. 3. Probabilistic generative model

friend document. Thus we need another parameter to select one influencing user v
from friends set L(u).
The key dierences from the existing topic models are that the interval of time and
the relationship of users are combined in our model. The mixture distribution over
topics is influenced by both word co-occurrences and the documents time stamp. A
per-topic Beta distribution generates the documents time stamp. The interval of two
related documents t is drawn according to the negative exponential distribution with
the parameter . Figure 3 shows the graphical structure of the model, which explains
the process of generative each word and time stamp in posts. The key dierences of the
generative process are illustrated in the algorithm 1.

Algorithm 1. Probabilistic generative model

1: for all influenced users documents du D do
2: Draw a friend mixture u = p(v|u)|L(u) dirichlet( ) restricted to v followed by u;
3: Draw an innovation topic mixture du = p(z|du ) dirichlet( );
4: Draw the proportion between tokens associated with friends and those associated with
5: the innovation topic mixture du = p(s = 0|du ) beta( , );
6: for all tokens i do
7: toss a coin sdu ,i bernoulli(du );
8: if sdu ,i = 0 then
9: Draw a friend fdu ,i multi(u );
10: Draw a topic zdu ,i multi( fdu ,i ) from the friends document topic mixture;
11: Draw a word wdu ,i multi(zdu ,i ) from zdu ,i specific word distribution;
12: Draw an interval of time tdu,i Exponential(zdu ,i );
13: Get a time stamp tdu,i = tdu,i + t fdu ,i ;
14: else
15: Draw the topic zdu ,i multi(du ) from the innovation topic mixture;
16: Draw a word wdu ,i multi(zdu ,i ) from zdu ,i specific word distribution;
17: Draw a time stamp tdu ,i Beta(zdu ,i );
18: end if
19: end for
20: end for
 Influence Strength Measurement for Microblogs 377

5 Parameter Estimation via Gibbs Sampling

We employ Gibbs sampling to perform approximate inference. We list the partial update
equations for each variable as below which are same as [2] [3].
Nd, f,s(di , fi , 0) + 1 N f ,z ( fi , zi ) + N f,z,s ( fi , zi , 0) + 1
p( fi | f i , di , si = 0, zi , )
Nd,s (di , 0) + L(di ) 1 N f ( fi ) + N f,s ( fi , 0) + T 1
(1)

N f ,z ( fi , zi ) + N f,z,s ( fi , zi , 0) + 1 Nd,s (di , 0) + 1

p(si = 0| si , di , fi , zi , )
N f ( fi ) + f f,s ( fi , 0) + T 1 Nd (di ) + + 1
(2)

Nd,z,s (di , zi , 1) + 1 Nd,s (di , 1) + 1

p(si = 1| si , di , zi , ) (3)
Nd,s (di , 1) + T 1 Nd (di ) + + 1

The f i represents the friend assignments for all tokens except wdi . And the si repre-
sents the coin assignments for all tokens except wdi . In the update equations, N() is
the function which counts occurrences of a configuration. For example, Nd, f,s (1, 2, 0)
denotes the number of tokens in user 1 that are assigned to friend 2, where the coin
result s is 0.

Next, we need to calculate the p(zi | z i , wi , ti , si = 0, fi , ) and p(zi | z i , wi , ti , di , si =

1, fi , ), where z i represents the topic assignments for all tokens except wdi .
Based on the generation process, we can get the posterior probability of the whole
data set by integrating out the distributions because the model uses conjugate priors.

"p(w, w , z, z , t, t , f, s| , , , , , , , ) =

" p(w, w |z, z , )p(| )p(t, t |, z, z )p(t|, z ) (4)

" p(z, z |s, f, , )p(|" )p( | )
p( f |, L)p(| , L) p(s|)p(|)

Using the chain rule, we can obtain the conditional probability conveniently. According
to p(zi , zi ) = p(z) and the bayesian inference. We can infer as follows.
p(w,w ,z,z,t,t , f,s| , , , , ,,,)
p(zi | z i , wi , ti , si = 0, fi , ) =
" p(w,w , z i ,"z i ,t,t , f,s| , , , , ,,,)

p(w,w |z,z ,)p(| )p(t,t |,z,z ) p(z,z |s, f,,)p(| )p( | )
"
"
p(w,w | z i , z i ,)p(|
" )p(t,t |, z i , z i ) " z i , z i |s, f,, )p(| )p( | )
p(
p(t|,z ) p(w,w |z,z ,)p(| ) p(z,z |s, f,,)p(| )p( | ) (5)
" "
p(t|, z i ) p(w,w | z i , z i ,)p(| ) p( z i , z i |s, f,,)p(| )p( | )
2
1 1 z 1
(1ti ) zi ti i Nw,z (wi ,zi )+Nw ,z (wi ,zi )+ 1 N f ,z ( fi ,zi )+N f ,z,s ( fi ,zi ,0)+ 1
B(1z ,2z )
Nz (zi )+Nz (zi )+V 1 N f ( fi )+N f ,s ( fi ,0)+T 1
i i

p(zi | z i , wi , ti , di , si = 1, fi , ) is derived analogously.

p(zi | z i , wi , ti , di , si = 1, fi , )
(1ti ) zi ti i
2
1 1 z 1
Nw,z (wi ,zi )+Nw ,z (wi ,zi )+ 1 Nd,z,s (di ,zi ,1)+ 1
(6)
B(1zi ,2zi )
Nz (zi )+Nz (zi )+V 1 Nd,s (di ,1)+T 1
378 Z. Ding et al.

After the Gibbs process, we can estimate of the negative exponential distribution.
Y
1#
= 1/ tiy (7)
Y y=1

The Y represents the number of all time intervals in the i topic. And the tiy represents
the yth time intervals in the i topic.
Moreover, we will also obtain the sampled coin si , friends fi , and topic zi for each
word, and the influence strength can be then estimated by Eq.(8), which are averaged
over the sampling chain after convergence. A denotes the length of the sampling chain.
(d,z, f,0)n
Nd,z, f ,s !
A nz exp(nz tin ) +
1 # i=1
Iu (v|z) = d ( f |z) = (8)
A n=1 Nd,z,s(d, z, 0)n + |L(d)|

The ti represents the shortest interval of two same tokens assigning to the same topic
between the user u and the user v.
The equations reflect our assumptions in a statistical way. It indicates that if one user
u copes more words of user v on topic z, then v has a stronger influence on u w.r.t. topic
z. Moreover, if the user u copes ideas of user v faster (has a shorter interval of time)
on topic z, then v has a stronger influence on u w.r.t. topic z. In addition, we can also
illustrate the evolutionary trend of the influence strength for two pair users according to
the time stamp of each user in microblogs.

6 Indirect Influence Strength

The above probabilistic model only measures the direct influence strength, but does not
consider indirect influence strength. In reality, influence can propagate through the net-
work. Liu et al [2] [3] and Guo et al [4] studied the propagation of influence macroscop-
ically. They first measured the direct influence strength between users, then combined
these influence strengths to measure indirect influence strength. However, the micro-
level mechanisms of influence propagation have been ignored by them.

,QIOXHQFH ,QIOXHQFH
SURSDJDWLRQ SURSDJDWLRQ
1%$
%U\DQW +RZDUG %U\DQW %U\DQW
/RV$QJHOHV +RZDUG
/DNHUV IROORZLQJ IROORZLQJ +RZDUG
1DVK 1DVK
*DVRO t 1h *DVRO t 5h

8VHU$ 8VHU% 8VHU&

Fig. 4. Influence propagation through words

In fact, we can study the influence propagation more microscopically according to

the propagation of words. Figure 4 gives an example of the propagation of words. User
C copes two words from the user A through the user B in the topic of basketball. Ac-
cording to the statistical way in the above subsection, we can infer the indirect influence
strength as follows.
 Influence Strength Measurement for Microblogs 379

A
1 # Nd,z,s, f1, f2 , (d, z, 0, f1, f2 , )n +
Iu ( f1 , f2 , |z) = (9)
A n=1 Nd,z,s(d, z, 0)n + |L(d)|
Iu ( f1 , f2 , |z) represents the indirect influence of user f2 on u through the user f1 .
Nd,z,s, f1, f2 , (d, z, 0, f1 , f2 , ) represents the number of the same tokens in user u that are
assigned to both the user f1 and the user f2 . For example, Nd,z,s, f1, f2 , (u, z, 0, f1, f2 , ) = 2
in the figure 4. Analogously, we can get indirect influence strength for further hops.
Moreover, the influence of ideas decreases with the process of propagation. Fortu-
nately, the interval of time can be used to illustrate the process of influence decrease.
Usually, the influence of ideas decreases with the development of the time. If the inter-
val of time is longer, the influence of ideas will decrease significantly. Otherwise, if the
interval of time is shorter, the influence of ideas will decrease insignificantly. We take
advantage of the function of negative exponential distribution to illustrate the process
of influence decrease quantitatively.
Nd,z,s, f1 , f2 , (d,z,0, f1 , f2 )n
!
A nz exp(nz tin ) +
1 # i=1
Iu ( f1 , f2 , |z) = (10)
A n=1 Nd,z,s(d, z, 0)n + |L(d)|
Also, the ti represents the interval of two tokens between two indirect users. For ex-
ample, ti (A, C) = 6h in the figure 4.
Next, we will analyze the multi-path influence propagation in microblogs. In the
figure 2, we can find that the user 4 is influenced indirectly by the user 1 through two
users (the user 2 and the user 5). In order to measure the indirect influence strength
through multi-path, we give an aggregation function to combine multi-path indirect
influence strength.
Iu ( f2 |z) = max(Iu ( f11 , f2 , |z), Iu ( f12 , f2 , |z)) (11)
The u is influenced indirectly by the f2 through two users and f11 f12 . We refer the highest
multi-path indirect influence strength as the final indirect influence strength.

7 Experiments
7.1 Dataset
For the purpose of this study, a set of Twitter data about Chinese-based twitters who
have published at least one Chinese tweet from 2011 04 15 to 2011 07 15 was
prepared as follows. About 0.26 million users and 2.7 million tweets were collected
through the API of Twitter. We begin by describing the dataset we are trying to mine.
As the figure 5 shows, the distributions of users followers and tweets are approximately
power-law, implying that the vast majority of users have a small quantity of followers
and tweets, while a small fraction have a large number of followers and tweets.
In order to measure the influence strength with our model, we extracted tweets and
users who published at least 10 tweets, then extracted their network relationships. Due
to the power-law distributions of users tweets, 3334 users and 168526 network re-
lationships(edges) were extracted from the data set. Our model is trained with hyper
parameters = 0.01, = = 0.1, = 0.1, = 1.0 and 50 topics.
380 Z. Ding et al.

5 5
10 10

The number of users

The number of users

0 0
10 10
0 2 4 6 0 2 4 6
10 10 10 10 10 10 10 10
The number of followers The number of tweets
(a) The distribution of users followers (b)The distribution of users tweets

Fig. 5. The distributions of users followers and tweets

7.2 Evaluation
Predicting User Behaviors by Direct Influence Strength. We apply the derived in-
fluence strength to help predict user behaviors and compare the prediction performance
with other methods. The behavior is defined as whether a user retweets a friends mi-
croblog. To do this, we rank the pairs of users by influence strength and measure the
area under the ROC curve (AUC) based on the feature values for the ranked pair (e.g.,
1 if a user is retweeted by other user, 0 otherwise). We compare the rankings using
influence strength to several alternative rankings (Named our model as TNIM-Time
Network Influence Model):
1) The influence strength between users is calculated as the Kullback-Leibler (KL)
divergence of user distributions over topics (Named as KLD).
1 1
Iu (v) = = ! (12)
KL(u, v) i
ziu log zzui
v
0<iT

2) The influence strength between users is calculated as the number of common friends
(Named as NCF).
|N(u) N(v)|
Iu (v) = (13)
|N(u) N(v)|
3) The influence strength between users is calculated as the methods proposed by Liu
et al [2] [3] (Named as NIM-Network Influence Model).
A
1 # Nd,z, f,s(d, z, f, 0)n +
Iu (v) = (14)
A n=1 Nd,z,s (d, z, 0)n + |L(d)|
In order to predict user behaviors, we divide the data set into two parts. We learn the
influence strength by the data set from 2011 04 15 to 2011 06 30. And we pre-
dict user behaviors by the data set from 2011 07 01 to 2011 07 15. We select
four users and their friends to evaluate our model. The ROC curve of dierent methods
for four users are shown in the figure 6.
Then, we select 1000 users and their friends to compute the average AUC (Area
Under Curve) of above four methods. The results are shown in the figure 7.
Experimental results show that the prediction performance of our model is the best.
The average AUC (Area Under Curve) of our model is 0.0647 higher than the method
of NIM, and is 0.0876 higher than the method of KLD. The prediction performance of
NCF is the worst. We can infer it is not too accurate to predicate the retweets by the
number of common friends and they do not always have common topics in virtual life.
 Influence Strength Measurement for Microblogs 381

1 1
ROC curve of user2
ROC curve of user1

True positive rate (Sensitivity)

True positive rate (Sensitivity)
0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
False positive rate (1Specificity) False positive rate (1Specificity)
1 1
ROC curve of user3 ROC curve of user4

True positive rate (Sensitivity)

True positive rate (Sensitivity)

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
False positive rate (1Specificity) False positive rate (1Specificity)
TNIM Random NIM KLD NCF

Fig. 6. The ROC curve of four users

Correlation Analysis by Direct Influence Strength. In this subsection, we will ana-

lyze the correlation of above three methods and our model by ranking the pairs of users
according to the influence strength. Spearman Rank Correlation is used to analyze the
correlation of two rankings.
! 2
6 (ri1 ri2 )
0<i<n
=1 (15)
n3 n
We select 1000 users and rank their friends according to the influence strength. Then,
we compute the correlation of our model and above three methods according to the
Spearman Rank Correlation. The average correlations of 1000 users are shown in the
figure 8.
Experimental results show that the influence strength ranking of our model (TNIM)
is more correlative with NIM. However, our model is less correlative with NCF. Re-
sults indicate that some users with higher influence strength by the number of common
friends may be friends in real life and they form relationships because they know each
other in real life, but they have little interaction in virtual life.

Predicting User Behaviors by Indirect Influence Strength. In order to evaluate the

prediction performance of indirect influence strength, the behavior is defined as whether
a user retweets a users microblog by multi-hops. For example, if a user A is influenced
by the user B indirectly through the user C, we will estimate whether the relationship
of retweets is as follows: A RT @C RT @B....
Because it is dicult to get indirect influence strength by the methods of the Kullback-
Leibler (KL) divergence and the number of common friends, we only compare our
model with the method proposed by Liu et al [2] [3]. Also, we select 1000 users and
382 Z. Ding et al.

0.8 0.6
0.75
TNIM
0.75 0.5
NIM
0.7

Spearman Rank Correlation

0.4 TNIM
0.7
The average AUC

NIM

The average AUC

0.65
0.3
0.65

0.2 0.6
0.6

0.1
0.55 0.55

0
TNIM TNIM TNIM NIM NIM KLD
0.5 and and and and and and 0.5
TNIM NIM KLD NCF KLD NCF NCF
NIM NCF KLD h=1 h=2 h=3

Fig. 7. The average AUC Fig. 8. The correlation of our Fig. 9. The average AUC of
model and other methods indirect influence strength

their indirect friends to compute the average AUC (Area Under Curve). Experimental
results are shown in the figure 9, where h denotes the hop of indirect influence.
Experimental results show that the prediction performance of our model is better
than the method proposed by Liu et al [2] [3]. Moreover, we find the prediction perfor-
mance of our model drops more slowly with the improvement of the hop. The indirect
influence strength proposed by Liu et al [2] [3] is impacted by the hops of network. And
all indirect influence strengths are computed by Iu (w) Iw (v), where is the concate-
nation function, e.g., multiplication and minimum value. However, the interval of time
is ignored by them. Usually, the up-to-date tweets are recommended to followers, and
followers are influenced more easily by these up-to-date tweets. It is not too accurate
to compute the indirect influence strength, not considering the interval of time. In fact,
Though the hop of network is farther, the interval of multi-hop may be shorter and the
tweet is retweeted more easily by other users. So, the prediction performance of the
method proposed by Liu et al [2] [3] drops faster with the improvement of the hop.

8 Conclusion

A novel time-aware probabilistic generative model is proposed to estimate influence

strength, considering the time interval, the relationship of following, and the post con-
tent. Then, we take advantage of the interval of time and the multi-path influence prop-
agation to to derive the indirect influence strength between nodes. Experimental results
show that our model has better prediction performance than other methods, in despite
of direct and indirect influence strength.
In the future work, we will take advantage of the influence strength to mine influencer
more accurately in microblogs. Also, we can employ the influence strength to viral
marketing and trigger a large cascade of influence more accurately.

References
1. Kwak, H., Lee, C., Park, H., Moon, S.: What is twitter, a social network or a news media? In:
Proc. of the 19th International World Wide Web Conference, WWW 2010, Raleigh, USA, pp.
591600 (April 2010)
2. Liu, L., Tang, J., Han, J., Jiang, M., Yang, S.: Mining topic-level influence in heterogeneous
networks. In: Proc. of the 19th ACM Conference on Information and Knowledge Manage-
ment, CIKM 2010, Toronto, Ontario, Canada, pp. 199208 (October 2010)
 Influence Strength Measurement for Microblogs 383

3. Liu, L., Tang, J., Han, J., Yang, S.: Learning influence from heterogeneous social networks.
Data Mining and Knowledge Discovery 25(3), 511544 (2012)
4. Guo, Z., Zhang, Z., Zhu, S., Chi, Y., Gong, Y.: Knowledge discovery from citation networks.
In: Proc. of the 9th IEEE International Conference on Data Mining, ICDM 2009, Miami,
Florida, USA, pp. 800805 (December 2009)
5. Dietz, L., Bickel, S., Scheer, T.: Unsupervised prediction of citation influences. In: Proc.
of the 24th International Conference on Machine Learning, ICML 2007, Corvallis, Oregon,
USA, pp. 233240 (June 2007)
6. Xiang, R., Neville, J., Rogati, M.: Modeling relationship strength in online social networks.
In: Proc. of the 19th International World Wide Web Conference, WWW 2010, Raleigh, USA,
pp. 981990 (April 2010)
7. Gilbert, E., Karahalios, K.: Predicting tie strength with social media. In: Proc. of the 27th
International Conference on Human Factors in Computing Systems, CHI 2009, Boston, USA,
pp. 211220 (April 2009)
8. Kahanda, I., Neville, J.: Using transactional information to predict link strength in online
social networks. In: Proc. of the 3rd International AAAI Conference on Weblogs and Social
Media, ICWSM 2009, San Jose, California, USA, pp. 7481 (May 2009)
A New Similarity Measure Based on Preference
Sequences for Collaborative Filtering

Tianfeng Shang1,2 , Qing He1 , Fuzhen Zhuang1 , and Zhongzhi Shi1

1
The Key Laboratory of Intelligent Information Processing, Institute of Computing
Technology, Chinese Academy of Sciences, Beijing, 100190, China
2
Graduate School of Chinese Academy of Sciences, Beijing, 100049, China
{shangtf,heq,zhuangfz,shizz}@ics.ict.ac.cn

Abstract. Collaborative filtering is one of the most popular techniques

in recommender systems, and the key point is to find similar users and
items. There are already some similarity measures, such as vector co-
sine similarity and Pearsons correlation coecient, and so on. How-
ever, in some cases, what recommender systems get are not the ratings,
but preference sequences of users on a series of items. For this type of
data, those traditional similarity measures may fail to meet the practical
application requirements. In this paper, a similarity measure based on
inversion is proposed for preference sequences naturally. Based on the
Inversion similarity measure, some structural information of user prefer-
ence sequences is analyzed. By merging average precision and weighted
inversion into similarity computation, a new similarity measure based on
preference sequences is proposed for collaborative filtering. Experimental
results show that the proposed similarity measure based on preference
sequences outperforms the common similarity measures on the datasets
with continuous real numbers.

Keywords: Recommender System, Collaborative Filtering, Similarity

Measure, Preference Sequences, Weighted Inversion, Average Precision.

1 Introduction

Recently, CF-based recommender systems have gradually become one of the

most important recommender systems. Nowadays, they are widely used and
make great success in many applications [2], especially in e-commerce web sites,
such as www.amazon.com, www.netflix.com, www.ebay.com, et al. Besides, lots
of researchers pay more and more attentions to CF techniques of these recom-
mender systems. Generally speaking, CF techniques mainly include three cate-
gories [10]: memory-based algorithms [7], model-based algorithms [5], and hybrid
CF algorithms [9].
In memory-based CF algorithms, they firstly compute similarity between users
or items on the user-item rating matrix, and then generate predictions or rec-
ommendations for objective users. So similarity computation is one critical step
of CF algorithms. At present, the most commonly used ones are vector cosine

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 384391, 2013.
c Springer-Verlag Berlin Heidelberg 2013

 A New Similarity Measure Based on Preference Sequences 385

similarity and Pearson correlation coecient [3]. Vector cosine similarity takes
each user or item as a vector of ratings, and computes the cosine angle formed
by the rating vectors. Pearson correlation coecient measures the degree of two
users or items linearly relate to each other. They have both made great success
in many practical CF applications.
However, in some cases, what recommender systems get are not the ratings,
but preference sequences of users on a series of items. For this type of data,
those traditional similarity measures may fail to meet the practical application
requirements. In this paper, from the point view of user behavior, a similarity
measure based on inversion (Inversion) is proposed for preference sequences nat-
urally. Based on the Inversion similarity measure, some structural information
of user preference sequences is analyzed. By merging average precision (AP) and
weighted inversion into similarity computation, a new similarity measure based
on preference sequences is proposed for collaborative filtering. Finally some ex-
perimental results show that the proposed similarity measure based on preference
sequences can get better performance than the common similarity measures in
preference sequences datasets.
The rest of this paper is organized as follows. Section 2 introduces some pre-
liminary knowledge, including overview of CF recommender systems and simi-
larity measures. Section 3 gives out some analysis on structural information of
user preference sequences, and a new similarity measure based on preference
sequences is proposed for collaborative filtering. Some preliminary experimen-
tal results and evaluations are shown in Section 4. Finally, Section 5 presents
conclusions and future work.

2 Preliminary Knowledge

2.1 Overview of CF Recommender Systems

In CF recommender systems, the fundamental assumption is that if users u and

v have similar ratings on some items, they will also give similar ratings on other
items [10]. Based on the above assumption, the basic form of CF recommender
systems can be described as follows. In a typical CF recommender system, there
are two lists: one is list of M users, {u1 , u2 , ..., uM }, and the other is list of
N items, {i1 , i2 , ..., iN }. The users ratings on items can either be explicit in-
dications, such as real numbers or like/dislikes, or implicit indications, such as
purchases or click-throughs. All the ratings constitute a user-item rating matrix,
usually a very sparse matrix.

2.2 Similarity Measures in CF Recommender Systems

Similarity computation is one of the most important phases of CF recommender

systems. For user based CF recommender systems, the co-rating items that the
two users u and v have both rated should firstly be found out, and then select the
appropriate method to compute the similarity between users u and v. It is similar
386 T. Shang et al.

for item based CF recommender systems. At present, there are many dierent
methods for similarity computation [1], but the most commonly used are vector
cosine similarity (Cosine) and Pearsons correlation coecient (Pearson).

3 A New Similarity Measure Based on Preference

Sequences

In this section, the ratings of one user are seen as a preference sequence and
inversion is employed as a similarity measure of the preference sequences. By
analyzing some problems of the similarity measure based on inversion and some
structural information of preference sequences, a new similarity measure based
on preference sequences for collaborative filtering is proposed by merging average
precision and weighted inversion into similarity computation. Here the similarity
measure based on inversion is firstly introduced as follows.

3.1 Similarity Measure Based on Inversion

From the point view of behavioral psychology, ones behavior on some items (such
as high or low score, purchase or not, interested or disgusted, et al.) can be seen
as a preference sequence, thus the similarity between two users can be measured
by the two preference sequences of their behaviors [6,8]. In some sense, it may
get better performance than the common vector cosine similarity and Pearson
correlation coecient.
Based on the above analysis, a similarity measure based on inversion can be
proposed for collaborative filtering, and thus similarity computation between two
users can be converted to inversion computation of their preference sequences.
First, some definitions are given as follows.
Definition 1 (Inversion) [4]: Suppose that P = (p1 , p2 , , pn ) is a sequence
containing N dierent real numbers, < pi , pj > (i, j = 1, 2, , n, i = j) is called
a pair, the number of pairs at how many times pi > pj occurs in sequence P is
called its inversion, written as t(P ) and shown as Formula (1).
n
" n
"
t(P ) = f (pi > pj ) (1)
i=1 j=i+1

where, f () is an indicator function.

Definition 2 (Inversion of two preference sequences): Suppose that P = (p1 , p2 ,
, pn ) and Q = (q1 , q2 , , qn ) are two preference sequences containing N dif-
ferent real numbers respectively, and there is permutation relationship between
them. The inversion of sequence Q can be calculated by taking P as benchmark.
In detail, sort P in ascending order to get P = (p1 , p2 , , pn ), meanwhile, ac-
cording to the permutation relationship between P and Q, Q = (q1 , q2 , , qn )
can be gotten. Thus, the inversion of sequence Q is called the inversion from Q
to P , written as t(Q P ).
 A New Similarity Measure Based on Preference Sequences 387

According to the above definitions and theorems, similarity between two se-
quences can be calculated by normalizing the inversion between them, as shown
in Formula (2).
2 t(Q P )
Sim(P, Q) = 1 (2)
|P | (|P | 1)
According to Formula (1) and Formula (2), the time complexity of similarity
measure based on inversion is O(n2 ). However, according to merge-sort algo-
rithm [4], similarity measure based on inversion can reduce its time complexity
to O(n log n).

3.2 A New Similarity Measure Based on Preference Sequences

In the above section, similarity measure based on inversion tries to understand
user behaviors by preference sequences. However, the similarity measure based
on inversion may not give the best performance in some cases.
Generally speaking, for high eciency and high accuracy of recommender
systems, not all the items are considered, but just the items that users are more
interested in should be considered in practical recommender systems, that is,
items with higher preference should be taken more attention to. To address the
above problems, we can take position information into account during similarity
computation. According to our research, two kinds of position information should
be considered as follows.
In similarity measure based on inversion, inversion values on dierent posi-
tions are given the same weight. Obviously, this approach is not the most
rational in practical recommender systems. One simple improved approach
is that each inversion value can be weighted by divided by its position re-
spectively. Similar with Formula (1), it can be written as Formula (3).
"n #n
j=i+1 f (pi > pj )
W Inv(P ) = (3)
i=1
i
Besides, there is another interesting phenomenon that some elements of
preference sequences are equal to elements of benchmark sequence on the
corresponding positions. Take preference sequence P = (1, 2, 6, 5, 4, 3) and
benchmark sequence O = (1, 2, 3, 4, 5, 6) as example, the 1st and 2nd ele-
ments of P are equal to elements of O on the corresponding positions re-
spectively, and this phenomenon will impact on similarity between P and
O. However, in similarity measure based on inversion, this phenomenon is
ignored. One way to improve the similarity computation can resort to av-
erage precision (AP) [3], which is an important measure for evaluating the
performance of information retrieval systems, to merge position informa-
tion into similarity computation. Specifically, construct an auxiliary sequence
P = (0, 0, 3, 1, 1, 3), and take element 0 as recall at its position, thus AP
of P can be calculated by Formula (4) to improve the similarity between
objective sequence P and benchmark sequence O.
388 T. Shang et al.

#N
i=1 (p(i) rel(i))
AP (P ) = (4)
N
where N is the length of sequence, p(i) is the precision at cut-o i in the sequence
P , and rel(i) is an indicator function.
According to Formula (3) and Formula (4), two kinds of position information
of each element in preference sequences can be taken into the similarity com-
putation, and a new similarity measure based on preference sequences can be
proposed for collaborative filtering, named similarity measure based on Weighted
Inversion and Average Precision (WIAP), as shown in Formula (5).
Sim(P, O) = (1 W Inv(P )) AP (P ) (5)
where O is the benchmark sequence, W Inv(P ) and AP (P ) are calculated accord-
ing to Formula (3) and Formula (4) respectively. Finally the proposed similarity
measure WIAP can be described as Algorithm 1.

Algorithm 1. Similarity Measure by Merging Weighted Inversion and Average

Precision (WIAP)
Input: Benchmark preference sequence O and objective preference sequence P , each sequence with
N dierent nature numbers from 1 to N .
Output: Sim(P, O), similarity between P and O.

1. Sort O in ascending order under the limit of keeping the permutation relationship between P
and O;
2. Initialize W I = 0, AP = 0, c = 0;
3. For i = 1 to N 1
4. Calculate inv as inversion of P (i);
5. W I = W I + inv / i;
6. If P (i) == i
7. c = c + 1;
8. AP = AP + c / i;
9. End If
10. End For
11. If P (N ) == N
12. c = c + 1;
13. AP = AP + c / N ;
14. End If
15. Normalize W I into [0, 1];
16. Calculate Sim(P, O) = (1 W Inv(P )) AP (P ) and return it;

In Algorithm 1, Step (1-2) are some data preprocessing and initialization.

Step (3-14) are to calculate of W Inv(P ) and AP (P ) according to Formula (3)
and Formula (4), where Step (4-5) are to calculate W Inv(P ), Step (6-9) and
Step (11-14) are to calculate of AP (P ). Step (15) is to normalize W Inv(P ) into
[0,1]. Step (16) is to calculate the final similarity measure WIAP according to
Formula (5).

4 Experiments
In this section, some experiments are conducted on four similarity measures (Co-
sine, Pearson, Inversion and WIAP) to measure the performance of the proposed
 A New Similarity Measure Based on Preference Sequences 389

similarity measure based on preference sequences. The performance of all the

similarity measures are measured by Recall, which is one of the most popular
metrics in recommender systems. In information retrieval, recall is the fraction
of the documents that are relevant to the query that are successfully retrieved,
and it can be given by Formula (6).
)
|{relevant documents} {retrieved documents}|
Recall = (6)
|{relevant documents}|
User-based top-N recommendation algorithm [10] is chosen for prediction and
recommendation computation in the experiments. It firstly identifies the k most
similar users to the objective user. After the k most similar users have been
discovered, their corresponding rows in the user-item matrix are aggregated to
identify a set of items, C, purchased by the group together with average of their
weighted ratings. By sorting each row of set C according to their weights, user-
based CF techniques then recommend the top-N items of set C to the objective
user. In practice, about 10 20 items (such as movies, music, books, et al) can
shown in one web page, so in the experiments, we set N to 15, and set k to 10.

4.1 Experiments on Synthetic Dataset

In this section, some experiments are conducted on the synthetic dataset which
is a 1,000 1,000 matrix generated by Matlab program. To verify the good
suitability of the proposed similarity measure based on preference sequences, the
synthetic matrix is randomly sampled to construct five matrices with dierent
matrix filling ratios from 20% to 100%. The experimental results on the five
matrices are shown in Fig.1.

1
Cosine
Pearson
0.9 Inversion
WIAP

0.8

0.7
Recall

0.6

0.5

0.4

0.3
20% 40% 60% 80% 100%
Filling Ratio

Fig. 1. Recall of Synthetic Dataset on Dierent Filling Ratios

From Fig.1, it can be seen that the proposed similarity measure WIAP outper-
forms all the other similarity measures Cosine, Pearson and Inversion on dierent
matrix filling ratios. In particular, similarity measure WIAP almost recalls all
390 T. Shang et al.

the items on the full matrix. Besides, recall of WIAP can increase rapidly along
with the rise of matrix filling ratio.

4.2 Experiments on Benchmark Datasets

In this section, some experiments are conducted on three benchmark datasets:

Movielens dataset, Jester dataset and BookCrossing dataset. Similar with ex-
periments on synthetic dataset, each benchmark dataset is randomly sampled
to construct five matrices with dierent sampling ratios from 20% to 100%. The
experimental results on three benchmark datasets are shown in Table 1.

Table 1. Recall of Benchmark Datasets on Dierent Sampling Ratios

Sampling Ratio
Dataset Measure
20% 40% 60% 80% 100%
Cosine 0.1444 0.2106 0.3400 0.3853 0.4790
Pearson 0.3343 0.3723 0.4296 0.4473 0.4671
Movielens
Inversion 0.2893 0.2715 0.4195 0.3829 0.3942
WIAP 0.3949 0.3725 0.4311 0.4469 0.4987
Cosine 0.5735 0.5059 0.4715 0.4621 0.4356
Pearson 0.6235 0.5695 0.5805 0.5790 0.5644
Jester
Inversion 0.6012 0.5108 0.4977 0.5037 0.4867
WIAP 0.7302 0.8138 0.9220 0.9647 0.9814
Cosine 0.9222 0.5204 0.2587 0.1566 0.1127
Pearson 0.9970 0.9508 0.8618 0.6873 0.5167
BookCrossing
Inversion 0.9970 0.9529 0.8512 0.6564 0.4664
WIAP 0.9970 0.9534 0.8715 0.6863 0.5128
Cosine 0.5467 0.4123 0.3567 0.3347 0.3424
Pearson 0.6516 0.6309 0.6240 0.5712 0.5161
Average1
Inversion 0.6292 0.5784 0.5895 0.5143 0.4491
WIAP 0.7074 0.7132 0.7415 0.6993 0.6643
1
Average represents average recall of each measure on three benchmark datasets.

In Table 1, it contains four blocks where the front three blocks are original
recalls of three benchmark datasets and the last block is the average recalls of
them. In all the 15 cases of the front three blocks, the proposed similarity mea-
sure WIAP can get much better performance than other similarity measures in
12 cases. In cases of 80% of Movielens dataset and 80% and 100% of BookCross-
ing dataset, Pearson is just a little bit better than WIAP. Actually, the values
of Movielens dataset and BookCrossing dataset are just few discrete numbers
which may be not suitable to be seen as preference sequences. Thus, WIAP may
not get best performance in some cases of them. For the Average block of
Table 1, it can be seen that WIAP can get much better performance than other
similarity measures in all the cases. Besides, for BookCrossing dataset, there
is an interesting phenomenon that on some sparser matrices, all the similarity
measures get higher recalls. That is because the matrices are so sparse that there
are not enough books to be recommended, and the similarity measures can recall
the very few books with higher proportion. Generally speaking, it can be con-
cluded that WIAP outperforms three other similarity measures when measuring
preference sequences.
 A New Similarity Measure Based on Preference Sequences 391

5 Conclusions

In this paper, some structural information of user preference sequences in recom-

mender systems is analyzed, and a new similarity measure based on preference
sequences, named WIAP, is proposed for collaborative filtering by merging AP
and weighted inversion into similarity compuation. Experimental results show
that the proposed similarity measure WIAP trends to get better performance on
the datasets with continuous real numbers. However, there are still many further
research issues. For example, how to improve the performance of WIAP on the
datasets with few discrete numbers may be a valuable research point.

Acknowledgments. This work is supported by the National Natural Science

Foundation of China (No.61175052, 60975039, 61203297, 60933004, 61035003), Na-
tional High-tech R&D Program of China (863 Program) (No.2012AA011003), Na-
tional Program on Key Basic Research Project (973 Program) (No.2013CB329502).

References
1. Ahn, H.: A new similarity measure for collaborative filtering to alleviate the new
user cold-starting problem. Information Sciences 178(1), 3751 (2008)
2. Alsaleh, S., Nayak, R., Xu, Y., Chen, L.: Improving matching process in social
network using implicit and explicit user information. In: Du, X., Fan, W., Wang,
J., Peng, Z., Sharaf, M.A. (eds.) APWeb 2011. LNCS, vol. 6612, pp. 313320.
Springer, Heidelberg (2011)
3. Chowdhury, G. Introduction to modern information retrieval. Facet Publishing
(2010)
4. Cormen, T., Leiserson, C., Rivest, R., Stein, C.: Introduction to algorithms, 3rd
edn. MIT Press and McGraw-Hill (2009)
5. Hofmann, T.: Latent semantic models for collaborative filtering. ACM Transactions
on Information Systems (TOIS) 22(1), 89115 (2004)
6. Jung, S., Hong, J., Kim, T.: A statistical model for user preference. IEEE Trans-
actions on Knowledge and Data Engineering 17(6), 834843 (2005)
7. Linden, G., Smith, B., York, J.: Amazon. com recommendations: Item-to-item
collaborative filtering. IEEE Internet Computing 7(1), 7680 (2003)
8. Park, M.-H., Hong, J.-H., Cho, S.-B.: Location-based recommendation system us-
ing bayesian users preference model in mobile devices. In: Indulska, J., Ma, J.,
Yang, L.T., Ungerer, T., Cao, J. (eds.) UIC 2007. LNCS, vol. 4611, pp. 11301139.
Springer, Heidelberg (2007)
9. Su, X., Greiner, R., Khoshgoftaar, T., Zhu, X.: Hybrid collaborative filtering al-
gorithms using a mixture of experts. In: Proceedings of the IEEE/WIC/ACM In-
ternational Conference on Web Intelligence, pp. 645649. IEEE Computer Society
(2007)
10. Su, X., Khoshgoftaar, T.: A survey of collaborative filtering techniques. In: Ad-
vances in Artificial Intelligence 2009, p. 4 (2009)
 Visually Extracting Data Records
from Query Result Pages

Neil Anderson and Jun Hong

School of Electronics, Electrical Engineering and Computer Science

Queens University Belfast, Belfast, UK, BT7 1NN
{nanderson423,j.hong}@qub.ac.uk

Abstract. Web databases are now pervasive. Query result pages are dynamically
generated from these databases in response to user-submitted queries. Automat-
ically extracting structured data from query result pages is a challenging prob-
lem, as the structure of the data is not explicitly represented. While humans have
shown good intuition in visually understanding data records on a query result
page as displayed by a web browser, no existing approach to data record extrac-
tion has made full use of this intuition. We propose a novel approach, in which
we make use of the common sources of evidence that humans use to understand
data records on a displayed query result page. These include structural regularity,
and visual and content similarity between data records displayed on a query result
page. Based on these observations we propose new techniques that can identify
each data record individually, while ignoring noise items, such as navigation bars
and adverts. We have implemented these techniques in a software prototype, rEx-
tractor, and tested it using two datasets. Our experimental results show that our
approach achieves significantly higher accuracy than previous approaches.

1 Introduction and Motivation

Web sites that rely on structured databases for their content are now ubiquitous. Users
retrieve information from these databases by submitting HTML query forms. Query
results are displayed on a web page, but in a proprietary presentation format, dictated
by the web site designer. We call these pages, query result pages. Figure 1 shows a
typical query result page from HMV.com. On this page each DVD is presented as a
data record. In turn, each data record contains a collection of data items; including the
title, description, release date, availability, price and customer rating, etc. as well as
an image of the product. Any item on the page that does not belong to a data record
is called a noise item, for example, a menu button or an advert. A query result page is
designed for a human to read rather than a computer to process, thus there is no standard
way to automatically extract structured data from the page.
Automatic data extraction is the process of extracting automatically a set of data
records and the data items that they contain from a query result page. Such structured
data can then be integrated with data from other data sources and presented to the user in
a single cohesive view in response to their query. For instance, there is great commercial
demand for comparison shopping search engines. A user may wish to buy a DVD; a
comparison shopping search engine can extract data from many different online stores,

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 392403, 2013.
c Springer-Verlag Berlin Heidelberg 2013
 Visually Extracting Data Records from Query Result Pages 393

Fig. 1. Query Result Page from HMV.com with visual blocks highlighted

integrate the data and display it to the user. Other practical applications include flight
and hotel booking sites, financial product comparisons, property sales and rentals.
In this paper, we focus on the problem of data record extraction, that is, to identify
the groups of data items that make up each data record on a query result page.
Data records on a query result page display regularity in their content, structure and
appearance. They exhibit structural and visual similarities: that is, they form visual pat-
terns which are repeated on the page. This is because data records on the same site
are often presented using the same template. The displayed data is in an underlying
database, and as the data items for each record are retrieved from the database (in re-
sponse to the users query), the same template is used each time to present the record.
Much of the existing work that deals with the extraction of data records is based on the
theme of identifying repeated patterns. The common premise is to find and use the re-
peated patterns of data records. The main differences between the existing approaches
are where they look for these patterns and how they use them in data extraction.
Early approaches [1,4] identify repeated patterns in the HTML source code of multi-
ple training pages from a data source in order to infer a common structure or template,
and use it to extract structured data from new pages from the same source. Other ap-
proaches [3,15] identify repeated patterns in the source code of a query result page
in order to directly extract data records and align data items in them. However, these
approaches are all limited by the rapidly increasing complexity of source code. For in-
stance, the widespread use of Javascript libraries, such as jQuery, can make source code
structurally much more complex; thus these approaches start to fail.
Later approaches [9,16,18,19] use the tag tree representation of a web page to iden-
tify repeated patterns in the subtrees of the tag tree. This representation is useful be-
cause it provides a hierarchal representation of the source code. However, the tag tree
was designed for the browser to use when displaying the page, and unfortunately does
not accurately resemble the structure of the data records on the displayed page. The use
of scripts and other runtime features contribute further to the differences between the
structure of the tag tree and the dynamically displayed web page.
394 N. Anderson and J. Hong

The state-of-the-art approaches [11,14,17] identify visually-repeated patterns using

additional information displayed on a rendered page. However, they all succumb to the
same limitation: they still rely on direct access to either the source code or the tag tree.
The approach in [10] is the first that uses primarily visual features for data record ex-
traction. However, this approach has several other limitations caused by the granularity
of its input data and how it deals with noise items on a web page.
In addition, many of the current approaches [10,14,16,17] start by identifying a sec-
tion of the page, referred to as the data-rich section, which contains all of the data
records. However, correctly identifying the data-rich section can be problematic. There
is a risk that some data items may be omitted from, or unwanted noise items incorrectly
included in, the data-rich section. Furthermore, it is possible to identify the wrong sub-
section of the page entirely.
In this paper, we propose a novel visual approach to data record extraction, which
shows a strong correlation with human intuition. Our approach is guided by how most
humans expect query result pages to be visually presented. This also influences web site
designers, who often present pages to meet human expectations.
A common convention is to place noise items, such as menus or adverts, around the
periphery of the page, reserving the centre of the page for the data records. Further-
more, data items in each data record are displayed together, forming what is known as a
locality constraint. As shown in Figure 1, query result pages follow these conventions.
To make it easy for a human reader to perceive data items in each data record as a
group, designers display the data items repeatedly in the same relative position in each
data record. Furthermore, designers also reserve the same horizontal space (or width)
for each data record. Originally they did so because human readers disliked pages that
required horizontal scrolling [7], therefore, the data records on the page must have a
uniform width so they would not cause the page to exceed the screen width. Indeed this
accepted design constraint contributes greatly to the regular visual appearance of data
records. Technical constraints also have a part to play. Each data record on the page is
displayed using the same template code, which, when rendered, reinforces a repeated
visual appearance and therefore a similar composition for each data record.
A human reader uses the regularity of the visual-repeated pattern and structure that
appears when the data records are displayed on the page to infer semantic relationships
between data items and can therefore group the data items into data records.
In summary, our main contributions in this paper are as follows: First, we remove
the requirement to identify the data-rich section of a query result page. We propose a
novel visual approach which identifies, as a seed block, a single data item, which is a
basic content block in a data record. We then implement our observations on visual and
structural regularity to group together only the data items in each record. Second, our
visual approach directly accesses a rendering engine to retrieve positional information
and visual features of each item on the page, avoiding the need to interpret increasingly
complex HTML source code and tag trees.
The rest of the paper is organised as follows. The fundamentals behind our proposed
approach are presented in Section 2. Solutions for identifying a record block for each
data record on the query result page are described in Section 3 and Section 4. Experi-
mental results are reported in Section 5 and finally, Section 6 concludes the paper.
 Visually Extracting Data Records from Query Result Pages 395

2 Visual Data Record Extraction

In this section, we first introduce the visual block model (VBM) to represent a rendered
web page, which allows us to access the positional information and visual properties
of each item on the page. Next, we present our method to measure the visual similarity
between the visual blocks in the model, followed by our definitions of the spatial rela-
tionships between these visual blocks. Finally, we present an overview of our approach.

2.1 Visual Block Model

The visual block model of a query result page is a product of the tag tree and the Cas-
cading Style Sheet (CSS) of the page. A layout engine generates visual blocks for each
node in the tag tree, according to the instructions contained in the CSS. This process,
called rendering, draws a rectangular box around the minimum boundary of each visi-
ble node on the page. We refer to each rectangular box as a visual block. The position
of each visual block is represented by its four borders in the four directions on the two-
dimensional plane. The plane has its origin at the top-left of the page, with the x-axis
running from left to right and y-axis running from top to bottom.
Some of the visual blocks in the VBM, such as those outlined in blue in Figure 1,
represent the structural components of the query result page. These blocks, which we
call container blocks, can be thought of as the scaffolding that holds the structure of
the page together. Each of these visual blocks contains at least one other visual block,
each of which we call a child block. The larger visual blocks shown in Figure 1, such
as those that appear to surround each data record, contain many other visual blocks.
The remaining visual blocks in the VBM, such as those shown in green in Figure 1,
which we call basic blocks, represent the individual labels and data items displayed on
the query result page. These blocks do not contain any other visual blocks. A container
block for a data record typically contains a number of these basic blocks.
For each visual block, we obtain 23 visual properties that can be used to define the
visual appearance of the visual block. We choose the properties that have historically
enjoyed good cross-browser representation and are the most widely used by web site
designers to define the visual appearance of query result pages, for example, fontWeight.
Our approach uses these properties to determine the visual appearance of each block and
is flexible; additional visual properties from the CSS specification can be incorporated
very easily. We use the WebKit layout engine [8] to render query result pages, however,
our approach is independent of any specific engine.
In contrast to the VIPS Algorithm [2], used in other visual approaches, our VBM
makes no prior assumptions regarding the organisational hierarchy of a query result
page, it simply provides the visual properties for each displayed item. Furthermore, our
VBM has finer granularity of visual blocks than the VIPS algorithm. For instance, it can
represent two consecutive visual blocks containing two visually distinct texts whereas
the VIPS cannot segment the two texts into two visual blocks.

VBM vs. Tag Tree. While the VBM and the tag tree are related, they are not equiva-
lent. The tag tree is a complex representation of the HTML code of the page, created
for the browser to interpret and is only part of the information required to render the
396 N. Anderson and J. Hong

page as the designer intended. We choose to use the VBM in preference to the tag tree
for data record extraction for a number of reasons. First, nodes that are close together
on the tag tree may be spatially far apart on the displayed page and vice-versa. By con-
sidering only the visual blocks in the VBM, our approach can see the result page in
the same way that a human can. Crucially, this is how the page was designed: inferred
relationships, such as a group of data items that form a data record, are much easier
to identify in a visual context. Second, our approach is insulated from developments in
coding practices and standards. Our approach relies on the rendering engine, accord-
ingly we are at liberty to make use of the best engine without the need to adapt our
VBM.

2.2 Similarity between Visual Blocks

Our approach decides if two visual blocks have visual, width or content similarity.
Definition 1. Visual Similarity: Two visual blocks, A and B, are visually similar if all
of the visual properties of both blocks are the same. Let PA = {Pa1 , Pa2 , ..., Pan } be a
set of visual properties of A, and PB = {Pb1 , Pb2 , ..., Pbn } be a set of visual properties
of B. The visual similarity between A and B, Sim(A, B), is defined as follows:
!
1 if Pai = Pbi , i = 1, 2, ..., n;
Sim(A, B) =
0 Otherwise.

For example, as shown in green in Figure 1, the visual blocks that contain the DVD
title in both records on the query result page have the same visual proprieties and are,
therefore, visually similar.
Definition 2. Width Similarity: Two visual blocks have similar widths if the width
properties for both blocks are within a threshold of 5 pixels of each other.
For example, as shown in blue in Figure 1, the visual block that contains all of the data
items of the first record, is the same width as the visual block that contains all of the
data items of the second record. We say these blocks have similar widths.
Our approach also needs to decide if two blocks have block content similarity, that
is, they have similar sets of child blocks. Our observation is that two record blocks
have a high degree of block content similarity. This is because both record blocks are
created from the same template, repeated for each record on the page. For example,
the two record blocks, as shown in Figure 1, contain a large number of visually similar
blocks. In contrast, our observation is that there is little block content similarity between
a record block and a noise block. For example, the container block for the sort by
options, which is the same width as, and directly above, the record blocks in Figure 1,
does not share any visually similar child blocks with the record blocks.
We use a variant of the Jaccard index [13] to measure block content similarity be-
tween two visual container blocks. The index ranges between 0 and 1, where 1 means
that the two blocks are identical, and 0 means they have nothing in common. In our
approach, we consider that each container block contains a set of child blocks. We can
then measure block content similarity between two container blocks by a similarity
index between two corresponding sets of visual child blocks.
 Visually Extracting Data Records from Query Result Pages 397

Definition 3. Block Content Similarity: Two visual container blocks have similar block
contents if they have a similarity index above a preset threshold.

For example, as shown in blue in Figure 1, the visual block that contains all of the data
items of the first record has a large number of child blocks that have the same visual
appearance as those child blocks in the visual block that contains all of the data items
of the second record. We say these blocks have block content similarity. In Section 4.2,
we formalise our usage of the similarity index.

2.3 Spatial Relationships between Visual Blocks

We now define a number of spatial relationships between visual blocks. The structural
regularity of data records on a query result page can be recognised by identifying these
spatial relationships between data items in data records.
Definition 4. Contains: A visual block contains another if all of the borders of the later
are inside those of the former.

For example, as shown in Figure 2, block F contains blocks A, B, C, D and E.

Fig. 2. Examples of Spatially Related Vi- Fig. 3. A Clockwise Ulam Spiral Encoun-
sual Blocks tering a Basic Block

2.4 Data Record Extraction

Each data record on the page is represented by a visual block, which contains all of
the contents of the data record and nothing else. We have completed a survey of 600
query result pages and found that in over 98% of cases there is a single visual block that
exactly contains each data record. The goal of our approach is to identify this block for
each data record on a query result page. We call such visual blocks, record blocks.
Our approach starts by identifying a single basic visual block that is very likely to
be one of the basic blocks of a data record, we call this the seed block. The seed block
is contained in a set of larger blocks, which we call the candidate record blocks, as
only one of them is the record block for the data record that the seed block is in. Our
approach must select the correct candidate record block as the record block.
398 N. Anderson and J. Hong

3 Seed Block Selection

The goal of seed block selection is to identify a single basic visual block from the VBM
which is part of a single data record. Let us look at the organisation of a query result
page. Since the Western reading order is from top to bottom, and from left to right,
it follows that the data records should start in the top left of the page. However, most
web pages have common navigation menu and header structures that appear around the
edges of the page. Thus, the starting point of the data records can be shifted down and to
the right. As human readers expect this convention, they start looking for data records
in this area of the page. A study on web usability by eye tracking [12] confirms that the
highest priority area for content is between the centre and the top left of the page.
Our approach starts at the centre of the page, furthest from the noise blocks at the
edges and closest to the highest priority area for data records. We trace a clockwise
Ulam Spiral [5], as shown in Figure 3, which naturally grows from the centre towards
the top left of the page. This is the area of the page most likely to contain data records.
The Ulam Spiral was specially selected as it covers the largest possible proportion of
the highest priority area, before it reaches the edges of the page. A simple plane between
the centre of the page and the top left corner of the page has, on the other hand, the
potential to miss the basic blocks belonging to sparsely populated data records. Instead
it could quickly reach the edge of the page and select as the seed a basic block belonging
to a noisy feature such as a left menu.
The exponential growth of the spiral combined with its direction of travel (clockwise)
ensures that it shows bias to the area between the centre and the top left of the page
thereby covering more of the highest priority area than is possible for a simple plane
cover. As shown in Figure 3, the spiral terminates when it first encounters a basic block.
This block is taken as the seed block.

4 Data Record Selection

The goal of data record selection is to identify a set of container blocks from the VBM,
one block for each of the data records on the query result page.
The seed block is contained inside a number of container blocks, each of which pro-
vides a structure on the page. Examples of these container blocks are shown in Figure 1,
highlighted in blue. By isolating only the container blocks in which the seed block is
actually contained, our approach identifies the set of candidate record blocks, as shown
in Figure 1, highlighted in red. We observe that one of these candidate record blocks is
the record block for the data record and furthermore, this visual block has the similar
width to the record block for each of the records on the same query result page.

4.1 Getting Candidate Record Blocks and Clustering Container Blocks

The seed block is a basic block, which is contained inside one or more of the container
blocks. As shown in Figure 1, the seed block, which was selected in the previous step, is
contained inside four container blocks, highlighted in red. As one of these blocks is the
record block, all four are taken as the candidate record blocks. Next, our approach filters
 Visually Extracting Data Records from Query Result Pages 399

the set of all container blocks on the page, discarding any block that is not the same
width as one of the candidate record blocks. Our approach uses a one-pass algorithm to
cluster the filtered container blocks into a strict partition based on block width. This step
creates a number of clusters, one of which contains the record blocks. In our example,
the algorithm would create four clusters, one for each of the candidate record blocks.

4.2 Measuring Block Content Similarity

Our approach uses a similarity measure to determine if two container blocks have sim-
ilar block contents. Assume that block A contains a set of child blocks {a1 , a2 , ..., am }
and block B contains a set of child blocks {b1 , b2 , ..., bn }.
It is reasonable to expect that a container block may contain more than one child
block with the same visual properties. For example, a data record on a car sales web
site may contain an individual child block for each feature of the car, such as the engine
size, number of doors and fuel type. These child blocks could share the same visual
properties. Accordingly, our approach uses a multi-set representation for each container
block. This generalisation of the notion of a set, in which members may appear more
than once, allows our approach to represent the child blocks of a container block.
Our approach uses a one-pass algorithm to cluster each of the child blocks contained
in a set into a strict partition based on their visual similarity. The function Sim, defined
in Definition 1, is used for this purpose. Two child blocks, a and a , are clustered to-
gether if Sim(a, a ) = 1. So a set of child blocks, A, is clustered into a strict partition
Ac = {A1 , A2 , ..., Am }.
For instance, assume that container block A contains four child blocks, a1 , a2 , a3
and a4 . Two child blocks, a1 and a2 , are visually similar, while a3 and a4 are visually
distinct. The corresponding multi-set, A, is clustered into a set of subsets, Ac , where
each subset in Ac represents a single cluster:

Ac = {{a1 , a2 }, {a3 }, {a4 }} (1)

Select one child block from each cluster in Ac as its representative, so we have a set of
representative child blocks, Ax , for Ac :

Ax = {x1 , x2 , ..., xm } (2)

Given two sets of visual blocks, A and B, we use our similarity measure to find the
block content similarity between A and B, defined as follows:
|A B|
SimBlockContent(A, B) = (3)
|A B|
We define an indicator function for Ax as follows:

1Ax (xi ) = |Ai | xi Ax Ai Ac , f or i = 1, 2, ..., m (4)

Assume that B is clustered into Bc and By = {y1 , y2 , ..., yn } is a set of representative

child blocks for Bc . We define an indicator function for By as follows:

1By (yj ) = |Bj | yj By Bj Bc , f or j = 1, 2, ..., n (5)

400 N. Anderson and J. Hong

We then have:
m "
" n
|A B| = min{1A(xi ), 1B (yj )}|Sim(xi , yj ) = 1 (6)
i=1 j=1

m "
" n
|A B| = max{1A (xi ), 1B (yj )}|Sim(xi , yj ) = 1 + |A B| + |B A| (7)
i=1 j=1

where
m "
" n
|A B| = 1A (xi )|Sim(xi , yj ) = 0
i=1 j=1

and
m "
" n
|B A| = 1B (yj )|Sim(xi , yj ) = 0
i=1 j=1

If SimBlockContent(A, B) is above a preset threshold, A and B are considered to

have similar block contents.

4.3 Selecting Record Blocks

Only one of the candidate record blocks represents the container blocks that provide the
structure for each data record on the page. The other candidate record blocks represent
container blocks that are used to provide structure to other areas of the page. By select-
ing the candidate record block, which has content similarity to the maximum number
of container blocks, our approach identifies the blocks for each data record.
However, our example demonstrates a scenario where two candidate record blocks
both correspond to the maximum number of container blocks. In this case the candidate
record blocks C and D in Figure 1 both have similar contents to the same number of
container blocks, as both blocks are present for each displayed record. In the event
of a tie for maximum blocks, our approach selects the candidate record block which
represents the widest container block, as these container blocks provide the structure to
the whole data record (for example, container block labelled D), rather than the internal
structure provided by the smaller blocks (for example, container block labelled C).

5 Experimental Results

We have implemented our algorithms for data record extraction in a software prototype
called rExtractor. In this section we first describe the two datasets that we use to eval-
uate the accuracy of our algorithms, and then discuss the performance metrics used to
interpret the results of these experiments. We compare the performance of our proto-
type, rExtractor, with that of ViNTs [17], a state of the art extraction system available
on the web (our experimental analysis makes use of the online prototype), which is
based on both visual content features and HTML tag structures.
 Visually Extracting Data Records from Query Result Pages 401

5.1 Datasets
In our experiments we use two datasets, DS1 and DS2. DS1 is used to compare our
method with ViNTs [17]. As these is no standard dataset for experimental analysis of
data record extraction techniques, DS1 comprises of web sites taken from the third party
list of web sites contained in Dataset 3 presented in [17]. The web pages contained in
Dataset 3 were downloaded from the Internet in 2004, as a result they are stale, that
is, they are not representative of the layout and appearance of modern web pages. From
a visual standpoint, modern pages are more vivid than older pages: they contain more
images and much more noise in the form of adverts and menus for example. Therefore,
for each web site from Dataset 3 we visited the same web site and downloaded five
modern equivalent query result web pages. DS1 comprises 50 web pages from 50 web
sites, one web page per web site. In total DS1, contains 752 data records.
DS2 contains a total of 500 web pages from 100 web sites, five per site. It is based
on Dataset 1 and Dataset 2 in [17], as well as the list of web sites in [6]. As with the
previous dataset, the source web pages used to compile DS2 are stale. Accordingly, for
each web site included in DS2 from Dataset 1 and Dataset 2, we visited the same web
site and downloaded a modern equivalent query result page. The web sites in [6] are
listed by their query interface, for each web site included in DS2, we used the query
interface to generate five modern query result pages. The web sites in our datasets are
drawn from a number of domains, including Books, Music, Movies, Shopping, Proper-
ties, Jobs, Automobiles and Hotels. In total, DS2 contains 11,458 data records.

5.2 Performance Metrics

We use a number of performance metrics for our experiments.
Ground Truth. The set of data records from all of the web pages collected per web
site.
True Positives. These are data records extracted correctly from all of the web pages per
web site. Ideally, the true positives are the same as the ground truth for each web site.
False Positives. These are data records extracted incorrectly from all of the web pages
per web site. Ideally, the number of the false positives should be zero.
Precision. This is the number of true positives divided by the number of both true and
false positives per web site. The average precision across web sites is calculated by
averaging the precisions of individual web sites.
Recall. This is the number of true positives divided by the number of data records in
the ground truth per web site. The average recall is calculated by averaging the recalls
of individual web sites.

5.3 Experimental Analysis

To undertake the experimental evaluation and comparison of rExtractor and ViNTs, we
first extract manually the data records on each web page in DS1. Second, we run both
rExtractor and ViNTs on all of the pages in DS1 and record the correctly extracted
(true positives) and incorrectly extracted (false positives) data records on each web
402 N. Anderson and J. Hong

Table 1. Results for DS1 and DS2

DataSet Algorithm Avg. Precision Avg. Recall

1 rExtractor 95.5% 96.3%
1 ViNTs 46.3% 48.9%
2 rExtractor 97.0% 95.5%

page. Finally, we calculate the average precision and recall for all of the web sites in
DS1 for both ViNTs and rExtractor. Next, we undertake the same procedure for DS2
and rExtractor, to determine how rExtractor preforms on a larger dataset. The results
for both DS1 and DS2 are presented in Table 1.

5.4 Experimental Analysis

As we can see from Table 1, the performance of rExtractor is considerably better than
that of ViNTs. Our analysis of the results for ViNTs on DS1 found that ViNTs identified
a large number of false positive data records (427 in total). This contributed greatly to
their low precision score. By inspecting these false positives, we discovered that ViNTs
frequently selected the wrong sub-section of the page as the data-rich section section
(DRS). For example, their approach often selected a centrally located, or near centrally
located, page navigation menu (for instance, a large footer menu) and then extracted
each menu item as a false positive record. In these cases, it was then impossible for their
technique to extract the correct data records. Consequently, the number of true positive
data records they identified was also small, which contributed to their low recall score.
This is not a criticism of ViNTs. Their technique worked very well on the web pages
that were available when ViNTs was developed; rather it serves to highlight that modern
web pages are vastly different from older web pages.
Conversely, the seed block technique implemented by rExtractor selected correctly a
child block belonging to a data record in all of the test cases. This demonstrates that our
technique has no need to explicitly identify a DRS. Furthermore, as we can see from
Tables 1 and 2, rExtractor preforms extremely well on both DS1 and DS2 (more than
12,000 data records in total). The experimental results show that the techniques imple-
mented in rExtractor to identify candidate record blocks, measure block similarity and
select record blocks are all robust and highly effective. Therefore, it is more worthwhile
to investigate the cases where rExtractor failed to correctly extract data records.
First, a small number of data records (1.3% in total) were not displayed in record con-
tainer blocks. rExtractor relies on these blocks to define the record boundaries. Close
inspection of web pages containing these data records reveals they use out-dated page
design conventions and are very much the exception rather than the norm.
Second, our content similarity technique prevented the extraction of a small number
of data records. These records were determined to have not enough in common with
other records on the same page. For instance, designers use a different structural layout
and visual appearance to distinguish between a normal record and a featured record. In
the cases where rExtactor failed, it was often because a featured record was excluded.
We see this as an opportunity to develop further our content similarity technique.
 Visually Extracting Data Records from Query Result Pages 403

6 Conclusions
This paper presents a novel approach to the automatic extraction of data records from
query result pages. Our approach first identifies a single visual seed block in a data
record, and then discovers the candidate record blocks that contain the seed. Next the
container blocks on the page are clustered and a similarity measure is used to identify
which of these blocks have similar contents to each candidate record block. Finally, our
approach selects the cluster of visual blocks that correspond to the data records. We
plan to extend our approach so that the similarity threshold, used to determine if two
blocks have similar contents, is set automatically by a machine learning technique.

References
1. Arasu, A., Garcia-Molina, H.: Extracting structured data from web pages. In: SIGMOD Con-
ference, New York, NY, USA, pp. 337348 (2003)
2. Cai, D., Yu, S., Wen, J., Ma, W.-Y.: Extracting content structure for web pages based on
visual representation. In: Zhou, X., Zhang, Y., Orlowska, M.E. (eds.) APWeb 2003. LNCS,
vol. 2642, pp. 406417. Springer, Heidelberg (2003)
3. Chang, C.-H., Lui, S.-C.: Iepad: information extraction based on pattern discovery. In:
WWW Conference, New York, NY, USA, pp. 681688 (2001)
4. Crescenzi, V., Mecca, G., Merialdo, P.: Roadrunner: Towards automatic data extraction from
large web sites. In: VLDB Conference, San Francisco, CA, USA, pp. 109118 (2001)
5. Prime spiral (2012), http://mathworld.wolfram.com/PrimeSpiral.html
6. Tel-8 query interfaces (2004),
http://metaquerier.cs.uiuc.edu/repository/datasets/tel8/
7. Jakob nielsen - usable i.t (2002),
http://www.useit.com/alertbox/20021223.html
8. Webkit - layout engine, http://www.webkit.org/
9. Liu, B., Grossman, R., Zhai, Y.: Mining data records in web pages. In: SIGKDD Conference,
New York, NY, USA, pp. 601606 (2003)
10. Liu, W., Meng, X., Meng, W.: Vide: A vision-based approach for deep web data extraction.
IEEE Transactions on Knowledge and Data Engineering 22, 447460 (2010)
11. Miao, G., Tatemura, J., Hsiung, W.-P., Sawires, A., Moser, L.E.: Extracting data records from
the web using tag path clustering. In: WWW Conference, pp. 981990 (2008)
12. Nielsen, J., Pernice, K.: Eyetracking Web Usability, 1st edn., pp. 97110. New Riders (2010)
13. Real, R., Vargas, J.M.: The probabilistic basis of jaccards index of similarity. Systematic
Biology 45, 380385 (1996)
14. Simon, K., Lausen, G.: Viper: augmenting automatic information extraction with visual per-
ceptions. In: CIKM Conference, New York, NY, USA, pp. 381388 (2005)
15. Wang, J., Lochovsky, F.H.: Data extraction and label assignment for web databases. In:
WWW Conference, New York, NY, USA, pp. 187196 (2003)
16. Zhai, Y., Liu, B.: Web data extraction based on partial tree alignment. In: WWW Conference,
New York, NY, USA, pp. 7685 (2005)
17. Zhao, H., Meng, W., Wu, Z., Raghavan, V., Yu, C.: Fully automatic wrapper generation for
search engines. In: WWW Conference, New York, NY, USA, pp. 6675 (2005)
18. Zhao, H., Meng, W., Yu, C.: Automatic extraction of dynamic record sections from search
engine result pages. In: VLDB Conference, pp. 9891000 (2006)
19. Zhao, H., Meng, W., Yu, C.: Mining templates from search result records of search engines.
In: SIGKDD Conference, New York, NY, USA, pp. 884893 (2007)
 Leveraging Visual Features
and Hierarchical Dependencies
for Conference Information Extraction

Yue You1,2 , Guandong Xu3 , Jian Cao1 ,

Yanchun Zhang2,4 , and Guangyan Huang2
1
Department of Computer Science and Engineering,
Shanghai Jiao Tong University, China
2
Centre for Applied Informatics,
Victoria University, Australia
3
Advanced Analytics Institute,
University of Technology Sydney, Australia
4
University of Chinese Academy of Sciences, China
{yy71107103,cao-jian}@sjtu.edu.cn, [email protected],
{yanchun.zhang,guangyan.huang}@vu.edu.au

Abstract. Traditional information extraction methods mainly rely on

visual feature assisted techniques; but without considering the hierar-
chical dependencies within the paragraph structure, some important
information is missing. This paper proposes an integrated approach for
extracting academic information from conference Web pages. Firstly,
Web pages are segmented into text blocks by applying a new hybrid page
segmentation algorithm which combines visual feature and DOM struc-
ture together. Then, these text blocks are labeled by a Tree-structured
Random Fields model, and the block functions are dierentiated using
various features such as visual features, semantic features and hierarchi-
cal dependencies. Finally, an additional post-processing is introduced to
tune the initial annotation results. Our experimental results on real-world
data sets demonstrated that the proposed method is able to eectively
and accurately extract the needed academic information from conference
Web pages.

Keywords: Information Extraction, Visual Feature, DOM Structure,

Tree-structured Conditional Random Fields.

1 Introduction
As we all know, thousands of conferences, symposiums and workshops are held
all around the world every year. The conference Web site is a main and ocial
platform to share and post related conference information which can be accessed
by people everywhere. With the increase of conference Web pages, it becomes
a cumbersome and time-consuming job for researchers to collect conference in-
formation and keep track of the hot research topics. Furthermore, people some-
times are more interested in discovering the future trends of the research field,

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 404416, 2013.
c Springer-Verlag Berlin Heidelberg 2013

Leveraging Visual Features and Hierarchical Dependencies for Conference IE 405

analyzing the social networks of scholars and the inner relationships among these
conferences. Building a repository of conference information can satisfy all the
above requirements and many value-added services and applications can be de-
veloped based on the repository. The ArnetMiner1 system [1] is a good example of
academic social networking based on the publication information repository. In or-
der to automatically and eectively archive clean and high quality academic data,
it is essential to extract useful academic information from the conference Web
pages, which is more up-to-date, comprehensive and reliable than other sources.
Web information extraction is a classical problem, which aims at identifying
interested information from unstructured or semi-structured data in Web pages,
and translates it into a semantic clearer structure. During the past years, many
techniques such as DOM structure analysis [2], visual based methods [3,4], ma-
chine learning methods [3,4] have been devised for Web information extraction.
Likewise, this paper studies the problem of automatically extracting structured
academic information from conference Web pages.
In the area of conference Web page extraction, most traditional approaches
focused on the use of visual features in the Web page. When people design
conference Web pages, they usually follow some implicit rules about how they
structure certain types of information on the page. So, it is easy to understand
that visual features are in some sense more stable than content features. VIsion-
based Page Segmentation (VIPS) algorithm [5] was proposed by Deng Cai et al
to segment the Web page into text blocks. But the segmentation results of VIPS
on conference Web page sometimes are not satisfied due to the ignorance of the
hierarchical dependencies (see the example as shown in Figure 3), thus we aim
to address this segmentation problem by proposing a new hybrid approach com-
bines these two kinds of features to archive better segmentation results. In order
to better organize the segmented text blocks, classification or block labeling pro-
cedures are then employed, after which we can perform the refined information
extraction on the annotated labels. The state-of-the-art models for sequence la-
beling problems include Support Vector Machine (SVM) [6] and Conditional
Random Fields (CRFs) [7], which mainly rely on the sequence structure. Look-
ing into the conference websites, we observe that apparent hierarchical relations
exist between dierent information blocks. For example, the submission dead-
line information always appears directly after the phrase - Important Dates
in a bold style. This observation gives us a hint that we should make full use of
the hierarchical dependencies to archive a better annotation performance. Due
to the fact that the hierarchical relation in conference Web page forms a tree
structure, we therefore introduce a Tree-structured Conditional Random Fields
model [8,9] to do the annotation work.
In summary, we have the following contributions: (1) we propose a hybrid
page segmentation algorithm, which combines the vision-based segmentation
algorithm with the DOM-based segmentation algorithm; (2) we introduce the
Tree-structured Conditional Random Fields model [8,9] to annotate the text
blocks, which making use of visual features, content features and hierarchical
1
http://arnetminer.org/
406 Y. You et al.

dependencies concurrently; (3) we refine the annotation quality by introducing

some heuristic post-processing rules. Our experimental results on the real world
data sets verify that the proposed integrated approach is highly eective for
extracting academic information from conference Web pages.
The rest of the paper is organized as follows. Related works are reviewed in
Section 2. We overview our approach in Section 3. Page segmentation algorithm
is discussed in Section 4. Text block annotation is described in Section 5. Ex-
periment and evaluation results are reported in Section 6. Finally, in Section 7
we summarize our work and make some conclusions.

2 Related Work
In this section, we present a survey in two aspects: (1) Classification/sequence
labeling problem and (2) Web information extraction.
First, our work is related to a classification or sequence labeling problem in
some sense. Identifying the useful information from Web pages is sometimes
equivalent to annotating the text blocks with dierent predefined labels. There
have been a lot state-of-the-art approaches proposed for this problem, such as
Bayes Network [10], SVM [6], Conditional Random Fields [7] and so on. But
these approaches usually depend on individual features or linear-dependencies in
a sequence of information, while in Web information extraction the information
can be two-dimensional [4] or hierarchically depended [3].
Second, our work belongs to the area of Web information extraction, which
receives a lot of attentions during past years. This work could be categorized
into two types: (1) Template level wrapper induction systems. Several au-
tomatic or semi-automatic wrapper learning approaches based on the templates
have been proposed. Typical representatives are RoadRunner [11], EXALG [12]
and so on. (2) Visual feature assisted techniques. In contrast, the dier-
ent display styles of dierent parts in a Web page provides an additional means
for segmenting content blocks. For example, [13,3] reported that visual features,
such as width, height, font, etc. are useful in page segmentation.
The limitations of existing Web information extraction methods are:
Some rule-based Web information extraction techniques are not scalable for
the heterogeneity of dierent conference Web pages.
Previous studies may be eective for sequence of text blocks as a linear
model. But they are not suitable for conference Web information extraction,
where the information embedded in the page always follows a tree structure.
Traditional methods can only extract information from a single Web page,
but can not integrate the useful information of a conference that are located
in multiple webpages.

3 Overview of Our Approach

The academic information of a specific conference is distributed within a set
of pages of the conference Web site. In general, we are primarily concerned
with three types of academic information: (1) Information about conference
Leveraging Visual Features and Hierarchical Dependencies for Conference IE 407

events (e.g. conference name, time, location, submission deadline and submis-
sion URL. (2) Information about conference topics (e.g. call for papers and
topics of interests). (3) Information about related people and institutes
(e.g. chairs, program committee, authors, companies and universities).

Fig. 1. Overview of our approach

This paper proposes and implements a Web information extraction system

which aims at extracting academic information from conference Web pages. Fig-
ure 1 shows the framework of our approach. We first use a hybrid algorithm to
segment the page into text blocks, then construct a Tree-structured Conditional
Random Fields [8,9] to annotate each text block based on visual features and
content features. Further processing including verification and some heuristic
rules are used to improve the initial annotation results. Finally, we integrate the
whole conference information based on the annotation results.

4 Page Segmentation Algorithm

The first task for web information extraction is to find a good representation for
Web pages. A good representation can make information extraction easier and
more accurate. Through a survey on a large amount of conference Web pages, we
find that although dierent conference Web sites adopt dierent design styles and
layouts, most pages are composed of some common function blocks. Therefore,
we first use the VIPS algorithm [5] to segment web pages into text blocks. VIPS
algorithm makes use of page layout features such as font, color, and size to
construct a vision-tree for a page.
Figure 2 illustrates the layout structure and the vision-based content structure
of the Call for Papers page of VLDB 2011. At the first level, the original web
page has three objects, namely, visual blocks V B1 V B3 and two separators
1 2 , as specified in Fig.2(b). Then we can further construct a sub-content
structure for each sub-web page. For example, V B2 has two o spring objects:
V B2 1 V B2 2 and one separator 12 . It can be further analyzed recursively.
VIPS can obtain good segmentation results for most Web pages, but it may
lose some important information when dealing with some conference Web pages.
For example, the red blocks in Figure 3 shows parts of the segmentation results
of the Important Dates page in VLDB 2011. We can find that it loses the blue
blocks, which contains important submission deadline information. The block
408 Y. You et al.

(a) (b)
Fig. 2. The layout structure and vision-based structure of a conference Web page

lost happens because VIPS algorithm is dependent on visual features of the

page elements too much. From Figure 3, we can find that the display style of
the red blocks adjacent to the lost part is the same (in this case, they are all in
bold), while the lost blocks are displayed in a dierent style (in this case, they
are in non-bold). In other words, VIPS ignores the adjacent blocks, which are
not displayed consistently. The lost of these blocks is intolerable.

Fig. 3. The segmentation result of Important Dates page in VLDB 2011.

After investigating lots of conference Web pages, we find that the most useful
text blocks appear in tag <p>, <table>, <ul>, <h>, <a> or in text node,
while these nodes are always located at the same level of the DOM tree and
adjacent to each other. Therefore, we define these six types of HTML nodes as

Table 1. Rules for filtering content nodes

Rules Description
Rule1: for <p> tag, save node
Rule2: for <h> tag, save node
Rule3: for <ul> tag, if its sibling nodes belong to Content Nodes, save all child <li> nodes
Rule4: for <table> tag, if it has no <table> children, and its sibling nodes belong to
Content Nodes, save all child <tr> nodes
Rule5: for text nodes, there are two cases:
1) if parent is <a> tag, and its sibling nodes belong to Content Nodes, save node
2) if parent is not tag define HTML display information (eg. <style>), save node
Rule6: for <br/> tag, save the node
Rule7: if the all the conditions above are not met, continue deep-first traversal
Leveraging Visual Features and Hierarchical Dependencies for Conference IE 409

Table 2. Rules for redividing and reorganizing nodes

Rules Description
Rule1: for tag <h>, <li>, <tr>, self form a block
Rule2: for tag <p>, redevise it into multiple blocks according to tag <br/>in the node
Rule3: for tag <a>or text node, reorganize multiple continuous nodes into one block
according to tag <br/>

Algorithm 1. Hybrid Page Segmentation Algorithm

Input: a conference Web page P .
Output: an array of segmented texts ST [ ].
1: ST1 [ ] V IP S(P ) {Segment the pages by VIPS}
2: ST1 [ ] RemoveN ontextBlock(ST1 [ ])
3: ST2 [ ] DOMP S(P ) {Analyzing the page by DOM structure}
4: ST [ ] Combine(ST1 [ ], ST2 [ ]) {Combine the results of two algorithms}
5: Function Combine(ST1 [ ], ST2 [ ])
6: for each ST1 [i] in ST1 [ ] do
7: if N otF indInST2 (ST1 [i]) then
8: continue
9: else if LostBlocks(ST1 [i]) then
10: ProcessLost(ST2 [ ], ST1 [i])
11: else
12: ST [ ].add(ST1 [i])
13: end if
14: end for

Content Nodes. A DOM-based page segmentation algorithm has two phases: (1)
performing a depth-first traversal of the DOM tree, and saving useful Content
Nodes based on rules showed in Table 1. (2) re-dividing and reorganizing saved
nodes into semantic blocks based on rules shown in Table 2.
According to the analysis above, we propose a hybrid page segmentation al-
gorithm combining the VIPS and the DOM-based algorithm. Algorithm 1 shows
the detail of the algorithm. First, it segments an input web page into visual
blocks using the VIPS algorithm (line 1). Second, it removes the noise blocks
(navigation blocks, copyright blocks, etc.) using some heuristic rules (line 2).
Third, it segments the input web page using the DOM-based algorithm (line
3). Finally, it combines the text blocks produced by the VIPS and DOM-based
algorithm to archive a more complete segmentation result (line 4).

5 Text Block Annotation

In this section, we use Tree-structured Conditional Random Fields (TCRFs) [8,9]
to annotate the text blocks outputted by the segmentation algorithm described
in Section 4.

5.1 Problem Definition

The input of our problem is a set of Web pages related to a conference, each
Web page can be parsed into a sequence of text blocks, namely B. And the
output is the structured academic information (as defined in Section 3) about
the conference. In order to achieve eective Web information extraction, we first
annotate the sequence of text blocks according to the display style and content
information of text block. The label space L is defined in Table 3.
410 Y. You et al.

Table 3. Definition of Label Space L

Labels Description
Title (TI): Refers to the text blocks shown in bold and larger fonts.
Topic (TO): Describes the topics of interest information of the conference.
List no text Text blocks displayed in list, but do not contain useful information.
(LNT):
Date Item (DI): This label describes a time table of a conference process.
Date (D): This label represents a specific time in the Web page.
Item (I): It usually appears simultaneously with Date label, and it describes a process of
the conference. Continuous Date category and Item category equal to a Date
Item category in semantics, both describe the timing of a conference.
Committee Per- It refers to the text blocks describing information about organizers or program
son (CP): committee. This label always includes name, aliation, country and so on.
Yes Text (YT): It refers to the text blocks of natural language, and it may contain useful
information.
No Text (NT): It refers to the text blocks of natural language, but do not contain useful
information.

Definition 1. (Text block annotation): Given a sequence of text blocks b

corresponding to a page, let b = {b0 , b1 , , bn } be the features of all the blocks
and each bi is a feature vector of one block, let l = {l0 , l1 , , ln } be one
possible label assignment of the corresponding blocks and each li L defined
above. The goal of conference Web data extraction is to compute a maximum
posteriori (MAP) probability of l and extract data from this assignment l :

l = argmax p(l|b). (1)

Based on the above definition, the main goal is to calculate p(l|b). Here, we
introduce TCRFs [8,9] to compute it, which will be described in the following.

5.2 Tree-Structured Conditional Random Fields

Conditional Random Fields (CRFs) [7] is a probabilistic model proposed by John

Laerty to segment and label sequence data. For understanding the formula, we
define the following notations: X is a random variable over text blocks outputted
by segmentation algorithm, and Y is a random variable over corresponding an-
notations. All components Yi of Y are assumed to range over label space L.
CRFs construct a conditional model p(Y |X) based on a given set of features
from text blocks and corresponding annotations. The definition of CRFs is given
below, for more details please refer to the original paper [7].
Definition 2. (Conditional Random Fields [7]): Let G = (V, E) be a
graph and Y is indexed by the vertices of G, i.e. (Yv )vV . Then (X, Y ) is a
conditional random field, when conditioned on X, Yv obey the Markov property
with respect to the graph: p(Yv |X, Yw , w = v) = p(Yv |X, Yw , w v), where
w v means that there is an edge between w and v in G.

Tree-structured Conditional Random Fields [8,9] is a particular case of

CRFs, whose graphical structure is a tree. TCRFs can model the parent-child
and sibling dependencies, while the simplest Linear-chain CRFs [7] cant do.
Leveraging Visual Features and Hierarchical Dependencies for Conference IE 411

We define (yp , yc ) as a parent-child dependency, (yc , yp ) as a child-parent de-

pendency, and (ys , ys ) as a sibling dependency. A TCRFs model has the form:

1 # #
p(y|x) = exp j tj (e, y|e , x) + k sk (v, y|v , x) ,
Z(x)
e{E pc ,E cp ,E ss },j vV,k
(2)
where x is a block sequence, y is annotation result, and y|e and y|v denotes
y associated with edge e and vertex v in the tree; Epc is set of (yp , yc ), Ecp
is set of (yc , yp ), and Ess is set of (ys , ys ); tj and sk are feature functions;
j and k are parameters corresponding to tj and sk respectively, and will be
estimated from the training data; Z(x) is the normalization factor.

5.3 Conference Web Page Annotation

Text Block Tree Representation. In order to apply TCRFs to conference
Web page annotation, we construct a tree representation of the text blocks in
a Web page. After investigating lots of typical top conference Web sites such as
VLDB, CIKM, KDD etc. we find that almost all the Web pages follow the rule
that the useful information displaying below titles are more obvious than other
blocks. For example, the time of a conference event always follows an Important
Dates block which belongs to Title (TI) label. Based on this observation, we
can construct a text block tree to represent a Web page. The root of the tree
represents the whole page. Each inner node represents the block displayed in a
more highlighted style (usually in a larger font and in bold style). Each leaf node
represents the block display below the inner node. Figure 4 shows the text block
tree representation of the Web page corresponding to Figure 2(a). In addition
to the parent-child dependencies, we also model the sibling dependencies in the
text block tree. For example, topics of a conference are always located close to
each other, and always in the same level of the text block tree.

Fig. 4. The text block tree of the Web page showed in 2(a)

Features in Annotation Model. After studying characteristics of each la-

bel defined in label space L, two kinds of features are chosen to represent a
text block, e.g. visual features and content features. The visual features include
FontSize, FontWeight, FontColor, StartWithH and so on. The content features
specify whether the text block contains month, country, institute, topic keyword
and other keywords. These features belong to the vertex features. We also take
412 Y. You et al.

edge features into account. We use f (bp , bc ), f (bc , bp ), and f (bs , bs ) to denote
the current text blocks parent-child dependency, child-parent dependency and
sibling dependency, respectively.

Parameter Estimation. Parameter estimation is to determine the parameters

= {1 , 2 , ; k , k+1 , } from the training data D = (x(i) , y (i) )
along with empirical distribution p(x,
y). Actually, we optimize the log-likelihood
function of the conditional model p(y|x, ):
#
L = (i) , y (i) ) log p (y(i) |x(i) )).
p(x (3)
i

We use f to represent both edge feature function and vertex feature function; c
to represent both edge and vertex; and to represent two types of parameters:
and . Thus, the derivative of log-likelihood function with respect to parameter
j associated with clique index c is:

( )
L # # (i) (i)
## (i) (i)
= fj (c, y(c) , x ) p(yc |x )fj (c, y(c) , x ) , (4)
j i c y c

(i)
where y(c) is the annotation result of clique c in x(i) , and y(c) ranges over
annotation results to the clique c. p(y(c) |x(i) ) requires the computation of the
marginal probabilities. We utilize the Tree Reparameterization (TRP) algorithm
[14] to compute the approximate probabilities of the factors. The function L
can be optimized by several techniques; but in this paper, we adopt gradient
based Limited-memory BFGS method [15], which outperforms other optimiza-
tion techniques for linear-chain CRFs [16].

Post-processing. The text blocks annotation results may directly aect the
quality of the final conference academic information extraction result. In order
to improve the quality of the annotation results, the post-processing is essential.
Known from the conference Websites, the text blocks can be divided into five
groups according to their display characteristics and contents, i.e., Title, Data
Item/Date/Item, Topic/List no text, Committee Person, Yes Text/No Text. We
summarize some heuristic rules for post-processing for dierent label groups,
and repair text blocks with wrongly annotated labels from the following two
aspects automatically. (1) A block has special features but cannot be annotated
correctly. For example, given a text block Camera Ready Papers Due: Thursday,
August 11, 2011, which have typical DI features. However, this block is classified
as YT. For this situation, it can be repaired by identifying some typical combina-
tion of features. This method is suitable for text blocks with clear features, such
as DI/D/I, TO and CP. (2) A text block is classified as a category but it does
not contain corresponding features. For example, text blocks about submission
instruction are classified as YT. Although it contains a lot of words, it usually
does not contain any people name, date or location. Therefore, we can check its
Leveraging Visual Features and Hierarchical Dependencies for Conference IE 413

feature values ContainMonth, ContainURL and Country Keyword to determine

whether it is a YT category. Due to the space limitation, we omitted the technical
details here.

6 Experiments

6.1 System Implementation

The system is mainly implemented in Java, and the Web page segmentation
module is implemented in C#. We also use MALLET2 , a Java-based package
for machine learning applications. Our experiments are conducted on a PC with
2.93GHz CPU, 4GB RAM and Windows 7. Our system can process a conference
Web page in average of 0.80 seconds, which contains 0.18 seconds for segmenting,
0.62 seconds for annotating and 0.002 seconds for post-processing.

6.2 Data Sets and Evaluation Measures

We collect our conference Web page data by meta-search. First, we prepare a

list of seed top conference names in computer science field. Then we throw each
name as a query to Google and get a list of urls. Using some heuristic rules, we
choose the url which is most likely to be the home page of the conference Web
site, then we request all the Web pages of that conference Web site. Finally, we
collect 50 academic conference Web sites in computer science field, which has 283
web pages and 10,028 labeled text blocks. In order to evaluate our approach, 10
students are invited to manually annotate all the text blocks as the ground truth
of labeling. We use the five-fold cross-validation method to evaluate our system,
that is, one fold of conference Web sites are training data and the rest sites are
test data. We use Precision, Recall and F-Measure as evaluation measures.

6.3 Experimental Results and Discussions

The object of the first experiment is to compare our hybrid segmentation al-
gorithm with the baseline VIPS algorithm. To verify the eectiveness of our
algorithm, we call these two algorithms for conference Web pages respectively,
and take the manually segmented results as the ground truth. If a text block
outputted by an algorithm is the same as the human segmented one, we call
it a correct block. Table 4 shows the count number of correct text blocks of a
conference Web site using the two segmentation algorithm respectively. From
the results, we can find that our hybrid segmentation algorithm can get more
correct text blocks than the VIPS segmentation algorithm. Especially for confer-
ence DASFAA-11, VIPS get a poor result because the visual separators between
two text blocks is not so obvious, while our hybrid segmentation algorithm gets
a satisfactory result.
2
http://mallet.cs.umass.edu/
414 Y. You et al.

Table 4. Number of correct text blocks using two segmentation algorithms

Conference Name VIPS Hybrid Al- Ground
gorithm Truth
VLDB-11 185 226 230 +18%
AIED-11 180 195 195 +8%
CGI-11 190 198 201 +4%
CIKM-11 283 292 292 +3%
ICDE-11 220 237 241 +7%
ICDE-10 490 540 540 +9%
DASFAA-11 10 318 318 +97%
EURO-PAR-11 510 515 515 +1%
SIGMOD-11 145 170 174 +14%
SOSP-11 180 190 190 +5%

Table 5. Annotation Comparison of Dierent Models

Bayes Network SVM CRFs HY+TCRFs
Label
Prec Recal F-ms Prec Recal F-ms Prec Recal F-ms Prec Recal F-ms
TI 0.93 0.91 0.92 0.90 0.89 0.89 0.89 0.91 0.90 0.94 0.97 0.95
TO 0.76 0.65 0.70 0.87 0.80 0.83 0.91 0.83 0.87 0.93 0.92 0.92
DI 0.80 0.92 0.86 0.83 0.70 0.76 0.82 0.90 0.86 0.87 0.93 0.90
D 0.96 0.77 0.85 0.88 0.60 0.71 0.95 0.57 0.71 0.98 0.88 0.92
I 0.28 0.53 0.37 0.78 0.70 0.74 0.82 0.72 0.76 0.90 0.82 0.86
CP 0.96 0.83 0.89 0.96 0.80 0.87 0.97 0.88 0.93 0.97 0.90 0.93
YT 0.68 0.60 0.64 0.50 0.62 0.55 0.47 0.60 0.53 0.80 0.81 0.80
LNT 0.44 0.66 0.53 0.58 0.78 0.67 0.76 0.87 0.81 0.90 0.94 0.92
NT 0.75 0.78 0.76 0.75 0.88 0.81 0.77 0.85 0.80 0.85 0.87 0.86
Average 0.73 0.74 0.72 0.78 0.75 0.76 0.82 0.79 0.80 0.90 0.89 0.90

The aim of the second experiment is to demonstrate the eectiveness of

combined TCRFs model. We have mentioned in Section 3, this work also ap-
proaches classification or sequence labeling problem, so we introduce Bayes Net-
work [10], SVM [6], CRFs [7] as baselines for comparison. Table 5 shows the
P, R, F measures of dierent labels in the label space L. We name our ap-
proach as HY+TCRFs, which indicates the combination of hybrid segmentation
algorithm and TCRFs model. From the numeric results in Table 5, we find
that HY+TCRFs always performs much better than the other models, because
it takes both the visual and dependent features into account. It can also be
found that by using HY+TCRFs, the TO/LNT, D/I/DI categories achieve a
notable improvement than other labels. It is explainable because these labels
have stronger dependencies in child-parent relation and sibling relation than
other labels.
The last group of experiments as shown in Table 6 is the comparison between ini-
tial annotation results and the results after post-processing. It has demonstrated
that post-processing generally improves the quality of the annotation results.
Looking close to the results, we observe that for label TI/TO/DI/D/CP/LNT
they all have more than 0.95 F-Measure, and NT has average 0.90 F-Measure. The
explanation is that they have apparent and specific features on not only display
style but also semantic level. The post-processing can improve the initial anno-
tation easier than other labels. We also observe that YT/NT/I have a relatively
low F-Measure for two reasons: (1) it is hard to eectively extract the features
of YT/NT labels; (2) dierent conferences have dierent descriptions for an Item
label, which has no constant format.
Leveraging Visual Features and Hierarchical Dependencies for Conference IE 415

Table 6. Comparison of before and after post-processing

Before After
Label
Prec Recall F-ms Prec Recall F-ms
TI 0.94 0.97 0.95 0.98 0.99 0.98
TO 0.93 0.92 0.92 0.99 0.99 0.99
DI 0.87 0.93 0.90 0.96 0.99 0.97
D 0.98 0.88 0.92 0.98 0.94 0.96
I 0.90 0.82 0.86 0.94 0.87 0.90
CP 0.97 0.90 0.93 0.98 0.93 0.95
YT 0.80 0.81 0.80 0.82 0.80 0.81
LNT 0.90 0.94 0.92 0.96 0.97 0.96
NT 0.85 0.87 0.86 0.89 0.91 0.90
Average 0.90 0.89 0.90 0.94 0.93 0.94

7 Conclusions
This paper has proposed a hybrid approach to extract academic information
from conference Web pages. Particularly, dierent from existing approaches, our
method combined visual feature and hierarchical structure analysis. Meanwhile,
we also developed heuristic-rule based post-processing techniques to further im-
prove the annotation results. Our experimental results based on the real world
data sets have shown that the proposed method is highly eective and accurate,
and it is able to achieve average 94% precision and 93% recall.

Acknowledgments. This work is partially supported by China National Sci-

ence Foundation (Granted Number 61073021, 61272438, 61272480), Research
Funds of Science and Technology Commission of Shanghai Municipality (Granted
Number 11511500102, 12511502704), Cross Research Fund of Biomedical Engi-
neering of Shanghai Jiao Tong University (YG2011MS38).

References
1. Tang, J., Zhang, J., Zhang, D., Yao, L., Zhu, C., Li, J.Z.: Arnetminer: An expertise
oriented search system for web community. In: Semantic Web Challenge. CEUR
Workshop Proceedings, vol. 295 (2007)
2. Sun, F., Song, D., Liao, L.: Dom based content extraction via text density. In:
SIGIR, pp. 245254 (2011)
3. Zhu, J., Nie, Z., Wen, J.R., Zhang, B., Ma, W.Y.: Simultaneous record detection
and attribute labeling in web data extraction. In: KDD, pp. 494503 (2006)
4. Zhu, J., Nie, Z., Wen, J.R., Zhang, B., Ma, W.Y.: 2d conditional random fields for
web information extraction. In: ICML. ACM International Conference Proceeding
Series, vol. 119, pp. 10441051 (2005)
5. Cai, D., Yu, S., Wen, J.R., Ma, W.Y.: Block-based web search. In: SIGIR, pp.
456463 (2004)
6. Duan, K.-B., Keerthi, S.S.: Which is the best multiclass SVM method? An empir-
ical study. In: Oza, N.C., Polikar, R., Kittler, J., Roli, F. (eds.) MCS 2005. LNCS,
vol. 3541, pp. 278285. Springer, Heidelberg (2005)
7. Laerty, J.D., McCallum, A., Pereira, F.C.N.: Conditional random fields: Proba-
bilistic models for segmenting and labeling sequence data. In: ICML, pp. 282289
(2001)
416 Y. You et al.

8. Bradley, J.K., Guestrin, C.: Learning tree conditional random fields. In: ICML, pp.
127134 (2010)
9. Tang, J., Hong, M., Li, J., Liang, B.: Tree-structured conditional random fields for
semantic annotation. In: Cruz, I., Decker, S., Allemang, D., Preist, C., Schwabe,
D., Mika, P., Uschold, M., Aroyo, L.M. (eds.) ISWC 2006. LNCS, vol. 4273, pp.
640653. Springer, Heidelberg (2006)
10. Heckerman, D.: A tutorial on learning with bayesian networks. In: Holmes, D.E.,
Jain, L.C. (eds.) Innovations in Bayesian Networks. SCI, vol. 156, pp. 3382.
Springer, Heidelberg (2008)
11. Crescenzi, V., Mecca, G., Merialdo, P.: Roadrunner: Towards automatic data ex-
traction from large web sites. In: VLDB, pp. 109118 (2001)
12. Arasu, A., Garcia-Molina, H.: Extracting structured data from web pages. In: SIG-
MOD Conference, pp. 337348 (2003)
13. Song, R., Liu, H., Wen, J.R., Ma, W.Y.: Learning block importance models for
web pages. In: WWW, pp. 203211 (2004)
14. Wainwright, M.J., Jaakkola, T., Willsky, A.S.: Tree-based reparameterization for
approximate inference on loopy graphs. In: NIPS, pp. 10011008 (2001)
15. Xiao, Y., Wei, Z., Wang, Z.: A limited memory bfgs-type method for large-scale
unconstrained optimization. Computers & Mathematics with Applications 56(4),
10011009 (2008)
16. Sha, F., Pereira, F.C.N.: Shallow parsing with conditional random fields. In:
HLT-NAACL (2003)
 Aggregation-Based Probing
for Large-Scale Duplicate Image Detection

Ziming Feng, Jia Chen, Xian Wu, and Yong Yu

Computer Science Department, Shanghai Jiao Tong University

Dongchuan Road, Minhang District, Shanghai, China
{fengzm,chenjia,wuxian,yyu}@apex.sjtu.edu.cn

Abstract. Identifying visually duplicate images is a prerequisite for a

broad range of tasks in image retrieval and mining, thus attracts heavy
research interests. Many ecient and precise algorithms are proposed.
However, compared to the performance duplicate text detection, the re-
call for duplicate image detection is relatively low, which means that
many duplicate images are left undetected. In this paper, we focus on
improving recall while preserving high precision. We exploit hash code
representation of images and present a probing based algorithm to in-
crease the recall. Dierent from state-of-the-art probing methods in im-
age search, multiple probing sequences exist in duplicate image detection
task. To merge multiple probing sequences, we design an unsupervised
score-based aggregation algorithm. The experimental results on a large
scale data set show that precision is preserved and the recall is increased.
Furthermore, our algorithm on aggregating multiple probing sequences
is proved to be stable.

1 Introduction

The explosion of web image scale inevitably brings overwhelming number of du-
plicate images. According to a recent study [18], more than 8.1% of web images
have more than ten near duplicates and more than 20% of web images have
duplicate images. Without proper processing, such duplicates can cause redun-
dancies in web search results and reduces user experiences. Therefore, detecting
duplicate images from a large image collection becomes a hot research topic.
As defined in work[17,9], the duplicate image detection task is to find all
visually duplicate image groups from an image collection. According to the def-
inition, it is a cluster style task rather than a search style task. The input is a
gigantic collection of images and the output is a cleaned version of that image
collection(all duplicate ones are grouped into one cluster).
To cope with large scale, eciency is the core consideration in designing algo-
rithms. State-of-the-art approaches usually exploit hash-code representation of
images [9,17] and put images with the same hash-code together. Since duplicate
images could be represented with dierent hash codes, such approaches achieve
high precision but relatively low recall.

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 417428, 2013.
c Springer-Verlag Berlin Heidelberg 2013

418 Z. Feng et al.

In this paper, we propose to improve recall while preserving high precision.

We design a probing style algorithm to look for more duplicate images. Dierent
from single probing sequence in image search [13], multiple probing sequences
exist in duplicate image detection. This raises a new challenge: how to merge
multiple probing sequences. In this paper, we formulate this challenge into a
rank aggregation problem. Under such formulation, each probing sequence is
first represented as a partial ordered list and then aggregated into an unified
probing sequence for execution.
In experiments, we compare the Principal Components Analysis (PCA) code
and Locality sensitive hashing (LSH) code on millions of images in this task. It
could reflect that PCA-code is more suitable in this task. Then we demonstrate
the proposed approach with PCA-code. In developing the rank aggregation on
probing sequences, we introduce three dierent distance measurements and com-
pare them with the baseline. All these three measurements are proved to increase
the recall while preserving high precision.
The rest of the paper is organized as follows. Section 2 discusses related work;
Section 3 introduces the framework of the algorithm. Section 4 generates the
PCA code; Section 5 proposes the aggregation-based probe algorithm. Section 6
provides the experimental results and section 7 concludes this paper.

2 Related Work
2.1 Duplicate Image Detection
Plenty of works [14,19,9,17,22] are devoted to duplicate image detection. Accord-
ing to dierent feature representations of an image, these works can be classified
into two categories.
Local Method: Typical works[3,9,22,16] are bases on SIFT[12]. The min-hash
[9] is used to speed up the process. The advantage of adopting local features
is that they are invariant to the scale, rotation, position and ane transform
of images. While the drawback is burden on computational resources con-
sumption, which makes it hard to build a system to handle the the billions
of images.
Global Method: [10,15,17] are based on global features. Typical global fea-
tures include color histogram [20], gray value [17] and edge histogram [23].
PCA is useful in duplicate task[6]. Wang et al.[17] proposed a PCA-based
hashing method depending on the global features. In their work, each image
is divided into k k regular grid elements and the corresponding mean gray
values are extracted to get simple but eective visual features. However, it
focuses on the precision but ignores the recall. In our work, we propose an
ecient rank-based probe method to increase the low recall.

2.2 Rank Aggregation

A lot of relevant works on rank aggregation methods has been published
[4,8,5,2,21,11]. In general, there are two main types: score-based aggregation
 Aggregation-Based Probing for Large-Scale Duplicate Image Detection 419

and rank-based aggregation. Recent work focuses on rank-based methods [4,5],

which means the order of the list is significant but the actual score of each element
in the list is not. There are some common metrics for rank-based aggregation
such as Canberra distance[7], Spearman footrule distance [4] and Kendalls tau
distance [4].
In our task, the probing score to a given code represents the probability of
the existence of duplicate images corresponding to it. Thus we use a score-
based rank aggregation method to handle the problem. The existing score-based
methods are often based on supervised learning [21,2]. Due to the large volume of
web images, its very expensive to acquire manual labeling. Therefore we adopt
an unsupervised approach. Dierent from Klementiev[8] , which measures the
agreement between the rankers, we design some distance measurements naturally
from the existing rank-based measurement and directly optimize them.

3 Framework
In this section, we introduce the framework for the duplicate image detection
task, which contains two steps:
At first, all images in the entire data set will be represented by gray value
and indexed according to their PCA-code; Then we use the PCA codes to divide
the image set into smaller groups, the images in each group share the same
PCA code. We refer to each group as a bucket in the rest of this paper. In the
generation of PCA-code, similar images may vary a lot in their PCA codes. The
slight shift of the features may cause totally dierent code, thus put into dierent
buckets. To improve the performance, the PCA-codes representing similar image
features should be identified and related. The traditional similar metric is the
hamming distance, which means the PCA-codes with small hamming distance
will be grouped together. But it suers a low eciency to probe all images in
the dataset.
To increase the eciency, for each bucket, we generate a probing sequence.
The probing sequence can be regarded as similarity measurement of the buckets.
The later element in the sequence means the less probability that contains the
duplicate image. If we can generate the probing sequence, it could eectively
and eciently find the duplicate image groups.
Recently, a multi-probe method [13] is proposed to generate the probing se-
quence for LSH. Since the Gaussian distribution assumption of PCA is equal to
that of LSH in [13], we can apply [13] to the PCA-code. But [13] method relays
on the query as the input parameter, however in duplicate image detection, there
are more than one images in each bucket, therefore, we cannot apply [13] directly.
We develop a two-step algorithm to handle the problem. First, for each image
in the bucket, we generate the basic probing sequence. Second, we apply the
rank aggregation algorithm to generate the merged probing sequence. Because
the length of the candidate list is limited, it is a partial list rank aggregation.
420 Z. Feng et al.

/SGMK9KZ

,KGZ[XK )UJK
9VGIK 9VGIK
,KGZ[XK+^ZXGIZOUT )UJK-KTKXGZOUT

3KXMKJ6XUHOTM9KW[KTIK (GYOI6XUHOTM9KW[KTIK
   
   
   

   

 'MMXKMGZOUT   
   
*[VROIGZOUT*KZKIZOTM

*[VROIGZKJ/SGMK-XU[VY

Fig. 1. Framework of our duplication detection system

Since we target to cope with large scale web images, its hard and expensive
to acquire the labeled data, thus it is hard to use supervised rank aggregation
method. We turn to an unsupervised score-based method which is displayed in
Figure 1.

4 PCA-Code Component
In this section, we introduce feature extraction and PCA hash-code generation.
For feature extraction, an image is divided into a k k grid, where k is set to 8
from previous work [17]. The mean gray value feature is extracted from each grid
as a robust image representation for duplicate detection. It can tolerate most
image transformations such as resizing, recoloring, compression, and a slight
amount of cropping and rotation. Each feature element is calculated by the
Equation (1)
1 " (i)
Fi = grayj i = 1, 2 . . . k 2 , (1)
Ni
(i)
where grayj is the j th gray value in grid i and Ni is the number of pixels in
grid i. So an image could be represented by a k 2 1 vector.
Before code generation, the raw features should be compressed into a compact
representation with noises reduced. The PCA hash-code algorithm uses a PCA-
based hash function to map a high-dimensional feature space to a low-dimensional
 Aggregation-Based Probing for Large-Scale Duplicate Image Detection 421

feature space. The PCA model is used to compress the features extracted from the
image database. After the dimension reduction, visual features are transformed
to a low-dimensional feature space. Then the hash code H for an image is built
as follows: each dimension is transformed to one if its value in this dimension is
greater than mean value, and zero otherwise. This is summarized in Equation (2).
!
1 vi meani
H(i) = (2)
0 vi < meani

where the meani is the mean value of the dimension i. The flowchart of PCA-
code component is shown in Figure 2.

Fig. 2. Illustration for hash-code generation component

5 Aggregation-Based Probe Algorithm

The input of our aggregation-based probe algorithm is a set of images in hash
code and feature representation. The output is duplicate image groups(code
groups). The algorithm traverses all the buckets. For a given bucket, it generates
a merged probing sequence of buckets and then gathers them with the given
bucket into a group. Then these buckets will not be probed again.
The merged probing sequence is generated in two steps.

Step 1: Basic Probing Sequence Generation. In this step, a probing se-

quence of buckets for each image in the starting bucket is generated. The top-K
elements are selected into the sequence based on a score function. The score func-
tion measures the probability of containing duplicate images. A smaller score
means larger probability of finding a duplicate image.

Step 2: Merged Probing Sequence Generation. In this step, the dierent

basic probing sequences generated from the starting bucket are merged. The
merging is formulated as a partial rank aggregation problem as shown in Table 1.
For elements in basic probing sequence, it is assigned with a probability value.
We treat probability values as scores in our solution. Additionally, given that
there is no labeled data available, we use an unsupervised approach in solving
the aggregation problem.
422 Z. Feng et al.

Table 1. Terminology Mapping Table

Terms in duplicate detection Terms in rank aggregation

Image Ranker
Basic probing sequence Partial ordered list
Merged probing sequence Result list

5.1 Step 1: Basic Probing Sequence Generation

At the beginning, each image in collection are encoded by PCA-code. And all
of them will generate a probing sequence by a score function. From [13], the
score function defined for one image in starting bucket hq to bucket hj is as in
Equation (3) "
scoreLSH (hq , hj ) = fk2 (3)
k
where f is the compressed feature value of the query image and is the set of
bits which are dierent between buckets hq and hj .
Then we apply this equation to PCA-code. Since we will compare the com-
pressed feature values between dierent dimensions in PCA-code, we normalize
them by X X . Considering the assumption that each dimension obeys the
Gaussian distribution, the normalized value conforms to a standard Gaussian
distribution. Then the score function of PCA is as in Equation (4)
" f k k 2
scoreP CA (hq , hj ) = ( ) (4)
k
k

where the k is the mean value of dimension k and k means the standard
deviation of dimension k. It is used to estimate the probability that images in hj
duplicate the query image. A lower score means that hj is more likely to contain
duplicate images. For each image in the starting bucket, we get a basic probing
sequence based on scoreP CA .

5.2 Step 2: Merged Probing Sequence Generation

Before modeling the score-based rank aggregation problem, we do some pre-
processing to turn the partial ordered list into a full ordered list. We combine all
the elements in partial ordered lists, to get a set T . Then we calculate scores
for all the elements in T for each image in the starting bucket. It becomes a
full ordered list problem.
Now we define some symbols to give a mathematical formulation for the score-
based rank aggregation problem. We assume that the V i represents the i th
full ordered list and V is the result list. Finally, we cast the score-based rank
aggregation problem into the optimization problem shown in Equation (5).
"
min dis(V i , V ) (5)
i

The dis(V , V ) measures the distance between V i and V .

i
 Aggregation-Based Probing for Large-Scale Duplicate Image Detection 423

In the formulation, dierent distance measurement dis(V i , V ) leads to dif-

ferent result list V . The following three measurements are designed in this
problem.
1. Footrule distance
Motivated by the Footrule distance[4], the score-based distance measurement
could be "
dis(V i , V ) = |Vji Vj | (6)
j

It naturally describes the dierence between the two rankers (The result list
V also could be seen as a ranker).
2. Canberra Distance
But in the footrule distance, we dont take into account that the large
elements in the list will have an larger weight than the small elements (i.e.
|Vji Vj | is not weighted). The weighted footrule distance is the Canberra
Distance[7] .
" |Vji Vj |
dis(V i , V / ) = (7)
j
|Vji + Vj |

Adding the denominator to the equation prevents the unbalanced estimation

problem.
3. Discounted Canberra Distance
In our problem, if we take the fewer elements to achieve the same recall, it
will be more ecient and eective. Then we can get a conclusion that the
top-K elements in V are more important than others. So we add a rank
adjustment factor to the Canberra Distance. This is the discounted Canberra
Distance in the following equation,
" |Vji Vj | 1
dis(V i , V ) = (8)
j
|Vji + Vj | 1 + lg P osVj

where the P osVj means the rank of element j in the V . The adjustment
in this equation eectively increases the weight of the element at the top of
the list.

Optimization Method. In this section, we show the details of optimization

method for dierent distance measurements in the score-based rank aggregation.
1. In footrule distance, the optimal score of each element in T is the median
value of the scores given by the rankers.

Vj = median(Vj1 . . . . . . VjN ) (9)

Where N is the number of the rankers.

2. The problem of Canberra Distance can be solved by a general gradient
descent algorithm. From Equation (8) we know that the elements in T are
separated from each other. So we can calculate the optimal value of each element
424 Z. Feng et al.

independently. The formulation for the gradient descent step is listed in the
following equation.
" Vji 2
Vj = Vj {I(Vji Vj ) 2+
i (Vj + Vji )
(10)
Vji 2
I(Vji < Vj ) 2}
(Vj + Vji )

Where I(x) is the Indicator function and is the learning rate(which is usually
setting as 0.01). The smallest k elements in V are the result list.
3. Because the rank adjustment factor of each element in T has relativity, the
problem of Discounted Canberra Distance cannot # be handled by the previous al-
gorithm. We add a restriction that the V = i Wi V i . Then we follow the idea
from the existing lambdarank optimization algorithm[1] (which approximately
optimizes NDCG as its object) and this problem can be optimized directly by
a approximate gradient descent algorithm. The gradient process is listed in the
following equation.
1 ""
Wk = WK {I(Vji Vj ) Vjk
1 + lg P osVj i j
(11)
2 Vji i k
2 Vji
2 + I(V j < Vj ) Vj 2 }
(Vj + Vji ) (Vj + Vji )

where is the learning rate (which is usually setting to 0.01) and I(x) is the

Indicator function. The P osVj will be updated after each iteration process. Then
we use the W to calculate the V and the smallest k elements in V are the result
list.

6 Experiment
6.1 Experimental Setting
The dataset contains 1,003,440 images crawled from the web. It is split into
two parts, the ground truth and the distractor set. The ground truth consists
239 groups of 1260 images. It is labeled manually. The distractor set is selected
randomly from the web with images in the ground truth removed. The size of the
distractor set is 1,002,180. All the images are resized to 160 pixels on their larger
side. We use 100,000 images to train a stable PCA for dimension reduction.

6.2 Evaluation Measurements

To measure the performance of the methods, we use the same metrics as [17]:
(#CorrectImageP airs)
P recision = 100% (12)
(#DetectedImageP airs)
 Aggregation-Based Probing for Large-Scale Duplicate Image Detection 425

(#CorrectImageP airs)
Recall = 100% (13)
(#GroundT ruthImageP airs)

The DetectedImageP air denotes image pairs in a detected group. The

CorrectImageP airs denotes the detected image pairs belong to the same group
of the ground truth. The GroundT ruthImageP airs represents the image pairs
in the same group of the ground truth.

6.3 Comparison of PCA-Code and LSH

In this experiment, the PCA-code is compared to LSH on duplicate image de-
tection tasks. To the best of our knowledge, the result of such comparison has
not been shown before. The length of PCA-code and LSH are set as 40. Figure
3 shows the P-R curves of the PCA-code and LSH. The buckets in both codes
are grouped by hamming distance. It can be seen that the curve of the LSH is
under the curve of the PCA-code. It means that with the same precision, we can
get a higher recall performance. So PCA-code is more eective than LSH. From
another perspective, the curve of LSH declines rapidly. On the contrary, the
curve of PCA declines slowly. It means that the buckets divided by PCA-code
containing duplicate images are close to each other. And the buckets containing
un-duplicate image are not close.
The reason is analyzed as follows. PCA model could be seen as a particular
list of projections of the original feature space. Therefore it will perform better
than LSH, whose projection is randomly generated.

Fig. 3. Comparison on P-R curve of PCA-code and LSH

426 Z. Feng et al.

6.4 Comparison of Dierent Aggregation-Based Probe Method

In this experiment, we show the improvement of our aggregation-based probe
method on three distances measurements. Then we evaluate the algorithms
stability on the partial ordered list length. The starting bucket is fixed for all
experiments.
In evaluating the improvement of our aggregation-based probe method, we use
two baselines. The first baseline is the PCA-code, which means the codes are di-
vided by Hamming distance. Comparison with this baseline show the improvement
of probing methods over non-probing ones. The second baseline is M C4 [4], a state-
of-the-art rank-based rank aggregation method, which is the best method in [4].
Comparison with this baseline shows the improvement of our score-based rank ag-
gregation methods to the rank-based rank aggregation methods.
Table 2 shows the performance of our aggregation-based probe method when
the length of partial ordered list is fixed in probing. In this experiment, the
length of result list is equal to the length of the partial ordered list.

Table 2. Comparison of Recall/Precision Performance of Various Methods. The

Length means the length of the partial ordered list. P CA, F Dst, CaDst, DCaDst and
M C4 means the PCA-code, the median value, Canberra distance OPT, Discounted
Canberra distance OPT and Markov-4 aggregation method.

Score-base Rank Aggregation

Length PCA MC4
FDst CaDst DCaDst
16 NA 0.577/0.998 0.577/0.998 0.577/0.998 0.750/0.998
40 0.556/0.998 0.775/0.992 0.775/0.992 0.775/0.994 0.813/0.994
50 NA 0.783/0.991 0.783/0.991 0.784/0.994 0.827/0.993
100 NA 0.83/0.985 0.83/0.985 0.831/0.989 0.894/0.985
780 0.813/0.981 0.92/0.917 0.92/0.917 0.92/0.923 0.93/0.92

In the table, all the rank aggregation methods improve the recall and main-
tain the precision. The F Dst and M C4 methods show similar performance and
the CaDst is slightly better than either of those on precision performance.The
improvement of DCaDst on recall performance is higher than all the others.
When we fix the number of buckets at 40, Recall/P recision of Baseline is
only 0.556/0.998 but DCaDis reaches 0.813/0.994, improving the recall by 40%.
The other methods are around 0.775/0.993, which is better than the baseline but
lower than DCaDst.When we fix the recall score at 0.813, Baseline probes 780
buckets and DCaDst probes only 40 buckets. DCaDst probes far fewer buckets
to reach the same recall. The other methods probe more than 50 buckets. This
good performance of DCaDst can be explained by its design principle: it focuses
on the top-K items of the result list and optimizes them. This design is helpful
for optimizing the top-k items of the result list. When the length of result list is
shorter, the performance is more better than the others. So the DCaDst could
probe less buckets than the other methods to reach the same performance.
We conclude that DCaDst is better than the other methods for this task. It
eectively reduces the number of the buckets to be probed and increase the recall
 Aggregation-Based Probing for Large-Scale Duplicate Image Detection 427

while preserving the precision performance. It also shows that the score function
in Equation 3 is a reasonable inference, which could reflect the probability of the
bucket containing the duplicate images.
Table 3 shows the stability of our aggregation-based probe algorithm on three
distances when the partial list length is changed.

Table 3. The recall performance with respect to dierent settings. The LR means the
length of the result list and the LP means the length of the partial ordered list.

LR LP FDst CaDst DCaDst

40 0.775 0.775 0.813
40
80 0.775 0.775 0.813
50 0.783 0.784 0.827
50
100 0.783 0.784 0.827

From Table 3, we can see that the recall performance is stable as the result
list length fixed, which indicated that our proposed methods are robust as the
length of partial ordered list varies. It is a exciting property in our task. So that
we can set the length of the partial ordered list as length of the result list while
maintaining the performance. It could reduce the computation and easy to apply
to the next application.
To sum up, the experiments show that our aggregation-based probe algo-
rithm can eectively increase the recall performance while maintaining the high
precision with high robustness.

7 Conclusion and Future Work

In this paper, we propose a probing based method to detect duplicate images from
a large image collection. Compared to previous works, the proposed approach re-
serves the advantages in high precision and at the same time improves the rela-
tively low recall. To increase the recall, the proposed method generate the probing
sequence to find the duplicate images more intelligently. Dierent from state-of-art
probing methods for image search, multiple probing sequences exist in duplicate
image detection. To unify multiple probing sequences, we present an unsupervised
score-based rank aggregation algorithm. As shown in experimental results on mil-
lions of images, relatively low recall is increased and high precision is preserved.
In future work, we plan to construct a larger dataset and explore more im-
age representations to analyze the content of image. Also we will discuss the
convergence of the gradient descent and more rank aggregation methods.

References
1. Burges, C.J.C., Ragno, R., Le, Q.V.: Learning to rank with nonsmooth cost func-
tions. In: NIPS, pp. 193200 (2006)
2. Chen, S., Wang, F., Song, Y., Zhang, C.: Semi-supervised ranking aggregation. In:
CIKM, pp. 14271428 (2008)
428 Z. Feng et al.

3. Chum, O., Philbin, J., Zisserman, A.: Near duplicate image detection: min-hash
and tf-idf weighting. In: BMVC (2008)
4. Dwork, C., Kumar, R., Naor, M., Sivakumar, D.: Rank aggregation methods for
the web. In: WWW, pp. 613622 (2001)
5. Fagin, R., Kumar, R., Sivakumar, D.: Ecient similarity search and classification
via rank aggregation. In: SIGMOD Conference, pp. 301312 (2003)
6. Huang, Z., Shen, H.T., Shao, J., Zhou, X., Cui, B.: Bounded coordinate system
indexing for real-time video clip search. ACM Trans. Inf. Syst., 27(3) (2009)
7. Jurman, G., Riccadonna, S., Visintainer, R., Furlanello, C.: Canberra distance on
ranked lists. In: Ranking NIPS 2009 Workshop, pp. 2227 (2009)
8. Klementiev, A., Roth, D., Small, K.: An unsupervised learning algorithm for rank
aggregation. In: Kok, J.N., Koronacki, J., Lopez de Mantaras, R., Matwin, S.,
Mladenic, D., Skowron, A. (eds.) ECML 2007. LNCS (LNAI), vol. 4701, pp. 616
623. Springer, Heidelberg (2007)
9. Lee, D.C., Ke, Q., Isard, M.: Partition min-hash for partial duplicate image discov-
ery. In: Daniilidis, K., Maragos, P., Paragios, N. (eds.) ECCV 2010, Part I. LNCS,
vol. 6311, pp. 648662. Springer, Heidelberg (2010)
10. Li, Y., Jin, J., Zhou, X.: Video matching using binary signature. In: Intelligent
Signal Processing and Communication Systems, pp. 317320 (December 2005)
11. Liu, Y., Liu, T.-Y., Qin, T., Ma, Z., Li, H.: Supervised rank aggregation. In: WWW,
pp. 481490 (2007)
12. Lowe, D.G.: Distinctive image features from scale-invariant keypoints. International
Journal of Computer Vision 60(2), 91110 (2004)
13. Lv, Q., Josephson, W., Wang, Z., Charikar, M., Li, K.: Multi-probe lsh: Ecient
indexing for high-dimensional similarity search. In: VLDB, pp. 950961 (2007)
14. Ponitz, T., Stottinger, J.: Ecient and robust near-duplicate detection in large
and growing image data-sets. In: ACM Multimedia, pp. 15171518 (2010)
15. Qamra, A., Meng, Y., Chang, E.Y.: Enhanced perceptual distance functions and in-
dexing for image replica recognition. IEEE Trans. Pattern Anal. Mach. Intell. 27(3),
379391 (2005)
16. Valle, E., Cord, M., Philipp-Foliguet, S.: High-dimensional descriptor indexing for
large multimedia databases. In: CIKM, pp. 739748 (2008)
17. Wang, B., Li, Z., Li, M., Ma, W.-Y.: Large-scale duplicate detection for web image
search. In: ICME, pp. 353356 (2006)
18. Wang, X.-J., Zhang, L., Liu, M., Li, Y., Ma, W.-Y.: Arista - image search to
annotation on billions of web photos. In: CVPR, pp. 29872994 (2010)
19. Wang, Y., Hou, Z., Leman, K.: Keypoint-based near-duplicate images detection
using ane invariant feature and color matching. In: ICASSP, pp. 12091212 (2011)
20. Zhang, D., Chang, S.-F.: Detecting image near-duplicate by stochastic attributed
relational graph matching with learning. In: ACM Multimedia, pp. 877884 (2004)
21. Zhao, X., Li, G., Wang, M., Yuan, J., Zha, Z.-J., Li, Z., Chua, T.-S.: Integrating
rich information for video recommendation with multi-task rank aggregation. In:
ACM Multimedia, pp. 15211524 (2011)
22. Zhou, W., Lu, Y., Li, H., Song, Y., Tian, Q.: Spatial coding for large scale partial-
duplicate web image search. In: ACM Multimedia, pp. 511520 (2010)
23. Zhu, J., Hoi, S.C.H., Lyu, M.R., Yan, S.: Near-duplicate keyframe retrieval by
nonrigid image matching. In: ACM Multimedia, pp. 4150 (2008)
 User Interest Based Complex Web Information
Visualization

Shibli Saleheen and Wei Lai

Faculty of Information and Communication Technologies,

Swinburne University of Technology,
Hawthorn, Victoria, Australia
{ssaleheen,wlai}@swin.edu.au

Abstract. Web graph allows end user to visualize web information and
its connectivity. But due to its huge size, it is very challenging to visualize
web information from the enormous cyberspace. As a result, web graph
lacks simplicity while the user is searching for information. Clustering
and filtering techniques have been employed to make the visualization
concise but it is often not accurate to the expectation of user because
they do not utilize user-centric information. To deliver the information
concisely and precisely to the end users according to their interest, we
introduce personalized clustering techniques for web information visu-
alization. We propose a system architecture, which considers the user
interests while clustering & filtering, to make the web information more
meaningful and useful to the end users. By the implementation of the ar-
chitecture, an experimental example is provided to reflect our approach.

Keywords: Clustering, Personalization, Visualization.

1 Introduction

It has become very dicult for end users to get their desired information quickly
and accurately from the gigantic World Wide Web. Research attempts to help
the users manage and control the information over the internet by presenting
user-focused information [4] [6] [11]. However, the existing technology is not ade-
quate enough to provide visual maps users expect to guide their web experience.
As a consequence, the next research challenge is to provide focused informa-
tion to the end users for quick and easy understanding of the web information
and its pattern. To provide an easy way to explore information according to
the need of user, techniques of information retrieval and visualization have been
introduced [7]. Information visualization presents the abstract data to amplify
human cognition and unfold the hidden relationships among the data. The pur-
pose of information visualization is not only to create interesting pictures for the
users but also to communicate information clearly and eectively [3] [7].
A web graph, which represents the web information with its internal con-
nectivity revealed, is a very useful way to visualize various aspects of the web

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 429436, 2013.
c Springer-Verlag Berlin Heidelberg 2013

430 S. Saleheen and W. Lai

information. Nevertheless, a huge web graph is less likely to create a good im-
pression for the end user compared to a relatively smaller one with information
relevant to the users interests. Therefore, reducing the size of the web graph is
very important from the end user viewpoint. Filtering and clustering techniques
from dierent aspects such as structure, context, have been applied on the web
graph to reduce the size. Besides the structure and content based clustering
techniques, compromise the direct users needs is expected for personalized visu-
alization. First of all, users are dierent and their needs vary. Most of the work
accomplished for personalization of web data is done for a group of people who
can be addressed as similar minded [13]. But it is not obvious that every user of
a specific group behaves similarly all the time. Secondly, short term interests or
preferences of a particular user change over time or in dierent context.
To provide the above personalization in web information visualization, we
propose to model every user separately. Instead of providing information di-
rectly in the traditional way, we introduce user profile driven clustering for web
information personalization. We strongly believe that this can classify the web
information into several categories and yield an improved visualization for the
end user as the information is clustered according to the individuals interest.
The rest of the paper is organized as follows: Section 2 presents related works;
Section 3 describes the proposed architecture for personalized visualization; Sec-
tion 4 presents the experiment; and finally, Section 5 concludes the paper.

2 Related Work
Work has been done in dierent directions to make the representation of web
information more eective to the end users. Some work deals with information
searching through search engines [14] [1], i.e., on search results. For personal-
ization, researchers consider the user search preference learned from the click
history or browsing history [9]; history of similar minded users [13]; location
metadata of the user for personalized search [2]; and other personalizing page
ranking techniques [11]. In the case of navigating in the web space, clustering
techniques plays important roles for reducing the complexity of the web graph.
Structure based clustering can be found in the works of Huang et al. [6] and
Rattigan et al. [12]. They provide structure based clustering techniques to re-
duce the web graph size. Graph clustering approaches like these only consider
the structural aspect of the graph without considering the content of the docu-
ments. Besides structure, Gao [4] integrates content in the process of clustering.
She first clusters using structure and then clusters using content only on the sub-
clusters created in the first phase. As a consequence, this two-phased clustering
heavily depends on the structure. Inclusion of structure in the visualization cre-
ates scopes for non-similar documents to be linked and hence put in the same
cluster. Our work avoids this possibility by considering only content based sim-
ilarities while relating documents. In contrast to previous research approaches,
this work is personalized as we consider the users point of view.
 User Interest Based Complex Web Information Visualization 431

Fig. 1. System Architecture

3 System Architecture
Our proposed system to resolve the issue of personalized visualization uses data
from user profile as preliminary knowledge-base to assist in filtering, cluster-
ing and hence final visualization. While the user uses the system, the system
automatically updates the existing user profile or creates a new one where ap-
propriate to reflect users latest focusing trends. Finding appropriate scope for
involving the user profiles in the system is very crucial for a personalized system.
Scopes of the user profiles for personalized searching from electronic information
space have been documented in [10]. In our system, after collecting information,
we include the user profile while analysing the web information, i.e., in filtering
& clustering, and finally we visualize the information.
Figure 1 depicts the architecture of the system for personalization of informa-
tion visualization. The key modules of this model are:
i) Result Analyser module (composed of Data Collection, Filtering and
Clustering components) collects the web data based on users request, filters
out irrelevant data, and finally generates a clustered web graph by using the
user profile for a handy and compact visualization.
ii) Visualization module is responsible for visualizing the web graph to pro-
vide a good overview; to help the user to find out the expected information;
and to create scopes for the user to interact with the information.
iii) Feedback Analyser module gets the user interaction data from the vi-
sualization module, then analyses the data, and lists possible updates.
iv) Profiler module gets the updates from Feedback Analyser module and
consults with the existing user profile to keep user profile up-to-date to
reflect users latest interests in order to provide more relevant output after-
wards. This module also gets direct inputs from the user where appropriate,
to construct or modify the user profile.

3.1 Data Collection and Filtering

The first component of our architecture is the web crawler which collects infor-
mation to represent URLs by their corresponding keywords for the initial web
graph. To ensure quick processing and get adequate & useful information, our
web crawler does not revisit the faulty, unreachable pages and pages containing
432 S. Saleheen and W. Lai

very less information rather discards them. We filter out all other html tags, any
data except text such as image, audio and video to keep page information simple.
Once crawling is finished, we have the original text data for further processing.
We convert the words to their base forms(for example: running to run) to get
the actual frequencies of words and remove stop words from the content.
After filtering the stop words, we have set of words, still big in size, to represent
a URL. To reduce the size, we apply well known TF-IDF to get the terms of
every URL weighted and then sort them. We take a fixed percentage of top
weighted terms as keywords to make further steps time & cost eective.
Now we have a set of documents in the corpus to be visualized by the web
graph. Every document is composed of a set of (Keyword, W eight) pairs. All
the distinct keywords of these documents form the total set of keywords. At this
point, it is very easy to define the documents as vectors of weighted keywords.
Hence, for the i th document, document vector di as < ti1 , ti2 , ti3 , ..., tin >
where tij is the weight j th keyword where 1 <= j <= n and n is the number
of total distinct keywords in the corpus.
We calculate the edges connecting the nodes for the initial graph based on the
content similarity of the nodes. If we consider two documents, the more keywords
they have common, the more they are connected. The similarity between them
is computed by cosine similarity1 . If the similarity score CSim(dy , dz ) is greater
than a threshold value then we create an edge between the documents dy and
dz . Finally, we have an undirected initial graph for next processing.

3.2 User Profiling

To optimize retrieval accuracy, we clearly need to model the user accurately and
personalize the information retrieval according to each individual user, which is
a dicult task. It is even harder for a user to precisely describe what the infor-
mation need is. Unfortunately, in the real world applications, users are usually
reluctant to spend time inputting additional personalized information which can
provide relevant examples for feedback.
Representation of user profiles has been classified into three categories in [5].
The simplest one to represent and maintain is keyword-based user profile where
the keywords are either automatically extracted from the web documents during
a users visit or through direct input by the user. Each keyword can represent
a topic of interest and weight can be associated with it to represent the degree
of interest. The other two categories are semantic network profile which deals to
address the polysemy problem and concept based user profile which consists of
concepts & reveals the relationship between the concepts.
In our system architecture, we adopt the keyword based user profiles for sim-
plicity to represent a user. So the user profile consists of a set of weighted key-
words, i.e., (Keyword, W eight) pairs. User profiles are maintained in two phases
by the Profiler component. In the creation phase, the user data are collected
directly from the user via a user interface. Here the user is allowed to answer
1
http://en.wikipedia.org/wiki/Cosine_similarity
 User Interest Based Complex Web Information Visualization 433

some predefined questions for the creation of basic user profile. After collect-
ing the basic data from the user, the system will collect more user data from
the internet according to the users authorization. The maintaining or updating
phase occurs mostly whenever the user uses the system. Profiler updates the
preferences based on user interaction data passed from the Feedback Analyser.

3.3 Clustering
We want to cluster the total corpus into dierent clusters where the documents
will be clustered by biased similarity values based on the user profile. We modify
the document vectors computed in section 3.1 to represent both user interests
and main topics of the document. For this we need to replace the existing doc-
ument terms with the keywords of the user profile where appropriate.
We update the total number of keywords of the corpus by adding the user
profile keywords in the keyword set. If the number of new distinct keywords is m,
the total number of keywords becomes (n + m). Now, we can write the modified
document vector for di as: < Ti1 , Ti2 , ..., Tin , Ti(n+1) , Ti(n+2) , ..., Ti(n+m) >.
We calculate the Ti(n+j) ; where 1 < j <= m; in the following way. We mea-
sure the similarity value of cj , where cj is a keyword from user profile, with
each keyword of the di . We use one of the wordNet similarity measures for this
purpose. There are six similarity measures applied on wordNet, three measures
including Lin [8] are based on information content of the least common sub-
sumer(LCS). Lin yields slightly higher correlation with human judgements than
others, which actuates us to use it instead of others.
If cj is similar to document keyword kil ; where 1 < l <= n; meaning that
(SIM (kl , cj ) Til ) > , we consider that document di is containing keyword
cj instead of the original document keyword. So, we remove the kil by making
the weight Til as 0. We repeat it for every cj of user profile for each di . As a
particular cj can be similar with multiple keywords of a document by dierent
values, we take the maximum similarity value as the weight of that cj .
Hence, the new vector to represent a document, di , reflects the user interests
and the topics of the original content as well. We apply k-means algorithm for
clustering, where the number k and initial means are calculated from the cosine
similarity values of the document vectors. We select di as an initial mean if the
degree of di is greater than the sum of the average degree of the web graph and
a predefined threshold.

3.4 Visualization and Feedback Analysis

For displaying the final view to the user, we have developed the visualization
component based on popular JGraph2 framework. The visualization interface
has three panes. The pane on the left side of the interface produces tree view of
the web graph clusters. In the middle of the interface the web graph is displayed.
To distinguish between the cluster and the URL nodes we use rectangular and
oval shapes, respectively. Beside navigating, we have also added browser support
2
http://www.jgraph.com/jgraph.html
434 S. Saleheen and W. Lai

Table 1. Keywords representing URL nodes

URL Keywords
1 - justice, evacuation, screencasts, wireless, concept
2 - centre, management, laboratory, college, group, resource
3 - directory, bing, yahoo, google, hospitality
... ... ...
6 - bachelor, flexibility, convenience, delivery
7 - postgraduate, information, city, night, suite
... ... ...
19 - command, wireless, sheet, layout

for the user to visit the web pages in the right pane. The user can- zoom in/out;
re-arrange the web graph by drag and drop; delete an inappropriate node or
cluster; and create new clusters.
As the user interacts with the visualization interface, it sends the browsing,
click and arrangement information generated from the actions accomplished by
the user to the feedback analyser component. Browsing feature is the primary
contributing feature in profile updating. When user browses a page, the keywords
with their weights along with the dwelling time are sent to the feedback analyser
module, which updates the weights of existing interests and adds the new ones.
While adding the new interests in the user profile, the interest and corresponding
weight-value to be added are shown to the user for any amendment.

4 Experiment
In this section, we provide an experimental example. It begins with showing a
simple web graph. For this example we analyse a university web site and set the
crawl style property as BFS whose termination is set to 1. We get 19 nodes while
crawling and name them according to their appearance number in the crawler.
We get the sets of keywords for representing the nodes after applying techniques
from 3.1, which are shown in Table 1. Figure 2a shows the original web graph.
We get two clusters and eight other nodes which is shown in Figure 2b after
applying content based clustering described in [4].
We also have two user profiles constructed earlier based on two dierent
users. We consider one user profile, U Pa , represents an international student
who looks for opportunity of postgraduate studies in computer science related
fields whether the other, U Pb , is of a local student focusing on undergraduate
studies in business, particularly interested in Hawthorn campus. The corre-
sponding user profiles are given in Table 2a as (Interest, W eight) pairs.
First, we take the user profile U Pa and generate the clusters based on the
methodology described in section 3.3. We get three clusters and eight unclustered
nodes. Second, applying the same process on the initial graph for the user profile
U Pb , we get four clusters and seven unclustered nodes as well. Table 2b shows
how the nodes are assigned in the clusters for all three cases. With no user
 User Interest Based Complex Web Information Visualization 435

(a) Initial web graph (b) Content based clustered graph

Fig. 2. Traditional Visualization

Table 2. User Profiles and Generated Clusters

(a) User Profiles (b) Generated Clusters
Profile Interest Weight Profile Cluster Label Nodes
Supervisor 0.5 bachelor-inquiry 3, 5, 6, 13, 17, 18
-
Scholarship 0.5 wireless-justice-itunes 1, 14, 15, 16, 19
U Pa Visa 0.3 computer 10, 14, 15
Computer 0.8 U Pa supervisor-international 2, 4, 8, 16
International 0.9 job-study 6, 7, 9, 13
Business 0.8 bachelor-business 6, 18
Bachelor 0.9 management-hawthorn 2, 4, 7
U Pb U Pb
Hawthorn 0.5 job-news 9, 11, 13
Management 0.7 wireless-justice-itunes 1, 14, 16, 19

(a) Clustered graph using U Pa (b) Clustered graph using U Pb

Fig. 3. Personalized Visualization

436 S. Saleheen and W. Lai

profile applied, the two clusters are calculated based on their content similarity.
Three and four clusters have been constructed for the user profiles U Pa & U Pb
and are shown in Figure 3a & 3b respectively. From these two visualization
it is noticed that clustering has been heavily influenced by the user profiles.
Rather displaying the same web graph of Figure 2b, our system shows the web
information to dierent users in dierent ways according to their interests.

5 Conclusion and Future Work

This paper demonstrates that the user interest profiles can be used to improve
the visualization of complex web information. Based on the experimental re-
sult, it is noticed that the generated clusters reflect the user interests eectively
meaning that the web graph is personalized. Further studies need to be done to
design domain specific ontology to get the web information in a structured way
and ontology for the user profile as well. Then a combination of both ontologies
in the system architecture will be investigated.

References
1. Agrawal, R., Gollapudi, S., Halverson, A., Ieong, S.: Diversifying search results. In:
WSDM, pp. 514 (2009)
2. Bennett, P.N., Radlinski, F., White, R.W., Yilmaz, E.: Inferring and using location
metadata to personalize web search. In: SIGIR, pp. 135144 (2011)
3. Card, S.: Information visualization. In: The Human-Computer Interaction Hand-
book: Fundamentals, Evolving Technologies, and Emerging Applications (2007)
4. Gao, J., Lai, W.: Visualizing blogsphere using content based clusters. In: Web
Intelligence and Intelligent Agent Technology, pp. 832835 (2008)
5. Gauch, S., Speretta, M., Chandramouli, A., Micarelli, A.: User profiles for person-
alized information access. In: Brusilovsky, P., Kobsa, A., Nejdl, W. (eds.) Adaptive
Web 2007. LNCS, vol. 4321, pp. 5489. Springer, Heidelberg (2007)
6. Huang, X., Eades, P., Lai, W.: A framework of filtering, clustering and dynamic
layout graphs for visualization. In: ACSC, pp. 8796 (2005)
7. Keim, D.A., Mansmann, F., Schneidewind, J., Ziegler, H.: Challenges in visual data
analysis. In: Information Visualization, pp. 916 (2006)
8. Lin, D.: An information-theoretic definition of similarity. In: ICML, pp. 296304
(1998)
9. Matthijs, N., Radlinski, F.: Personalizing web search using long term browsing
history. In: WSDM, pp. 2534 (2011)
10. Micarelli, A., Gasparetti, F., Sciarrone, F., Gauch, S.: Personalized search on the
world wide web. In: Brusilovsky, P., Kobsa, A., Nejdl, W. (eds.) Adaptive Web
2007. LNCS, vol. 4321, pp. 195230. Springer, Heidelberg (2007)
11. Qiu, F., Cho, J.: Automatic identification of user interest for personalized search.
In: WWW, pp. 727736 (2006)
12. Rattigan, M.J., Maier, M., Jensen, D.: Graph clustering with network structure
indices. In: ICML, pp. 783790 (2007)
13. Smyth, B., Balfe, E., Boydell, O., Bradley, K., Briggs, P., Coyle, M., Freyne, J.: A
live-user evaluation of collaborative web search. In: IJCAI, pp. 14191424 (2005)
14. Wang, X., Zhai, C.: Learn from web search logs to organize search results. In:
SIGIR, pp. 8794 (2007)
 FIMO: A Novel WiFi Localization Method

Yao Zhou, Leilei Jin, Cheqing Jin , and Aoying Zhou

Shanghai Key Laboratory of Trust Worthy Computing,

Software Engineering Institute
East China Normal University, Shanghai 200062, China
[email protected], [email protected], {cqjin,ayzhou}@sei.ecnu.edu.cn

Abstract. With the development of technology and the proliferation

of mobile computing devices, peoples need for pervasive computing is
rapidly growing. As a critical part of pervasive computing, Location
Based Service (LBS) has drawn more and more attention. Localization
techniques that report the real-time position of a moving object are key
in this area. So far, the outdoor localization technologies (i.e, GPS) are
relatively mature, while the indoor localization technologies are still un-
der improvement. In this paper, we propose a novel WiFi localization
method, called FIMO (FInd Me Out). In this method, we take the in-
stability of signal strength and the movement of objects into considera-
tion when determining the location based on Fingerprint. Experimental
results show the proposed method is capable of estimating a moving
objects location precisely.

Keywords: LBS, localization, indoor, WiFi.

1 Introduction
For a long time, people expect the electronic devices around them can have
perception towards the environment.In early 2012, IDC (Internet Data Center)
predicted that around 1.1 billion smart connected devices would be sold in 2012,
and this number would even be doubled in 20161. As one of the important ap-
plications of ubiquitous computing system, localization application has drawn
more and more attention. In general, these systems often need to provide infor-
mation or services according to users current location. For example, a customer
who is shopping in a new mall may need to know where his /her favorite shop
is and how to reach it. Advertisements can be recommended to users based on
their locations etc..
Nowadays, outdoor localization technologies, such as GPS, have become quite
popular in daily life. Recent years also witness the increasing demand for indoor
localization. Several indoor localization techniques have been developed in recent
years, including Bluetooth [3], RFID [8] and WiFi [1]. Bluetooth and WiFi based
approaches rely on Received Signal Strength (RSS) for localization, while RFID
based approaches determine the location of the moving object by reading the

Corresponding author.
1
See http://www.199it.com/archives/29224.html

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 437448, 2013.
c Springer-Verlag Berlin Heidelberg 2013

438 Y. Zhou et al.

active RFID tag. WiFi localization has many advantages, such as ubiquitous
coverage, scalability, no additional hardware required, extended range, no line
of sight restrictions and free measurement etc.
We can see that WiFi is a good choice for localization in indoor space. How-
ever, the accuracy of basic approaches is low due to the instability of WiFi signal
strength. Some projects combined WiFi with other devices such as RFID [5] [13],
Bluetooth [2] [12] to improve the accuracy. Although the accuracy is improved,
additional devices are required, and will cause extra burden to the setup process.
Besides, very few approaches have taken advantage of continuous monitoring of
users locations and considered the features of WiFi signal strength itself.
In this paper, we propose a novel approach to estimate a moving objects
location, by taking into consideration the instability of signal strength and the
movement of objects. This method is based on the Fingerprint approach in
estimating the location of a moving object.
The remainder of this paper is organized as follows. In Section 2, we survey
related work in localization technologies. In Section 3, we discuss our research
methodology. In Section 4, we present and analyze our approach. Section 5 re-
ports a comprehensive experimental study. Finally, we present our conclusions
and provide some future research directions in Section 6.

2 Related Work
The existing work on indoor localization can be divided into two main groups:
signal propagation approaches and fingerprinting approaches.

2.1 Signal Propagation (SP) Approaches

Signal propagation approaches tend to as exponential attenuation models for
WiFi signals and use the path loss2 to determine location based upon distance
from the Access Points (APs), whose locations are assumed known [10]. Al-
though the signal propagation models may work well in open space, they cannot
be used directly in indoor environment since there exists many kinds of ob-
structions. In other words, we need to consider additional path loss caused by
the physical obstructions between the APs in indoor localization and make the
models much more complex than in the open area. The model proposed by Seidel
and Rappaport [11] is the most popular one till now. In this model, the walls
made of dierent materials are assigned dierent attenuation factors for better
precision. SP models can therefore be generalized to locations without available
reference data. Furthermore, it only requires storage for the location of each AP
and a simple description of its signal attenuation. However, since WiFi signals
are unstable, attenuation is almost never radially symmetric. In addition, other
complex factors, such as multipath, dierent construction materials, and various
objects in the building, ever decrease the precision of the above methods.
As a conclusion, signal propagation approaches are easy to implement, but
the performance is low.
2
See http://en.wikipedia.org/wiki/Path_loss
 FIMO: A Novel WiFi Localization Method 439

2.2 Fingerprinting (FP) Approaches

FP approaches ignore attenuation and instead compute likelihoods from location-
specific statistics compiled from reference data. FP needs to create a signal
strength map in o-line manner in order to start localization. When a user sub-
mits a location request, his current signal strength values are compared with
those in the database. The system then compute a most probable location for
the user. Although FP techniques require reference data to setup, reference data
is only needed to be collected once in a new building and the accuracy of FP is
better than SP.
FP approach was developed by Microsoft Research RADAR [1]. In their paper,
they proposed two ways to match signal strength values. One is to return the
position with the nearest neighbor in signal space; the other is to return the
average of the coordinates of k neighbors in signal space as an estimate. In
general, it was found that WiFi signal strength at the same location changes
with time, temperature and objects around it. Thus, it is hard to achieve good
accuracy based on pure matching of signal strength.
Most existing methods estimate the users location only by using of the cur-
rent signal strengths. Ho et. al. proposed a method to retrain the system when
the user finds that the estimation location is wrong [7]. N. Hernandez et. al.
proposed a WiFi localization system based on fuzzy techniques. It is only suit-
able for the small-scale variations which happen when the user moves over a
small distance (in the range of wavelength) [6]. Baniukevic et. al. proposed an
indoor localization algorithm by combining WiFi with Bluebooth [2]. It uses
WiFi signal strength to localization and adjusts the coordinates of the location
using Bluetooth at reference points. The methods proposed by D. L. Lee and Q.
X. Chen have considered the movement of objects, but these methods can only
identify the room where the user is in [4] [9].
Our algorithm is a type of Fingerprint approach. Existing approaches deter-
mine the location according to signal strength received directly. However, the
signal strength is unstable in the indoor environments. Meanwhile, few of them
consider the continuous movement of the user. Our approach records the signal
strength received at the mobile user continuously, eliminates the locations that
are unlikely to be reachable from the previous location by considering the walk
speed and then determines the users location in the possible locations by match-
ing the signal strength with weights (which are got by considering the instability
of signal strength. To be discussed in Section 4.2).

3 Research Methodology
In this section, we describe the experimental testbed and data collection.

3.1 Experimental Testbed

Our experimental testbed is located at the ground floor of a 3-storey building.
Figure 1 shows the layout of the floor (765.43 m2 ). Five Access Points (APs) are
440 Y. Zhou et al.

Fig. 1. Map of the floor where the experiments were conducted. The stars show the
locations of the 5 access points. The black dots denote locations were empirical signal
information was collected.

deployed at dierent locations, denoted as AP1 , AP2 , AP3 , AP4 and AP5 . Each
AP is a Totolink N300R Router. In our tests, a mobile host is a laptop computer
equipped with Intel WiFi Link 5100 AGN running Windows 7. The APs provide
overlapping coverage over the entire floor together.

3.2 Data Collection

Data collection phase is critical in this research methodology. We use Native
WiFi API3 to record the signal strength values. The reference data set and the
testing data sets are created as follows:

reference data set: We select 71 reference points, as shown in Figure 1. At

each reference point, we repeatedly recorded the signal strength for 40 times,
and used the average as the signal strength of this point.
testing data set 1: This data set contains the signal strength values received
by the user when he was walking uniformly from the location near AP5 to
the location near AP1 .
testing data set 2: This data set contains the signal strength values received
by the user when he was walking with variable velocity from the location
near AP5 to the location near AP1 .

In the first data set, each record is described as (x, y, SS), where x and y

represent the location, and SS is an array of signal strengths for five APs. In

the last two sets, the form of the each record tuple is (t, x, y, SS), where t
represents the timestamp to record the signal strength, and the definition of the
rest attributes are as same as the above definition.
3
The Native Wifi API is supported by Windows platform. It automatically sets com-
ponent configures, connects to, and disconnects from wireless networks. Native Wifi
can store profiles on the networks it interacts with in the form of XML documents.
 FIMO: A Novel WiFi Localization Method 441

4 Our Algorithm

In this section, we describe our algorithm in detail. Ideally, we expect the signal
strength of an AP remains almost stable with time going on. Figure 2 shows the
signal strength levels at dierent distances. We can find that almost all of the
signal strength values received 1m away from the AP are the same, while the
stand deviation of the signal strength values received 10m away from the AP is
around 3%. When the distance reaches 35m, the stand deviation can be around
6%. From this, we conclude that the signal strength keeps fluctuating all the
time and the ones with higher value is more stable. We also test the stability
of signal strengths received by a moving object, as shown in Figure 2. In this
figure, a user walks from AP5 to AP1 slowly, we can observe that the signal
strengths would have greater fluctuation when the user is moving. How to use
these unstable signal strengths to localization is challenging to us.

1m 10m 25m 35m AP1 AP2 AP3 AP4 AP5

110 100
100
80
signal strength(%)

signal strength(%)

90
80 60
70 40
60
20
50
40 0
1 11 21 31 41 51 61 71 2 22 42 62 82 102 122 142 162 182
time(s) time(s)

Fig. 2. Signal Level at dierent distances Fig. 3. The signal strength values re-
ceived the user along work

Figure 2 and 3 show that the signal strength would reflect distance more
accurately if the fluctuation of the signal strength could be smoothed. Besides,
in Figure 2, its easy to see that stronger signal strength are more stable. And
the degree of fluctuation of the signal strength changes with time. So we can be
convinced that the signal strength with dierent values and the signal strength
received at dierent time should have dierent credibility.
Due to these particular features, we propose an approach, called FIMO (FInd
Me Out), for indoor localization. FIMO continuously records signal strength
values from each AP at the users location at fixed intervals when the user
is walking around. When the user submits a location request, FIMO will pro-
cess the data and then determine the location of the user using Algorithm 1.
FIMO backtracks the signal strengths received by the user which are recorded
in the database and put them into ssArray (Line 2). It smoothes the signal
strength values in ssArray chronologically and puts the Smoothed Values into
smoothArray (Line 3), as discussed in Section 4.1. After calculating the weight of
442 Y. Zhou et al.

every signal strength value using the Original Value (The signal strength values
that havent been smoothed) (Line 5), as discussed in Section 4.2, FIMO returns
the location to the user by matching the reference points (Line 7-14, Section 4.4)
in the possible range (Line 6, Section 4.3).

Algorithm 1. FInd Me Out (FIMO)

Require: l: threshold of the records tracked back; n: the number of APs; userID:the
ID of the moving object;
Ensure: location of userID: loc
1: Initialize isF irst true and minSim 0;
2: Set ssArray as the latest l records received by userID;
3: smoothArray SM OOT H(ssArray);
4: // smooth the signal strength values in chronological order;
5: Set w1 ...wn based on SSArray;
6: Set rpArray as reference points which locate at the possible range;
7: for each RPc rpArray do

8: simc SIM (
w , smoothArray.get[0].SS, RPc .SS);
9: // compute the similarity value between the signal strengths received right now
and reference points;
10: if (isF irst == true||simc < minSim) then
11: minSim simc ;
12: loc RPc ;
13: end if
14: end for
15: return loc;

4.1 Smoothing the Signal Strength

Since the signal strength is unstable, we smooth the signal strength values be-
fore matching them. As shown in Figure 3, all of values have dierent degree of
fluctuation. We need to smooth the fluctuation while keeping the trend of signal
strength. A moving average4 is commonly used with time series data to smooth
short-term fluctuations and highlight longer-term trends or cycles. So we calcu-
late the moving averages of all values to smooth their fluctuations. Figure 4(a)
shows that this method is eective.
However, there is a special situation when the fluctuation is not continuous,
just as shown in Figure 4(b) . If we just calculate the moving averages, the
smoothed values may be inaccurate. Under the situation, we detect the outliers
and deal with them through the following process before calculating the moving
averages of all values.

1. Detecting the outliers: The system scans the signal strength values ac-
cording to the time sequence. Assume that represents the threshold of
normal value, and the system acquires four sequential values, a, b, c and d.
If (b a > and b c > and c d ) or (a b > and c b >
4
See http://en.wikipedia.org/wiki/Moving_average
 FIMO: A Novel WiFi Localization Method 443

(a) The fluctuation (b) The outliers

Fig. 4. Smoothing the signal strength

and d c ), b will be considered as an outlier. All of a, b, c and d are

measured in percentage. As the threshold for the outlier, is set to be 20%
in our experiments. The system will regard the value as an outlier (as shown
in Figure 4(b)) and proceed to step 2.
2. Dealing with the outliers. The system replaces the outlier b with the
average of a and c. Figure 4(b) shows the result of smoothing.

AP1 AP2 AP3 AP4 AP5

100

80
signal strength(%)

60

40

20

0
2 22 42 62 82 102 122 142 162 182
time(s)

Fig. 5. The signal strength values received by the user along work after smoothing

Figure 5 shows the signal strengths after smoothing. Compared with Figure
3 which is before smoothing, it seems that the quality of the data has improved
obviously after smoothing. We call the values after smoothing the Smoothed
Value and the values having not been smoothed the Original Value.

4.2 Calculating the Weights

Figure 2 shows stronger signal strength tends to be more stable. To a certain

extend, stronger signal strength has higher credibility. There are a lot of moving
objects in indoor spaces, and the number and behaviors of them are constantly
changing. These will cause the fluctuation of signal strength. Significant fluc-
tuation will have material impact upon the credibility of the signal strength.
444 Y. Zhou et al.

Therefore, even at the same location, dierent signal strength has dierent cred-
ibility, and we should give dierent weights to dierent signal strength when
matching the signal strength in the database. Considering the two factors af-
fecting the credibility of signal strength mentioned above, we propose a method
to calculate the weights. The rule is to give the signal strength with the value
which is higher and more stable higher weight. We can use the Original Value
(oi ) of APi s signal strength to represent the credibility of APi s signal strength.
Then it comes to the question of how to measure the stability of signal
strength. It is generally known that the sample standard deviation shows how
much variation from the average (or expected value). A low sample standard
deviation indicates that the data points tend to be very close to the mean, vice
verse. However, the average value is not the expected value because the signal
strength varies with distance, so we cannot simply use the sample standard de-
viation here. Although the trend of the signal strength should be described as
curve function, we just focus on the stability of signal strength in a short time
(such short time is deemed as a time window), and then it can be seen as linear.
We regard the time period containing four continuous signal strength values as
a time window in our experiments, and the length of the time window can be
changed. If the length is too long, the items could not be fit with the linear
function. And if it is too short, the linear function fitting out could not rea-
sonably reflect the trend of the changes of signal strength. The method of least
squares5 is a standard approach to the approximate solution of overdetermined
systems. Least Squares means that the overall solution minimizes the sum of the
squares of the errors made in the results of every single equation. So we use Least
Squares to find out the linear function which could describe the distribution of
signal strength in a short time. Then we can get the expected signal strength
value (ei,j ) received from APi at time tj with the linear function.
Using the expected signal strength values, the sample standard deviation of
several signal strength values (i ) in the same period can be calculated with
Equation (1) below.

@
A W
A 1 "
i = B (oi,j ei,j )2 (1)
W 1 j=1

where W indicates the length of the time window, and oi,j is the Original Value
of APi at time tj .
To sum up, the weight of every signal strength value of APi would be wi
coming from Equation (2).

oi 1
wi = (2)
i maxk (ok /k )

5
See http://en.wikipedia.org/wiki/Least_squares
 FIMO: A Novel WiFi Localization Method 445

4.3 Considering the Walk Speed

The points on the circle centered at AP would have the same signal strength in
free space. And the user might receive the same signal strength values at dierent
locations. Therefore, the users position might be located at a location which is
far away from the actual one. In order to avoid this situation, we should consider
the behavior of the user. In the indoor space, walking is the most common way
of moving and the speed will has an upper limit. Thus, we could use the speed of
the user and the last location of the user recorded to limit the possible candidate
locations.
Specific studies have found pedestrian walking speed ranging from 1.25 m/s to
1.32 m/s for older individuals to 1.48 m/s to 1.51 m/s for younger individuals6 .
The highest normal pedestrian walk speed is about 1.5 m/s. Of course, with
speed testing equipment, the accuracy of the speed can be improved. Here we
use 1.5 m/s as the pedestrian walk speed to limit the candidates. We take the
last location of the user as the center of circle and the maximum distance the
user could walk from the last location as radius. The locations in the circle are
the candidates. But the last location of the user recorded in database might not
be the actual location which the user was, so we should extend the range of
candidate locations properly because each record interval is the same. We use
2 times of the maximum distance as the radius of circle to limit the candidate
locations. The process of calculation can be described as r = 2vT , where r is
the radius of the circle containing candidate locations, v is the pedestrian walk
speed (here we use 1.5m/s) and T is the time from the last record to this record.
The reference points located in the circle with the radius of r, which is centered
at the last location of the user are the candidate locations of the users current
location. Thus, the location of the user must be one of the candidate locations.

4.4 Determining the Location

Similar with the base Fingerprint approach, we use the reference point with the
most similar signal strength values as the users location. We use the Equation
(3) to compute the degree of similarity.
n
" 7 ;
simc = wi (si rc,i )2 (3)
i=1
where simc is the similarity between the signal strength values received by the
user and the signal strength values of the cth reference point which is a candidate
location, n is the number of APs which can be detected by the users device, wi is
the weight of APi s signal strength value, si is the Smooth Value of APi s signal
strength value and rc,i is the APi s signal strength value of the cth reference
point.
The reference point with the signal strength closest to that received by the
user (simc is the minimum) would be returned to the user as the estimated
location .
6
See http://en.wikipedia.org/wiki/Walking
446 Y. Zhou et al.

9 10
FIMO 14 FIMO FIMO
8 FIMO w/o candidates Fingerprint (NN) Fingerprint (KNN)
7 12 8
Erorr distance(m)

Erorr distance(m)

Erorr distance(m)
6 10
6
5 8
4
6 4
3
4
2 2
1 2
0 0 0
1 2 3 4 5 6 7 8 9101112131415161718192021 1 2 3 4 5 6 7 8 9101112131415161718192021 1 2 3 4 5 6 7 8 9101112131415161718192021
ID of query point ID of query point ID of query point

(a) (b) (c)

Fig. 6. The query points of the first user

8
FIMO 16 FIMO 18 FIMO
7 FIMO w/o candidates Fingerprint (NN) 16 Fingerprint (KNN)
14
6
Erorr distance(m)

Erorr distance(m)

Erorr distance(m)
12 14
5 12
10
4 10
8
8
3 6 6
2 4 4
1 2 2
0 0 0
1 2 3 4 5 6 7 8 9101112131415161718192021 1 2 3 4 5 6 7 8 9101112131415161718192021 1 2 3 4 5 6 7 8 9101112131415161718192021
ID of query point ID of query point ID of query point

(a) (b) (c)

Fig. 7. The query points of the second user

5 Experiments

We have conducted series of experiments on the testbed described in Section 3.1.

The testing data sets are also introduced in Section 3.2. We use four dierent
approaches to locate the user. Two approaches are ours, one is FIMO introduced
in Section 4, the other is FIMO without considering the walk speed. We use the
Fingerprint using the nearest neighbor (Fingerprint(NN))and the Fingerprint
using the multiple nearest neighbors (Fingerprint(kNN)) introduced in Section
2.2 as comparisons.
Figure 6 and 7 show the results of localization. Here, x-axis represents the ID
of querying requirement, and the y-axis represents the error. In Figure 6(b)-(c)
and 7(b)(c), we can observe that FIMO is much better than the Fingerprint(NN)
and Fingerprint(kNN). In FIMO, considering the walk speed is the only step that
might requires other device or estimation. In Figure 6(a) and 7(a), we can see the
FIMO without considering the walk speed of the user already has good accuracy,
but the FIMO considered the walk speed could generate better results.
Table 1 shows the sums of squares among the groups7 (SSA) in Figure 6 and 7.
The statistics suggests that the accuracy of FIMO has improved significantly com-
paring with the Fingerprint(NN) and Fingerprint(kNN). The FIMO is proved to
have the best accuracy with respect to all the parameters (i.e. the average, medium
and maximum values,as shown in Table 2). The errors in distances of all query
points are less than 7m, and the points erred within 3m are more than 70%.
7
SSA is a measure of how much the means dier from one another. Its conceptualized a
little dierently, because it is thought of as the variation of each mean from the mean
of the total sample.
 FIMO: A Novel WiFi Localization Method 447

Table 1. Sum of Squares Among the Groups (SSA)

FIMO vs. FIMO w/o FIMO vs. Fingerprint FIMO vs. Fingerprint
Candidates (NN) (kNN)
1 3.79 78.71 43.69
2 0.64 136.69 86.33

Table 2. The average, median and maximum of the error distances

FIMO w/o Can- Fingerprint

FIMO Fingerprint (NN)
didates (kNN)
average 2.17m 2.59m 5.34m 4.62m
median 1.95m 1.99m 4.04m 3.61m
maximum 6.42m 7.64m 15.33m 16.66m

6 Conclusions and Future Works

In this paper, we presented FIMO, a WiFi-based localization method. In this

approach, the system records the signal strength values the user receives while
he is moving, and determines the users location by backtracking the signal
strength values and the users movement. Experimental reports show that the
performance is significantly improved over existing methods.
In this paper, our work merely applies to specific environmental. We are con-
sidering to expand the application to more complicated situations, such as mul-
tiple floors or models of access points etc. Since FIMO is based on fingerprint
approach, a lot of data is required to be collected for reference points in ad-
vance. If the signal strength values of some reference points could be estimated,
the setup would be less time consuming. The curve fitting might be a practicable
method. We believe that it is a direction worth exploring in the future.

Acknowledgments. The research of Cheqing Jin is supported by the National

Basic Research Program of China (Grant No. 2012CB316200), the Key Program
of National Natural Science Foundation of China (Grant No. 60933001), National
Natural Science Foundation of China (Grant No.61070052). The research of Aoy-
ing Zhou is supported by National Science Foundation for Distinguished Young
Scholars (Grant No. 60925008), and Natural Science Foundation of China (No.
61021004).

References
1. Bahl, P., Padmanabhan, V.N.: Radar: An in-building rf-based user location and
tracking system. In: Proc. of INFOCOM, pp. 775784 (2000)
2. Baniukevic, A., Sabonis, D., Jensen, C.S., Lu, H.: Improving wi-fi based indoor po-
sitioning using bluetooth add-ons. In: Proc. of MDM, pp. 246255. IEEE Computer
Society, Washington, DC (2011)
448 Y. Zhou et al.

3. Bargh, M.S., de Groote, R.: Indoor localization based on response rate of bluetooth
inquiries. In: Proc. of MELT, pp. 4954. ACM, New York (2008)
4. Chen, Q., Lee, D.-L., Lee, W.-C.: Rule-based wifi localization methods. In: Proc.
of EUC, pp. 252258. IEEE Computer Society, Washington, DC (2008)
5. Chen, Y.-C., Chiang, J.-R., Chu, H.-H., Huang, P., Tsui, A.W.: Sensor-assisted
wi-fi indoor location system for adapting to environmental dynamics. In: Proc. of
MSWiM, pp. 118125. ACM, New York (2005)
6. Hern andez, N., Herranz, F., Ocana, M., Bergasa, L.M., Alonso, J.M., Magdalena,
L.: Wifi localization system based on fuzzy logic to deal with signal variations. In:
Proc. of ETFA, pp. 317322. IEEE Press, Piscataway (2009)
7. Ho, W., Smailagic, A., Siewiorek, D.P., Faloutsos, C.: An adaptive two-phase ap-
proach to wifi location sensing. In: Proc. of PERCOMW, pp. 452456. IEEE Com-
puter Society, Washington, DC (2006)
8. Jin, G.-Y., Lu, X.-Y., Park, M.-S.: An indoor localization mechanism using active
rfid tag. In: Proc. of SUTC, pp. 4043. IEEE Computer Society, Washington, DC
(2006)
9. Lee, D.L., Chen, Q.: A model-based wifi localization method. In: Proc. of InfoScale,
pp. 40:140:7. ICST (2007)
10. Letchner, J., Fox, D., Lamarca, A.: Large-scale localization from wireless signal
strength. In: Proc. of the National Conference on Artificial Intelligence, pp. 1520.
The MIT Press (2005)
11. Seidel, S.Y., Rapport, T.S.: 914 MHz path loss prediction model for indoor wireless
communications in multi-floored buildings. In: Proc. of Antennas Propagation.
IEEE Trars (1992)
12. Wang, R., Zhao, F., Luo, H., Lu, B., Lu, T.: Fusion of wi-fi and bluetooth for
indoor localization. In: Proc. of MLBS, pp. 6366. ACM, New York (2011)
13. Yeh, L.-W., Hsu, M.-S., Lee, Y.-F., Tseng, Y.-C.: Indoor localization: Automati-
cally constructing todays radio map by irobot and rfids. In: Proc. of Sensors, pp.
14631466 (2009)
 An Algorithm for Outlier Detection
on Uncertain Data Stream

Keyan Cao, Donghong Han, Guoren Wang, Yachao Hu, and Ye Yuan

College of Information Science & Engineering, Northeastern University, China

Key Laboratory of Medical Image Computing (NEU), Ministry of Education
[email protected]

Abstract. Outlier detection plays an important role in fraud detection,

sensor net, computer network management and many other areas. Now
the flow property and uncertainty of data are more and more apparent,
outlier detection on uncertain data stream has become a new research
topic. Firstly, we propose a new outlier concept on uncertain data stream
based on possible worlds. Then an outlier detection method on uncer-
tain data stream is proposed to meet the demand of limited storage and
real-time processing. Next, a dynamic storage structure is designed for
outlier detection on uncertain data stream over sliding window, to meet
the demands of limited storage and real-time response. Furthermore, an
ecient range query method based on SM-tree(Statistics M-tree) is pro-
posed to reduce some redundant calculation. Finally, the performance
of our method is verified through a large number of simulation exper-
iments. The experimental results show that our method is an eective
way to solve the problem of outlier detection on uncertain data stream,
and it could significantly reduce the execution time and storage space.

Keywords: Outlier detection, uncertain data stream, possible world.

1 Introduction
Uncertainty is inherent in data collected in various applications, such as sensor
networks, marketing research, and social science [7]. Since the intrinsic dierences
between uncertain and deterministic data, existing data mining algorithms on
deterministic data are not suitable for uncertain data. Such data present in-
teresting research challenges for a variety of data set and data mining applica-
tions [1] [4] [6] [7] [9] [10] [11]. Recent years, more and more attention is paid to
uncertain data mining.
Outlier detection is considered as an important data mining task, aiming at
discovery of the elements that are significantly dierent from other elements,

This research are supported by the NSFC (Grant No. 61173029, 61025007, 60933001,
75105487 and 61100024), National Basic Research Program of China (973, Grant
No. 2011CB302200-G), National High Technology Research and Development 863
Program of China (Grant No. 2012AA011004) and the Fundamental Research Funds
for the Central Universities (Grant No. N110404011).

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 449460, 2013.
c Springer-Verlag Berlin Heidelberg 2013

450 K. Cao et al.

or does not conform to the expected normal behavior, or conforms well to a

defined abnormal behavior [3] [5] [8]. Now the flow property and uncertainty of
data more and more apparent, outlier detection on uncertain data stream has
become a new research topic.
To the best of our knowledge, no existing algorithm considers distance-based
outlier detection on uncertain data stream that uncertain tuple consist of a set of
possible instances. In this paper, the algorithm is proposed for outlier detection
on uncertain data stream over sliding window that uncertain data tuple consist
of a set of instances. In summary, the major contributions of this work are as
follows:
The first algorithm for outlier detection over sliding window on uncertain
data stream that are tuple level uncertainty is designed.
A dynamic storage structure is designed for outlier detection on uncertain
data stream over sliding window, to meet the demands of limited storage and
real-time response.
We propose an ecient range query method based on SM-tree(Statistics
M-tree), it could reduce some of unnecessary calculations thereby improve the
implementation eciency of the algorithm.
We also implement a series of experiment evaluations, showing that our
methods are eective and ecient.
This paper is organized as follows. In the next section, we will introduce the
related work. After that, we introduce semantic of outlier on uncertain data
stream in section 3. Section 4 proposes the method for outlier detection on
uncertain data stream. In section 5, we will present the experimental results.
Section 6 concludes the paper.

2 Related Work
In this section, we illustrate related work in two fields: (1) outlier detection on
uncertain data; (2) outlier detection on uncertain data stream.

2.1 Outlier Detection on Uncertain Data

The first method for outlier detection on uncertain data was proposed in [2].
Uncertain data object was represented by a probability density function. The
algorithm used a probabilistic definition of outliers in conjunction with density
estimation and sampling. However, their work only focused on detecting outlier
objects without considering outlier instances.
Another outlier detection method for uncertain data was presented in [12].
If the existing probability of neighbor object was very low, then the centered
object was a distance-based outlier. Their method only concerned tuple level
uncertainty, i.e. probabilistic dimension call confidence, so it was not suitable
for property level uncertainty.
Bin Jiang et al. [7] started with a comprehensive model considering both
uncertain objects and their instances. They assumed that the uncertain objects
 An Algorithm for Outlier Detection on Uncertain Data Stream 451

with similar properties tended to have similar instances, the normal instances of
each uncertain object using the instances of objects with similar properties.
The algorithms mentioned above work are all in static fashion. This means
that the algorithm must be executed from scratch if there are changes in the
underlying data objects, leading to performance degradation when operate on
uncertain data stream.

2.2 Outlier Detection on Uncertain Data Stream

The first method for outlier detection on uncertain data stream was proposed in
[13]. They studied the semantic of outlier detection on probabilistic data stream
and proposed a new definition of distance-based outlier over sliding window.
However, this method only concern the tuple level uncertainty, not the attribute
level, so it is not suitable for the uncertain data model studied in this paper.
In this research, we propose a new algorithm for outlier detection on uncertain
data stream which is described by a set of discrete instances over sliding window.
Our techniques meet the demand of uncertain data stream for running time and
storage requirements. Therefore, our study extends to this new research field.

3 Semantic of Outlier on Uncertain Data Stream

This section serves a two-fold purpose: firstly to introduce the uncertain outlier
definition, and secondly to explain the meaning of the symbol. The traditional
outlier definition on certain data is as follows:
Definition 1. (Neighbors) R is a user specified threshold, R 0. xi and xj
are two data objects, if the distance between them is no longer than R, then xi is
a neighbor of xj . Let n(xi , R) denotes the number of neighbors that data object
xi has.
Definition 2. ((K, R)-Outlier) Let R 0, K 0, be a user-specified thresh-
old. A data object xi is marked as an outlier, if the number of neighbors that xi
has is less than K, i.e. n(xi , R) < K.
This definition is only suitable for determine data. For uncertain data set, we
need to consider all possible worlds to determine whether the uncertain data
tuple is an outlier in the current window.
It is assumed that uncertain data stream consist of uncertain tuples x1 , x2 ,
, xi , , xn , arriving at time stamp t1 , t2 , , ti , , tn , for any i < n,
ti < tn . Each uncertain tuple has m possible instances and existence #m probability,
i.e. xi = {(x1i , px1i ), (x2i , px2i ), , (xji , pxj ), , (xm
i , p m
xi )}, j=1 x = 1.
p j
i i

Definition 3. ((R, K)-Outlier Possible World ) We define a possible world

is a (R, K)-outlier possible world of uncertain tuple xi , if the tuple xi is an
outlier in this possible world. Let Nxi is a set of (R, K)-Outlier Possible World
452 K. Cao et al.

that xi has. sum(xi ) is the probability

# summation of ((R, K)-Outlier Possible
World probability in Nxi , sum(xi ) = wi Nx Pwi .
i

Definition 4. ((R, K, )-Outlier) We define that an uncertain tuple xi to be

a (R, K, )-outlier, if sum(xi ) is at lest .
The set of outliers is denoted by O(R, K, ), the set of the candidates is denoted
by C(R, K, ), the candidates is non-outlier in current window, it is possible
become outlier. The set of safe inliers (never the outliers) denoted by I(R, K, ).
Outliers set and candidates set are stored, safe inliers set neednt be stored.
Table 1 summarizes of the most frequently used symbols throughout the paper,
along with interpretation.

4 Outlier Detection on Uncertain Data Stream

In this section, we provide a new method for outlier detection on uncertain
stream. We start by describing the framework for outlier detection, and then
introduce storage structure and incremental method. If we want to get the outlier
on uncertain data stream, we can not use the outlier detection method on static
database, because of the most straightforward scratch, towards updating the
results. It is expected to be very computationally expensive. A wise choice is to
design incremental algorithm, it processes only the changes based on previous
result to produce the new result for current window. Next, we will introduce our
methods.

Table 1. Frequently Used Symbols

symbol Interpretation
K the number of neighbors parameter
R the distance parameter for the outlier detection
W the windows size
S the set of uncertain tuples in the current window
slide the window slide length
xji the j-th instance of uncertain tuple xi
p xj the existing probability of instance xji
i
ppxj the probability that instance xji is an outlier
i
sum(xi) the probability summation of possible worlds that xi is an outlier in
n(xji , R) the number of neighbors that instance xji has
Sx j the succeeding neighbors set of instance xji has
i
Pxj the preceding neighbors set of instance xji has
i
Nxi the set of (R,K)-outlier possible worlds that xi has
x*i the store structure of xi after updated
xarr . the newly arrived tuple
xexp . the latest expired tuple
 An Algorithm for Outlier Detection on Uncertain Data Stream 453

4.1 Framework
At first, we introduce the storage structure, in order to meet the demand for
increment processing. We define the storage structure, which is designed to cap-
ture key information of outlier during the course of uncertain data stream. For
example, we list partial uncertain tuple in uncertain data stream:
x1 {[(1, 2), (0.7)], [(1, 1), (0.3)]}, x2 {[(2, 1), (0.4)], [(3, 1), (0.6)]},
x3 {[(2, 2), (0.5)], [(2, 3), (0.5)]}, x4 {[(4, 4), (0.8)], [(3, 3), (0.2)]}
There are the value of the four tuples from x1 to x4 . x1i and x2i are two instances
of tuple xi , px1i and px2i are exist probability of x1i and x2i respectively. let R = 2,
K = 2, range query results for every instance in the current window is showed
in Figure 1(range query is find neighbors for instance). We neaten the range
query result on the two principles: (1) The instance of one tuple can not appear
in the range query result of the same tuples instance, i.e., in the range query
result of x11 , x21 can not exist, and delete x21 from the range query results of x11 .
(2) If there are at least two instances of the same uncertain tuple in the range
queries result of a instance, then merge these instances existing probability of
the same uncertain tuple, i.e., x13 and x23 are in the range queries result of x11 , the
probability of x13 and x23 are added, px13 + px23 = px3 . The result after neatening
is more conducive to analyze in figure 2.
Since the tuple is uncertain data, the probability of an uncertain tuple being
an outlier is probability summation of its every instance being an outlier as
Definition 5.
Definition 5. (Uncertain Tuple Outlier Probability) We assume that the
uncertain tuple xi consist of m instances ((x1i , px1i ), (x2i , px2i ), , (xji , pxj ),
i
j
, (xm
i , pxi )), 1 j m, Pxji is the probability of instance xi . The probability
m

of uncertain tuple being an outlier is#equal to the probability summation of each

m
instance being an outlier, i.e. ppxi = j=1 ppxj
i

By the above Definition 5, we can calculate the probability of x1 being an outlier,

which is equal to the probability summation of x11 and x21 being outlier, i.e. px1
= ppx11 + ppx21 . It converts judging whether an uncertain tuple as an outlier to
get the probability of each instance being an outlier.
Definition 6. (Instance Outlier Probability) The probability of the instance
xji being an outlier is denoted by ppxj , ppxj =pxj pn(xj ,R)<k , pn(xj ,R)<k denotes
i i i i i

probability that the number of neighbor instance xji has is less k, 0 pn(xj ,R)<k 1.
i

In figure 2, the range query result of x11 contain x12 and x3 , the probability of
x11 being an outlier ppx11 is 0.7 (1 0.4) = 0.42. The probability of x21 being
an outlier ppx21 is 0.3 (1 0.5) = 0.15. Then we can get the probability of x1
being an outlier is equal to px1 = ppx11 + ppx21 =0.42+0.15=0.57.
Definition 7. (Non-outlier Instance) Let sxi denote the set of instances that
tuple xi has. We determine the instance is a non-outlier instance, if the instance
is not an outlier in any possible world. syxi is a set of outlier instances that tuple
454 K. Cao et al.

Fig. 1. Result of range query without neaten Fig. 2. Result of range query after
neaten

xi has, and snxi is a set of non-outlier instance that tuple xi has.(Since these two
sets do not overlap and cover the complete objects set, i.e., syxi
) snxi = sxi , syxi
n
sxi = .#If the probability summation of non-outlier instance is greater than
1 , i.e. xj sxn pxj > 1 , then the tuple is non-outlier.
i i i

Through analysis of the results, we can get the outlier in the current window,
and then we will design a storage structure. Since we are dealing with uncer-
tain data in stream environment, the arrival time of uncertain data will aect
the result. Firstly, we divide the range query results of every instance into two
sets, preceding neighbors and succeeding neighbors set, denoted by Pxj and Sxj
i i
respectively.
Definition 8. (Non-outlier Instance Probability) The non-outlier instance
probability increase, along with the neighbor number n(xji , R) of an instance
xji increase. If the exist of a part neighbors, non-outlier instance probability is
greater than 1 , then xji is non-outlier instance.
According to the above definition, we only need to store the neighbors that
make the tuple become non-outlier(according to the time order), instead of stor-
ing all neighbors of each instance. The tuple in Pxj is eliminated before the
i
tuple xi . It may make tuple xi become outlier from candidate set, therefore,
we must consider the expires time of tuple in candidate set. The tuples in Pxi
is stored with time information and probability. The tuples in the Sxi can not
be expired before the tuple xi was deleted, so we only store the probability
information. For example, in figure 3, we depict an example in two dimen-
sional, for K = 3, = 0.7. Let the subscripts denote the order of arrival of
these tuples. We focus on tuple x18 , the range query result of instance x118 is
{x16 (0.4), x210 (0.6), x115 (0.5), x216 (0.6), x117 (0.7), x219 (0.8)}, the range query result of
instance x218 is {x23 (0.6), x116 (0.4), x120 (0.7), x222 (0.8)}. Range query result based
on time, be divided into two sets to store, i.e., Sxj and Pxj . In order to reduce
18 18
the storage space to meet the demand of uncertain data stream, we only store
the necessary information. According arrival time of each neighbor, to find the
existence of the neighbors make tuple x18 is non-outlier, i.e., pxi (no) > 1 .
 An Algorithm for Outlier Detection on Uncertain Data Stream 455

When x22 , x20 , x19 , x17 and x16 exist, px18 (no) = px118 (no) + px218 (no) = 0.3136 >
1 , then x18 is non-outlier. We only store structure of x18 as follows:

x118 (0.8){(16, 0.6), (17, 0.7)}{0.8}

x218 (0.2){(16, 0.4)}{0.7, 0.8}

4.2 Incremental Processing

In this paper, we use the count-based window which always maintains the W
most recent uncertain tuples. The uncertain tuples maintained by the sliding
window are termed active objects. When the tuple leaves the window we think
that the tuple expires. Along with the arrival of the new uncertain tuples, the
old tuples disappear from the window.
In the section, we introduce the incremental processing for outlier detection
on uncertain data stream. Whenever the window slid, the new uncertain tuple
arrives, and the old uncertain tuple is moved out from the window. In order to
reduce calculation, we first delete the expired tuple, and then receive a new un-
certain tuple. When the old tuple is deleted, it may make the tuple in candidate
set become outlier. The impact that a new tuple arrives includes two classifica-
tions: (1) The impact to a new uncertain tuple from existing uncertain tuples.
(2) The impact to exist uncertain tuples from a new uncertain tuple. It might
make the outlier become candidate tuple, or make the candidate tuples become
a safe inlier.
When a new tuple arrives, the range query is done according the probability
descending of each instance. According to Definition 6, the important tuples are
found and stored in the data structure. Range query could be done according the
time of tuple. For instance, when tuple xi arrivals, we first determine whether
the instance of xi1 is neighbor of instance of xi , and then determine xi2 , in
turn forward according to the time of arrival, until finding the tuples that make
xi is non-outlier. These instances could be store in store structure of xi . The
pseudo code of these operations is given in Algorithm 1.
When the tuple is expired, it is removed from the window, if it is in O(R, K, )
or C(R, K, ). Tuple storage structure is updated, that the store structure con-
tain expired tuple instance. If the tuple is in O(R, K, ), we only update the
storage structure. If the tuple is in C(R, K, ), update the storage structure,
and then determine whether the tuple become an outlier. The pseudo code of
these operations is given in Algorithm 2.
In figure 3, we assume that a new tuple x28 arrives, instance x128 is neighbor
of x118 and its existing probability is 0.9, then according Definition 6, when x28 ,
x22 , x20 , x19 and x17 exist, x18 is non-outlier but not a safe one. Then we update
the store structure of x18 as follows:

x118 (0.8) : {(17, 0.7)}, {0.8, 0.9}

x218 (0.2) : {}, {0.7, 0.8}
456 K. Cao et al.

Algorithm 1. Arrival (tuple xarr , now)

Require: xnow : the arrival tuple, now: the current time window
1: make a range query for each instance of tuple xarr , let the set of object
returned;
2: for each instance xji do
3: add xarr to Sxi , update xi storage structure
4: if xi O(R, K, ) then
5: if x*i is candidate then
6: add x*i to C(R, K, )
7: end if
8: if x*i is safe inlier then
9: delete store structure of x*i from window
10: end if
11: end if
12: if xi C(R, K, ) then
13: if x*i is an safe inlier then
14: add x*i to I(R, K, )
15: end if
16: end if
17: end for
18: construct xarr storage structure
19: if xarr is an outlier then
20: add xarr to O(R, K, )
else
21: add xarr to C(R, K, )
22: end if

Some time later, when tuple x35 arrivals, instance x235 is the neighbor of x118 ,
and its probability is 0.6, we can see, the existence of x35 , x28 , x22 , x20 and
x19 make x18 become non-outlier, and they are all in set Sx18 , so x18 is a safe
inlier. When x18 become a safe inlier along with window slid, we deleted store
structure of x18 .

4.3 Improve the Eciency of Range Queries

In the previous section, we introduce the method of range queries from back to
front according arrival time. To be able to improve the eciency of range query,
we use the SM-tree(Statistics M-tree) to range query.
The formation of SM-tree that each instance in current window is a left node,
and the radius of the leaf node is R is showed in figure 4. Parent node con-
tains statistic information of child nodes, M u and M l are min and max time
of M node respectively. When a new tuple arrives, the range query is done for
each instance of the new tuple. We assume that the new instance in the range
M1 and M2 , because M1u = t4 , M1l = t3 , M2u = t1 , M2l = t1 , i.e., M2u < M1l ,
then first calculate the relationship between leaf nodes of M1 . If we get the all
 An Algorithm for Outlier Detection on Uncertain Data Stream 457

Algorithm 2. Delete (tuple xi .exp, now)

Require: xexp : the departing tuple, now: the current time window
1: if xexp in C(R, K, ) or O(R, K, ) then
2: remove xexp from C(R, K, ) or O(R, K, )
3: end if
4: for each tuple xi which storage structure contain expired tuple xexp do
5: update xi storage structure
6: if xi C(R, K, ) then
7: if x*i is an outlier then
8: add x*i to O(R, K, )
9: end if
10: end if
11: end for

instances that make the new tuple become non-outlier in the range of M1 , it
isnt necessary to consider the leaf node of M2 . It could reduce the unnecessary
calculations with SM-tree to improve eciency of range queries.

m1 m2 m3

t14 t15 t16 t17

Fig. 3. Example of tuple store structure Fig. 4. SM-tree

5 Performance Evaluation
We conducted a series of experiments on synthetic data set and real data sets
to evaluate the performance of the algorithms proposed in this paper, namely
Uncertain Outlier Detection on Uncertain Data Stream (denoted by UOD), and
Uncertain Outlier Detection based on M-tree(denoted by UOD-M). For com-
parison, we implemented Naive Outlier Detection based on Examining Possible
Worlds method named (denoted by NUOD).
All the algorithms are implemented in visual C 2010, and experiments are
carried out on a PC with a Core i3-3.3 GHz CPU and 4G of main memory.

5.1 Running Time

We study the performance of the proposed methods by varying several of impor-
tant parameters such as the window size W, the number of outliers, the number
of required neighbors K, the distance R and the probability threshold .
458 K. Cao et al.

40 40 40

30 30 30
Running Time(s)

Running Time(s)

Running Time(s)
20 20 20

10 10 10
NUOD NUOD NUOD
UOD UOD UOD
UOD-M UOD-M UOD-M
0 0 0
20 40 60 80 100 20 40 60 80 100 20 40 60 80 100
Window Size(K) Window Size(K) Window Size(K)

(a) SynDiscrete (b) Network stream intru- (c) Forest cover

sion

Fig. 5. Running time vs. window size

30 30 30
NUOD NUOD NUOD
UOD UOD UOD
UOD-M UOD-M UOD-M
25 25 25
Running Time(s)

Running Time(s)

Running Time(s)
20 20 20

15 15 15

10 10 10
0.5 1 1.5 2 2.5 3 0.5 1 1.5 2 2.5 3 0.5 1 1.5 2 2.5 3
Outlier(% W) Outlier(% W) Outlier(% W)

(a) SynDiscrete (b) Network stream intru- (c) Forest cover

sion

Fig. 6. Running time vs. number of outliers (%W )

60 60 60
Running Time(s)

Running Time(s)

Running Time(s)

40 40 40

20 20 20

UOD UOD UOD

UOD-M UOD-M UOD-M
0 0 0
20 40 60 80 100 20 40 60 80 100 20 40 60 80 100
K value K value K value

(a) SynDiscrete (b) Network stream intru- (c) Forest cover

sion

Fig. 7. Running time vs. k value

60 60 60

50 50 50
Running Time(s)

Running Time(s)

Running Time(s)

40 40 40

30 30 30

20 20 20

10 10 10
UOD UOD UOD
UOD-M UOD-M UOD-M
0 0 0
10 15 20 25 30 10 15 20 25 30 10 15 20 25 30
R value R value R value

(a) SynDiscrete (b) Network stream intru- (c) Forest cover

sion

Fig. 8. Running time vs. R value

 An Algorithm for Outlier Detection on Uncertain Data Stream 459

6 6 6
Running Time(s)

Running Time(s)

Running Time(s)
4 4 4

2 2 2

UOD UOD UOD

UOD-M UOD-M UOD-M
0 0 0
0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8
Probability Threshold Probability Threshold Probability Threshold

(a) SynDiscrete (b) Network stream intru- (c) Forest cover

sion

Fig. 9. Running time vs. value

First, we test the performance of the algorithm for varying values of W from
104 to 105 . Figure 5 certify the running time of all algorithms increased when
increasing the length of the sliding window. The performance of UOD-M and
UOD algorithm performs better than NUOD.
Figure 6 depicts the results that number of outliers varying in range [0.1%, 3%].
As expected, UOD-M algorithm performs better than UOD algorithm.
In figure 7, we show that the performance of NUOD, UOD-SM and UOD
algorithm. We vary K from 1 to 100. Figure 7 show that the running time of
UOD-M and UOD decreased when increasing the value of K.
Figure 8 show that running time of UOD is much higher than UOD-M for
varying R from [10, 50]. Figure 9 demonstrate that the running time when in-
creasing the probability threshold . We vary value from 0.1 to 0.9. Figure 9
show that of both algorithms decreasing when increasing value.

5.2 Memory Consumption

Figure 10 show the memory consumption of three data sets. The consumed mem-
ory corresponds to the memory needed to store the information for uncertain
tuples in current window. From the figure 10, we can see the required amount
of memory is only a small fraction of the total memory available in modern
machines.

10 10 10
Memory Requirements(MBytees)

Memory Requirements(MBytees)

Memory Requirements(MBytees)

8 8 8

6 6 6

4 4 4

2 2 2
UOD UOD UOD
UOD-M UOD-M UOD-M
0 0 0
10 20 30 40 50 10 20 30 40 50 10 20 30 40 50
Window Size(K) Window Size(K) Window Size(K)

(a) SynDiscrete (b) Network stream intru- (c) Forest cover

sion

Fig. 10. Memory requirements vs. window size

460 K. Cao et al.

6 Conclusion

In this paper we studied the problem of outlier detection on uncertain data

stream, which the uncertain tuple consist of a set of instances and existing prob-
ability. We proposed outlier definition based on possible worlds, and then the
method used to outlier detection on uncertain data stream is proposed. A dy-
namic storage structure was designed. We have conducted extensive experiments
to evaluate the relative performance of the algorithms. The results showed that
the algorithm can eectively detect outliers.
We will continue to study the problem based on outlier detection on uncertain
data stream. Next steps of our work are taking more complex uncertain data,
such as outlier detection on multi-dimensional uncertain data stream.

References
1. Aggarwal, C.C.: On density based transforms for uncertain data mining. In: ICDE,
pp. 866875 (2007)
2. Aggarwal, C.C., Yu, P.S.: Outlier detection with uncertain data. In: SDM, pp.
483493 (2008)
3. Assent, I., Kranen, P., Baldauf, C., Seidl, T.: Anyout: Anytime outlier detection
on streaming data. In: VLDB, pp. 228242 (2012)
4. Burdick, D., Deshpande, P.M., Jayram, T.S., Ramakrishnan, R., Vaithyanathan,
S.: Olap over uncertain and imprecise data. In: VLDB, pp. 970981 (2005)
5. Chandola, V., Banerjee, A., Kumar, V.: Outlier detection: A survey. ACM Com-
puting Surveys (2007) (to appear)
6. Cheng, R., Kalashnikov, D., Prabhakar, S.: Evaluating probabilistic queries over
imprecise data. In: SIGMOD, pp. 551562 (2003)
7. Jiang, B., Pei, J.: Outlier detection on uncertain data: Objects, instances, and
inferences. In: ICDE, pp. 422433 (2011)
8. Kontaki, M., Gounaris, A., Papadopoulos, A., Tsichlas, K., Manolopoulos, Y.:
Continuous monitoring of distance-based outliers over data streams. In: ICDE,
pp. 135146 (2011)
9. Sarma, A.D., Benjelloun, O., Halevy, A., Widom, J.: Working models for uncertain
data. In: ICDE, p. 7 (2006)
10. Singh, S., Mayfield, C., Prabhakar, S., Shah, R., Hambrusch, S.: Indexing uncertain
categorical data. In: ICDE, pp. 616625 (2007)
11. Tao, Y., Cheng, R., Xiao, X., Ngai, W.K., Kao, B., Prabhakar, S.: Indexing multi-
dimensional uncertain data with arbitrary probability density functions. In: ICDE,
pp. 922933 (2005)
12. Wang, B., Xiao, G., Yu, H., Yang, X.: Distance-based outlier detection on uncertain
data. CIT 1, 293298 (2009)
13. Wang, B., Yang, X., Wang, G., Yu, G.: Outlier detection over sliding windows for
probabilistic data streams. Journal of Computer Science and Technology 25(3),
389400 (2010)
 Improved Spatial Keyword Search
Based on IDF Approximation

Xiaoling Zhou, Yifei Lu, Yifang Sun, and Muhammad Aamir Cheema

University of New South Wales, Australia

{xiaolingz,yifeil,yifangs,macheema}@cse.unsw.edu.au

Abstract. Spatial keyword search is a widely investigated topic along

with the development of geo-positioning techniques. In this paper, we
study the problem of top-k spatial keyword search which retrieves the
top k objects that are most relevant to query in terms of joint spatial and
textual relevance. Existing state-of-the-art methods index data objects
in IR-tree which supports textual and spatial pruning simultaneously,
and process query by traversing tree nodes and associated inverted files.
However, these search methods suer from vast number of times of ac-
cessing inverted files, which results in slow query time and large IO cost.
In this paper, we propose a novel approximate IDF-based search algo-
rithm that performs nearly twice better than existing method, which are
shown through an extensive set of experiments.

1 Introduction
There has been a strong trend of incorporating spatial information into keyword-
based search systems in recent years for the support of geographic queries. It
is driven mainly by two reasons: highly user demands and rapid advances in
Geo-positioning and Geo-coding technologies. User interests include finding the
nearby restaurants, tourist attractions, entertainment services, public transport
and so on, which renders the spatial keyword queries important. Studies [13] have
shown that about 20% of web search queries are geographical and exhibit local
intent. In addition, both the various advanced Geo-positioning technologies such
as GPS, Wi-Fi and cellular geo-location services (oered by Google, Skyhook,
and Spotigo), and dierent geo-coding technologies which enable annotating of
web content with positions accelerate the evolution of spatial keyword search.
Spatial keyword queries take a set of keywords and a location as parameters
and return the ranked documents that are most textually and spatially relevant
to the query keywords and location. Usually, distance is used to measure spatial
relevance and document similarity functions such as TF-IDF weighting [7] are
used to measure textual relevance. A fair amount of investigations have been
conducted aiming at this issue, such as the algorithm of solving kNNs (k-nearest
neighbor queries) using R-tree [12], loose combinations of inverted file and R*-
tree [2] [6] [15], and signature augmented IR -tree [4]. Whereas limitations exist

This research was supported by ARC DP130103401 and DP130103405.

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 461472, 2013.
Springer-Verlag Berlin Heidelberg 2013
462 X. Zhou et al.

in those works, they either cannot support pruning simultaneously according to

both keyword similarity and spatial information, or become infeasible when it
comes to handling ranked queries.
One of the most promising and popular ones is IR-tree index based method.
This particularly designed index structure enables simultaneous spatial pruning
and textual filtering, and also an incremental searching and ranking process.
Most importantly, it supports query specific IDF(Inverse Document Frequency)
values computation for determining document textual scores, which is more rea-
sonable than traditional IDF values used in purly textual search. Essentially,
IDF value measures the selectivity(or rarity) of a query keyword. In purly tex-
tual search cases, for a word w, IDFw is computed as the logrithm of Collection
Size (|D|) divided by Document Frequency of w (DFw ). While in spatial key-
word search cases, it is inappropriate to keep using this tranditional IDF values
as the word distribution among documents within query spatial range might be
completely dierent with word distribution in whole documents set. Therefore,
the query specific IDF value used in spatial keyword search is computed as the
logrithm of Total Number of Documents within Query Range (Ds )divided by
Word DF within Query Range (dfw,Ds ). However, in order to determine IDF
value of each query keyword for later calculating textual score of each object,
IR-tree based searching method suers from multiple times of accessing in-
verted files attached with leaf nodes of the index tree, which deteriorates search-
ing performance substantially, espectially when their inverted files are stored
on disk.
Based on the above observation, in this paper, we propose a new searching al-
gorithm which does not access inverted files when computing IDFs thus results
in an IDF range for each query keyword, and then retrieve top-k documents
based on range scores computed using IDF ranges. Our approximate IDF-based
method reduces query search time considerably to nearly half of the original
search method while without compromising accuracy of final ranked results. We
conduct a comprehensive set of experiments that clearly show the search perfor-
mance gain of our proposed method in comparison with the original algorithm
under a wide variation of various query or index parameters. Some essential pre-
liminaries related to spatial keyword search are presented in Section 2. Section 3
details our proposed approximate IDF-based algorithm. Section 4 provides re-
sults of experiments over a wide range of parameter settings and the last section
concludes the work.

2 Preliminaries
2.1 Problem Definition
We consider a database consists of N documents, each associated with a spatial
location. A query Q also consists of a set of keywords (denoted as WQ ), and a
location (denoted as Q.loc).
 Improved Spatial Keyword Search Based on IDF Approximation 463

Given a query range SQ and a scoring function score(Q, d), a top-k spatial
keyword search returns top k scoring documents from the database.
The scoring function should take into consideration both the textual rele-
vance of the document to the query as well as the spatial proximity between
the locations of the query and the document. Hence, the final function can be
represented as score(Q, d) = scoretext (Q, d)+ (1 )scorespatial (Q, d), where
[0, 1] is a tuning parameter.

Spatial Relevance. We only need to consider documents within the search range
as specified by SQ (these doucments are also known as spatially qualified ). For
spatially qualified documents, we give its proximity score based on the euclidean
distance between its location and the query location, i.e., scorespatial (Q, d) =
1 dist(d.loc,Q.loc)
distmax , where distmax is the maximum possible distance between the
query and any spatially qualified document.

Textual Relevance. We adopt the popular Vector Space Model to measure the
textual relevance between the document and the query. In this paper, we con-
sider approaches using the query-specific IDF, which is defined as idfw,D,SQ =
Ds
log dfw,D . Note that Ds is number of spatially qualified documents rather than
s
that of the entire collection of documents, and dfw,Ds refers to number of spa-
tially qualified documents that contain w, therefore, idfw,D,SQ reveals query
words distributions among spatially qualified documents rather than in the
entire collection. Finally, the textual score is calculated as: scoretext (Q, d) =
#
wWQ (tfw,d idfw,D,SQ ). We say a document d is textually qualified to a given
query Q i d contains at least one query keyword of Q.

Notations. We use the following notations in the rest of the paper. The root of an
IR-tree is r. The number of documents within SQ as Ds . Number of documents
within SQ that contain w as dfw,Ds and query specific idf value of w as idfw,Ds .
Similarly, for a tree node e, dfw,De refers to the number of spatially qualified doc-
uments stored underneath e that contain w. In later sections, unless explictly
stated, when saying dfw or idfw , we refer to dfw,Ds and idfw,Ds respectively.
Finally, the candidate set produced from IDF calculation step as B.

2.2 IR-Tree Index Structure

IR-tree [11] is one of the most eective and ecient query processing methods
for top-k spatial keyword search. Its eectiveness largely comes from the query-
specific IDF. In order to facilitate query processing, a novel index, IR-tree, is
proposed, which is essentially a clever combination of R-tree [5] and inverted
files [16]. Document locations are indexed in an R-tree enhanced as follows:
each internal node of the R-tree is associated with textual summary informa-
tion about documents in the subtree, including (1) the count of documents
in the subtree, and (2) the max-TF and DF for each words appearing among
documents in the sub-tree.
464 X. Zhou et al.

each leaf node is linked to an inverted file indexing all the documents belonging
to the nodes.
An example IR-tree is shown in Figure 1.

2.3 The Existing Search Algorithm for IR-Tree

The query processing algorithm in [11] consists of two steps, IDF calculation
and top-k document retrieval.

Algorithm 1. IDFCalculation (r, WQ , SQ )

1 T.push(r); Ds 0;
2 while T is not empty do
3 e T.pop();
4 if e.MBR is fully contained in SQ then
5 Ds Ds + |De |;
6 if w WQ such that dfw,De > 0 then
7 forall w WQ do dfw dfw + dfw,De ;
8 B.add(e);

9 else if e is intersect with SQ then

10 if e is an internal node then
11 forall child node c of e do T.push(c);
12 else /*e is a leaf node */
13 S1 ; S2 ;
14 forall d e such that d.loc SQ do S1 S1 {d};
15 Ds Ds + |S1 |;
16 if w WQ such that dfw,De > 0 then
17 Ie the inverted file for entry e;
18 forall w WQ do
19 forall d Ie [w] S1 do
20 dfw dfw + 1;
21 S2 S2 {d};

22 forall d S2 do
23 B.add(d) ; /* Also keep tfw,d information about d */

Ds
24 return {log dfw
}, w WQ

IDF Calculation. We show the pseudo-code of IDF calculation in Algorithm 1.

It is basically a top-down traversal of the IR-tree, and accumulates Ds and dfw s
during the process.

Top-k Document Retrieval. This step performs best-first search on the IR-tree,
and it terminates as soon as top-k objects have been determined. Initially, all
the entries in B and their estimated maxscores are pushed into max-heap H.
 Improved Spatial Keyword Search Based on IDF Approximation 465

The algorithm then iteratively pop an entry e in H until either H is empty or

top-k results have been determined. If e is a document, it is added to R; if e is a
non-leaf node, it pushes all its children and their scores to H; otherwise, e must
be a leaf node, it pushes all its documents and their scores into H.

4 Root root maxTF:2, DF:3 Spatial TF

TFw1 TFw2
Score w3
Sq
5 N3 docSum: 8 N1 maxTF:0, DF:0 root 1 2 2
N1 N2 N3 w1 N1 0.4 0 2
3
N2 maxTF:2, DF:1
N2 1 2 2
N3 maxTF:1, DF:2 N3 0.35 1 0
1 N1 N2 N3 d1 0.7 2 1
root maxTF:2, DF:3 .
N2 d2 0.6 0 2
docSum: 3 docSum: 2 docSum: 3 N1 maxTF:2, DF:1 .
2 d3 0.35 1 0 .
d6 d7 d8 d3 d4 d5 w2
6 d1 d2 N2 maxTF:2, DF:2 d4 0 1 0

N3 maxTF:0, DF:0 d5 0 0 0
N1
d6 0.2 0 2
Inverted File Inverted File Inverted File
7 8 ... ... ... d7 0 0 0
N1 N2 N3
d8 0 0 0

Example Data location map IR-tree(1):Tree Part IR-tree(2):Node Summary Info Example Data statistics

Fig. 1. Example Data statistics

Example 1. Consider the example IR-tree in Figure 1. Document statistics and

spatial scores are given. Assume k = 1, = 0.5, WQ = (w1, w2). The algorithm
works as follows:
In the IDF calculation step, IR-tree is traversed from root, node N2 is con-
tained in SQ , so it is added into candidate set B, Ds is computed as 2, dfw1 = 1
and dfw2 = 2. Nodes N1 and N3 are leaf nodes that partially overlap with
Q, therefore their associated inverted files are retrieved to calculate accu-
rate dfw1 = 2, dfw2 = 3, and d3 and d6 are found to be both spatially and
textually qualified and are added to B. At the end of this step, we have
B = {N 2, d3, d6}, idfw1 = log10 24 = 0.301, and idfw2 = log10 34 = 0.125.
In the TopKRetrieval step, H is initialized as {(N 2, 0.926), (d3, 0.326),
(d6, 0.225)} by adding all entries in B and their scores(for nodes, the scores
calculated are their estimated maxscore; for docs, they are accurate score)
to H. Then, node N2, the head entry in H is popped and expanded so that
its children d1 and d2 are added to H. Now H = {(d1, 0.714), (d2, 0.425),
(d3, 0.326), (d6, 0.225)}.
Since d1s score is higher than the next highest score, d1 is returned as the
top-1 answer, and the algorithm stops.

3 Approximate IDF-Based Searching Algorithm

One performance issue with the above query processing algorithm on IR-tree is
that generally more than 70% of the search time is used for IDF calculation. For
example, the shaded bars in Figures 2 indicate the IDF calculation time.
After analyzing the IDF calculation algorithm, Lines 17-23 of Algorithm 1 are
usually the most time-consuming part. This is because the algorithm needs to
access inverted files for each intersecting leaf node. These inverted files are typ-
ically proportional to the total size of document collection, and are much larger
466 X. Zhou et al.

than spatial information and other index overhead associated with each docu-
ment. In addition, due to the sheer size of the inverted files, they are typically
stored on the disk, and many random I/Os are needed to retrieve them.
Therefore, our goal is not to access all of the inverted files while still retrieving
the top-k results. We instead approximate the IDF values by a range. We can
find conditions that guarantee the correctness of top-k answers even with such
approximate IDF values.

3.1 Framework
Our idea is to leave out Lines 17-23 of Algorithm 1, and compute a DF range
[DFwmin , DFwmax ] for each query word w without traversing inverted lists.
Consider the previous example. When node N1 is traversed, we find it is an
intersecting leaf node, instead of scanning its inverted files, we know that under
N1, there are dfw,DN 1 docs contain w, but not sure if all of them are spatially
qualified. Thus, we can bound DFw as [DFw , DFw + dfw,DN 1 ]. Subsequently, we
Ds Ds
can bound the IDF values for each query keyword as [log DFw +df w,DN 1
, log DF w
].
Once we have the IDF range for every query keyword, the min and max scores
of an IR-tree node or a document can be computed using the IDF ranges and
the scoring function presented in section 2. Note that spatial score of an IR-tree
node can also be bounded by computing the min and max distance between the
MBR of the node and the query location.
Denote minScore of d as (d) and maxScore of d as (d). It is obvious that
if (d1 ) > (d2 ), then we know d1 s score is better than d2 and we can safely
rank d1 before d2 .

3.2 IDF Range Computation

The main dierence of this algorithm with Algorithm 1 is the case when current
entry e is a leaf node(line 12). Instead of retrieving the inverted file of e, only
locations of child objects stored in e are scanned so as to update Ds (lines 14-15),
then e is added into candidate set simply if it is textually qualified, and replace
lines 17-23 by DFwmax DFwmax + dfw,De to accumulate min and max DFs for
each query keyword.
A subtlety here is to additionally record intersecting leaf node if it is textu-
ally qualified, so that later when two objects have very similar score ranges and
cannot be ordered exactly, we need to expand these unchecked intersecting leafs
and update query word DFs to get tighter object score ranges (add a line under
line 16: ILFs.add(e), where ILF s denotes a list of intersecting leaf nodes).

3.3 Top-k Docs Retrieval

After getting candidate set, IDFs ranges, top-k documents are retrieved. As
there may be much more than k qualified documents in candidate set, we use
the incremental search algorithm which stops as long as k docs are identified.
 Improved Spatial Keyword Search Based on IDF Approximation 467

Algorithm 2. TopKRetrieval (WQ , SQ , k, , B, ILFs, minDF s, maxDF s)

1 newM inDF s minDF s, newM axDF s maxDF s; R ;
2 for e B do H.push(e, (e), (e));
3 while |H| > 0 and |R| < k do
4 e H.pop();
5 if e is node then
6 ExpandTreeNode(e, H);
7 else
8 if |H| = 0 then
9 R R {e};
10 else
11 e H.top();
12 if (e) (e ) then
13 R R {e};
14 else
15 if e is node then
16 e H.pop(); H.push(e);
17 ExpandTreeNode(e , H);
18 else
19 H.push(e);
20 expand a proportion(10-100%) of ILFs to update DF s;
21 compute newM inIDF s, newM axIDF s ;
22 update entry scores in H;

23 return R

Algorithm 3 outlines top-k document retrieval process. Firstly, for each entry
e in B, we estimate its (e) and (e) , and put entry (e, (e) , (e)) into a
max-heap H which sorts its entries in descending order of their maxScores.
Secondly, we iteratively check the head entry (e, (e), (e)) of H to identify
top-k documents until H is empty. If e is a node, expand this node by adding its
children and their scores into H. If it is a document and H is not empty, the next
top entry(n) in H needs to be examined and compared with e in order to deter-
mine whether e can be added into R.This leads to two cases: 1) Scores of e and
n are comparable, i.e. (e) is greater than or equal to (n), we can simply add
e into R; 2) Their scores are not comparable, i.e. (e) is less than (n), we can-
not determine directly whether e is the most relevant document in H, thus extra
process is required. If n is a node, push e back to H and expand node n, whereas
if n is also a document, which means the current DFs ranges are too rough to de-
termine exact top-k documents, more intersecting leaf nodes need to be expaned
to update min and max DFs of query keywords, IDF values are recomputed and
entry scores in H are also updated to a tighter bound for exact ranking.
ExpandTreeNode(n, H) process examines all the child nodes (cn) of the node
to be expaned (n) and add (cn, (cn), (cn))s into H for later processing if n
468 X. Zhou et al.

is an internal tree node. If n is a leaf node which is within query range, tex-
tually qualified documents (d) which is implied by that they exist in at least
one of the inverted lists of query keywords associated with n are examined
and (d, (d), (d))s are added into H. However, if n is an intersecting leaf
node(n.MBR is interect with SQ ), not only documents stored underneath n
have to be checked whether they are textually qualified, they also have to be
checked if spatially qualified so that they can be added into H and minDFs and
maxDFs are also updated accordingly.
Lemma 1 (Correctness). Algorithm 2 correctly retrieves the top-k result.
Example 2. Using the same example in Figure 1, the proposed algorithm works
as below:
IDFRangeComputation procedure traverses IR-tree to find contained node
N2, and intersecting leaf nodes N1 and N3, thus updates Ds = 4, ILFs=(N1,
N3), B =(N1, N2, N3), and dfw1,Ds = [1, 3], dfw2,Ds = [2, 3], and computes
idfw1,Ds = [0.125, 0.602], idfw2,Ds = [0.125, 0.301].
TopKRetrieval first constructs initial H = {(N2, 1.403, 0.75)(N1, 0.501, 0.325)
(N3, 0.476, 0.238)}. Then node N2 is expanded to get newH={(d1 , 1.103,
0.538)(d2 , 0.601, 0.425)(N1, 0.501, 0.325)(N3, 0.476, 0.238)}
Two documents are found at the top of heap, but with incomparable scores.
Therefore, more intersecting leaf nodes have to be expanded. Assume 50% of
|ILFs|is expanded at one time, then in this case, node N1 is expanded. dfw1,Ds
is not changed, newdfw2,Ds = [3, 3], so newidfw2,Ds = [0.125, 0.125]. The new
updated H = {(d1 , 1.015, 0.538)(N3, 0.476, 0.238)(d2 , 0.425, 0.425)(N1, 0.325,
0.325)}
d1 is returned as its (d1 ) > (N 3). Search stops.

4 Experiments
4.1 Experiments Setup
We select two publicly available datasets.
Trec is from the TREC-9 Filtering Track1 . We use TREC as text dataset.
POI-Italy contains gas station and car parking locations of Italy2 . for spatial
data.
There are totally 239580 documents in Trec dataset. We randomly combine two
documents to form a longer document, and 100000 of these combined documents
are selected as test data set. We assign each combined document a unique spatial
location selected randomly from POI-Italy dataset. All the x and y coordinates
data in POI-Italy dataset are normalised into range [0, 40000]. Correspondingly,
we randomly select 1 to 8 words from the vocabulary of the documents to form
textual part of queries and assign random rectangle as the spatial part. We show
some statistics in Table 1.
1
http://trec.nist.gov/data/t9_filtering.html
2
http://poi-osm.tucristal.es/
 Improved Spatial Keyword Search Based on IDF Approximation 469

In the experiments, we investigate performance gains of proposed search algo-

rithms and influences of 5 factors on search performance shown below in table 2.
The underlined values are default settings if not specified.
SQ Query Spatial Range.
Score ratio, which measures the relative importance of textual relevance to
spatial relevance.
k The number of documents to be returned.
|WQ | The number of query keywords.
|D| Test size, the total number of documents indexed.
The index structure implemented is IR-tree Index demonstrated in Section 2.2
using the public R-Tree Library at http://jsi.sourceforge.net. Two search
algorithms are implemented, which are denoted as IRTree (described in sec-
tion 2.3) and ApproIDF (our proposed algorithm).

4.2 Search Performance

In the experiments, we compare the search performance of two search algorithms

over wide variation of dierent query and index factors.
Vary Query Word Number |WQ |: In Figure 2(a), we compare the experi-
mental results of the two search algorithms in terms of total query search time
against query word numbers. The shaded area represents IDF computation time
of the two algorithms. It is obvious that our proposed algorithm outperforms
the original search algorithm, and consumes nearly half of the search time of
original algorithm. This trend remains the same while the number of words are
increasing. Another observation is that the IDF computation time of ApproIDF
is reduced substantially as expected, which becomes less than one thirds of that
of original search algorithm. Finally, the obvious trend shows that total search
time increases linearly as query word number increases. This is due to the in-
crease of textually qualified documents since our implementation supports OR
semantic keyword search, which means a document is qualified as long as it
contains at least one of the query keywords.

Table 1. Trec Dataset Statistics

Test Size Total Distinct Words Total Words Average Document Size
100000 241297 35883968 358

Table 2. Experiment Parameter Settings

Parameters Values
Query range 10, 100, 1000, 10000
Score ratio 0, 0.25, 0.5, 0.75, 1
Top k 1, 10, 20, 50, 100
Query word number 1, 2, 4, 8
Test size 10000, 20000, 40000, 60000, 80000, 100000
470 X. Zhou et al.

Vary Score Ratio : Figure 2(c) presents search perfomance of varying weight
ratio between textual and spatial relevance. It can be seen that search time stays
steadily for the two algorithms when varies from 0 to 1 except that when is
set to 0(only consider spatial score), approIDF performs particularly good than
any other cases. The reason is that when is 0, the ordering of documents only
depends on spatial score, which renders IDF values useless no matter they are
exact or approximate. Therefore, top k documents can be selected directly and
quickly using accurate spatial score. In addition, the overall trend is the same
as above, ApproIDF performs much better than original search algorithm in all
ratio settings. Here,pre-checking of can be added before search process starts
as an optimization to deal with special cases such as is 0 or 1.
Vary Top K k: Figure 2(d) demonstrates the impact of number of requested
documents(k). It is shown that search time of the two algorithms increase slighly
as k increases. As both of the search processes are incremental, which means they
stop as soon as top-k docs have been retrieved without having to go through all
candidate set, when k becomes larger, more scores of documents need to be com-
puted and ranked. Besides, our proposed algorithm is still more ecient than
IR-tree search algorithm.
Vary Query Range SQ : Next, eects of query spatial range on search perfor-
mance are evaluated. The log scale search time is summarized in Figure 2(b).
When query range is small(10-1000), total search time for both algorithm is
less than 0.04 millisecond, whereas when search range increases to 10000, to-
tal search time grows drastically to around 0.63ms for IR-tree and 0.33ms for
ApproIDF. The is mainly due to increase of candidate size when query range
grows. There are nearly 10% (9231) of documents are within query range when
SQ is set to 10000. Overall, ApproIDF outperforms IR-tree search algorithm,
especially obvious when query range is large.
Vary Test Size |D|: As shown in Figure 2(e), search time grows up linearly for
both algorithms along the increasing of test size (number of documents indexed).
It is mainly due to the increase of candidate size when test size becomes larger
as shown in Figure 2(f).

5 Related Works

Numerous works have been done in this field. Some works use inverted index for
textual retrieval and separate index such as Quad-tree or Grid index for spatial
filtering [2] [10]. Currently the most widely used index structure is hybrid index,
and many variations exist besides the IR-tree structure [3] [11] we already men-
tioned, such as Nave Hybrid Index [9], R*-tree [1] [8], KR*-tree [6], IR2 -tree [4],
and W-IR-tree [14].
Nave Hybrid Index [9] combines document location to every word in the
document to be a new word, and build index on them. This method clearly
wastes lots of space and does not support flexible spatial relevance computation.
 Improved Spatial Keyword Search Based on IDF Approximation 471

1 1
IRTree approidf IRTree approidf

0.8

0.1
Time (ms)

Time (ms)
0.6

0.4
0.01

0.2

0 0.001
1 2 4 8 10.0 100.0 1000.0 10000.0
Number of Words Query Range

(a) Search Performance, versus |WQ | (b) Search Performance, versus SQ (Log Scale)

1 1
IRTree approidf IRTree approidf

0.8 0.8
Time (ms)

Time (ms)
0.6 0.6

0.4 0.4

0.2 0.2

0 0
0.0 0.25 0.5 0.75 1.0 1 10 20 50 100
Score Ratio Number of Top K

(c) Search Performance, versus (d) Search Performance, versus K

1
IRTree approidf Candidate Size

0.8 10000
Candidate Size
Time (ms)

0.6

0.4

0.2 1000

0
10000 20000 40000 60000 80000 100000 10000 20000 40000 60000 80000 100000
Test Size Test Size

(e) Search Performance, versus |D| (f) Candidate Size versus |D|

Fig. 2. IDF Computation Time and Total Search Performance Comparison

R*-tree based method use loose combinations of an inverted file and an R*-
tree [6] [15]. These works usually suer from the inability to prune search space
simultaneously according to keyword similarity and spatial information. IR2 -
tree [4] augments the R-tree with signatures enables keyword search on spatial
data objects that each have a limited number of keywords. This technique suers
from substantial I/O cost since the signature files of all words need to be loaded
into memory when a node is visited. Also, signature approach is useless and
infeasible when it comes to handling ranked quires. W-IR-tree [14] is essen-
tially an IR-Tree combined with word partitioning technique, which partitions
keywords before constructing IR-tree. W-IBR-tree is simply a variation of W-IR-
Tree by replacing inverted files attached to each node by inverted bitmap, which
reduces the storage space and also saves I/O during query processing. This tree
structure combines both keyword and location information, better matches the
semantics of top-k spatial keyword queries and performs better than IR-Tree.
However, it does not consider textual similarity as a ranking factor as it treats
an object as either textual relevant (containing all the keyword queries) or not.
472 X. Zhou et al.

6 Conclusion and Future Work

In this paper, we introduce a new approximate IDF-based searching algorithm
to deal with top-k spatial keyword search problem based on IR-tree index struc-
ture. We evaluate the performance of our proposed algorithm against the original
one [11] via wide variation of experiment factors, and the experimental results
demonstrate that our proposed search algorithm performs nearly twice better
than the original one. The only drawback of IR-tree index structure is that
it consumes substantial amount of storage space whose improvement can be a
promising work in the future.

References
1. Beckmann, N., Kriegel, H.P., Schneider, R., Seeger, B.: The R*-Tree: An ecient
and robust access method for points and rectangles. In: SIGMOD Conference, pp.
322331 (1990)
2. Chen, Y.Y., Suel, T., Markowetz, A.: Ecient query processing in geographic web
search engines. In: SIGMOD Conference, pp. 277288 (2006)
3. Cong, G., Jensen, C.S., Wu, D.: Ecient retrieval of the top-k most relevant
spatial web objects. In: PVLDB, vol. 2(1), pp. 337348 (2009)
4. Felipe, I.D., Hristidis, V., Rishe, N.: Keyword search on spatial databases. In:
ICDE, pp. 656665 (2008)
5. Guttman, A.: R-trees: A dynamic index structure for spatial searching. In:
SIGMOD Conference, pp. 4757 (1984)
6. Hariharan, R., Hore, B., Li, C., Mehrotra, S.: Processing spatial-keyword (SK)
queries in geographic information retrieval (GIR) systems. In: SSDBM, p. 16 (2007)
7. Hiemstra, D.: A probabilistic justification for using tf x idf term weighting in
information retrieval. Int. J. on Digital Libraries 3(2), 131139 (2000)
8. Hjaltason, G.R., Samet, H.: Distance browsing in spatial databases. ACM Trans.
Database Syst. 24(2), 265318 (1999)
9. Jones, C.B., Abdelmoty, A.I., Finch, D., Fu, G., Vaid, S.: The SPIRIT spatial
search engine: Architecture, ontologies and spatial indexing. In: Egenhofer, M.,
Freksa, C., Miller, H.J. (eds.) GIScience 2004. LNCS, vol. 3234, pp. 125139.
Springer, Heidelberg (2004)
10. Lee, R., Shiina, H., Takakura, H., Kwon, Y.J., Kambayashi, Y.: Optimization of
geographic area to a web page for two-dimensional range query processing. In:
WISEW 2003, pp. 917. IEEE Computer Society, Washington, DC (2003)
11. Li, Z., Lee, K.C.K., Zheng, B., Lee, W.C., Lee, D.L., Wang, X.: IR-Tree: An e-
cient index for geographic document search. IEEE Trans. Knowl. Data Eng. 23(4),
585599 (2011)
12. Roussopoulos, N., Kelley, S., Vincent, F.: Nearest neighbor queries. In: SIGMOD
Conference, pp. 7179 (1995)
13. Sanderson, M., Kohler, J.: Analyzing geographic queries. In: Workshop on
Geographic Information Retrieval SIGIR (2004)
14. Wu, D., Yiu, M.L., Cong, G., Jensen, C.S.: Joint top-k spatial keyword query
processing. IEEE Trans. Knowl. Data Eng. 24(10), 18891903 (2012)
15. Zhou, Y., Xie, X., Wang, C., Gong, Y., Ma, W.Y.: Hybrid index structures for
location-based web search. In: CIKM, pp. 155162 (2005)
16. Zobel, J., Moat, A.: Inverted files for text search engines. ACM Comput.
Surv. 38(2) (2006)
 Efficient Location-Dependent Skyline Retrieval
with Peer-to-Peer Sharing

Yingyuan Xiao1, Yan Shen1, Hongya Wang2, and Xiaoye Wang1

1
Tianjin Key Laboratory of Intelligence Computing and Novel Software Technology, Key
Laboratory of Computer Vision and System, Tianjin University of Technology, 300384, China
2
Donghua University, 201620, Shanghai, China

Abstract. Motivated by the fact that a great deal of information is closely re-
lated to spatial location, a new type of skyline queries, location-dependent sky-
line query (LDSQ), was recently presented. Different from the conventional
skyline query, which focuses on non-spatial attributes of objects and does not
consider location relationship (i.e., proximity), LDSQ considers both the prox-
imity and non-spatial attributes of objects. In this paper, we explore the problem
of efficient LDSQ processing in a mobile environment. We propose a novel
LDSQ processing method based on peer-to-peer (P2P) sharing. Extensive expe-
riments are conducted, and the experimental results demonstrate the effective-
ness of our methods.

Keywords: location-based services, location-dependent skyline query, valid

scope, peer-to-peer sharing.

1 Introduction

Owing to the ever growing popularity of mobile devices and rapid advent of wireless
technology, information systems are expected to be able to handle a tremendous
amount of service requests from the mobile devices. The majority of these mobile
devices are equipped with positioning systems, which gave rise to a new class of mo-
bile services, i.e., location-based services (LBS). LBS enable users of mobile devices
to search for facilities such as restaurants, shops, and car-parks close to their route. In
general, mobile clients send location-dependent queries to an LBS server from where
the corresponding location-related information is returned as query results.
The skyline query is one very important query for users decision making. The
conventional skyline query focuses on non-spatial attributes of objects and does not
consider proximity. Considering both the proximity and non-spatial attributes of
objects, a new type of skyline queries, location-dependent skyline query (LDSQ),
was recently presented [1, 2]. Unlike the conventional skyline query, whose result is
based only on the actual query itself and irrelevant to the location of the mobile
client, the query result of an LDSQ is closely relevant to the location of the mobile
client, i.e., the location of the mobile client is a parameter of the LDSQ. To answer
LDSQs, mobile clients need to connect to the corresponding LBS server through a

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 473481, 2013.
Springer-Verlag Berlin Heidelberg 2013
474 Y. Xiao et al.

GSM/3G/Wi-Fi service provider and the LBS server is responsible for processing
each LDSQ. With the ever growing number of mobile devices, the sharply increas-
ing query loads are already resulting in network congestion and the performance
bottle of the LBS server. Aim to the above scalability limitations, in this paper, we
explore to leverage P2P data sharing to answer LDSQs.
The remainder of this paper is organized as follows. We review the related work in
Section 2. In Section 3, we first formulate the problem studied in this work, and then
describe the reference infrastructure for processing LDSQ. Section 4 presents the
proposed processing approach. We evaluate the proposed approaches through com-
prehensive experiments in Section 5. Finally, Section 6 concludes this paper.

2 Related Work

Skyline query has received considerable attention due to its importance in numerous
applications. The skyline operator is first introduced into the database community by
Borzsonyi et al. [3]. Since then a large number of methods have been proposed for
processing skyline queries in a conventional set. These methods can be classified in
two general categories depending on whether they use indexes or not. The popular
indexes-based methods mainly include index [4], nearest neighbor [5], branch-and-
bound skyline [6] and ZBtree [7]. The typical methods without indexes are block
nested loop [3], divide and conquer [3], sort first skyline [8], linear elimination sort
for skyline [9] and sort and limit skyline algorithm [10]. Furthermore, a number of
interesting variants of the conversional skyline, like multi-source skyline [11], distri-
buted skyline [12] and etc., have been addressed.
The abovementioned methods only focus on non-spatial attributes of objects and
do not consider location relationship. Motivated by the popularity of LBS and the
increased complexity of the queries issued by mobile clients, Zheng et al. [1] first
address the problem of LDSQ which is the most relevant to our work. LDSQ enables
clients to query on both spatial and non-spatial attributes of objects. In [1], the con-
cept of valid scope is introduced to address the query answer validation issue and two
query algorithms to answer LDSQ, namely brute-forth and -Scanning, are proposed.
Furthermore, [2] explores how to efficiently answer LDSQ in road networks where
the important measure is the network distance, rather than their Euclidean distance.
The main disadvantage of the methods present in [1, 2] is lack of scalability and thus
they cannot avoid network congestion and the performance bottle of the LBS server
caused by the ever growing number of mobile devices.

3 Preliminary

In this section, we first formally define LDSQ and related concepts, and then describe
the system infrastructure for supporting LDSQ.
 Efficient Location-Dependent Skyline Retrieval with Peer-to-Peer Sharing 475

Notations and Definitions. Let S be the set of spatial objects stored at the LBS server
and every object o S has both spatial location attributes denotes by L(o) and a set of
non-spatial preference attributes denoted by P = {p1, p2, , pm}. We use o. pi (1 i
m) to represent the os value for the i-th non-spatial attribute and d(L(o), L(q)) to
represent the distance between o and query point q where d(.) denotes a distance me-
tric obeying the triangle inequality. Let t u denote object t dominates object u with
respect to non-spatial preference attributes. We use SK(S) to denote the skyline over S,
i.e., SK(S) = { S S s.t. k o}. In the following, we formally define the LDSQ
and related concepts.
Definition 1 ( ). Given two spatial objects t, u and a query point q, we say t u
with respect to q, if t u with respect to non-spatial preference attributes and d(L(t),
L(q)) d(L(u), L(q)), formally, t u d(L(t), L(q)) d(L(u), L(q)).

Definition 2 (Location-dependent Skyline). Given a set of spatial objects S and a

query point q, we call LSK(S, q) = { S| S s.t. k o} the location-dependent
skyline of q over S.
Definition 3 (LDSQ). An LDSQ is to find the location-independent skyline over a
given set of spatial objects with respect to a given query point.

Definition 4 (Valid Scope). The valid scope of an LDSQ with respect to a query
point q, denoted by VS(S, q), is defined as a spatial region wherein all query points
will receive an identical result, i.e., for any two query points q1, q2, if L(q1) VS(S, q)
and L(q2) VS(S, q), then LSK (S, q1) = LSK (S, q2) = LSK (S, q).
The introduction of valid scope provides the chance of sharing query results among
neighboring mobile clients.

System Infrastructure. Our system infrastructure is a LBS system based on GSM or

3G and supports anytime, anywhere data access and processing using a portable size
wireless mobile device. Mobile client in general communicates with a Mobile Base
Station (MBS) to carry out any activities such as transaction and information query.
MBS has a wireless interface to establish communication with mobile clients and can
serve a large number of mobile clients. The power of a MBS defines its communica-
tion region, which we refer to as a cell. One or more MBSs are connected with a Base
Station Controller (BSC), which coordinates the operations of MBSs when com-
manded by the Mobile Switching Center. The LBS server is connected to a BSC
through a high-speed wired network and provides location based information services
for mobile clients. Mobile clients equipping with GPSs can communicate with MBSs
using wireless channels. A mobile client can freely move among cells and transpa-
rently accesses the spatial database residing at the LBS server. Benefiting from the
increasing deployment of new P2P wireless communication technologies, there are
short-range networks (MANETs) that allow ad hoc connections with neighboring
mobile clients. This enables mobile clients to communicate with neighboring peers in
an ad hoc manner for data sharing [13].
476 Y. Xiao et al.

4 LDSQ Processing Approach

In this section, we propose the sharing-based location-dependent skyline query ap-

proach (SLSQ) to improve the preceding -Scanning algorithm present in [1]. Differ-
ent form the existing LDSQ processing algorithms, in which mobile clients directly
send the LDSQ requests and their current locations to the LBS server and the LBS
server is responsible for answering each LDSQ, SLSQ first tries to get the query result
for each mobile client from the cached data by its neighboring mobile clients. Only
when these cached data do not match the LDSQ request, are the corresponding LDSQ
request and location information transmitted to the LBS server through the neighbor-
ing MBS. To leverage query results cached in the neighboring peers, SLSQ needs
each mobile client to cache its last LDSQ result and the corresponding valid scope.
Assume that R denotes the LDSQ request issued by a mobile client MC, L(MC)
denote the current location of MC, peeri for i=1, 2, , k denotes a single-hop mobile
client of MC where k represents the number of its single-hop mobile clients, LSK(S,
peeri) denotes the result of last LDSQ issued by peeri, LSK(S, MC) represents the
result of last LDSQ issued by MC, VS(S, peeri) denotes the valid scope of the last
LDSQ with respect to peeri and VS(S, MC) denotes the valid scope of the last LDSQ
with respect to MC. LSK(S, peeri) and VS(S, peeri) are cached in peeri, and LSK(S,
MC) and VS(S, MC) are cached in MC. Specifically, the processing strategy of SLSQ
is described as follows:
(1) MC checks whether L(MC) falls within VS(S, MC). If yes, MC returns LSK(S,
MC) and the algorithm ends; otherwise MC broadcasts the message <R, L(MC)> to
all its single-hop peeri for i=1, 2, , k and waits the result from peeri.
(2) For each peeri (1ik), once receiving <R, L(MC)> from MC, it checks wheth-
er L(MC) falls within VS(S, peeri). If yes, peeri sends the message Yes to MC; oth-
erwise it sends the message No to MC.
(3) If MC receives Yes from a peeri within the specified deadline Tdeadline, it
sends the message OK to the peeri and waits the result from the peeri; otherwise it
forwards <R, L(MC)> to the LBS server and waits the result from the LBS Server.
Here, Tdeadline is calculated by the formula: Tdeadline = tissue 2tdelayslack, where tissue is
the time at which the message <R, L(MC)> is broadcast by MC, tdelay denotes the
largest network delay between two neighboring mobile clients based on MANETs,
and slack represents the slack factor.
(4) Once receiving OK from MC, the peeri sends LSK(S, peeri) and VS(S, peeri)
to MC; otherwise the LBS server calls the algorithm present in [1] to compute LSK(S,
MC) and VS(S, MC) and sends them to MC, if it receives <R, L(MC)> from MC.
(5) MC caches the result received from the peeri or the LBS server and meanwhile
returns LSK(S, peeri) (or LSK(S, MC)).
Based on the above processing strategy, we formalize SLSQ in the following Algo-
rithms 1-3. Algorithms 1-3 are respectively executed at mobile clients, those single-
hop peeri and the LBS Server.
 Efficient Location-Dependent Skyline Retrieval with Peer-to-Peer Sharing 477

Algorithm 1. SLSQ -MC (R, L(MC))

Input: LDSQ request (R) and the location of MC (L(MC))
Output: the result of the LDSQ
1: If CacheData L(MC) CacheData.VS(S, MC)
2: return CacheData.LSK(S, MC)
3: else
4: Broadcast the message <R, L(MC)> to all its single-hop mobile clients peeri, set Tdeadline
and wait response messages from them;
5: if receiving the message Yes from a peeri within the specified deadline Tdeadline then
6: Send the message OK to the peeri and wait the results from the peeri;
7: Once receiving the result <LSK(S, peeri), VS(S, peeri)> from the peeri, CacheData.
LSK(S, MC) LSK(S, peeri) and CacheData.VS(S, MC) VS(S, peeri);
8: return CacheData. LSK(S, MC);
9: else
/ * all its single-hop peers send No or MC does not receive any Yes within Tdeadline */
10: Forward <R, L(MC)> to the LBS server through its MBS and wait the result from the
LBS server;
11: Once receiving the result <LSK(S, MC), VS(S, MC)> from the LBS server, CacheData.
LSK(S, MC) LSK(S, MC) and CacheData.VS(S, MC) VS(S, MC);
12: return CacheData. LSK(S, MC);

In Algorithm 1 and the following Algorithm 2, CacheData denotes the cache ca-
pacity allocated by MC or peeri for caching its last LDSQ result and the correspond-
ing valid scope, CacheData.LSK(S, MC) represents the space of CacheData used to
caching the last LDSQ result and CacheData.VS(S, MC) denotes the remaining space
of CacheData used to caching the corresponding valid scope.

Algorithm 2. SLSQ - peeri (R, L(MC))

Input: LDSQ request (R) and the location of MC (L(MC))
Output: <LSK(S, peeri), VS(S, peeri)> or null
1: if CacheData L(MC) CacheData.VS(S, peeri)
2: Send Yes to MC, set Toverdue and wait the message OK from MC;
3: else
4: Send No to MC and return;
5: if receiving the message OK from MC within Toverdue then
6: Send <LSK(S, peeri), VS(S, peeri)> to MC and return;
7: else /* time out */
8: return;

In Algorithm 2, CacheData.VS(S, peeri) denotes the valid scope of the last LDSQ
cached in CacheData and Toverdue is calculated by the formula: Toverdue = tcur-
rent2tdelayslack, where tcurrent is the current system time of peeri, tdelay denotes the
largest network delay between two neighboring mobile clients based on MANETs,
and slack represents the slack factor.
478 Y. Xiao et al.

Algorithm 3. SLSQ -LBS (S, L(MC))

Input: the set of spatial objects (S) and the location of MC (L(MC))
Output: LSK(S, MC), VS(S, MC)
1: LSK(S, MC) -Scanning (S, L(MC));
2: VS(S, MC) -Scanning VS (S, LSK(S, MC));
3: return < LSK(S, MC), VS(S, MC)>;

In Algorithm 3, functions -Scanning() and -Scanning VS() are responsible for re-
spectively computing LDSQ and valid scope, whose detailed descriptions are present
in [1].

5 Performance Evaluation

In this section, we first describe the experimental settings and then present the simula-
tion results.
Experiment Settings. Our simulation experiments are based on the same data sets of
spatial objects as [1], namely School and NBA. In the simulation, mobile clients have
sufficient memory space to buffer their last query results. Like [1], we also assume
the LBS server has sufficient memory space to load all the candidate objects. In order
to reduce the randomness effect we average the results of the algorithms over 50
LDSQs. Table 1 lists the main simulation parameters.

Table 1. Main simulation parameters

Parameter Description Range / Default value

Sdataset The size of spatial object set 48k or 17k
|P| The number of non-spatial preference attributes 1~13 / 6
The P2P wireless transmission range of a mobile
R client based on MANETs 20~140(Meters) / 80
Pr The probability of MC receives Yes from a peeri 0.2~0.8 / 0.5
Nm The number of single-hop peers of a MC 2~10 / 5

Our simulator consists of two main modules: the MC module and the BSC module.
The MC module generates and controls the query requests of all MCs. Each MC is an
independent object that decides its movement autonomously. The BSC module man-
ages the spatial object set and meanwhile coordinates the operations of MBSs.
Simulation Results. First, we compare SLSQ with brute-forth and -Scanning in
CPU time. Fig. 1 shows how CPU time of the three methods varies with the number
of non-spatial preference attributes |P|, i.e., the performance of the three methods in
terms of CPU time, as a function of |P|. As shown in Fig.1, SLSQ outperforms the
other methods under two different data sets (School and NBA). This is because SLSQ
enables a large number of LDSQs to be answered directly from the results cached in
 Efficient Location-Dependent Skyline Retrieval with Peer-to-Peer Sharing 479

neighboring mobile clients and thus avoids a great majority of dominance relationsship
tests required by brute-forrth and -Scanning. We can also see from Fig. 1 thatt -
Scanning gets a distinct ad dvantage over brute-forth. The reason is that brute-foorth
blindly checks all objects and
a thus incurs the significant overhead of dominance reela-
tionship tests. The another observation from Fig. 1 is the CPU cost under School and
NBA data sets shows a littlee different trend. This is because School is an independdent
data set in non-spatial attrib
butes while NBA follows ant-correlated distribution in nnon-
spatial attributes, which cauuses the different overhead of dominance relationship teests.
Fig. 2 shows CPU time,, as a function of the P2P wireless transmission range R R. R
reflects the range of P2P wireless
w transmission between neighboring mobile clieents.
The bigger R usually meanss a larger number of single-hop peers of a mobile client.. As
we expect, SLSQ achieves a better CPU performance than brute-forth and -Scannning
and the performance gap beecomes wider with the increase of R. This is not surprissing
because SLSQ avoids a greaat majority of dominance relationship tests by P2P shariing.
Moreover, for SLSQ, the in ncrease of R enables each mobile client to have a higgher
DSQ by P2P sharing, but brute-forth and -Scanning are not
opportunity to fulfill its LD
influenced by R.

(a) CPU time (School) (b) CPU

C time (NBA) (a) CPU time (School) (b) CPU time (NB
NBA)
Fig. 1. CPU tim
me vs. |P| Fig. 2. CPU time vs. R

(a) Delay (School) (b) Delay (NBA) (a) Delay (School) (b) Delay (NBA))
Fig. 3. Network deelay vs. |P| Fig. 4. Network delay vs. R

Fig. 3 shows the perform

mance of the three methods in terms of network delay, aas a
function of |P|. As shown ini Fig.3, SLSQ outperforms the other methods in netw work
delay. This is because SLSQQ tries to fetch the query result for each mobile client frrom
the cached data by its neighboring peers and only when these cached data do not
match the LDSQ request, does
d SLSQ ask for help from the remote LBS server. T That
is, in many cases SLSQ answers LDSQs by directly fetching data cached in the neiigh-
boring peers. This enables SLSQ
S to avoid the overhead of routing, message forwaard-
ing and handoff caused by y connecting the remote LBS server. We can also see frrom
480 Y. Xiao et al.

Fig. 3 that the network delay grows as attribute dimensionality increases. This is be-
cause the volume of query result, which needs to be transmitted, grows with the in-
crease of attribute dimensionality.
Fig. 4 shows the performance of the three methods in terms of network delay, as a
function of R. As we expect, SLSQ has a distinct advantage over brute-forth and -
Scanning in network delay. The reason is that for many LDSQs, SLSQ avoids the
overhead of routing, message forwarding and handoff caused by connecting the re-
mote LBS server while the overhead is required by brute-forth and -Scanning for all
LDSQs. We can also see from Fig. 4 that brute-forth and -Scanning are not influ-
enced by R in network delay while the network delay of SLSQ decreases as R grows.
The reason is that brute-forth and -Scanning directly ask the LBS server to compute
LDSQs so do not involve the P2P wireless transmission between mobile clients, while
SLSQ leverages the P2P sharing. Thus, the increase of R enables SLSQ to have a
higher opportunity to fulfill its LDSQ by P2P sharing.

6 Conclusion

In this paper, we propose the sharing-based location-dependent skyline query ap-

proach (SLSQ). SLSQ tries to get the query result for each mobile client from the
cached data by its neighboring mobile clients. Only when these data cached do not
match the LDSQ request, are the corresponding LDSQ request and location informa-
tion transmitted to the LBS server. Extensive experiments are conducted, and the
results demonstrate the effectiveness of our methods. A promising direction for future
work is o consider on-air processing of LDSQ in road networks.

Acknowledgment. This work is supported by the Natural Science Foundation (NSF)

of China under Grant No. 61170174 and the NSF of Tianjin under Grant No.
11JCYBJC26700.

References
1. Zhang, B., Lee, K.W.C., Lee, W.C.: Location-dependent skyline query. In: Proc. of MDM,
pp. 148155. IEEE Press, Beijing (2008)
2. Xiao, Y., Zhang, H., Wang, J.: Skyline computation for supporting location-based services
in a road network. Information 15(5), 19371948 (2012)
3. Stephan, B., Donald, K., Konrad, S.: The skyline operator. In: Proc. of ICDE, pp. 421430.
IEEE Press, Heidelberg (2001)
4. Tan, K.L., Eng, P.K., Ooi, B.C.: Efficient progressive skyline computation. In: Proc. of
VLDB, Roma, pp. 301310 (2001)
5. Kossmann, D., Ramsak, F., et al.: Shooting stars in the sky: An online algorithm for sky-
line queries. In: Proc. of VLDB, pp. 275286 (2002)
6. Papadias, D., Tao, Y., et al.: Progressive skyline computation in database systems. ACM
Transactions on Database Systems 30(1), 4182 (2005)
7. Lee, K., Zheng, B., Li, H., Lee, W.-C.: Approaching the skyline in Z order. In: Proc. of
VLDB, pp. 279290 (2007)
 Efficient Location-Dependent Skyline Retrieval with Peer-to-Peer Sharing 481

8. Chomicki, J., Godfrey, P., Gryz, J., et al.: Skyline with presorting. In: Proc. of ICDE, pp.
717816 (2003)
9. Godfrey, P., Shipley, R., Gryz, J.: Maximal vector computation in large data sets. In: Proc.
of VLDB, pp. 229240 (2005)
10. Bartolini, I., Ciaccia, P., Patella, M.: Efficient sort-based skyline evaluation. ACM Trans-
actions on Database Systems 33(4), 145 (2008)
11. Deng, K., Zhou, X., Shen, H.: Multi-source skyline query processing in road networks. In:
Proc. of ICDE, Istanbul, Turkey, pp. 796805 (2007)
12. Xiao, Y., Chen, Y.: Efficient distributed skyline queries for mobile applications. Journal of
Computer Science and Technology 25(3), 523536 (2010)
13. Ku, W.S., Zimmermann, R.: Location-Based Spatial Query Processing in Wireless Broad-
cast Environments. IEEE Transactions on Mobile Computing 7(6), 778790 (2008)
 What Can We Get from Learning Resource
Comments on Engineering Pathway

Yunlu Zhang, Wei Yu , and Shijun Li

School of Computer, Wuhan University, Wuhan, China

[email protected]

Abstract. K-Gray Engineering Pathway (EP) is a digital library web-

site that could help users share their learning resources. Users can cat-
alog, and comment news, blogs, videos, books and papers after login as
registered users, and search without needing to login. We have already
had the ability to search over comments for all resources, and the most
commented resources are accessible on the K-12, higher education and
disciplinary pages, but for now the ranking is only based on the com-
ments number of each learning resources or the average rating.
To help users to evaluate the learning resources not only based on their
comments quantity, but also on their quality, we introduce the Analytic
Hierarchy Process (AHP). For AHP couldnt completely suit to our sit-
uation, to make it work well, we established a three layers model with
objective layer, criteria layer and alternative layer. Subjective expert and
objective quantitative information are used to build the judgement ma-
trices. The attributes of the comment,including total number, ratings,
length of its content, poster, posting time, are used to defined the ab-
solute importance, and further used to calculate the relative importance
and judgement matrices. High dimensional calculation of weightings and
consistence are avoided by adopting our estimation method. Experimen-
tal results and the comparisons with existing methods validate the eec-
tiveness of our proposed methods.

Keywords: learning resources, Analytic Hierarchy Process, digital li-

brary, guide user.

1 Introduction

K-Gray Engineering Pathway (EP),website on http://www.engineeringpathway.

com:8080/engpath/ep/Home, is an online digital library for multiple disciplines.
The photographic images on our homepage have been selected to capture the
excitement and diversity of engineering, computing education and practice. Dif-
ferent taglines are used for the dierent audiences within EP. The homepage and
general tagline is to Turn Ideas into RealityLearn, Connect and Create. The K-12
pages communicate aspirations for the future Shape the FutureDream, Design

Corresponding author.

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 482493, 2013.
c Springer-Verlag Berlin Heidelberg 2013

 What Can We Get from Learning Resource Comments on EP 483

and Do. The broadening participation tagline motivates professional develop-

ment and personal goals: Make a World of DierenceDream Big, Love What
You Do.
We have already done some data mining research based on our new EP in the
last two years, such as [1,2], in this work we want to try another method to mine
more useful information from our EP. For we can get search results by inputting
query keywords in our EP homepage, as shown in Figure 2, we input data min-
ing as search keyword, and we can get 26 learning resource objects. From the
results page, we can get a summary of each item, such as: title, resource, disci-
pline, and the source where they come from. If we click the link under the first
learning resource title, we get more details about it and add comment about it.

Fig. 1. Screen image of the Engineering Fig. 2. Search results by keyword data
Pathway mining

The 100 Most Commented is what we used in this paper. The contents of
comment should contain rating, comment title, comment description, author and
post time. How to use these dataset to answer the questions like Which learning
resource object should we choose or Which one is the best search result and
which one is exactly what I want ? That is the reason why we focus on using
these comments to evaluate the learning resources, and we can achieve this by
selecting the best alternative that matches all of the digital library users criteria.
Learning recourses evaluation by their comments is such a multiple, diverse
criteria decision-making problem with high complexity, so we draw lessons from
Analytic Hierarchy Process (AHP), a classical method presented by Saaty[3] and
be cited more than 1,000 times. We make some modifications on this method
and will give more details on the following sections.

2 Related Work
AHP can be used in many applications besides our web data mining area, such as
in ecosystem [4], emergency management [5], plant control [6] and so on so forth.
484 Y. Zhang, W. Yu, and S. Li

From those research works, it is clear that AHP has been widely applied in many
of the data management and decision making areas. To address the problem
encountered in our web data mining area, that is, digital learning resource mining
for EP, we summarize some related former research works as follows:
Some researchers choose a combination of fuzzy theory and AHP, such as Tao
et al. [8] , Uzoka et al. [9] and Feng et al. [10]. Tao proposed a decision model
by the application of AFS (axiomatic fuzzy set) theory and AHP method to get
the ranking order. They also provided the definitely semantic interpretations for
the decision results by their theory. Uzoka used fuzzy logic and AHP for medical
decision support systems. It is interpreting idea to introduce the semantic infor-
mation proposed by Tao, thus we try to add them in our EP comments mining
this time. However it causes high computational complexity, we will try this in
our future work in this aspect.
AHP also could be used with traditional data mining algorithm, such as Svo-
ray et al [11]. They used a data mining (DM) procedure based on decision trees to
identify areas of gully initiation risk, and compared it with AHP expert system.
Rad et al. [13] proposed an very inspiring idea by considering the problem of
clustering and ranking university majors in Iran. By eight dierent criteria and
177 existing university majors, they give the rank of all the university majors.
In our paper, we use five dierent criteria and 100 learning resources as ex-
perimental dataset, however, both subjective expert and objective quantitative
information are used to build the judgement matrices.
As for educational digital library applications, we focus on how to provide both
teachers and students enriching and motivational educational resources as the
former researchers do in [14] and [15] . Jing and Qingjun [16] provided teachers
and students a virtual knowledge environment where students and teachers enjoy
a high rate of participation. Alias et al. [17] described an implementation of
digital libraries that integrated semantic research. Mobile ad hoc networks are
becoming an important part of the digital library ecology . Social navigation can
be used to replicate some social features of traditional libraries, to extend digital
library portals and to guide portal users to valuable resources in [13].

3 Learning Recourses Automatic Evaluation

3.1 Learning Recourses Comments Hierarchy

There are three lays on our system, that is ,the objective layer, criteria layer, and
alternatives layer. The framework of the comments hierarchy and the content of
each layer is as shown in Figure 7.

Objective layer: Select the most valuable learning resource by comments;

Criteria layer: rating, comment title, comment description/comment content,
comment author type , posting time;
Alternatives layer: the top 100 most commented resources in one certain
discipline.
 What Can We Get from Learning Resource Comments on EP 485

Fig. 3. comments hierarchy

3.2 Judgement Matrices

There are one judgement matrix in criteria layer and five judgement matrices in
alternatives layer to be determined. We will establish them based on the expert
information and quantitative information respectively.

Judgement Matrices in Criteria Layer. Let matrix M R55 denote the

judgement matrix in criteria layer and Ci , i = 1, . . . , 5, denote the five criteria
(factors) in the criteria layer respectively. In matrix M , its element Mi,j means
the quantitative relative importance judgement of pairwise factors Ci and Cj ,
that is, Mj,i = 1/Mi,j . The standard relative importance scale is employed, as
shown in Table.1.

Table 1. The relative importance of the element Ci to Cj (Mi,j )

The scale The meaning

1 Ci and Cj have the same importance
3 Ci is slightly more important than Cj
5 Ci is obviously more important than Cj
7 Ci is strongly more important than Cj
9 Ci is extremely more important than Cj
2, 4, 6, 8 Compromise between the above judge

Judgement Matrices in Alternative Layer. Let matrices M i R100100 ,

i = 1, . . . , 5, denote the judgement matrix for criteria Ci in the alternative
layer and Ai , i = 1, . . . , 100, denote the top 100 most commented resources
respectively. To establish each matrix M i , we use the following two steps:
Calculate the Absolute Importance (AI i )
Based on the related information, such as the rating value and number of
comments for rating C 1 , calculate the absolute importance value for each
alternative under criterion Ci ;
486 Y. Zhang, W. Yu, and S. Li

Scale and calculate the Relative Importance (RI i )

Scale the absolute importance value to 1 to 9, and calculate the relative
importance to get the matrix M i ;
Specifically, we present the following methods to determine the judgement ma-
trices in the alternative layer
1. Rating matrix M 1
For the criterion of rating, the average rating value, varying from 1 to 5, and
the number of comments are used to determine the absolute importance of
the rating for each resource j = 1, . . . , 100 as follows:
Nj
1 "
AIj1 = rj,n + 1 Nj
Nj n=1

where Nj is the number of comments for resource j, rj,n is the rating value
given by comment n and 1 is a properly selected weighting for Nj . Suppose
1 1
AImax and AImin are the maximum and minimum absolute importance of
the top 100 most commented resources respectively, thus the relative impor-
tance value of resource j is given by the following linear scaling method:
8
RIji = i i (AIji AImin
i
)+1
AImax AImin
i i
We assume that AImax > AImin .The element in matrix M i can be given
i RIki
by Mk,j = RIji
Note that we will use the same linear scaling method and
definition of RIji for the following matrices, i = 2, . . . , 5, and omit them.

2. Comment title matrix M 2

For the criterion of comment title, we assume the longer the comment title is,
the more important it is, and also the number of comments are key factor for
the importance of resource. Thus the absolute importance of the comment
title for resource j = 1, . . . , 100 is defined by
Nj
1 "
AIj2 = LTj,n + 2 Nj
Nj n=1

where Nj is the number of comments for resource j, LTj,n is the length of the
comment title given by comment n and 2 is a properly selected weighting
for Nj . The relative importance RI 2 and comment title matrix M 2 can be
obtained by the same method in rating matrix.

3. Comment description matrix M 3

Similar to the previous analysis, we use the length of the comment and
the number of comments to define the absolute importance of the comment
description for each resource j = 1, . . . , 100 as follows:
 What Can We Get from Learning Resource Comments on EP 487

Nj
1 "
AIj3 = LCj,n + 3 Nj
Nj n=1

where Nj is the number of comments for resource j, LCj,n is the length of the
comment given by comment n and 3 is a properly selected weighting for Nj .

4. Author type matrix M 4

Since there are three types of authors, that is, anonymous , person and
organization. Here we assign dierent author type value to each of them:
Vanonymous = 1, Vperson = 2 and Vorganization = 3. Thus the absolute impor-
tance of the comment author type for resource j = 1, . . . , 100 is determined
as follows:
Nj
1 "
AIj4 = Vj,n + 4 Nj
Nj n=1

where Nj is the number of comments for resource j, Vj,n is author type value
of comment n and 4 is a properly selected weighting for Nj .

5. Posting time matrix M 5

We assume the more greater the average posting time is, the more important
the resource is. Thus the absolute importance of the posting time for each
resource j = 1, . . . , 100 is defined as follows:
Nj
1 "
AIj5 = (Tj,n Tc ) + 5 Nj
Nj n=1

where Nj is the number of comments for resource j, Tj,n is posting time

(day) of comment n, Tc is the current day and 5 is a properly selected
weighting for Nj .

3.3 Weighting

After getting the judgement matrix M , we obtain the weightings of the five
criteria in the second layer by calculating its normalized eigenvector wmax cor-
responding to the maximum eigenvalue max , that is,
5
"
M wmax = max wmax , wmax,i = 1.
i=1

Similarly, we get weightings of dierent alternative under criteria Ci , i = 1, . . . , 5,

i
as the normalized eigenvectors wmax of M i as follows:
100
"
M i wmax
i
= imax wmax
i
, i
wmax,j = 1.
j=1
488 Y. Zhang, W. Yu, and S. Li

Note that for M i , i = 1, . . . , 5, they are 100 100 dimension matrices, thus it
is time-consuming job to exactly calculate the eigenvalues and eigenvectors. To
obtain a high quality estimation in short time, we use the following approach :
M
1. Normalize each column in M , that is, Mi,j = "n i,jMi,j ;
#
i=1

2. Sum Mi,j in each row, that is, Mi = nj=1 Mi,j ;

3. Estimate the desired eigenvector w by normalize M i , that is, w = {wj }j=1,...,n ,

Mj
where wj "n
Mi
;
i=1
1 #n (Mw)i
4. The maximum eigenvalue can be estimated by n i=1 wi .

3.4 Consistency

A well-defined judgement matrix should be transitive (that is, if A is better than

B and B is better than C, then we have A is also better than C) and numerically
consistent (that is, if A = 2B and B = 2C, then we have A = 4C). To verify the
consistence of the judgement matrix M Rnn , Saaty [3] suggests to calculating
the following Consistency Ratio (CR):

CIM
CRM =
RIn

where CIM is so-called Consistence Index of matrix M , defined as, CIM =

max n
n1 and RIn is the so-called Random Index for n n matrix
Saaty Rule says that if CRM < 0.1, then matrix M is consistent; oth-
erwise, we need to adjust M until the condition satisfies.
However, for matrices M i , i = 1, . . . , 5, we assert that they are always consis-
tent because
i i RIki RIji i
Mk,j Mj,l = = Mk,l
RIji RIli

thus imax = n and CIM i = 0. Furthermore, the total consistency is also holds
because
wmax,1 CIM 1 + wmax,2 CIM 2 + + wmax,5 CIM 5
= 0 < 0.1.
wmax,1 RI100 + wmax,2 RI100 + + wmax,5 RI100

3.5 Implementation of AHP

In this subsection, we summarize the implementation of AHP for our problem.

1. Criteria Evaluation
Get judgment matrix M based on expert information, calculate its maximum
eigenvalue and corresponding eigenvector wmax , and check its consistence by
Saaty Rule (adjust it if necessary);
 What Can We Get from Learning Resource Comments on EP 489

2. Alternative Weightings Calculation:

For each criteria Ci , i = 1, . . . , 5, calculate its judgement matrix based on
method in subsection 3.2, and further calculate its maximum eigenvalue
i
wmax and corresponding eigenvector based on method in subsection 3.3;

3. Alternative Evaluation:
The final evaluation vj for alternative j is given as
5
"
i
vj = wmax,i wmax,j .
i=1

Then the top 10 learning resources will be returned to users.

4 Experiment
In this section, we will evaluate the proposed AHP method on the K-Gray
Engineering-Pathway library data. The proposed method will be experimentally
compared with other general methods, such average rating, average comment
number, to validate its eectiveness.

4.1 Experimental Set Up

The top 100 commented learning resources on computer science from EP home-
page are selected as the the test data. We collect the comment information, that
is, comment title, comment description, rating, author and posting time for each
learning resource.
To implement the proposed AHP method, first the following judgement ma-
trix M in criteria layer is obtained based on the expert information from the
questionnaire.

1 4 1 5 6
1 1 1
1 4
4 4 5
M =
11 4 1 6 5

1 1
1 1
5 6 2
1 5 1
6 4 5 2 1
5
The maximum eigenvalue max of M is 5.0859, the CI = max 51 = 0.0215 and
CI
CR = RI = 0.0192 < 0.1, thus it is consistent. The corresponding normal-
ized eigenvector is wmax = (0.381, 0.081, 0.379, 0.067, 0.092), which is also the
weightings of the five criteria.
To properly set parameters i , i = 1, . . . , 5, we calculate the average number
of comments, rating, comment title length, comment length, author value and
posting time value (4.31, 3.51, 38.3, 481.4, 2.51, 0.24). Note that i can be viewed
as a scale factor for the number of comments, so we set i according to the
above average value, as shown in Table. 2. Then we can obtain the eigenvectors
of M i , i = 1, . . . , 5, by the estimation method in subsection 3.3, and further get
the final evaluation of the top 100 commented resources.
490 Y. Zhang, W. Yu, and S. Li

Table 2. Parameters setting for AHP

i 1 2 3 4 5
i 0.8 8 10 0.5 0.05

4.2 Comparison and Discussion

The proposed AHP method is compared with the general average methods, such
as average rating, average number of comments methods.
We normalized the evaluation of the 100 resources by average rating (AR),
average comments length (ACL), average comment title length(ACTL), average
number of comments (ANC), average author value (AAV) and average posting
time value (APTV). Figure 4 - Figure 9 show the first 50 resource evaluation
value by AHP and the above average methods. From Figure 4-Figure 9, we
can see that the evaluation results obtained by AHP and ACL own the highest
similarity. Furthermore, AHP can be viewed as a mixed method of these average
method.

Fig. 4. Comparison with AR Fig. 5. Comparison with ACL

Fig. 6. Comparison with ANCL Fig. 7. Comparison with ATL

 What Can We Get from Learning Resource Comments on EP 491

Fig. 8. Comparison with AAV Fig. 9. Comparison with APTV

To make it clear, we define the standard deviation between AHP and method
as 2
"100
SD = (vjAHP vj )
j=1

where vjAHP , vjare the normalized evaluation value for resource j obtained
by AHP and method . Table. 3 reports the standard deviation for dierent
methods. ACL and AR have the smaller SD values compared with other methods.
It is consistent with the expert information and our proposed AHP method from
the judgement matrix M , we see that criteria rating and comment length have
the bigger weightings. Finally, we report the top 10 recommened resources by
AHP, ACL and AR based on the 100 most commented resources on computing
sciences.

Table 3. The squared deviation of dierent methods with AHP

ACL AR AAV ATL APTV ANCL

SD 0.034 0.041 0.050 0.057 0.059 0.060

Table 4. The top 10 recommended resources by dierent methods

AHP ACL AR
1 Alice 2.0 The Black Swan Alice 2.0
2 iWoz ACM K-12 Web-CAT
3 First Barbie Pair Programming Jeliot 3
4 Pair Programming Educating Engineers Award Courseware
5 The Black Swan Computer Museum RVLS
6 Computing ThinkCycle Greenfoot
7 Computational Geometry Initiatives World Without Oil
8 Language Media The Enigma
9 Achieving Dreams Scratch Girl Geeks
10 ACM K-12 Children Website Award Courseware
492 Y. Zhang, W. Yu, and S. Li

5 Conclusion and Future Work

The AHP opens up a wide variety of novel applications. So after introducing
the problem of automatic learning resources evaluation, we make an attempt to
solve it by using this method for reference. However the traditional method is not
satisfactory for our situation, the primary ones are: the ranking criteria contains
multiple variables, each criteria in each Hierarchy may has their own sub-criteria,
and the vector matrix for selecting items is huge. So we made some improvement
on AHP to solve our problem. By establishing three layers model with objec-
tive layer, criteria layer and alternative layer, we can provide our users a more
comprehensive ranking of our learning resources. By using subjective expert and
objective quantitative information, we can build the judgement matrices, then
we calculate the absolute importance, relative importance and finally the judge-
ment matrices with the quantified attributes of comments, such as the number
of comments, rating, the length of content, the author type and the posting time.

Acknowledgments. This work is supported by the National Natural Science

Foundation of China (NSFC-61272109). Thanks to Alice Merner Agogino.

References
1. Zhang, Y., Agogino, A.M., Li, S.: Lessons Learned from Developing and Evaluat-
ing a Comprehensive Digital Library for Engineering Education. In: JCDL 2012
Proceedings of the 12th ACM/IEEE-CS Joint Conference on Digital Libraries, pp.
393394 (2012)
2. Zhang, Y., Zhou, G., Zhang, J., Xie, M., Yu, W., Li, S.: Engineering pathway for
user personal knowledge recommendation. In: Gao, H., Lim, L., Wang, W., Li, C.,
Chen, L. (eds.) WAIM 2012. LNCS, vol. 7418, pp. 459470. Springer, Heidelberg
(2012)
3. Saaty, T.L.: Fundamentals of Decision Making and Priority Theory with the AHP,
2nd edn. RWS Publications, Pittsburgh (2000)
4. Koschke, L., Fuerst, C., Frank, S.: A multi-criteria approach for an integrated
land-cover-based assessment of ecosystem services provision to support landscape
planning. Ecological Indicators 21, 5466 (2012)
5. Ergu, D., Kou, G., Peng, Y.: Data Consistency in Emergency Management. Inter-
national Journal of Computers Communications & Control 7(3), 450458 (2012)
6. Forsyth, G.G., Le Maitre, D.C., OFarrell, P.J.: The prioritization of invasive alien
plant control projects using a multi-criteria decision model informed by stakeholder
input and spatial data. Journal of Environmental Management 103, 5157 (2012)
7. Tao, L., Chen, Y., Liu, X.: An integrated multiple criteria decision making model
applying axiomatic fuzzy set theory
8. Uzoka, F.-M.E., Obot, O., Barker, K.: An experimental comparison of fuzzy logic
and analytic hierarchy process for medical decision support systems. Computer
Methods and Programs in Biomedicine 103(1), 1027 (2011)
9. Feng, Z., Wang, Q.: Research on health evaluation system of liquid-propellant
rocket engine ground-testing bed based on fuzzy theory. Acta Astronautica 61(10),
840853 (2007)
 What Can We Get from Learning Resource Comments on EP 493

10. Svoray, T., Michailov, E., Cohen, A.: Predicting gully initiation: comparing data
mining techniques, analytical hierarchy processes and the topographic threshold.
Earth Surface Processes and Landforms 37(6), 607619 (2012)
11. Amin, G.R., Gattoufi, S., Seraji, E.R.: A maximum discrimination DEA method
for ranking association rules in data mining. International Journal of Computer
Mathematics 88(11), 22332245 (2011)
12. Rad, A., Naderi, B., Soltani, M.: Clustering and ranking university majors using
data mining and AHP algorithms: A case study in Iran. Expert Systems with
Applications 38(11), 755763 (2011)
13. Brusilovsky, P., Cassel, L., Delcambre, L., Fox, E., Furuta, R., Garcia, D.D., Ship-
man III, F.M., Bogen, P., Yudelson, M.: Enhancing digital libraries with social
navigation: The case of ensemble. In: Lalmas, M., Jose, J., Rauber, A., Sebastiani,
F., Frommholz, I. (eds.) ECDL 2010. LNCS, vol. 6273, pp. 116123. Springer,
Heidelberg (2010)
14. Fernandez-Villavicencio, N.G.: Helping students become literate in a digital,
networking-based society: a literature review and discussion. International Infor-
mation and Library Review 42(2), 124136 (2010)
15. Hsu, K.-K., Tsai, D.-R.: Mobile Ad Hoc Network Applications In The Library. In:
Proceedings of the 2010 Sixth International Conference on Intelligent Information
Hiding and Multimedia Signal Processing (IIHMSP 2010), pp. 700703 (2010)
16. Jing, H., Qingjun, G.: Prospect Application in the Library of SNS. In: 2011
Third Pacific-Asia Conference on Circuits, Communications and System (PACCS),
Wuhan, China, July 17-18 (2011)
17. Alias, N.A.R., Noah, S.A., Abdullah, Z., et al.: Application of semantic technology
in digital library. In: Proceedings of 2010 International Symposium on Information
Technology (ITSim 2010), pp. 15141518 (2010)
 Tuned X-HYBRIDJOIN
for Near-Real-Time Data Warehousing

M. Asif Naeem

School of Computing and Mathematical Sciences, Auckland University of Technology,

Private Bag 92006, Auckland 1142, New Zealand
[email protected]

Abstract. Near-real-time data warehousing defines how updates from

data sources are combined and transformed for storage in a data ware-
house as soon as the updates occur. Since these updates are not in
warehouse format, they need to be transformed and a join operator is
usually required to implement this transformation. A stream-based al-
gorithm called X-HYBRIDJOIN (Extended Hybrid Join), with a favor-
able asymptotic runtime behavior, was previously proposed. However,
X-HYBRIDJOIN does not tune its components under limited available
memory resources and without assigning an optimal division of memory
to each join component the performance of the algorithm can be subop-
timal. This paper presents a variant of X-HYBRIDJOIN called Tuned X-
HYBRIDJOIN. The paper shows that after proper tuning the algorithm
performs significantly better than that of the previous X-HYBRIDJOIN,
and also better as other join operators proposed for this application found
in the literature. The tuning approach has been presented, based on mea-
surement techniques and a revised cost model. The experimental results
demonstrate the superior performance of Tuned X-HYBRIDJOIN.

Keywords: Data warehousing, Tuning and performance optimization,

Data transformation, Stream-based joins.

1 Introduction
Near real-time data warehousing exploits the concepts of data freshness in tra-
ditional static data repositories in order to meet the required decision support
capabilities. The tools and techniques for promoting these concepts are rapidly
evolving. Most data warehouses have already switched from a full refresh [7] to
an incremental refresh policy [4]. Further the batch-oriented, incremental refresh
approach is moving towards a continuous, incremental refresh approach.
One important research area in the field of data warehousing is data transfor-
mation, since the updates coming from the data sources are not in the format
required for the data warehouse. Furthermore, a join operator is required to
implement the data transformation.
In traditional data warehousing the update tuples are buered in memory and
joined when resources become available [6]. Whereas, in real-time data warehous-
ing these update tuples are joined when they are generated in the data sources.

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 494505, 2013.
c Springer-Verlag Berlin Heidelberg 2013

 Tuned X-HYBRIDJOIN for Near-Real-Time Data Warehousing 495

One important factor related to the join is that both inputs of the join come from
dierent sources with dierent arrival rates. The input from the data sources is
in the form of an update stream which is fast, while the access rate of the lookup
table is comparatively slow due to disk I/O cost. It creates a bottleneck in pro-
cessing of update stream and the research challenge here is to minimize this
bottleneck by optimizing the performance of the join operator.
To overcome these challenges a novel stream-based join algorithm called X-
HYBRIDJOIN (Extended Hybrid Join) [1] was proposed recently by the author.
This algorithm not only addressed the issues described above but also was de-
signed to take into account typical market characteristics, commonly known as
the 80/20 sale Rule [2]. According to this rule 80 percent of sale focus on only 20
percent of the products, i.e., one observes a Zipfian distribution. To achieve this
objective, one component of the algorithm called disk-buer is divided into two
equal parts. The contents of one part of the disk buer (called the non-swappable
part) are kept fixed while the contents of the other part (called the swappable
part) are exchanged for each iteration of the algorithm. As the non-swappable
part of the disk buer always contains the most frequently used disk pages, most
stream tuples can be processed without invoking the disk. Although the author
presented an adaptive algorithm to adopt the typical market characteristics, the
components of the algorithm are not tuned to make ecient use of available
memory resources. Further details about this issue are provided in Section 3.
On the basis of these observations, a revised X-HYBRIDJOIN is proposed
with name Tuned X-HYBRIDJOIN. The cost model of existing X-HYBRIDJOIN
is revised and the components of the proposed algorithm are tuned based on that
cost model. As a result the available memory is distributed among all compo-
nents optimally and consequently it improves the performance of the algorithm
significantly.
The rest of the paper is structured as follows. The related work to proposed
algorithm is presented in Section 2. Section 3 describes problem statement about
the current approach. The proposed solution for the stated problem is presented
in Section 4. Section 5 presents the tuning of the proposed algorithm based on
the revised cost model. The experimental study is discussed in Section 6 and
finally Section 7 concludes the paper.

2 Related Work

Some techniques have already been introduced to process the join queries over
continuous streaming data [5]. This section presents only those approaches which
are directly related to the stated problem domain.
A stream-based algorithm called Mesh Join (MESHJOIN) [8] was designed
specifically for joining a continuous stream with disk-based data in an active
data warehouse. This is an adaptive approach but there are some research is-
sues related to inecient memory distribution among join components due to
unnecessary constraints and an inecient strategy for accessing the disk based
relation.
496 M.A. Naeem

R-MESHJOIN (reduced Mesh Join) [9] is a revised version of MESHJOIN

that focuses on the optimal distribution of memory among the join components.
The R-MESHJOIN algorithm introduces the new strategy for memory distribu-
tion among the join components by implementing real constraints. However, the
mechanism used for accessing the disk based relation is similar to MESHJOIN.
One approach to improve MESHJOIN has been a partition-based join algo-
rithm [10]. It uses a two-level hash table in order to attempt to join stream tuples
as soon as they arrive, and uses a partition-bases waiting area for other stream
tuples. In this approach the author keeps focus about the analysis of the stream
buer in terms of back log tuples rather than analysing the performance of the
algorithm.
A recent piece of work introduces a novel stream-based join called HYBRID-
JOIN (Hybrid Join) [11]. The key objective of this eort is to minimize the disk
overhead using an index-based approach and to deal with intermittency in the
stream. Although both issues are successfully addressed in this approach, it is
not optimal with respect to stream data with Zipfian distribution.

3 Preliminaries and Problem Definition

This section presents a working overview of X-HYBRIDJOIN along with the

research issue. In the field of real-time data warehousing X-HYBRIDJOIN is an
adaptive algorithm for joining the bursty data stream with disk-based master
data. Although the typical characteristics of market data are considered, optimal
settings for the available limited memory resources are not considered. Before
describing the problem it is first necessary to explain the major components of
X-HYBRIDJOIN and the role of each component. Figure 1 presents an abstract
level working overview of X-HYBRIDJOIN where m is the number of partitions
in the queue to store stream tuples and n is the number of pages in disk-based
master data R. Moreover, R is assumed to be sorted with respect to the access
frequency. The stream tuples are stored in the hash table while the join attribute
values are stored in the queue. The queue is implemented using a doubly linked-
list data structure to allow the random deletion of matching tuples. The disk
buer is used to load the disk pages into memory. To make ecient use of R by
minimizing the disk access cost, the disk buer is divided into two equal parts.
One is called the non-swappable part which stores a small but most frequently
used portion of R into memory on a permanent basis. The other part of the disk
buer is swappable and for each iteration it loads the disk page pi from R into
memory.
Before the join process starts X-HYBRIDJOIN loads the most frequently used
page of R into the non-swappable part of the disk buer. During the join process,
for each iteration the algorithm dequeues the oldest join attribute value from
the queue and using this value as an index it loads the relevant disk page into
the swappable part of the disk buer. After loading the disk page into memory
the algorithm matches each of the disk tuples available in both the swappable
and non-swappable parts of the disk buer with the stream tuples in the hash
 Tuned X-HYBRIDJOIN for Near-Real-Time Data Warehousing 497

Queue Queue

tm . . . . t3 t2 t1
tm ... t3 t2 t1

Hash table

Join Stream
Stream Hash Join
S Hash output function output
... Stream buffer
function
Stream buffer
...

... Disk buffer

Disk buffer Hash table for stream
Non- Swappable
swappable
Non-swappable part Non-swappable part
Size=l number of pages Size=k number of pages

Join window
Disk-based p1 p p p ... p
master data p2
R .. Master data (R)
pn Size of each p=k number of pages

Fig. 1. X-HYBRIDJOIN working Fig. 2. Memory architecture for Tuned X-

overview HYBRIDJOIN

table. If the tuple match, the algorithm generates the resulting tuple as output
and deletes the stream tuple from the hash table along with its join attribute
value from the queue. In the next iteration the algorithm again dequeues the
oldest element from the queue, loads the relevant disk page into the swappable
part of the disk buer and repeats the procedure.
X-HYBRIDJOIN minimizes the disk access cost and improves performance
significantly by introducing the non-swappable part of the disk buer. But in
X-HYBRIDJOIN the memory assigned to the swappable part of the disk buer
is equal to the size of disk buer in HYBRIDJOIN [11] and the same amount
of memory is allocated to the non-swappable part of the disk buer. In the
following it will be shown that this is not optimal. The problem considered in
this paper is to tune the size of both parts of the disk buer so that the memory
distribution among these two components is optimal. Once these two components
acquire optimal settings, based on that memory can be assigned to the rest of
join components.

4 Proposed Solution

As a solution for the above stated problem, a revised version of X-HYBRIDJOIN

called Tuned X-HYBRIDJOIN is proposed. This section presents the memory
architecture and a revised cost model for the proposed algorithm. Most im-
portantly, the tuning for the proposed algorithm is presented while the tuning
procedure is based on both a measurement strategy and the cost model.

4.1 Memory Architecture

The memory architecture that Tuned X-HYBRIDJOIN uses is shown in Fig-

ure 2. From the figure, Tuned X-HYBRIDJOIN includes the same number of
components as X-HYBRIDJOIN however, the memory size for each component
498 M.A. Naeem

is dierent to that in X-HYBRIDJOIN. In Tuned X-HYBRIDJOIN since mem-

ory is allocated to the each component after executing the tuning module and
therefore, each component is assigned an optimal size of memory. Particulary,
X-HYBRIDJOIN uses same memory for the both swappable and non-swappable
parts of the disk buer but after tuning it has been explored that the optimal size
of memory for the both components is dierent. The reason for it is presented
in Section 5.

4.2 Cost Calculation

This section revises the cost formulas derived in X-HYBRIDJOIN. The reason for
revising the cost model is that X-HYBRIDJOIN uses equal memory for both the
swappable and non-swappable parts of the disk buer, and therefore the formulas
do not apply for other relative sizes. Following the style of cost modeling used
for MESHJOIN, the cost for any algorithm is expressed in terms of memory and
processing time. Equation 1 describes the total memory used to implement the
algorithm while Equation 2 calculates the processing cost for w tuples.

Memory Cost. Since the optimal values for the sizes of both the swappable
part and non-swappable part can be dierent, it is assumed k number of pages for
the swappable part and l number of pages for the non-swappable part. Overall
the largest portion of the total memory is used for the hash table while a much
smaller amount is used for each of the disk buer and the queue. The memory
for each component can be calculated as given below:
Memory for the swappable part of disk buer (bytes)= k vP (where vP is the
size of each disk page in bytes).
Memory for the non-swappable part of disk buer (bytes)= l vP .
Memory for the hash table (bytes)= [M (k + l)vP ] (where M is the total
allocated memory and is memory weight for the hash table).
Memory for the queue (bytes)= (1 )[M (k + l)vP ] (where (1 ) is memory
weight for the queue).
The total memory used by the algorithm can be determined by aggregating the
above.

M = (k + l)vP + [M (k + l)vP ] + (1 )[M (k + l)vP ] (1)

Currently the memory reserved for the stream buer is not included because it
is small (0.05 MB was sucient in all experiments presented in this paper).

Processing Cost. This section presents the processing cost for the proposed
approach. The cost for one iteration of the algorithm is denoted by cloop and
express it as the sum of the costs for the individual operations. Therefore the
processing cost for each component is first calculated separately.
Cost to read the non-swappable part of disk buer (nanoseconds) = cI/O (l vP ).
Cost to read the swappable part of disk buer (nanoseconds)= cI/O (k vP ).
 Tuned X-HYBRIDJOIN for Near-Real-Time Data Warehousing 499

Cost to look-up the non-swappable part of disk buer in the hash table (nanosec-
onds) = dN cH (where dN = l vvR P
is the size of the non-swappable part of disk
buer in terms of tuples, vP is size of disk page in bytes, vR is size of disk tuple
in bytes, and cH is look-up cost for one disk tuple in the hash table).
Cost to look-up the swappable part of disk buer in the hash table (nanoseconds)=
dS cH (where dS = k vvR P
is the size of the swappable part of disk buer in terms
of tuples).
Cost to generate the output for w matching tuples (nanoseconds) = wcO (where
cO is cost to generate one tuple as an output).
Cost to delete w tuples from the hash table and the queue (nanoseconds)= w cE
(where cE is cost to remove one tuple from the hash table and the queue).
Cost to read w tuples from stream S into the stream buer (nanoseconds)= wcS
(where cS is cost to read one stream tuple into the stream buer).
Cost to append w tuples into the hash table and the queue (nanoseconds)= wcA
(where cA is cost to append one stream tuple in the hash table and the queue).
As the non-swappable part of the disk buer is read only once before execution
starts, it is excluded. The total cost for one loop iteration is:

cloop (secs) = 109 [cI/O (k vP ) + (dN + dS )cH + w(cO + cE + cS + cA )] (2)

Since in every cloop seconds the algorithm processes w tuples of stream S, the
performance or service rate can be calculated by dividing w by the cost for
one loop iteration.
w
= (3)
cloop

5 Tuning
The stream-based join operators normally execute within limited memory and
therefore tuning of join components is important to make ecient use of the
available memory. For each component in isolation, more memory would be bet-
ter but assuming a fixed memory allocation there is a trade-o in the distribution
of memory. Assigning more memory to one component means less memory for
other components. Therefore it needs to find the optimal distribution of memory
among all components in order to attain maximum performance. A very impor-
tant component is the disk buer because reading data from disk to memory is
expensive.
In the proposed approach tuning is first performed through performance mea-
surements by considering a series of values for the sizes of the swappable and
non-swappable parts of the disk buer. Later a mathematical model for tuning is
also derived from the cost model. Finally, the tuning results of both approaches
are compared to validate the cost model. The details about the experimental
setup are presented in Table 1.

5.1 Tuning through Measurements

This section presents the tuning of the key components of the algorithm through
measurements. In the measurement approach the performance is tested on
500 M.A. Naeem

Table 1. Data specification

Parameter value
Disk-based data
Size of R 0.5 million to 8 million tuples
Size of each tuple vR 120 bytes
Stream data
Size of each tuple vS 20 bytes
Size of each node in queue 12 bytes
Benchmark
Based on Zipfs law
Characteristics Bursty and self-similar

particular memory settings for swappable and non-swappable parts rather than
on every contiguous value.
The measurement approach assumes the size of total memory and the size
of R are fixed. The sizes for the swappable and non-swappable parts vary in
such a way that for each size of the swappable part the performance is measured
against a range of sizes for the non-swappable part. By changing the sizes for
both parts of the disk buer the memory sizes for the hash table and the queue
are also aected.
The performance measurements for varying the sizes of both swappable and
non-swappable parts are shown in Figure 3. The figure shows that the perfor-
mance increases rapidly by increasing the size for the non-swappable part. After
reaching a particular value for the size of non-swappable part the performance
starts decreasing. The plausible reason behind this behavior is that in the be-
ginning when the size for the non-swappable part increases, the probability of
matching stream tuples with disk tuples also increases and that improves the
performance. But when the size for the non-swappable part is increased further
it does not make a significant dierence in stream matching probability. On the
other hand, due to higher look-up cost and the fact that less memory is avail-
able for the hash table the performance decreases gradually. A similar behavior
is seen when the performance against the swappable part is tested. In this case,
after attaining the maximum performance it decreases rapidly because of an
increase in the I/O cost for loading the growing swappable part. From the mea-
surements shown in the figure it is possible to approximate the optimal settings
for both the swappable and non-swappable parts by finding the maximum on
the two-dimensional surface.

5.2 Tuning Based on Cost Model

A mathematical model for the tuning is also derived based on the cost model
presented in Section 4.2. From Equation 3 it is clear that the service rate depends
on the size of w and the cost cloop . To determine the optimal settings it is first
necessary to calculate the size of w. The main components on which the value
 Tuned X-HYBRIDJOIN for Near-Real-Time Data Warehousing 501

4
x 10

Frequency in stream
1.35

1.3
Service rate (tuples/sec)

1.25

1.2

1.15

1.1
950
Siz
. . .
eo 900 2000
f sw
app 1750
able 850
r (tuples)
part
of d
is 800
1500
of disk
buffe Part of master data Master Data on disk Part of master data
that is loaded into memory
k bu 1250 part that exists permanently in memory.
ffer
750 ap pable in the form of partitions.
(tup 1000 nsw
les) of no
Size

Fig. 3. Tuning using measurement ap- Fig. 4. A sketch of matching proba-

proach bility of R in stream

of w depends are:- size of the non-swappable part (dN ), size of the swappable
part (dS ), size of the master data (Rt ), and size of the hash table (hS ).
Typically the stream of updates can be approximated through Zipfs law with
a certain exponent value. Therefore, a significant part of the stream is joined
with the non-swappable part of the disk buer. Hence, if the size of the non-
swappable part (i.e. dN ) is increased, more stream tuples will match as a result.
But the probability of matching does not increase at the same rate as increasing
dN because, according to Zipfian distribution, the matching probability for the
second tuple in R is half of that for the first tuple and similarly the matching
probability for the third tuple is one third of that for the first tuple and so on [2].
Due to this property, the size of R (denoted by Rt ) also aects the matching
probability. The swappable part of the disk buer deals with the rest of the
master data denoted by R (where R = Rt dN ), which is less frequent in
the stream than that part which exists permanently in memory. The algorithm
reads R in partitions, where the size of each partition is equal to the size of the
swappable part of the disk buer dS . In each iteration the algorithm reads one
partition of R using an index on join attribute and loads it into memory through
a swappable part of the disk buer. In the next iteration the current partition in
memory is replaced by a new partition, and so on. As mentioned earlier, using
the Zipfian distribution the matching probability for every next tuple is less than
the previous one. Therefore, the total number of matches against each partition
is not the same. This is explained further in Figure 4, where n total partitions
are considered in R . From the figure it can be seen the matching probability for
each disk partition decreases continuously as one moves toward the end position
in R.
The size of the hash table is another component that aects w. The reason
is simple: if there are more stream tuples in memory, the number of matches
will be greater and vice versa. Before driving the formula to calculate w it is
first necessary to understand the working strategy of Tuned X-HYBRIDJOIN.
Consider for a moment that the queue contains stream tuples instead of just
join attribute values. Tuned X-HYBRIDJOIN uses two independent inner loops
under one outer loop. After the end of the first inner loop, which means after
502 M.A. Naeem

finishing the processing of the non-swappable part, the queue only contains those
stream tuples which are related to only the swappable part of R, denoted by R .
For the next outer iteration of the algorithm these stream tuples in the queue
are considered to be an old part of the queue. In that next outer iteration the
algorithm loads some new stream tuples into the queue and these new stream
tuples are considered to be a new part of the queue. The reason for dividing the
queue into two parts is that the matching probability for both parts of the queue
is dierent. The matching probability for the old part of the queue is denoted by
pold and it is only based on the size of the swappable part of R i.e. R . On the
other hand, the matching probability for the new part of the queue, known as
pnew , depends on both the non-swappable as well as the swappable parts of R.
Therefore, to calculate w it is first needed to calculate both these probabilities.
Therefore, if the stream of updates S obeys Zipfs law, then the matching
probability for any swappable partition k with the old part of the queue can be
determined mathematically as shown below.

dN #
+kdS
1
x
x=dN +(k1)dS +1
pk =
#
Rt
1
x
x=dN +1

Each summation in the above equation generates a harmonic series, which can
#k
1
be summed up using the formula x = ln k + + k , where is a Eulers
x=1
constant whose value is approximately equal to 0.5772156649 and k is another
1
constant which is 2k . The value of k approaches 0 as k goes to [3]. In this
1
paper the value of 2k is small and therefore, it is ignored.
If there are n partitions in R , then the average probability of an arbitrary
partition of R matching the old part of the queue can be determined using
Equation 4.
#
n
pk
1
pold = k=1 = (4)
n n
Now the probability of matching is determined for the new part of the queue.
Since the new input stream tuple can match either the non-swappable or the
swappable part of R, the average matching probability of the new part of the
queue with both parts of the disk buer can be calculated using Equation 5.

1
pnew = pN + pS (5)
n

where pN and pS are the probabilities of matching for a stream tuple with the
non-swappable part and the swappable part of the disk buer respectively. The
values of pN and pS can be calculated as below.
 Tuned X-HYBRIDJOIN for Near-Real-Time Data Warehousing 503

#
d N
1
#
Rt
1
x x
x=1 x=dN +1
pN = and pS =
#Rt
1 #Rt
1
x x
x=1 x=1

Assume that w are the new stream tuples that the algorithm will load into the
queue in the next outer iteration. Therefore,
The size of the new part of the queue (tuples)=w
The size of the old part of the queue (tuples)=(hS w)
If w are the average number of matches per outer iteration with both the swap-
pable and non-swappable parts, then w can be calculated by applying the bino-
mial probability distribution on Equations 4 and 5 as given below.

w = (hS w)pold (1 pold ) + wpnew (1 pnew )

After simplification the final formula to calculate w is described in Equation 6.

hS pold (1 pold )
w= (6)
1 + pold (1 pold ) pnew (1 pnew )

By using the values of w and cloop in Equation 3 the algorithm can be tuned.

5.3 Comparisons of Both Approaches

To validate the cost model the tuning results based on the measurement approach
are compared with those that are achieved through cost model.

Swappable Part: This experiment compares the tuning results for the swap-
pable part of the disk buer using both the measurement and cost model ap-
proaches. The tuning results of each approach (with 95% confidence interval in
case of measurement approach) are shown in Figure 5 (a). From the figure it is
evident that at every position the results in both cases are similar, with only
0.5% deviation.

Non-swappable Part: Similar to before, the tuning results of both approaches

for the non-swappable part of the disk buer are also compared. The results are
shown in Figure 5 (b). Again, it can be seen from the figure, the results in both
cases are nearly equal with a deviation of only 0.6%. This proves the correctness
of the tuning module.

6 Experimental Study
To strengthen the arguments an experimental evaluation of proposed Tuned X-
HYBRIDJOIN is performed using the synthetic datasets. Normally, in Tuned
X-HYBRIDJOIN kinds of algorithms, the total memory and the size of R are
504 M.A. Naeem

4 4
x 10 x 10
1.5 1.5
Measured Measured
Calculated Calculated
1.4 1.4
Service rate (tuples/sec)

Service rate(tuples/sec)
1.3 1.3

1.2 1.2

1.1 1.1

1 1
750 800 850 900 950 1000 1250 1500 1750 2000
Size of the swappable part of the disk buffer(tuples) Size of the nonswappable part of the disk buffer(tuples)

(a) Tuning Comparison for swappable (b) Tuning Comparison for non-swappable
part: based on measurements Vs based on part: based on measurements Vs based on
cost model cost model

Fig. 5. Comparisons of tuning results

5 4
10 x 10
4
Tuned XHYBRIDJOIN
Tuned XHYBRIDJOIN
XHYBRIDJOIN XHYBRIDJOIN
HYBRIDJOIN HYBRIDJOIN
3.5
RMESHJOIN RMESHJOIN
MESHJOIN MESHJOIN
3
Service rate (tuples/sec)

Service rate (tuples/sec)

2.5
4
10

1.5

3
10 0.5
0.5 1 2 4 8 50 100 150 200 250
Size of R (million tuples) Total memory (MB)

(a) Size of disk-based relation varies (on (b) Total allocated memory varies
log-log scale)

Fig. 6. Performance comparisons

the common parameters that vary frequently. Therefore, the experiments pre-
sented here compare the performance by varying both parameters individually.

Performance Comparisons for Dierent Sizes of R: This experiment com-

pares the performance of Tuned X-HYBRIDJOIN with the other related algo-
rithms for dierent sizes of R. Therefore it is assumed that the size of R varies
exponentially while the total memory budget remains fixed (50MB) for all values
of R. For each value of R the performance is measured separately. The perfor-
mance results of this experiment are shown in Figure 6 (a). From the figure it is
clear that for all settings of R the Tuned X-HYBRIDJOIN performs significantly
better than other approaches.

Performance Comparisons for Dierent Memory Budgets: Second ex-

periment analyses the performance of Tuned X-HYBRIDJOIN using dierent
memory budgets while the size of R is fixed (2 million tuples). Figure 6 (b)
 Tuned X-HYBRIDJOIN for Near-Real-Time Data Warehousing 505

depicts the performance results. From the figure it can be observed that for all
memory budgets the Tuned X-HYBRIDJOIN again performs significantly better
than all approaches. This improvement increases gradually as the total memory
budget increases.

7 Conclusions
This paper investigates a well known stream-based join algorithm called X-
HYBRIDJOIN. Main observation about X-HYBRIDJOIN is that the tuning
factor is not considered but it is necessary, particularly when limited memory
resources are available to execute the join operation. By omitting the tuning fac-
tor, the available memory cannot be distributed optimally among the join com-
ponents and consequently the algorithm cannot perform optimally. This paper
presents a variant version of X-HYBRIDJOIN called Tuned X-HYBRIDJOIN.
The cost model presented in X-HYBRIDJOIN is revised and the proposed algo-
rithm is tuned based on that revised cost model. To strengthen the arguments a
prototype of Tuned X-HYBRIDJOIN is implemented and the performance with
existing approaches is compared.

References
1. Naeem, M.A., Dobbie, G., Weber, G.: X-HYBRIDJOIN for Near-Real-Time Data
Warehousing. In: Fernandes, A.A.A., Gray, A.J.G., Belhajjame, K. (eds.) BNCOD
2011. LNCS, vol. 7051, pp. 3347. Springer, Heidelberg (2011)
2. Anderson, C.: The Long Tail: Why the Future of Business is Selling Less of More.
Hyperion (2006)
3. Milton, A., Irene, A.S.: Handbook of Mathematical Functions with Formulas,
Graphs, and Mathematical Tables. Ninth Dover printing, Tenth GPO printing,
New York (1964)
4. Labio, W.J., Wiener, J.L., Garcia-Molina, H., Gorelik, V.: Ecient resumption of
interrupted warehouse loads. SIGMOD Rec. 29(2), 4657 (2000)

5. Golab, L., Tamer Ozsu, M.: Processing Sliding Window Multi-Joins in Continuous
Queries over Data Streams. In: VLDB 2003, Berlin, Germany, pp. 500511 (2003)
6. Wilschut, A.N., Apers, P.M.G.: Dataflow query execution in a parallel main-
memory environment. Distrib. Parallel Databases 1(1), 103128 (1993)
7. Gupta, A., Mumick, I.S.: Maintenance of Materialized Views: Problems, Tech-
niques, and Applications. IEEE Data Engineering Bulletin 18, 318 (2000)
8. Polyzotis, N., Skiadopoulos, S., Vassiliadis, P., Simitsis, A., Frantzell, N.: Mesh-
ing Streaming Updates with Persistent Data in an Active Data Warehouse. IEEE
Trans. on Knowl. and Data Eng. 20(7), 976991 (2008)
9. Naeem, M.A., Dobbie, G., Weber, G.: R-MESHJOIN for Near-real-time Data
Warehousing. In: DOLAP 2010: Proceedings of the ACM 13th International Work-
shop on Data Warehousing and OLAP. ACM, Toronto (2010)
10. Chakraborty, A., Singh, A.: A partition-based approach to support streaming up-
dates over persistent data in an active datawarehouse. In: IPDPS 2009: Proceedings
of the 2009 IEEE International Symposium on Parallel & Distributed Processing,
pp. 111. IEEE Computer Society, Washington, DC (2009)
11. Naeem, M.A., Dobbie, G., Weber, G.: HYBRIDJOIN for Near-real-time
Data Warehousing. International Journal of Data Warehousing and Mining
(IJDWM) 7(4) (2011)
 Exploiting Interaction Features
in User Intent Understanding

Vincenzo Deufemia, Massimiliano Giordano,

Giuseppe Polese, and Luigi Marco Simonetti

Universit`
a di Salerno, Via Ponte don Melillo, Fisciano(SA), Italy
{deufemia,mgiordano,gpolese}@unisa.it, [email protected]

Abstract. Understanding user intent during a web navigation session is

a challenging topic, which is drawing the attention of many researchers
in this area. The achievements of such research goals would have a great
impact on many internet-based applications. For instance, if a search
engine had the capability of capturing user intents, it could better suite
the order of search results to user needs. In this context, the research has
mainly focused on the analysis of user interactions with Search Engine
Result Pages (SERPs) resulting from a web query, but most methods
ignore the behavior of the user during the exploration of web pages as-
sociated to the links of the SERP s/he decides to visit. In this paper we
propose a novel model that analyzes user interactions on such pages, in
addition to the information considered by other mentioned approaches.
In particular, captured user interactions are translated into features that
are part of the input of a classification algorithm aiming to determine
user informational, navigational, and transactional intents. Experimental
results highlight the eectiveness of the proposed model, showing how
the additional analysis it performs on visited web pages contributes to
enhance user intent understanding.

Keywords: user intent understanding, interaction features, web search.

1 Introduction

The success of internet applications is bound to the capability of search engines

to provide users with information meeting their expectations. This cannot be
guaranteed by merely analyzing the web structure as several existing search
engines do. Thus, nowadays, search engines need to incorporate the capability
of predicting user intents in order to adapt the order of search results so as to
meet their expectation. For this reason, user intention understanding (UIU) has
recently become an important research area.
Some approaches to user behavior analysis in web navigation have highlighted
the importance of analyzing user interactions with web pages in order to infer
their interest and satisfaction with respect to the visited contents [13]. Other
studies have investigated how user interactions with Search Engine Result Pages
(SERPs) can be exploited to infer user intent [49]. However, such methods

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 506517, 2013.
c Springer-Verlag Berlin Heidelberg 2013

 Exploiting Interaction Features in User Intent Understanding 507

for UIU limit their analysis to the results contained in a SERP, ignoring many
important interactions and contents visited from such results. On the contrary,
the aim of this paper is to show that UIU can be considerably improved by
performing additional analysis of user interactions on the web pages of a SERP
that the user decides to visit. In particular, we define a new model for UIU that
incorporates interactions analyzed in the context of SERP results and of the
web pages visited starting from them. The interaction features considered in the
model are not global page-level statistics, rather they are finer-grained and refer
to portions of web pages. This is motivated on the basis of the results of literature
studies, performed with eye-trackers, revealing that although a user might focus
on many sections composing a web page, s/he will tend to overlook portions
of low interest [10]. Thus, capturing interaction features on specific portions of
web pages potentially conveys a better accuracy in the evaluation of the user
actions. These and other features are evaluated in our approach by means of a
classification algorithm to understand user intents. In particular, to simplify the
classification process, we use a common taxonomy that defines three types of
queries: informational, navigational, and transational [11].
Experimental results highlight the eciency of the proposed model in query
classification, showing how the interaction features extracted from visited web
pages contribute to enhance user understanding. In particular, the proposed set of
features has been evaluated with three dierent classification algorithms, namely
Support Vector Machine (SVM) [12], Conditional Random Fields (CRF) [13], and
Latent Dynamic Conditional Random Fields (LDCRF) [14]. The achieved results
have been compared with those achieved with other subsets of features and demon-
strate that the proposed model outperforms them in terms of query classification.
Moreover, a further analysis highlights the eectiveness of the proposed features
for the classification of transactional queries.
The rest of this paper is organized as follows. In Section 2, we provide a
short review of related work. Then, we present the model exploiting interaction
features for UIU in Section 3. Section 4 describes experimental results. Finally,
conclusions are given in Section 5.

2 Related Work
A pioneer study by Carmel et al. [15] in the early 90s focusses on the analysis of
hypertext. In particular, the study highlights three navigation strategies: scan
browsing, aiming to inquire and evaluate information contained in the hypertext
based on interest, review browsing, aiming to integrate information contained
in the user mental context, and search-oriented browsing, aiming to look for
new information based on specific goals, and integrating them with information
collected in successive scans.
A further interesting study by Morrison et al. [16] aimed to classify types of
queries based on search intents. Such taxonomies are defined as formalizations of
three basic questions that the user asks him/herself before starting a search ses-
sion: why, how, and what to search. The taxonomy that is closest to user intent un-
derstanding is the one used in the context of a research called (method taxonomy),
508 V. Deufemia et al.

which aims to detect the following types of search activities: explore, monitor, find,
and collect. By monitoring user search activities, Sellen et al. extend previously
defined taxonomies, introducing transactional search types [17], defined as trans-
acting, and finalized to search purposes and to the use of web services on which
it is possible to better exploit user experiences while performing search activities.
In this context, the goal of the search is not to find information, rather services of
information manipulation: hence, the search engine becomes a means to migrate
towards a web service. Successively, new approaches to user intent understanding
have highlighted the necessity to introduce a clear classification of search queries
based on user intents. To this end, one of the former studies by Broder et al. [11]
refers to a simple taxonomy of search queries based on three distinct categories:
navigational queries to search for a particular web site, informational queries to
collect information from one or more web pages, and transactional queries to reach
a web site on which further interactions are expected.
In the last decade, all existing approaches have aimed at applying taxonomies
in a recognition and automatic classification context. One of the former method-
ologies by Kang et al. [4] aims at two types of search activities: topic relevance,
that is, searching documents for a given topic, of informational type, and home-
page finding, aiming to search main pages of several types of navigational web
sites. Starting from common information used by Information Retrieval (IR)
systems, such as web page content, hyperlinks, and URLs, the model proposes
methods to classify queries based on the two categories mentioned above.
The analysis and the comprehension of user interaction models during web
navigation is the basis of a predictive model on real case studies, proposed by
Agichtein et al. [6]. The model tries to elicit and understand user navigation
behaviors by analyzing several activities, such as clicks, scrolls, and dwell times,
aiming to predict user intention during web page navigation. The features this
study proposes to analyze are used to characterize the complex system of in-
teractions following the click on a result. Successively, a study by Guo et al. [9]
starts from the hypothesis that such interactions can help accurately inferring
two particular tightly correlated intents: search and purchase of products, a two
phase activity defined as search-purchase.
Starting from experimental studies on real user navigation strategies, Lee
et al. propose a model for the automatic identification of search goals based
on features, and restricted to navigational and informational queries [5]. These
studies have primarily revealed that most queries can be eectively associated
to one of two categories defined within the taxonomy, making it possible the
construction of an automatic identification system. Moreover, most queries that
are not eectively associable to a category are related to few topics, such as
proper nouns or names of software systems. This makes it possible to use ad-hoc
systems of features for all those unpredictable cases. The model proposes two
features: past user-click behavior that infers user intent from the way s/he has
interacted with the results in the past, and anchor-link distribution, which uses
possible targets of links having the same text of the query.
 Exploiting Interaction Features in User Intent Understanding 509

The strategies for user intent understanding described so far aim to classify
search queries exclusively using features modeled to describe several aspects of
search queries. A study by Tamine et al. proposes to analyze search activities
that the user has previously performed in the same context, aiming to derive
data useful for inferring the type of the current query [8]. The set of queries
already performed represents the query profile.

3 The Proposed Model

In this section we describe the proposed model and the features used for the
classification process.

3.1 Search Model: Session, Search, Interaction

Several approaches and models have been proposed to provide solutions to the
user intention understanding problem, but all of them mainly focus on the in-
teraction between users and the SERP. Additional interactions originating from
SERPs contents, such as browsing, reading, and multimedia content fruition are
not considered by the research community.
The proposed model aims to extend existing models, by analyzing user in-
teractions between users and web pages during a search session. Our aim is to
analyze not only the interactions between users and SERP, but also between
users and web pages reached by clicking on SERPs results. We believe that
data about interactions between users and web pages may be very useful to
clarify the intent of the user, because these interactions are driven by the same
motivation behind the initial search query. All user interactions with web pages
could be reduced to three main categories, which are the same we consider in
our model: session, search, and interaction.
Session. A session is a sequence of search activities aimed at achieving a goal.
When the first query does not provide the desired result, the user tries to grad-
ually approach the target, refining or changing search terms and keywords. All
these research activities constitute a session.
Search. A search activity is the combination of the following user actions: sub-
mission of a query to a search engine, analysis of search results, navigation on
one or more web pages inside them. During a search activity a user has a spe-
cific goal, generally described by the query itself. This goal is classifiable in a
taxonomy, as defined by previous studies [11, 18].
Interaction. Interaction is the navigation of a web page using a wide range of
interactions that include mouse clicks, page scrolling, pointer movements, and
text selection. Starting from these interactions, combined with features such as
dwell time, reading rate, and scrolling rate, it is possible to derive an implicit
feedback of users about web pages [2]. Moreover, several studies have proven the
usefulness of user interactions to assess the relevance of web pages [1, 3, 19, 20],
to classify queries, and to determine the intent of search sessions [9, 21].
510 V. Deufemia et al.

The proposed model introduces a new methodology for tracing interactions

between users and web pages. In particular, user interaction analysis is restricted
to a smaller portion of a web page: the subpages. Indeed, a web page consists
of several blocks of text, graphics, multimedia, and can have a variable length.
As shown by studies using eye-tracking [10], the user frequently adopts a E or
F reading pattern during browsing, excluding areas of little interest. Compared
to a global analysis of the entire page, the introduction of subpage-level analysis
gives higher accuracy in the assessment of the interaction.

3.2 Features

Interaction data extracted from user web navigation have been encoded into
features that characterize user behavior. We organize the set of features into the
following categories: query, search, interaction, and context.
Query. These features are derived from characteristics of a search query such
as keywords, the number of keywords, the semantic relations between them, and
other characteristics of a search or an interaction.
Search. These features act on the data from search activities such as: re-
sults, time spent on SERP, and number of results considered by the user. The
DwellTime is measured from the start of the search session until the end of
the last interaction originated by the same search session. The reaction time,
TimeToFirstInteraction, is the time elapsed from the start of the search session
and the complete loading of the first selected page. Other features dedicated
to interactions with the results are ClicksCount, which is the number of visited
results, and FirstResultClickedRank, determining the position of the first clicked
result.
Interaction. These features act on the data collected from interactions with
web pages and subpages, taking into account the absolute dwell time, the ef-
fective dwell time, all the scrolling activities, search and reading activities. The
DwellRate measures the eectiveness of the permanence of a user on a web
page, while the reading rate ReadingRate, measures the amount of reading of a
web page [2]. Additional interactional features are: ViewedWords, the number of
words considered during the browsing, UrlContainsTransactionalTerms, which
verifies if the URL of the page contains transactional terms (download, software,
video, watch, pics, images, audio, etc.), AjaxRequestsCount, which represents the
number of AJAX requests originated during browsing.
Context. These features act on the relationship between the search activities
performed in a session, such as the position of a query in the sequence of search
requests for a session.

4 Experiments

In this section we describe the dataset constructed for evaluating the proposed
approach and the results achieved with dierent classification algorithms.
 Exploiting Interaction Features in User Intent Understanding 511

In particular, after an overview on the evaluation metrics used and on the subsets
of features taken into consideration, experimental results are presented.

4.1 Experiments Configuration

The goal of this experimentation is to evaluate the eectiveness of the proposed

model for inferring user intention in web searches. In particular, we analyze the
performances of the model for classifying queries, based on the query taxonomy:
informational, navigational, and transactional.
Search activities have been accomplished by a pool of thirteen subjects (seven
men and six women), whose age ranges from 24 to 37 years old. Sixth subjects
had a master degree (5 in computer science and 1 in linguistics), 3 a bachelor
degree (1 in computer science, 1 in political science, and 1 in chemistry), and 4
a high school degree in accounting. The average experience in web browsing was
of 9.6 years.

Dataset. The dataset has been constructed by the thirteen participants to the
test. In particular, all the subjects were requested to perform a series of search
activities related to various topics and with dierent intentions. Each subject
was asked to determine his/her own goal in advance. After starting the session,
subjects performed a series of searches using the Google search engine, without
any limitation as to time or the results to visit. By following this protocol, we had
129 sessions and 353 web searches, which were subsequently manually classified
by relying on the intent of the user. Starting from web searches, 490 web pages
and 2136 sub pages were visited. All interactions were detected by the YAR
plug-in for Google Chrome/Chromium [2].

Evaluation Metrics. In order to evaluate the eectiveness of the proposed

model, we adopted the classical evaluation metrics of Information Retrieval:
precision, recall, and F1-measure. They are defined as follows:
" #correctly classif ied queries #category queries
Precision =
#classif ied queries #total queries
Category(i)

#correctly classif ied queries

Recall =
#total queries
P recision Recall
F1 measure = 2
P recision + Recall

Feature Subsets. In order to verify the eectiveness of the proposed features

in dierent combinations, we have grouped them into several subsets:

All : subset of all the proposed features query, search, interaction, and con-
text ;
512 V. Deufemia et al.

Query : subset of all the features related to queries;

Search: subset of all the features related to search and context ;
Interaction: subset of all the features related to interactions;
Query+Search: subset of the features derived as union from Query and
Search. The goal is to evaluate the eectiveness of query classification by
using the features considered in other studies [3, 6, 9];
Transactional : subset of all the features related to interactions over transac-
tional queries ViewWords, AjaxRequestsCount, ScrollingDistance, Scrolling-
Count, and UrlContainsTransactionalTerms. The goal here is to evaluate the
classification of transactional queries by adopting more specific features;
All Transactional : subset derived by the exclusion of the transactional
features from the set All. The goal here is to evaluate the eectiveness of the
classification of transactional queries by comparing results achieved with all
features to those achieved by excluding transactional features.

Classifiers. We considered three classifiers to evaluate the proposed feature

model: SVM [12], CRF[13], and LDCRF[14]. Support Vector Machine (SVM) is
a discriminative model, based on the structural risk minimization principle from
statistical learning theory [12], widely used in binary classification problems. In
particular, SVM maps the input data in a higher dimensional feature space,
and splits the data into two classes by computing a hyperplane that separates
them. CRF is a discriminative probabilistic model used for labeling sequential
data [13]. The model utilizes inter-dependent features to select the most probable
label sequence for one observation LDCRF extends CRF by incorporating hidden
state variables which model the sub-structure of input sequences [14]. It is able to
learn the transitions between input elements by modeling a continuous stream of
class labels, and the internal sub-structure by using intermediate hidden states.
In the context of query classification, SVM assumes that the queries in a user
session are independent, Conditional Random Field (CRF) considers the sequen-
tial information between the queries, and LatentDynamic Conditional Random
Fields (LDCRF) models the sub-structure of user sessions by assigning a disjoint
set of hidden state variables to each class label. They have been configured as
follows:
1. SVM. We used MSVMpack [22] as the SVM toolbox for model training and
testing. The SVM model is trained using a linear kernel and the parameter
C has been determined by cross-validation.
2. CRF. We used the HCRF library1 as the tool to train and test the CRF
model. For the experiments we used a single chain structured model and the
regularization term for the CRF model was validated with values 10k with
k = 1 . . . 3.
3. LDCRF. We used the HCRF library for training and testing LDCRF model.
In particular, the model was trained with 3 hidden states per label, and the
regularization term was determined by cross-validation to achieve the best
performance.
1
http://sourceforge.net/projects/hcrf/
 Exploiting Interaction Features in User Intent Understanding 513

4.2 Results
In order to simulate an operating environment, the set of queries made by users
was separated into two subsets, which included 60% and 40% of web searches,
and were used for training and testing the classifiers, respectively.
Taking into account informational queries (see Figure 1), it could be found
that best classifications were achieved by using the CRF classifier with the All set
of features, whereas worst performances were achieved by considering the sub-
sets of features Query and Search. Similar considerations hold for Search+Query.
These initial observations demonstrate the importance of the subset of interac-
tional features, which alone are able to achieve classification performances close
to those achieved by the All set of all features.

Fig. 1. Performances of the considered feature subsets for informational queries

Among the interactional set of features proposed above, there is a subset,

named Transactional, which was introduced to improve the classification of
transactional queries. The experimental results highlight that these features also
improve the classification of informational queries. In fact, the Search+Query
subset achieves performances that are closer to the subset All Transactional,
as shown in Figure 1. This means that the subset of interactional features, which
are not part of Transactional, are not eective for the classification of informa-
tional queries. The relevance of transactional features in this type of query is
confirmed by two other important experimental results. The first one regards
the application of the single subset Transactional, which is able to achieve per-
formances equivalent to the Search+Query and All Transactional subsets. The
second one is the performance degradation of the InteractionTransactional
features, which achieve results by far lower than the ones achieved through the
application of all the interactional features, and any other subset of features
considered individually.
Turning to navigational query results shown in Figure 2, there may be some
surprising findings. First of all, the set All of all the features does not provide the
best performances, which are instead provided by the subset All Transactional.
This highlights the problem of classifying navigational and transactional queries
due to their similarity. Even if slightly, performances of the model improve by
excluding transactional features. The ineciency of these features can also be
514 V. Deufemia et al.

Fig. 2. Performances of the considered feature subsets for navigational queries

seen by considering the subset Transactional, which provides performances worse

than with other subsets. Poor results were also achieved from subsets of features
considered individually, except for Interaction, which provides acceptable per-
formances.
Finally, let us analyze performances related to transactional queries (Figure
3). In this context, the set All confirms to be more eective than the other
subsets. What we want to highlight here is the ecacy of transactional fea-
tures. The subset Transactional, in terms of performances, is second only to the
subset of all features, but it is close to it. By excluding such features in the
subset All Transactional, performances degrade providing results comparable
to Query+Search. The ineciency of pure interactional features with respect
to transactional ones also arises by a direct comparison: The subset Interaction
achieves performances comparable to Transactional, but far better than Inter-
actionTransactional, which excludes transactional features from Interaction.

Fig. 3. Performances of the considered feature subsets for transactional queries

An Analysis of Transactional Features. A substantial limit of traditional

approaches described in the literature is represented by the classification of trans-
actional queries. In this paper we propose some features that aim to characterize
transactional queries by analyzing scrolling behaviors, ajax requests, read text,
 Exploiting Interaction Features in User Intent Understanding 515

and presence of transactional terms in queries or URLs. The goal here is to

compare the eectiveness of query classification by applying all the features
(All ) with respect to applying a subset of them that excludes the transactional
features (All Transactional ). Referring to transactional queries, Table 1 shows
that there is a conspicuous improvement of the set All, especially when using
SVM. To be noted the precision value of 0.93, by far higher than the precision
measured through all the classifiers for the subset All Transactional.

Table 1. Precision, Recall, and F1-measure of the classification of transactional queries

with the feature sets All and All Transactional

Class Model Precision Recall F1-measure

SVM All 0.93 0.76 0.84
SVM AllTransactional 0.72 0.62 0.67
CRF All 0.78 0.74 0.76
CRF AllTransactional 0.85 0.50 0.63
LDCRF All 0.77 0.68 0.72
LDCRF AllTransactional 0.65 0.65 0.65

Moreover, transactional features are also useful to improve the quality of the
classification process of informational queries (Table 2), whereas they have no
eect on navigational ones (Table 3).

Table 2. Precision, Recall, and F1-measure of the classification of informational queries

with the feature sets All and All Transactional

Class Model Precision Recall F1-measure

SVM All 0.83 0.90 0.88
SVM AllTransactional 0.80 0.84 0.82
CRF All 0.86 0.92 0.89
CRF AllTransactional 0.76 0.87 0.81
LDCRF All 0.84 0.83 0.84
LDCRF AllTransactional 0.80 0.78 0.79

Table 3. Precision, Recall, and F1-measure of the classification of navigational queries

with the feature sets All and All Transactional

Class Model Precision Recall F1-measure

SVM All 0.82 0.90 0.86
SVM AllTransactional 0.87 0.90 0.89
CRF All 0.84 0.70 0.76
CRF AllTransactional 0.90 0.87 0.88
LDCRF All 0.59 0.67 0.62
LDCRF AllTransactional 0.77 0.80 0.79
516 V. Deufemia et al.

5 Conclusions

In this paper we have proposed a new model for user intent understanding in web
search. Assuming that the interactions of the user with the web pages returned
by a search engine in response to a query can be highly useful, in this research
we aimed at defining a new model, based on the results returned by the search
engine, on the interactions of the user with them, and with the web pages visited
by exploring them. By examining these interactions, we have produced a set of
features that are suitable for determining the intent of the user. Each feature
involves a dierent level of interaction between the user and the system: query,
search, and web pages (interaction). To simplify and make the process of classifi-
cation more ecient, we also adopt a simple one level taxonomy: informational,
navigational, and transactional. In addition to the set of all the proposed fea-
tures, during the testing phase we also considered some subsets corresponding to
features of individual interactional contexts and to their union or dierence, in
order to evaluate the eectiveness of the classification with respect to traditional
models, and to interactional, or transactional features.
Experimental results have highlighted the eectiveness of query classification
when applying both features representing interactions on web pages and those
representing interactions in the context of queries and results. This also arise in
the classification of transactional queries, further highlighting the eectiveness
of interactional features, and more important, of transactional features.

References
1. Agichtein, E., Brill, E., Dumais, S.: Improving web search ranking by incorporat-
ing user behavior information. In: Proceedings of the International Conference on
Research and Development in Information Retrieval, SIGIR 2006, pp. 1926. ACM
(2006)
2. Deufemia, V., Giordano, M., Polese, G., Tortora, G.: Inferring web page relevance
from human-computer interaction logging. In: Proceedings of the International
Conference on Web Information Systems and Technologies, WEBIST 2012, pp.
653662 (2012)
3. Guo, Q., Agichtein, E.: Beyond dwell time: estimating document relevance from
cursor movements and other post-click searcher behavior. In: Proceedings of the
International Conference on World Wide Web, WWW 2012, pp. 569578. ACM
(2012)
4. Kang, I., Kim, G.: Query type classification for web document retrieval. In: Pro-
ceedings of the Conference on Research and Development in Informaion Retrieval,
SIGIR 2003, pp. 6471. ACM (2003)
5. Lee, U., Liu, Z., Cho, J.: Automatic identification of user goals in web search. In:
Proceedings of the International Conference on World Wide Web, WWW 2005,
pp. 391400. ACM (2005)
6. Agichtein, E., Brill, E., Dumais, S., Ragno, R.: Learning user interaction models
for predicting web search result preferences. In: Proceedings of the International
Conference on Research and Development in Information Retrieval, SIGIR 2006,
pp. 310. ACM (2006)
 Exploiting Interaction Features in User Intent Understanding 517

7. Jansen, B.J., Booth, D.L., Spink, A.: Determining the user intent of web search
engine queries. In: Proceedings of the International Conference on World Wide
Web, WWW 2007, pp. 11491150. ACM (2007)
8. Tamine, L., Daoud, M., Dinh, B., Boughanem, M.: Contextual query classification
in web search. In: Proceedings of International Workshop on Information Retrieval
Learning, Knowledge and Adaptability, LWA 2008, pp. 6568 (2008)
9. Guo, Q., Agichtein, E.: Ready to buy or just browsing? detecting web searcher
goals from interaction data. In: Proceedings of the International Conference on
Research and Development in Information Retrieval, SIGIR 2010, pp. 130137.
ACM (2010)
10. Nielsen, J.: F-shaped pattern for reading web content (2006),
http://www.useit.com/articles/f-shaped-pattern-reading-web-content/
11. Broder, A.: A taxonomy of web search. SIGIR Forum 36, 310 (2002)
12. Cortes, C., Vapnik, V.: Support-vector networks. Mach. Learn. 20, 273297 (1995)
13. Laerty, J.D., McCallum, A., Pereira, F.C.N.: Conditional random fields: Proba-
bilistic models for segmenting and labeling sequence data. In: Proceedings of the
International Conference on Machine Learning, ICML 2001, pp. 282289 (2001)
14. Morency, L.P., Quattoni, A., Darrell, T.: Latent-dynamic discriminative models
for continuous gesture recognition. In: Proceedings of IEEE Conference Computer
Vision and Pattern Recognition, CVPR 2007, pp. 18 (2007)
15. Carmel, E., Crawford, S., Chen, H.: Browsing in hypertext: a cognitive study. IEEE
Transactions on Systems, Man and Cybernetics, Part A: Systems and Humans 22,
865884 (1992)
16. Morrison, J., Pirolli, P., Card, S.K.: A taxonomic analysis of what world wide
web activities significantly impact peoples decisions and actions. In: Extended
Abstracts on Human Factors in Computing Systems, CHI EA 2001. pp. 163164.
ACM (2001)
17. Sellen, A., Murphy, R., Shaw, K.: How knowledge workers use the web. In: Ex-
tended Abstracts on Human Factors in Computing Systems, CHI 2002, pp. 227
234. ACM (2002)
18. Rose, D., Levinson, D.: Understanding user goals in web search. In: Proceedings of
the International Conference on World Wide Web, WWW 2004, pp. 1319. ACM
(2004)
19. Guo, Q., Agichtein, E.: Towards predicting web searcher gaze position from mouse
movements. In: Proceedings of the International Conference on Human Factors in
Computing Systems, CHI EA 2010, pp. 36013606. ACM (2010)
20. Kelly, D., Teevan, J.: Implicit feedback for inferring user preference: a bibliography.
SIGIR Forum 37, 1828 (2003)
21. Guo, Q., Agichtein, E.: Exploring mouse movements for inferring query intent.
In: Proceedings of the International Conference on Research and Development in
Information Retrieval, pp. 707708. ACM (2008)
22. Lauer, F., Guermeur, Y.: MSVMpack: a multi-class support vector machine pack-
age. Journal of Machine Learning Research 12, 22692272 (2011)
 Identifying Semantic-Related Search Tasks
in Query Log

Shuai Gong1,3 , Jinhua Xiong1 , Cheng Zhang1 , and Zhiyong Liu1,2

1
Institute of Computing Technology, Chinese Academy of Sciences
2
State Key Laboratory of Computer Architecture
3
University of Chinese Academy of Sciences
{gongshuai,zhangcheng2}@software.ict.ac.cn, {xjh,zyliu}@ict.ac.cn

Abstract. Users often submit multiple related queries in order to ac-

complish one search task. Identifying search tasks faces two challenges:
1) Search tasks are often intertwined and may span from seconds to days.
2) Queries triggered by semantic-related search tasks may share few com-
mon terms or clicked documents. To address the challenges, we exploit
semantic features of named entities to improve semantic-related search
tasks identification. A novel approach to learning the semantic-related
distance function between pair-wise queries is proposed. The approach
uses categories of named entities as regularization, which reinforces that
queries containing entities from the same category more probably belong
to one search task. Finally, semantic-related search tasks are identified
by the hierarchical agglomerative clustering algorithm with the learned
distance function. Experiments show significant improvement of our ap-
proach over corresponding state-of-the-art ones.

Keywords: Query log analysis, User intent, Search task, Named entity.

1 Introduction

Detecting user intents in queries is crucial for search engines to better satisfy
users information need. Usually users may issue multiple related queries to ac-
complish one search task [7], e.g. multiple queries magnolia junior high and
chino high school may belong to the same search task for looking for schools
in the City of Chino. Identifying search tasks could help search engines to bet-
ter understand user intents. Many works have been proposed to identify search
tasks, but long-term and semantic-related search task identification has not been
well studied. Search tasks are often intertwined [11], and may span from seconds
to days [2]. For example, the search task trip planning can span many days.
During this period the user may issue many queries of other search tasks. To our
best knowledge, existing technologies do not work well to identify such long-term
search tasks [3,9,8]. Because of the sparsity of clicks and lexical features of query
log, queries triggered by the same semantic-related search task may share few
common terms or clicked documents. Identifying long-term and semantic-related
search tasks is dicult. Firstly, it is not straightforward how to exploit semantic

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 518525, 2013.
c Springer-Verlag Berlin Heidelberg 2013

 Identifying Semantic-Related Search Tasks in Query Log 519

information. It is also dicult to combine properly both semantic and lexical

features between pair-wise queries. Secondly, search tasks may span widely, so
annotators have to check all the queries issued by the same user, in order to la-
bel sucient semantic-related queries for learning a distance function. Labeling
large-scale queries is expensive, therefore the distance function between pair-wise
queries has to be learned from a small set of labeled data. It is prone to suer
the over-fitting problem.
To address the above challenges, this paper exploits semantic information
properly and learns an optimal distance function between pair-wise queries.
Firstly, this paper extends respectively named entities and their contexts in
user queries with semantic information. In most cases, queries containing named
entities of the same category tend to belong to the same search task, this paper
exploits the category of named entities to regularize the learning of the dis-
tance function. Semantic-related search tasks, i.e. query clusters, are identified
by the hierarchical agglomerative clustering algorithm with the learned distance
function.
The rest of the paper is structured as follows. Related work is discussed
in Section 2. Then we formalize basic concepts of identifying semantic-related
search tasks in Section 3. In Section 4, we introduce our approach to identify-
ing semantic-related search tasks. In Section 5, we present our experiments on
commercial search log. Finally, we conclude our work in Section 6.

2 Related Work
The related work can be organized into the following two categories:
Query intent analysis has attracted substantial attention. User queries
are often classified into three high-level intents: navigational, informational and
transactional intents [4]. Recently, many works [6,12] attempted to exploit named
entities to better understand user intent. Guo et al. [6] exploited the category
of named entities to classify user queries. Yin et al. [12] built a hierarchical
taxonomy of the generic search intents based on named entities. Recently the
research on query intent of multiple queries such as query refinements [10] and
search tasks [9] have also attracted much attention [5,1,8,2].
Search tasks identification was proposed by Jones et al. [9] in the setting
of queries sessions detection [3]. Jones et al. [9] determined the boundary of
search tasks between chronological queries with a decision tree. Donato et al. [5]
determined whether pairs of chronologically issued queries targeting the same
search task. This approach could not deal with intertwined search tasks well.
Aiello et al. [1] clustered the identified search tasks of dierent users into more
general topics by the approach proposed in [5]. But this paper aims to discover
more semantic-related search tasks accomplished by the same user. Lucchese et
al. [8] clustered user queries in a time-split session. They exploited both content
features and semantic features in a straightforward way. The dierent impor-
tance of terms and dependency among terms in queries is ignored. The semantic
of queries might deform. Sadikov et al. [10] clustered refinements of a query ac-
cording to dierent user intents. But the intertwined search tasks could not be
520 S. Gong et al.

eectively identified. Boldi et al. [3] introduced the query-flow graph to represent
query log. This model was used to separate the chronological queries into sets of
related queries, named query chain, with three categories of features. Kotov et
al. [2] modeled the user behaviors of cross-sessions search tasks. This work also
showed the importance of identifying long-term search tasks.

3 Problem Formalization

This section introduces the distance function and semantic-related features used
for search task identification.

3.1 Distance Function

Q denotes all the queries issued by a user. Search task T is a subset of Q. E de-
notes the named entities set collected from collaborative knowledge bases, such
as Wikipedia. Cat denotes the categories set of named entities. The category of
a named entity could be denoted as cat(e). A user query qi Q can be formal-
ized as a bi-tuple as qi =< ei , ci >, where ei denotes the entity contained in the
query qi and ci denotes the remaining terms of that query. q =< qi , qj > denotes
any pairs of queries issued by a user. The function f : Q Q [0, 1] denotes
semantic-related distance between pair-wise queries. Given labeled semantic-
related user query pairs, finding out the distance function f becomes the follow-
ing optimal problem:
l
1"
f = arg min [ V (f, xi , yi ) + f 2 ] (1)
f HK l i=1

where V is the loss function over the labeled data, such as squared loss for Reg-
ularized Least Squares, Hk is an Reproducing Kernel Hilbert space of functions
with kernel function K, l is the size of labeled queries pairs, and is the coef-
ficient of regularization on the ambient space. This paper defines the distance
function as a linear function about feature vector xq between pair-wise queries.
The solution of the minimization problem will be discussed in Section 4.

3.2 Entity Feature

The features between pair-wise queries can be split into two parts: entity features
and context features. The entity features used in the distance function are:
Semantic Relevance of Named Entity. To get the semantic relevance feature
between pair-wise named entities, it is necessary to use some external semantic
sources, such as Wikipedia. Therefore we crawled and indexed the web pages
from Wikipedia, built a search engine. The relevance between a named entity
and all indexed documents constitutes the semantic feature vector for that entity.
 Identifying Semantic-Related Search Tasks in Query Log 521

The semantic relevance between pair-wise named entities is cosine similarity

between the semantic feature vectors.
Context Cluster Relevance of Named Entity. Named entities occurring in
analogous contexts tend to target relevant tasks. But considering spelling errors
in user queries, directly using contexts of named entities will make the feature
vector space very sparse. So the contexts of entities are clustered. Each element
of a context cluster feature vector for each named entity is P (clui |e). The context
clusters relevance between pair-wise named entities is cosine similarity of context
cluster feature vector of each named entity.

3.3 Context Feature

The context of a query is the remaining part of the query except the named
entity. The semantic and lexical features of context are introduced as follows:

Context Semantic Relevance. This paper also uses the vector model to build
semantic feature vector for each context. Dierent with [8], this paper submits
each context as a whole to the search engine. The semantic relevance between
pair-wise contexts is cosine similarity between the semantic feature vectors.
Context Lexical Distance. As illustrated in [9,8], the queries containing com-
mon terms may target relevant search tasks. This paper uses both Jaccard index
on tri-grams and normalized Levenstein distance to represent lexical distance
between pair-wise contexts respectively, as suggested by the works [9,3].

4 Identifying Semantic-Related Search Tasks

This section introduces our approach for long-term Semantic-related search

Tasks Identification (STI ). The STI approach consists of two major steps:
1) Learning Distance Function: In the first step, STI learns an optimal dis-
tance function. The distance function will be used to measure similarity between
pair-wise queries when clustering user queries in the second step.
2) Clustering Semantic-Related Queries: At the second step, STI uses
the hierarchical agglomerative clustering algorithm with the learned distance
function to cluster user queries. The resulting query clusters are semantic-related
search tasks.

4.1 Learning Distance Function

As discussed in Section 3, learning a distance function can be transformed into
solving an optimal problem (Eq. 1). Because it is dicult to label large scale
search tasks, learning the distance function will suer the over-fitting problem.
Therefore, the entity category regularization is proposed.
522 S. Gong et al.

Entity Category Regularization. In most cases, queries issued by a user

containing named entities of the same categories tend to be in the same search
task, such as queries Toyota radios and Ford parts probably constitute the
search task for maintaining a car. So regularizing a distance function with cat-
egories of named entities does make sense. That is to say the semantic distance
between pair-wise queries containing named entities of the same category may
be close. The regularization of categories of named entities can be formalized as
follows:
N
1 "
(f (qi )2 g(qi )) < (2)
N i=1
where g(q) denotes whether the named entities contained in the query pair q
belong to the same category. N is the number of pair-wise queries.

Distance Function Optimization. When minimizing the objection function

(Eq. 1), the constraint of Eq. 2 must be satisfied. By leveraging the Lagrange
multiplier and using sum of squared loss as the loss function, the object function
can be reformulated as follows:
l N
1" "
f = arg min [ (yi f (qi ))2 + f 2 + (f (qi )2 g(qi ))] (3)
f HK l i=1 N i=1

where l is the number of labeled pair-wise queries, N is the total number of pair-
wise queries, yi denotes the labeled value of semantic-related distance function,
is the coecient of regularization in ambient space, and is the coecient of
same categories regularization. As above mentioned, the distance function is a
linear combination of dierent features. So Eq. 1 can be reformulated as follows:
l
1" T Xw]

w = arg min[ (yi wT xi )2 + wT w + wT X (4)
w l i=1 N

where w is the parameters of the distance function. X is a N D matrix, the

denotes the feature vector between i-th pair-wise queries which
i-th row of X
containing named entities of the same category. D is the dimension of feature
vector between pair-wise queries. Setting the gradient of Eq. 4 to zero gives
l T
(X T X + lI + X X)w = X T Y (5)
N
where X is a l D matrix, which is usually named as design matrix. The i-th
row of X is the feature vector between the i-th labeled pair-wise user queries. I
is the identity matrix. Solving the Eq. 5 will get the optimal w .

4.2 Clustering Semantic-Related Queries

The hierarchical agglomerative clustering algorithm is used to cluster semantic-
related queries with the leaned distance function. As illustrated in [10], the
 Identifying Semantic-Related Search Tasks in Query Log 523

complete-link similarity of two sets of queries works best in query clustering

setting. Therefore this paper uses the complete link similarity to merge dierent
query clusters. The algorithm works by picking each pair of query sets that have
the highest value of complete-link similarity. If the similarity is greater than ,
then merges the selected query sets. This proceeds until no more similar query
sets are left. The resulting query clusters are semantic-related search tasks.

5 Experiments
5.1 Experiment Setup
This section discusses our experiment data set, two extended state-of-the-art
approaches to be compared with our approach STI and evaluation measures.

Data Set. The 2006 AOL query log is used for experiment, in order to compared
with related approaches tested by the same log. This query log consists of about
20 million of web queries issued by more than 650,000 users over 3 months. In
order to learn the distance function, 200 pairs of queries random selected are
labeled as semantic-related and 200 pairs of queries random selected are labeled
as not-semantic-related. Named entities are collected from Wikipedia with the
list of * query such as list of states in U.S..

Extended State-of-the-Art Approaches. 1) Query Chain Identifica-

tion Boldi et al. [3] proposed an approach to identify the chains of queries,
which are similar to search tasks. They used the logistic regression to learn the
distance function between pair-wise queries. To compare with STI, we modify
the approach, named as mQCI. 2) Search Task Discovery in Sessions The
approach proposed in [8] was used to address task-based session discovery prob-
lem (TSDP ) in time-split sessions. It can not be used to identify long-term and
semantic-related search tasks directly. Therefore we modified the approach to
enable that all the queries issued by the same user will be included in for search
tasks identification, named as mTSDP.

Search Task Analysis. To compare dierent approaches, we use User Anno-

tation and three measures (Time Span, Cohesion and Coverage) for evaluation.

5.2 Evaluation
The user annotation results are presented in Table 1. Each evaluator is required
to rate each identified search task as 1(Poor), 2(Bad), 3(Acceptable), 4(Good)
or 5(Perfect) and indicates optionally whether an identified search task is under
clustered or over clustered. According to the result of user rating score, the
dierences among the three approached are statistically significant, and STI
performs the best. The column of the ratio of under clustered search tasks shows
that STI yields the least one. As showed by Fig. 1, the coverage of identified
search tasks by STI is much higher than both mQCI and mTSDP. The column
of the ratio of over clustered search tasks shows that STI yields the largest one.
524 S. Gong et al.

We now look at the reason that is likely to underlie the observed results.
Both mTSDP and mQCI perform better to aggregate the misspelling queries,
such as 800flowers and 1-800flower. But STI tends to yield long-term
and more semantic-related search tasks than others, such as magnolia junior
high and chino high school, which are not in the results of both mQCI
and mTSDP. This is consistent with what Fig.1 and Fig. 2 shows.

Table 1. User Annotations Result

1.0
STI
mQCI
mTSDP

0.8
Method Average Under Over
Rating Clustered Clustered

0.6
Coverage
STI 4.4 7% 10%

0.4
mQCI 2.6 18.9% 5%
mTSDP 3.8 17% 7%
0.2
0.0

0.2 0.4 0.6 0.8

Clustering Threshold

Fig. 1. Average Coverage

1.0

STI
mQCI
mTSDP
300

0.8
Cohesion & Coverage
# of Search Task

0.6
200

0.4

STI Cov.
STI Coh.
100

no Reg. Cov.
no Reg. Coh.
0.2

0.18
0.12
0.0
0

0 5 10 15 20 25 0.0 0.2 0.4 0.6 0.8 1.0

Days Clustering Threshold

Fig. 2. Time Span Distribution Fig. 3. Category Regularization Impact

Fig. 3 shows the impact of the entity category regularization, which has been
discussed in Section 4.1. When learned the distance function with entity category
regularization, the balance point of coverage and cohesion curves is 50% higher
than that of without regularization. With regularization, the coverage of the
identified search tasks is always higher. And when > 0.6, the coverage and
cohesion of identified search tasks with regularization are both higher than those
of without regularization. The category regularization helps to overcome the
over-fitting problem and reinforces that the queries containing named entities of
the same category probably belong to the same search task.
 Identifying Semantic-Related Search Tasks in Query Log 525

6 Conclusion

In this paper, we propose an approach to identifying long-term and semantic-

related search tasks in query log. A distance function between pair-wise queries
from the semantic-related point of view is learned with the entity category reg-
ularization. Then, the hierarchical agglomerative algorithm is used to cluster
semantic-related queries with the learned function. We compared our approach
with two state-of-the-art ones. The result shows our approach outperforms the
others significantly.
Our work can be further studied in two aspects. Firstly, how to represent
a search task is important. A better representation is very helpful for users to
understand the search task. Secondly, how to use the identified search tasks to
improve user search experience can also be studied, e.g. to recommend more
semantic-related queries to the user.

Acknowledgment. This work was supported by Foundation for Innovative Re-

search Groups of the NSFC(Grant No. 60921002) and the National Key Tech-
nology R&D Program(Grant No. 2011BAH11B02).

References
1. Aiello, L.M., Donato, D., Ozertem, U., et al.: Behavior-driven Clustering of queries
into topics. In: Proc. of the 20th CIKM, pp. 13731382 (2011)
2. Kotov, A., Bennett, P.N., White, R.W., et al.: Modeling and analysis of cross-
session search tasks. In: Proc. of the 34th SIGIR (2011)
3. Boldi, P., Bonchi, F., Castillo, C., et al.: The query-flow graph: model and appli-
cations. In: Proc. of the 17th CIKM, pp. 609618 (2008)
4. Broder, A.: A taxonomy of web search. SIGIR Forum 36, 310 (2002)
5. Donato, D., Bonchi, F., Chi, T., et al.: Do you want to take notes?: identifying
research missions in Yahoo! search pad. In: Proc. of the 19th WWW (2010)
6. Guo, J., Xu, G., Cheng, X., et al.: Named entity recognition in query. In: Proc. of
the 32nd SIGIR, pp. 267274 (2009)
7. Ji, M., Yan, J., Gu, S., et al.: Learning search tasks in queries and web pages via
graph regularization. In: Proc. of the 34th SIGIR, pp. 5564 (2011)
8. Lucchese, C., Orlando, S., Perego, R., et al.: Identifying task-based sessions in
search engine query logs. In: Proc. of the 4th WSDM, pp. 277286 (2011)
9. Jones, R., Klinkner, K.L.: Beyond the session timeout: automatic hierarchical seg-
mentation of search topics in query logs. In: Proc. of the 17th CIKM (2008)
10. Sadikov, E., Madhavan, J., Wang, L., et al.: Clustering query refinements by user
intent. In: Proc. of the 19th WWW, pp. 841850 (2010)
11. Spink, A., Park, M., Jansen, B.J., et al.: Multitasking during web search sessions.
Information Processing and Management 42(1), 264275 (2006)
12. Yin, X., Shah, S.: Building taxonomy of web search intents for name entity queries.
In: Proc. of the 19th WWW, pp. 10011010 (2010)
 Multi-verifier:
A Novel Method for Fact Statement Verification

Teng Wang1,2 , Qing Zhu1,2, and Shan Wang1,2

1 Key Laboratory of the Ministry of Education for Data Engineering and Knowledge
Engineering, Renmin University of China, Beijing, China
2 School of Information, Renmin University of China, Beijing, China

{wangteng,zq,swang}@ruc.edu.cn

Abstract. Untruthful information spreads on the web, which may mislead users
or have a negative impact on user experience. In this paper, we propose a method
called Multiveri f ier to determine the truthfulness of a fact statement. The basic
idea is that whether a fact statement is truthful or not depends on the information
related to the fact statement. We utilize a popular search engine to collect the
top-n search results that are most related to the target fact statement. We then
propose support score to measure the extent to which a search result supports the
fact statement, and we make use of credibility ranking to rank the search results.
At last, we combine the support score and credibility ranking value of a search
result to evaluate its contribution to a fact statement determination. Based on the
contributions of the search results, we determine the fact statement. Our proposals
are evaluated by experiments and results show availability and high precision of
the proposed method.

Keywords: Fact statements, Credibility ranking, Support score.

1 Introduction
The Web has become one of the most important information sources in peoples daily
life, however, much untruthful information spreads on the web. Untruthful information
may mislead users or affect user experiences, therefore, it is urgent to determine whether
a piece of information is trustful or not. Information is mainly loaded by sentences.
Sentences that state facts, rather than opinions, are called fact statements[8]. Generally
speaking, fact statements are in two categories: positive fact statements and negative
fact statements. In this paper, we mainly focus on positive fact statements, since for
any negative fact statement, there is a positive one corresponding to it, and truthfulness
determination of a negative fact statement can be inferred through determination of the
positive one. In this paper, when we say a fact statement, we mean a positive one.
A fact statement is either trustful or untruthful, depending on wether the contents
it carries is a objective fact or not. When a user determine the trustfulness of a fact
statement, he/she will specify some part(s) of the fact statement he/she is not sure about.
The part(s) are called the doubt unit(s) [8]. After the doubt unit(s) is/are specified, the
fact statement can be regarded as an answer to a question. Based on the number of
correct answers of the corresponding questions, fact statements can be divided into two

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 526537, 2013.
c Springer-Verlag Berlin Heidelberg 2013
 Multi-verifier: A Novel Method for Fact Statement Verification 527

categories: unique-answer fact statement and multi-answer fact statement. For example,
Obama is the American President is a trustful fact statement. If the doubt unit is Obama,
it is an answer of who is American President. Since the question has only one correct
answer, the fact statement is an unique-answer fact statement.
Most of the existing works utilize a popular search engine to collect the search re-
sults of the target fact statement, and the truthfulness of the target fact statement is
determined by analyzing the search results, e.g., the amount and the sentiments of the
search results are considered as the dominant factors in some works. However, this is
not desirable for the following reasons: (i). The amount of the search results is not
always reliable in trustfulness determination. Since not all search results support the
target fact statement. (ii).The sentiments of the search results do not always represent
their sentiments on the target fact statement, since not every search results theme hap-
pens to be the fact statement. Fig.1 shows a search result about the fact statement War-
ren Moon was born in 1956. Obviously, the theme is not the fact statement. In other
works, the trustful fact statement is identified from the alternative fact statements of a
target fact statement. Here, the alternative statements are the fact statements which are
answers to the question corresponding to the target fact statement. There are three lim-
itations in these works: (i). The doubt unit must be specified, otherwise the alternative
fact statements cant be found. (ii). If the doubt unit includes much information, it is
difficult, sometimes even impossible to find the proper alternative fact statements. (iii).
This method is not fit for the trustfulness determination of multi-answer fact statements,
since only one fact statement is picked out as the trustful fact statement.

Fig. 1. A example of search result

In order to solve these problems, we propose a method, called Multi-verifier, to de-

termine the truthfulness of a fact statement, which is suitable for the truthfulness deter-
mination of multi-answer fact statements. In summary, the contributions of this paper
are summarized as follows.

We propose a novel method, called Multi-verifier, to determine the truthfulness of

a fact statement. Multi-verifier can be used for the truthfulness determination of
multi-answer fact statements.
We measure the support score of a search result to the target fact statement. We
propose a way based on grammatical relationships between words to extract the
words, which make sense for the truthfulness determination, from the search result
and measure the support score of the search result to the fact statement based on
the words.
We rank the search results on the target fact statement based on credibility. We
consider the importance and the popularity of the sources from which the search
results are derived and get the importance ranking and popularity ranking of the
search results. Through merging the two rankings, the credibility ranking of the
search results is captured.
528 T. Wang, Q. Zhu, and S. Wang

We run a set of experiments and evaluate the method in terms of reasonability and
precision. The results show availability and high precision of our proposals.
For the rest of this paper, we first summarize the related works. Then, we describe the
method in details. Experiments and analysis are shown in section 4. At last, we conclude
this paper.

2 Related Works
We conduct a brief review of existing works on the following aspects: web page credi-
bility and statement trustfulness.
Generally speaking, credible sources are likely to present trustful information. Some
researchers study credibility of web pages, and they utilize web page features to deter-
mine whether a web page is credible or not, e.g., page keywords, page tittle, page style,
etc[1][2][3]. However, the trustfulness of web page content hasnt been considered in
these works. Some web pages which meet these features may present incorrect facts.
Other studies focus on spam web pages detection based on link analysis and content
for filtering out low quality pages[4][5]. However spam web pages are not equal to un-
truthful web pages and non-spam web pages may present incorrect facts. These studies
cant be used in fact statement determination directly.
Some researchers have started discussing the truthfulness of individual statements
in very recent years. [6] introduces system Honto?search1.0, which doesnt deter-
mine an uncertain fact directly, but provides users with additional data on which users
can determine the uncertain fact. Authors believe that if an uncertain fact is trust-
ful, the sentiments of the search results on the uncertain fact should be consistent.
Honto?search2.0[7] is the improved version of Honto?search1.0. This system has
two function: comparative fact finding and aspect extraction. Given an uncertain fact,
the system collects comparative facts and estimates the validity of each fact. Just for
checking the credibility of these facts in details, necessary aspects about the uncertain
fact are extracted. Scoring the facts and the aspects, users can get more information and
determine the uncertain fact. Verify[8] can determine the truthfulness of the fact state-
ment directly. Given a target fact statement, Verify finds the alternative fact statements
of the fact statement, ranks these fact statements, and chooses the fact statement at the
highest position as the trustful fact statement. In [7][8], doubt unit(s) of the fact state-
ment should be specified. In addition, when the doubt unit includes much information,
it is difficult, sometimes even impossible to find the proper alternative(comparative) fact
statements. Especially, for multi-answer fact statements, they cant work well.

3 Proposed Solution
The goal of this paper is to determine whether a fact statement is trustful or not. Firstly,
we submit the target fact statement to a popular search engine, and get the top-n search
results; secondly, the support scores of the search results of the fact statement are mea-
sured; thirdly, the credibility ranking of the search results is captured; at last, based
on the support scores and credibility ranking, the truthfulness of the fact statement is
determined.
 Multi-verifier: A Novel Method for Fact Statement Verification 529

3.1 Notation
We use f s to denote the target fact statement. After deleting stop words in f s, the rest
are the keywords of f s, which is represented by K. R denotes the collection of the search
results of f s. We set R = {r1 , . . . , rn }(n 2) and ri =< ti , ui , si > (1 i n). Suppose
that ri is extracted from web page pi , ti and ui denote the title and url of pi respectively.
And si is the snippet related to f s in pi . Kri is the collection of words which belong to
both K and ri . ci is used to denotes the shortest consecutive sentences including Kri in
ri . sup(ri , f s) [0, 1] is used to describe the support score of ri to f s.

3.2 Support Score Measurement

Not all words and sentences in a search result make sense for the trustfulness deter-
mination of the target fact statement. The words and sentences in the search result,
which make sense for the truthfulness determination of the fact statement, are called
necessary words and necessary sentences respectively. Necessary words are extracted
from necessary sentences, but not every word in necessary sentences is necessary.
Fig.1 shows a search result on Moon was born in 1956. Obviously, Born Harold War-
ren Moon, November 18, 1956... is necessary. In the sentence, Los Angeles, CA are not
necessary words.
In order to measure the support score of a search result more accurately, we extract
the necessary words from the search result, and compute the corresponding support
score based on the necessary words. The process to measure the support score is com-
posed of two stages: necessary words detection and support score computation.
Necessary Words Detection. Given a search result ri , the sentences in ci are necessary
sentences, since ci is the shortest consecutive sentences including Kri in ri . We analyze
the relationships of words in ci and find the necessary words of ri . In this paper, we use
Stanford Parser to find the grammatical relationships of words in ci .
Stanford Parser1 is used to find the grammatical relationships of words in a
sentence[10]. The grammatical relationship format is d.name(d.dependent, d.governor).
For example, amod(great, president) is a grammatical relationship. Here, great is a de-
pendent and president play a role of governor. amod is the name of the relation. The
relation means great is the modifier of president. There are 52 grammatical relationships
in Stanford typed dependencies. We divide the grammatical relationships provided by
Stanford Parser into two categories based on their importance to the skeleton of a sen-
tence: essential grammatical relationships, represented by Re and unessential grammat-
ical relationship, represented by Ro .
Ni is used to denote the collection of necessary words in ri . Di = {di1 , . . . , dim }(1 <
m) denotes the collection of the grammatical relationships between words in ci . We take
the following steps to find Ni .
Capture the shortest consecutive sentence ci in ri ;
Get the collection of the grammatical relationships between the words in ci by
Stanford parser and set Ni to Kri .
Use heuristic rules on the collection of grammatical relationships to find Ni .

1 http://nlp.stanford.edu/software/stanford-dependencies.shtml
530 T. Wang, Q. Zhu, and S. Wang

The heuristic rules are as follows: (i). If di j (1 j m) is essential, di j .dependent and

di j .governor are put into Ni . (ii). If di j (1 j m) isnt essential and di j .dependent
Ni , di j .governor are put into Ni .
Algorithm 1 shows the procedure of finding the necessary word collection of a search
result. The input is the search result ri =< ti , ui , si >, the set of words which belong to
the keyword set of f s and the search result, represented by Kri , the essential grammat-
ical relationship collection Re and the unessential grammatical relationship collection
Ro . While the output is the necessary word collection Ni . Firstly, ci is captured. If ti
contains the all words of Kri , ci is set to ti ; if not, ci is the shortest consecutive sentences
containing Kri in si (Line1-7). Secondly, the collection of grammatical relationships
between words in ci , called Di , is found out(Line 8). Thirdly, Ni is set to Kri (Line 9).
Fourthly, retrieving Di and applying the heuristic rules, new necessary words are put
into Ni (Line 12-17). The algorithm repeat the fourth step, until no new necessary word
is found(Line 10-18).

Algorithm 1. Finding necessary words

Input : ri , Re , Ro , Kri
Output: Ni
1 ci = si ;
2 if ti includes the all words in Kri then
3 ci = ti ;
4 else
5 for each consecutive sentences si j in si and si j includes the all words in Kri do
6 if len(si j ) " len(ci ) then
7 ci = si j ;

8 Di =Stanford Parser(ci );
9 Ni = Kri ;
10 while LastNi = Ni do
11 LastNi = Ni ;
12 for each di j Di do
13 if di j .name Re then
14 Ni di j .governor; Ni di j .dependent;
15 if di j .name Ro then
16 if di j .dependent Ni then
17 Ni di j .governor;

18 return Ni ;

Support Score Computation. We compute sup(ri , f s) based on the distance between

ri and f s, represented by dis(ri , f s). dis(ri , f s) can be computed based on the distance
of Ni and K. Inspired by edit distance, we define the distance between two sets.
Definition 1 (Set distance). The distance of two sets is the number of insertion or
deletion operations required to transform a set into another.
 Multi-verifier: A Novel Method for Fact Statement Verification 531

For example, S1 = {a, b, c} and S2 = {b, c, d} are two set. If a is deleted from S1 and d
is inserted in S1 , S1 = S2 . So the set distance of S1 and S2 is 2.
We use Sdis(Ni , K) to denote the set distance of Ni and K. Here, K is the collection of
keywords of f s. In order to get accurate Sdis(Ni , K), the words in K or Ni are stemmed
before computing Sdis(Ni , K). However, Sdis(Ni , K) cant be used to measure dis(ri , f s)
directly. The reason is that f s can be express in a different way. Given a sentence f s1 ,
which is different from f s but express the meaning of f s. f s can be transformed into
f s1 by word change. We suppose there are some words in f s cant be changed from
f s to f s1. And if these words changes, the meaning f s cant be kept. Here we use Du
to denote the collection of these unchangeable words in f s. Obviously, Du K. Since
word matching is a factor considered in returning search results, Du is composed of
the nouns and numbers in f s. Based on K and Du, we give out an equation to compute
dis(ri , f s).
Sdis(Ni ,K)+|Du||K|
|K| if Sdis(Ni , K)+ | Du | | K |> 0
dis(ri , f s) = (1)

0 if Sdis(Ni , K)+ | Du | | K |<= 0

Inspired by inverse hyperbolic function, we use the following equation to compute

sup(ri , f s). The value of is determined in our experiments.
2dis(r , f s)
2dis(ri , f s) 1
i
if dis(ri , f s) > 0
+1
sup(ri , f s) = (2)

1 if dis(ri , f s) = 0

In support score computation, we do not consider whether the search result oppose the
fact statement or not. Since the focus is the positive fact statements in this paper and
there are few search results explicitly opposing the target positive fact statement. In
addition, for the search results implicitly opposing the fact statement, corresponding
support scores may be smaller based on the equations or the search results will be
considered neutral.

3.3 Credibility Ranking Capture

If a fact statement is supported by more credible search results, the fact statement is
more likely to be trustful. Thus, the credibility of the search results is an important
factor in fact statement determination. Usually, credible search results are likely to be
presented by credible web sites/pages. The credibility of search results can be measured
by the credibility of the web sites/pages presenting them. Moreover, reputable web
sites/pages are likely to be credible[1][2][9]. If a web site/page is popular and important,
it could be reputable. The credibility of the web site/page can be measured based on its
importance and popularity.

Importance Ranking. The search results of a given fact statement display in a certain
order. We consider the order as the importance ranking of the search results. The order
is brought by the search engine which is used to find the search results. Besides, the
532 T. Wang, Q. Zhu, and S. Wang

importance of the sources of the search results are considered by the search engine. For
a fact statement f s and a search result ri on f s, Srank is used to denote the importance
ranking of the search results on f s, and Sranki is the position of ri in Srank. Given an-
other search result r j on f s, if Sranki < Srank j , ri is more important than r j . Note that,
we do not adopt the popular Pagerank for the following reasons: (i). The Pagerank val-
ues of web sites/pages cant be derived. (ii). Although the Pagerank levels are available,
the range of Pagerank level is too small ([0, 10]) to satisfy our requirements.

Popularity Ranking. We use Arank to denote the popularity ranking, Aranki represents
the Arank value of ri . We capture the popularity ranking of search results based on
Alexa2 ranking. Alexa ranking is a traffic ranking of each site based on reaches and
page views. It can be used to measure popularity of web sites. Given a fact statement
f s and a collection of search results R on f s. Alexai is used to denote Alexa ranking
value of the site from which ri is derived. Based on Alexai (1 i n), we can get the
popularity ranking of the search results in R. Based on Alexa ranking, Aranki can be
derived by the following equations.

IntervalAmid = min((max(Alexai ) Alexam ),

(3)
(Alexam min(Alexai ))) 2/(n 1)

Alexai min(Alexai )
Amidi = 1 + [ ] (4)
IntervalAmid

IntervalArank = (max(Amidi ) 1)/(n 1) (5)

Amidi 1
Aranki = 1 + [ ] (6)
IntervalArank
Since Alexa ranking is for all web sites, given Alexai and Alexa j , the gap between them
may be very large. We use the above four equations to map Alexai into range [1, n]. We
first introduce the median of Alexai represented by Alexam and get the interval ranking
Amid by Equation.3 and Equation.4, just for keeping the distribution of Alexai (1 i
n) unchanged. Then, we linearly map Amidi into [1, n] and get Aranki by Equation.5
and Equation.6.

Credibility Ranking. We use Crank to denote the credibility ranking of the search
results on the same fact statement. Cranki is the position of ri in Crank. Based on Sranki
and Aranki , Cranki can be computed. The follow equation describes how to compute
Cranki based on Sranki and Aranki .
Cranki = w1 Sranki + w2 Aranki (7)

Here, Srank and Arank play the same important roles in getting Crank. We set wi and
w2 to 0.5.

2 https://www.alexa.com
 Multi-verifier: A Novel Method for Fact Statement Verification 533

3.4 Fact Statement Determination

The search results on a given fact statement are in two categories: positive search results
and neutral search results. S pos denotes the collection of the positive search results and
Sneu is the collection of the neutral search results. Given a search result ri , if Kri Du
(Du is the unchangeable word collection of f s), ri S pos ; otherwise, ri Sneu . We
define the contributions of ri as the ratio of sup(ri , f s) and Cranki . The Pos is the sum
of contributions of the search results in S pos. Similarly, Neu is the sum of contributions
of the search results in SNeu .
We believe if a fact statement is trustful, Pos should be larger than Neu . Here,
Pos /Neu is the key factor in fact statement determination. We make use of a threshold
. If Pos /Neu is larger than or equal to , f s is trustful; otherwise, f s is untrustful.

Algorithm 2. Fact statement determination

Input : R = {r1 , . . . , rn }, Crank, sup(ri , f s), Kri (i = 1, . . . , n), Du,
Output: trustful or untruthful
1 Neu = Pos = 0;
2 for each ri R do
3 if Kri Du then
sup(ri , f s)
4 Pos + = CRank i
;
5 else
sup(ri , f s)
6 Neu + = CRanki ;
Pos
7 if Neu then
8 return trustful;
9 else
10 return untruthful;

Algorithm 2 shows the procedure of determining a fact statement. The input of the
algorithm is the collection of search results on the target fact statement R, the credibility
ranking of the search results Crank, the support scores of the search results to the fact
statement sup(ri , f s)(1 i n), the set of words which belong to the keyword set of
f s and the search result, represented by Kri , the unchangeable word collection of the
fact statement Du, and the threshold . While the output is the result of determination
(trustful or untruthful). First, the contributions form positive search results and neutral
search results are worked out respectively (Line 1-6). Secondly, the ratio between pos-
itive contributions and neutral contributions are computed. At last, the algorithm can
decide whether the fact statement is truthful or not (Line 7-10).

4 Experiments
We run a set of experiments to evaluate Multi-verifier in terms of precision and avail-
ability. First, we detect the distribution of the search results including the target fact
534 T. Wang, Q. Zhu, and S. Wang

statements; secondly, we evaluate the availability of the method to capture credibility

ranking; thirdly, we evaluate the availability of the method to compute support score; at
last, we measure precision of Multi-verifier while varying threshold and the number
of adopted search results n.
We make use of a synthetic dataset, which can be generated by the method intro-
duced in [8]. The dataset is composed of 50 trustful fact statements and 50 untruthful
fact statements. These fact statements are fetched from Trec20073. In the trustful fact
statements, 30 are unique-answer ones and the rest are multi-answer ones. Yahoo boss
2.04 is used to collect search results. Here, for each fact statement, top-150 search re-
sults are collected. 11 experienced users, who are graduate students and experienced
Internet users, help to mark the search results. If the search result includes the meaning
of the corresponding fact statement, it is marked as 0, otherwise, it is marked as 1. The
experiments run on an Intel Core 2 Quad 2.66Hz, windows 7 machine with 2GB main
memory.

4.1 Distribution of Search Results

In this experiment, we detect the average percentage of the search results which include
the meanings of the target fact statements. For untruthful fact statements, there are few
search results including them. Thus, this experiment is about trustful fact statements.
We use Pf to represent the average percentage of the search results including the target
fact statements. f su denotes the trustful unique-answer fact statements, f sm represents
the trustful multi-answer fact statements. Fig.2 shows Pf values when n changes. Here,
n is the number of the search results adopted for a fact statement. It can be seen from
the figure, both Pf for f su and Pf for f sm decrease with the increase of n. For a multi-
answer fact statement, there are more than one correct answers to the corresponding
question. Thus Pf for f su is always larger than the Pf for f sm . In addition, when n
increases, more search results which dont include the target fact statements come out.

4.2 Evaluation of Credibility Ranking

By computing the spearman correlation value between importance ranking(Srank) and

popularity ranking(Arank), we discuss whether the two rankings can replace each other
or not. Fig.3 shows the average correlative value of Srank and Arank, when n changes. It
can be seen that, the average correlative value is always smaller than 0.63; when n = 40,
the average correlative value is at the peak; when n # 40, the average correlation value
decreases with the increase of n. Therefore, Srank and Arank can not replace each other.
In the experiments, distributions of credibility ranking (Crank) and popularity rank-
ing (Arank) are detected. Fig.4 shows distributions of credibility ranking and popular-
ity ranking. The horizontal axis represents the positions of search results in importance
ranking(Srank), and the vertical axis represents the average credibility rank values or
popularity rank values of the search results at corresponding positions . It can be seen in
Fig.4, the average Arank values increase with the increase of Srank. However, changes
3 http://trec.nist.gov
4 http://boss.yahoo.com
 Multi-verifier: A Novel Method for Fact Statement Verification 535

of average Crank values are not so sharply. It means that the Arank or Srank can not
replace Crank, and some search results at higher positions in Srank or Arank may be at
lower positions in Crank. It is consistent with our observation.

0.85 0.64 120

0.8 fs 0.63 Crank

Average ranking
Correlation value
u
90 Arank
0.75 fsm 0.62
0.7 0.61
f

60
P

0.65 0.6
0.6 0.59 30
0.55 0.58
0.5 0.57 0
10 30 50 70 90 110 130 150 10 30 50 70 90 110 130 150 0 30 60 90 120 150
Top n search results Top n seacrh results Rank of search results

Fig. 2. Average percent on n Fig. 3. Correlation of Srank Fig. 4. Distribution of Crank

and Arank and Rank

4.3 Evaluation of Support Score

We evaluate the availability of the way to measure the distance of a search result and
the target fact statement and ascertain the optimal values of by experiments.
Fig.5(a) shows the average distance for trustful unique-answer fact statements, when
n changes. rdsu represents the average distance for these search results which do not
included the meaning of the corresponding fact statement. rdsi denotes the average dis-
tance for these search results including the meaning of the target fact statements. It can
be seen from the figure, rdsi is smooth. With n increases, rdsu increases. rdsu is always
greater than rdsi . The difference between rdsu and rdsi is larger than 0.3. Fig.5(b) shows
the average distance for trustful multi-answer fact statements, when n changes. rdsu and
rdsi represent the same as them in Fig.5(a). In this figure, rdsu is always above rdsi and
the smallest difference is larger than 0.4. Fig.5(c) shows the average distance for un-
truthful fact statements, when n varies. In Fig.5(c), rdsu is always above rdsi and the
smallest difference is greater than 0.8. From the experiments, we can know whether a
search result includes the target fact statement can be reflected by the distance.
In computation of support scores, plays an important role (Equation.2). We believe
if the value of makes the difference between the average support scores of the search
results including target fact statements and that of the other search results reach a max-
imum, the value is optimal. Fig.6 shows the difference on n = 150, when changes.
From this figure, when = 1.3, the difference reach the peak(0.2). It means 1.3 is the
optimal value of .

4.4 Evaluation of Truthfulness Determination

During the truthfulness determination of fact statements, the number of the search re-
sults adopted for a fact statement n and the threshold influence the precision of the
truthfulness determination. In the experiments, we measure the precision of Multi-
verifier, based on various n and values. Since is more than one, the range of
is from 1 to 5.
Fig.7 shows the precision on n when = 1.5. From this figure, when the n = 30,
the precision reaches the peak(0.9). When n " 30, the precision is enhanced with the
536 T. Wang, Q. Zhu, and S. Wang

1.3 1.3 1.5

0.25

Difference of support scores

1.2 1.2
Average distance

Average distance

Average distance
1.2
1.1 0.2
1.1
1 0.9
rd 1 rdsi
0.9 si 0.15
rd 0.6 rdsi
rdsu 0.9 su
0.8
0.3 rd 0.1
0.8 su
0.7
0.6 0.7 0 0.05
10 30 50 70 90 110 130 150 10 30 50 70 90 110 130 150 10 30 50 70 90 110 130 150 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9
Top n search results Top n search results Top n search results Vlaue of

(a) unique-answer (b) multi-answer (c) untruthful

Fig. 5. Average distance on n Fig. 6. Average differ-
ence on

0.92 1

0.9
0.9
Precision

Precision

0.88

0.86 0.8

0.84
0.7
0.82

0.8 0.6
10 30 50 70 90 110 130 150 1 1.5 2 2.5 3 3.5 4 4.5 5
Top n search results (=1.5) Value of (n=30)

Fig. 7. Precision on n Fig. 8. Precision on Fig. 9. Precision on n and

increase of n. The reason is that more search results are positive in top 30 search results,
and more positive search results are considered with the increase of n. When n # 30, the
precision decreases with the increase of n. The reason is that more neutral search results
are considered with increase of n when n # 30. Fig.8 shows the precision on when
n = 30. From this figure, When = 1.5, the precision reaches the greatest value(0.9).
When " 1.5, the precision is enhanced with the increase of . Since, when is
smaller, some untruthful fact statements are regarded as trustful ones. However, when
is greater, some trustful fact statements are regarded as untruthful ones. Thus, when
# 1.5, the precision decreases with the increase of . Fig.9 shows the precision on
and n. From this figure, the overall trend is that the precision increases and then
decreases as the increase of n and . Especially, when = 1.5 and n = 30, the precision
reach the peak(0.9).
The experiments show the method is available and accurate. Since the dataset in-
cludes many multi-answer fact statements, we can see that the method is also available
for multi-answer fact statements determination.
Since Honto?search can not determine a fact statement directly, it is not necessary
to run experiments to compare it with our method. Although Verify can directly de-
termine a fact statement, it is based on alternative fact statements and not suitable for
multi-answer fact statements. Thus, it is not necessary to compare it with our method.

5 Conclusion and Future Work

In this paper, we propose a method called Multi veri f ier to determine the truthfulness
of a fact statement. In Multi veri f ier, the search results on the fact statement,whose
truthfulness is needed to be determined, are found by a popular search engine. And the
 Multi-verifier: A Novel Method for Fact Statement Verification 537

support scores and the credibility ranking are considered in determining the truthfulness
of the fact statement. The experiments shows the method is powerful. However, our
focus is domain-independent fact statements, and the domain knowledge isnt used. In
the future, we will focus on the domain-dependent fact statements, we think that usage
of domain knowledge can bring higher precision. In addition, the quality of related
information is important for the truthfulness determination. We will try to find a method
to enhance the quality of related information on a fact statement.
Acknowledgments. This work is partly supported by the Important National Science
& Technology Specific Projects of China (Grant No.2010ZX01042-001-002), the Na-
tional Natural Science Foundation of China (Grant No.61070053), the Graduates Sci-
ence Foundation of Renmin University of China(Grant No.12XNH177) and the Key
Lab Found (07dz2230) of High Trusted Computing in Shanghai.

References
1. McKnight, D.H., Kacmar, J.: Factors and effects of information credibility. In: 9th Interna-
tional Conference on Electronic Commerce, pp. 423432. ACM Press, NewYork (2007)
2. Schwarz, J., Ringel Morris, M.: Augmenting web pages and search results to scport credibil-
ity assessment. In: International Conference on Human Factors in Computing Systems, pp.
12451254. ACM Press, NewYork (2011)
3. Lucassen, T., Schraagen, J.M.: Trust in Wikipedia: How Users Trust Information from an
Unknown Source. In: 4th ACM Workshop on Information Credibility on the Web, pp. 19
26. ACM Press, New York (2010)
4. Gyongyi, Z., Garcia-Molina, H., Pedersen, J.: Combating Web Spam with TrustRank. In:
30th International Conference on Very Large Data Bases, pp. 576587. Morgan Kaufmann,
San Fransisco (2004)
5. Ntoulas, A., Najork, M., Manasse, M., Fetterly, D.: Detecting spam web pages through con-
tent analysis. In: 15th International Conference on World Wide Web, pp. 8392. ACM Press,
NewYork (2006)
6. Yamamoto, Y., Tezuka, T., Jatowt, A., Tanaka, K.: Supporting Judgment of Fact Trustworthi-
ness Considering Temporal and Sentimental Aspects. In: Bailey, J., Maier, D., Schewe, K.-
D., Thalheim, B., Wang, X.S. (eds.) WISE 2008. LNCS, vol. 5175, pp. 206220. Springer,
Heidelberg (2008)
7. Yamamoto, Y., Tanaka, K.: Finding Comparative Facts and Aspects for Judging the Credi-
bility of Uncertain Facts. In: Vossen, G., Long, D.D.E., Yu, J.X. (eds.) WISE 2009. LNCS,
vol. 5802, pp. 291305. Springer, Heidelberg (2009)
8. Li, X., Meng, W., Yu, C.: T-verifier: Verifying Truthfulness of Fact Statements. In: 27th
International Conference on Data Engineering, pp. 6374. IEEE Press, New York (2011)
9. Fogg, B.J.: Persuasive Technology: Using Computers to Change What We Think and Do.
Morgan Kaufmann, San Francisco (2002)
10. Marneffe, M., Maccartney, B., Manning, C.: Generating typed dependency parsers from
pharse structure parses. In: The International Conference on Language Resources and Eval-
uation, ELRA, Luxembourg (2006)
 An Ecient Privacy-Preserving RFID
Ownership Transfer Protocol

Wei Xin, Zhi Guan, Tao Yang , Huiping Sun, and Zhong Chen

Institute of Software, EECS, Peking University, Beijing, China

MoE Key Lab of High Confidence Software Technologies (PKU)
MoE Key Lab of Network and Software Security Assurance (PKU)
{xinwei,guanzhi,ytao,sunhp,chen}@infosec.pku.edu.cn

Abstract. RFID technology is increasingly become popular in supply

chain management. When passing tags on to the next partner in the sup-
ply chain, ownership of the old partner is transferred to the new partner.
In this paper, we first introduce some existing RFID tag ownership trans-
fer protocols, then give the security and privacy requirements for such
kind of protocols, finally, we propose a novel RFID tag ownership trans-
fer protocol which supports constant-time authentication, and eectively
protects the privacy of the old tag owner and the new tag owner.

Keywords: RFID, Ownership transfer, security, privacy, supply chain.

1 Introduction
Radio Frequency Identification (RFID) technology represents a fundamental
change in the information technology infrastructure. RFID has many applica-
tions for both business and private individuals. Several of these applications will
include items that change owners at least once in its lifetime. The swapping and
resale of items is a practice that is likely to be popular in the future, and so
any item that depends on RFID for function or convenience should be equipped
to deal with change of ownership. Ownership transfer presents its own set of
threats, and therefore demands the attention of security researchers.
Generally speaking, a secure ownership transfer protocol should follow the
following assumptions: The old owner should not be able to access the tag after
the ownership transfer has taken place; The new owner should be able to perform
mutual authentication with the tag after the ownership transfer has taken place.
In this paper, we concentrate on designing a RFID ownership transfer proto-
col with high eciency and robust security and privacy properties. Our protocol
consists of three sub-protocols: an authentication protocol, an ownership trans-
fer protocol, and a secret update protocol. The rest of the paper is organized as
follows. In Section 2, we introduce a brief overview of some ownership transfer
protocols in RFID systems. Section 3 presents the security and threat model.
Section 4 describes the proposed ownership transfer protocol. Sections 5 demon-
strates security and privacy analysis. And finally, Section 6 concludes.

Corresponding author.

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 538549, 2013.
c Springer-Verlag Berlin Heidelberg 2013

 An Ecient Privacy-Preserving RFID Ownership Transfer Protocol 539

2 Related Work

One of the earliest protocols addressing ownership transfer with Trusted Third
Party (TTP) was proposed by Saito et al.[1]. The goal of the protocol is to
transfer ownership from one entity R1 to another R2 . The protocol steps are
summarized in Fig. 1. An adversary can commit de-synchronization attacks by
blocking the message from R2 to the tag. T T P and R2 have the new key, while
the tag keep the previous one. This can be prevented by T T P and R2 storing the
previous key. Unfortunately, in this case, since R2 knows s1 , backward security
is violated.

R1 R2
secret s1 secret s2 TTP
s1 secret s
s1 , s2

Es (s1 , s2 )

Tag
secret s, s1
Es (s1 , s2 )
if s1 is valid
update s1 to s2

Fig. 1. Saitos three party model

Kapoor and Piramuthu proposed a lightweight and secure ownership transfer

protocol with TTP [2]which was shown in Fig. 2. In order to provide backward
unlinkability, the new owner R2 does not have access to the key of the previous
R1 . In addition, R2 does not update the shared key unless he receives an ac-
knowledgment from T in order to solve de-synchronization issues. However, as
the scheme relies on symmetric primitives, it suers from linear time authenti-
cation and is not suited for resource limited tags.
Song put forward an ownership transfer protocol without TTP [3] in 2008.
The protocol consists of two sub-protocols: an ownership transfer protocol P1 ,
and a secret update protocol P2 . P1 transfers ownership from one owner R1
to another R2 and the second stage R2 updates the secret. The protocol P1
was shown in Fig. 3. The protocol suers from linear complexity problems that
tag authentication is linear in the number of tags. Another vulnerability is de-
synchronization problems when an adversary blocks the last communication M 3
between R2 and the tag.
Elkhiyaoui et al. proposed an ownership transfer protocol without TTP [4](see
Fig. 4) achieving constant time authentication and issuer verification. Elkhiyaouis
protocol violates tag owner privacy that any legitimate owner which has been the
owner of a specific tag is able to trace it by computing pairing of e(vi , g2 ).
540 W. Xin et al.

Tagi R1 R2 TTP
ti , s 1 , s 2 r1 , s1 r2 , s2 ti , rj , si

NP ||EN P ti s1
(s2 )
NT ||Hti NT (s2 NP )
Er1 (s1 )
NP ||Er2 NP (s2 r2 )
Hr2 (s2 NP )
NR2 ||Es 2 (NR2 )
NT ||HNT s2 (NR2 s2 )

Fig. 2. Protocol of Kapoor and Piramuthu

R1 R2 Tag
t, s, t , s t, s t, s
r1
r2 {0, 1}l
M1 t r2
M1 , M2 M2 ft (r1 r2 )
RT , r1 , M1 , M2
find (t,s) s.t.
M2 =ft (r1 M1 t)
r2 = M1 t
M3 s (r2 2l )
s s
t t
s (s1 4l ) r1
(t 4l ) r2
t h(s) M3 , (t, s),Info (t, s),Info
M3
s M3 (r2 2l )
if h(s) == t
t h((s 4l ) r1
(t 4l ) r2 )

Fig. 3. Protocol of Song

Tag R2 R1

N
NT , i , ci
N, NT , i , ci , Av
ref R1 , ref V Tag authenti
i+1 , ci+1 Issuer verification

Fig. 4. Protocol of Elkhiyaoui

3 Security and Threat Model

3.1 Privacy Issues
RFID privacy is one of the areas which are most discussed in recent years. There
are some common notions of privacy in RFID system:
1. Anonymity: an adversary A should not be able to disclose the identity of
tags he reads from or writes into.
 An Ecient Privacy-Preserving RFID Ownership Transfer Protocol 541

2. Untraceability: an adversary A should not be able to trace or track the

person(product) attached with tags by a fake reader. Untraceability contains
a special case called backward untraceability, which implies that an adversary
A should not be able to to trace past transactions between an owner and a
tag, even capturing the tags secret information.
A great number of privacy-preserving RFID protocols have been proposed in the
literature. Tag anonymity corresponds to anonymization protocol, one trivial ap-
proach is proposed by Sarma named hash-lock[5], using meta-ID instead of real
ID. Untraceability implies that the tag response dierently upon receiving the
same request. The approach to reach untraceability is either based on the psedu-
dorandom numbers or the random noise [6,7,8]. The Ohkubo-Suzuki-Kinoshita
protocol (OSK) [9,10] made backward privacy possible, the scheme relies on the
use of two one-way hash functions. Variants of the OSK scheme proposed in [11],
making it resistant to replay attacks. An further improvement to the OSK [12]
was proposed by Berbain using a pseudorandom number generator and a uni-
versal family of hash functions, moreover, they provided a provable secure proof
under standard model. Recently, Ma [13] gave a simpler one of [12] using tags
only equipped with pseududorandom generator.
In addition to the common notions of privacy, there are two extra privacy
issues dedicated for ownership transfer protocols listed as follows:
1. New owner privacy: Once ownership of a tag has been transferred to a new
owner, only the new owner should be able to identify and control the tag.
The previous owner of the tag should no longer be able to identify or trace
the tag.
2. Old owner privacy: When ownership of a tag has been transferred to a new
owner, the new owner of a tag should not be able to trace past interactions
between the tag and its previous owner.

3.2 Attacks on Ownership Transfer Protocols

An adversary A of the system is a malicious entity whose aim is to perform
some attacks, either through the wireless communications between readers and
tags (e.g., eavesdropping), or on the RFID devices themselves (e.g., corruption
a device and getting all the information stored on it). We mainly consider the
following attacks:
1. Replay attacks: Attackers intercept a valid response emitted by a tag and
retransmit it to a legitimate reader.
2. Man-in-the-middle attacks: In this case, an attacker is able to insert or mod-
ify messages exchanged by legitimate principals without detection.
3. Eavesdropping: Attackers listen passively to messages exchanged by legiti-
mate principals and are able to decode them.
4. Denial-of-service (DoS): The attacker disturbs or impedes communications
between principals.
542 W. Xin et al.

5. De-synchronization: Attackers makes the owner and the tag no longer be

able to successfully communicate with each other because they are using
dierent keys.

4 RFID Protocol for Tag Ownership Transfer

Our protocol works under the following assumptions. A T T P maintains a secure
database of information for all the readers and the tags that it owns, a reader
have significantly greater processing ability than a tag. Each tag has a rewritable
memory that may not be tamper-resistant, can generate pseudorandom numbers.
We use Songs protocol and Elkhiyaouis protocol as our main references,
Songs scheme suers from linear complexity problems that tag authentication
is linear in the number of tags, we address it by using Elgamal re-encryption
mechanisms introduced by Elkhiyaoui. However, Elkhiyaouis protocol have the
vulnerabilities that any legitimate owner which has been the owner of a specific
tag is able to trace it. We solve this problem by updating the index I in ownership
transfer process. In addition, we use only pseudorandom number generator at
the tag side which is well suited for resource limited passive RFID tags. To
sum up, we try to design a ownership transfer protocol satisfying the following
properties:
1. Constant-time for tag ownership transfer.
2. Old owner privacy, new owner privacy.
3. Issuer verification.
4. Resist common RFID attacks such as replay attacks, de-synchronization etc.

4.1 Pseudorandom Number Generators

A pseudorandom generator is a deterministic algorithm that receives a short
truly random seed and stretches it into a long string that is pseudorandom.
Stated dierently, a pseudorandom generator uses a small amount of true ran-
domness in order to generate a large amount of pseudorandom.
Let g : {0, 1} {0, 1}2 and g = (g1 , g2 ). g1 and g2 respectively map the
input of g to the first n output bits of g and the last n output bits of g. We take
g as the pseudorandom generator in the tag.

4.2 Protocol Description

We introduce a novel RFID ownership transfer protocol with high eciency
and robust security and privacy properties. Our protocol consists of three sub-
protocols: an authentication protocol, an ownership transfer protocol, and a
secret update protocol. Before introducing the authentication process and own-
ership transfer process, we first do some initiation works including setup and tag
Initialization.
Setup: The issuer I outputs (q, G1 , G2 , GT , g1 , g2 , e), where G1 , G2 , GT are
subgroups of prime order q, g1 and g2 are random generators of G1 and G2
 An Ecient Privacy-Preserving RFID Ownership Transfer Protocol 543

respectively, and e : G1 G2 GT is a bilinear pairing. I chooses x Zq and

computes g2x . Is secret key is sk = x and his public key is g2x . The system
supplies each owner with his secret key sk = i and his public key pk = g1i .
Tag Initialization: A tag T is initialized by the issuer I. I picks a random
number t Fq . Using a cryptographic hash function H : Fq G1 , I computes
h = H(t), then computes u0 = 1 and v0 = hx , and stores Ie0 = (u0 , v0 ) into the
tag, I also choose a random s R {0, 1} as the initialized state for pseudoran-
dom number generator in the tag. Before accepting the tag, the reader checks
the equation: e(h, g2x ) = e(I, g2 ), in which I = hx . If true, it implies that T is
actually issued by I.
Authentication Protocol: Our authentication protocol is based on Mas pro-
posal [13] using only pseudorandom number generator as its cryptographic prim-
itive, we extend the protocol from one way tag authentication protocol to mutual
authentication protocol. According to the proof of Ma, the authentication proto-
col satisfies backward security. We also use idea of re-encryption in Elkhiyaouis
protocol. The whole process is summarized in Fig. 5.

R Tag
I, h, s, ID, pk, sk Ie , s
NR
k=g2 (s)
M1 =g1 (k NR )
Ie , M1 s=g1 (s)
I=Decsk (Ie )
find (I, h, s, ID) by I s.t.
M1 =g1 (g2 (s) NR )
e(h, g2x ) = e(I, g2 )

Ie = Encpk (I)
s=g1 (s)
M2 = g1 (s NR )
Ie s, M2
check M2 =g1 (s NR )
if true,Ie = Ie

Fig. 5. Authentication Process

1. The reader R initiates the protocol by sending a random number to the tag.
2. Upon receiving a challenge NR from R, the tag T derives two values, a new
secret k = g2 (s) and a new state s = g1 (s). After updating the state to g1 (s),
T sends Ie and M1 = g1 (k NR ) to R.
3. On receipt of the messages, R first decrypts Ie using his secret key as
I=Decsk (Ie ), then find (I, h, s, ID) in database using index I and satisfying
both M1 =g1 (g2 (s)NR ) and e(h, g2x ) = e(I, g2 ). If M1 =g1 (g2 (s)NR ) is not
true, for j = 1 to , R judges whether M1 equals g1 (g2 (g1j (s)) NR ). If true,
update s to g1j (s), otherwise, reject. If e(h, g2x ) = e(I, g2 ) is not true, reject
as well. After that, R encrypts I by his public key as Ie = Encpk (I) and up-
dates the state of T to s=g1 (s). In our proposal, we use Elgamal encryption
544 W. Xin et al.

primitive to encrypt I, so Ie is dierent each time according to the security

properties of the scheme. Finally R sends Ie s and M2 = g1 (s NR ) to T .
4. When T receives Ie s and M2 , T first verifies whether M2 = g1 (s NR ),
if true, then update Ie to Ie s s. Otherwise, T do not update Ie .

Ownership Transfer Protocol: In this protocol, the new owner of a tag first
requests the T T P for ownership of the tag. If the request is valid, the T T P then
transfers all the information related to the tag to the new owner via a secure
channel. The ownership transfer protocol is based on our authentication proto-
col, since the authentication protocol satisfies backward privacy, the ownership
transfer protocol follows old owner privacy innately. However, we have to up-
date the tag state to protect new owner privacy. In addition, the index I must be
changed in order to achieve untraceability. The protocol steps are summarized
in Fig. 6 and described as follows:

1. At first, the old owner R1 sends IDR1 to the new owner R2 through a secure
communication channel(omitted in Fig. 6 for better representing).
2. The new owner R2 initiates the protocol by sending a random number NR
to the tag.
3. Upon receiving a challenge NR from R2 , the tag T derives two values, a new
secret k = g2 (s) and a new state s = g1 (s). After updating the state to g1 (s),
T sends Ie and M1 = g1 (k NR ) to R2 .
4. R2 then forwards Ie , M1 = g1 (k NR ), NR , IDR1 and IDR2 to T T P . In
our protocol, we use issuer as our T T P .
5. On receipt of the messages, T T P first decrypts Ie using R1 s secret key as
I=DecskR1 (Ie ), then find (I, h, s, ID) in database using index I and satisfy-
ing both M1 =g1 (g2 (s) NR ) and e(h, g2x ) = e(I, g2 ). If M1 =g1 (g2 (s) NR )
is not true, T T P tries to find g1j (s) that M1 equals g1 (g2 (g1j (s)) NR ) for
j = 1 to and update s to g1j (s). After that, T T P picks a new random
number t Fq and computes h = H(t ), u1 = g1r1 , v1 = hx g1r1 ( is the
secret key of R2 ) and let I = hx , Ie = (u1 , v1 ). Then, T T P updates the
state of T to s=g1 (s). Finally T T P sends inf o(ID), s, I , h , Ie , M2 to R2 .
6. R2 stores inf o(ID), s, I , h and forwards Ie s and M2 to T .
7. When T receives Ie s and M2 , T first verifies whether M2 = g1 (s NR ),
if true, then update Ie to Ie s s. Otherwise, T do not update Ie .

Secret Update Protocol: The main purpose of secret update protocol is to

establish new secrets to protect the new owners privacy. More specifically, we
need to update the state s of the tag, our protocol is also based on authentication
protocol, the whole process is summarized in Fig. 7 and described as follows:

1. The reader R initiates the protocol by sending NR and M1 =g1 (g2 (s) NR )
to the tag T .
2. Upon receiving a challenge NR and M1 from R, the tag T computes k = g2 (s)
and verify M1 =g1 (k NR ). If true, computes M2 =g2 (k NR ) and update
the tag state to s=g1 (s NR ), then sends Ie k and M2 to R.
 An Ecient Privacy-Preserving RFID Ownership Transfer Protocol 545

TTP
R2 Tag
I, h, s, I , h
I , h , s, ID, pk, sk Ie , s
ID, x, g2x
NR
k=g2 (s)
M1 =g1 (k NR )
Ie , M1 s=g1 (s)
NR , Ie , M1 , IDR1 , IDR2
I=DecskR1 (Ie )
find (I, h, s, ID) by I s.t.
M1 =g1 (g2 (s) NR )
e(h, g2x ) = e(I, g2 )

generate new I , h
Ie = (u1 , v1 )
s=g1 (s)
M2 = g1 (s NR ) inf o(ID), s, I , h , Ie , M2
Ie s, M2
check M2 =g1 (s NR )
if true,Ie = Ie

Fig. 6. Ownership transfer Protocol

R Tag
I, h, s, ID, pk, sk Ie , s

generate NR
NR , M 1
M1 =g1 (g2 (s) NR )
k=g2 (s)
verify M1 =g1 (k NR )
M2 =g2 (k NR )
s=g1 (s NR )
Ie , M2
I=Decsk (Ie )
find (I, h, s, ID) by I s.t.
M2 =g2 (g2 (s) NR )
e(h, g2x ) = e(I, g2 )

s=g1 (s NR )

Fig. 7. Secret Update Protocol

3. When R receives Ie k and M2 , R first get Ie by computing Ie k g2 (s),

then decrypts Ie using his secret key as I=Decsk (Ie ), then find (I, h, s, ID) in
database using index I and satisfying M2 =g2 (kNR ) and e(h, g2x ) = e(I, g2 ).
If true, update the tag state s to s=g1 (s NR ).

5 Security Analysis
5.1 Security Issues
In this section, we mainly discuss some attacks that are commonly used by ad-
versaries against RFID systems and present how our authentication protocol to
546 W. Xin et al.

resist these attacks, including Replay (RA) attacks, man-in-the-middle (MITM)

attacks, backward traceability (BT) attacks, Denial of Service (DoS) attacks. In
addition, we also do the analysis on how to protect new owner privacy and old
owner privacy in our ownership transfer protocol. The security analysis of secret
update protocol is just as authentication protocol, so we do not analyze it here
due to repetitiveness.
Replay Attack. All communicating messages between the reader and the tag
always contain a dynamically generated fresh random number. Adversaries are
unable to pass the authentication scheme simply by replaying messages.
Man-in-the-Middle Attack(MITM). Since adversaries are unable to launch
replay attacks and they do not have the tag state s, so they cannot pass our
authentication scheme by masquerading as a reader or tag in an MITM attack.
Backward Traceability. Backward security refers to the scenario that even if
the secret of T is compromised, the adversary cannot trace past transactions. In
our protocol, backward security is guaranteed because the state s of T is updated
each time using PRNG, according to the conclusion of Berbain et al. [14], the
previous states of T are not computable.
DoS/Synchronization Problem. We address the DoS problem in the follow-
ing way: Suppose that an adversary prevents message M1 from reaching R. Then
T will update its state s, R will not. However, R knows the previous value of
s, and can use g1j (s) to recover synchronization with T , j is limited to chains of
length . Our protocol is more eective than songs proposal, because we use
index I for locating the tuple in stead of searching it in the whole database.
Note that the secret update protocol is also vulnerable to de-synchronization
attacks if the adversary prevent M2 from reaching R. However, R can compute
g1 (s NR ) itself to recover synchronization with T .
Old Owner Privacy. Old owner privacy implies that the new owner should not
be able to trace past interactions between the tag and its previous owners. That
property is guaranteed by backward traceability of the authentication protocol
that the state s of T is updated in each session, and the previous states of T are
not computable..
New Owner Privacy. New owner privacy implies that the old owner can-
not trace the tag. We use Elgamal re-encryption mechanisms introduced by
Elkhiyaoui to encrypt the tag index I in each session, However, in Elkhiyaouis
scheme, I keeps unchanged which can be used for adversary to trace the tag.
Therefore, in our proposal, the T T P changed I when ownership transfer hap-
pens. In addition, we use secret update protocol to change the state of the tag
into a new one, which cannot be inferred by the old owner if he does not obtain
the values of NR queried by the new owner. Under such assumption, the privacy
of new owner is protected.
We compare our proposed scheme with other ownership transfer schemes in
terms of the aspects as mentioned above, which is shown in Table 1.
 An Ecient Privacy-Preserving RFID Ownership Transfer Protocol 547

Table 1. Security features comparison

Security
Schemes RA MITM DoS BT NP OP
Saito [1] X O O X O X
Kapoor [2] O O O O O O
Kulseng [15] O X X X * X
Osaka [16] O O X X * X
Fouladgar [17] X X O X O X
Song [3] X O X O * O
Elkhiyaoui [4] O O O X X X
ours O O O O * O
O: success
X: failure
*: success under an assumption

Table 2. Performance comparison

Schemes Tag Reader TTP/Database

Saito TE 0 TE
Kapoor 2TE +2TH +2TRNG 3TE +2TH +TRNG 3TE +2TH +2TRNG
Kulseng 4TRNG +2TH 3TH +3TRNG TH +5TRNG
Osaka 2TH TRNG 3TE +TH
Fouladgar 2TH +TRNG 0 2TH
Song 3TH +TRNG TRNG 2TH
2TE +3TH +TRNG
Elkhiyaoui TRNG +3TH 0
+7TP O + 2Texp
ours 4TRNG TRNG 2TE + 4TRNG +2TP O

5.2 Performance

In this section, we analyze our protocol in computation costs and proves our
scheme is viable for lightweight passive RFID tags. We will mainly focus on the
performance characteristics of ownership transfer protocol. TE denotes the time
for one encryption or decryption; TRN G is the time to generate a random num-
ber; TH is the time for one hash function; TP O stands for the time of pairing
operation; and Texp denotes the time of exponentiation operation. Table 2 com-
pares the performance characteristics of our ownership transfer protocol with
the other proposed schemes.
To notice that, we do not consider database searching process in the Table
2. Actually, Osakas protocol, Fouladgars protocol and Songs protocol do not
support index, so the computation costs are much more than what we have
listed in table 2. Although Kulsengs protocol supports index, the index may
cause de-synchronization problem (index updates asynchronously between the
reader and the tag), the reader still has to search the whole database when
it happens. We adopt Elkhiyaouis scheme to use re-encryption primitive to
548 W. Xin et al.

encrypt the same index in each session. It eliminates de-synchronization problem,

meanwhile, increases the computation loads on the reader. Saitos protocol and
Kapoors protocol use symmetric encryption/decryption method at the tag side
which is not suitable for resource limited RFID tags.

6 Conclusion

In this paper, we presented a novel RFID ownership transfer protocol to ad-

dress security and privacy issues related to RFID ownership transfer in sup-
ply chains. Our proposal provides following advantages: First, we use Elgamal
re-encryption mechanisms to solve linear complexity problem that authentica-
tion takes a linear time in the number of tags. Second, at the tag side, we use
only pseudorandom number generator as its cryptographic element which is well
suited for lightweight passive RFID tags. Third, we protect old owner privacy
by adopting backward security authentication protocol and preserve new owner
privacy by using secret update protocol. Our analysis show that our proposed
protocol provides stronger security robustness and higher performance eciency
in comparison with existing solutions.

Acknowledgments. This work is partially supported by the HGJ National

Significant Science and Technology Projects under Grant No. 2012ZX01039-004-
009, Key Lab of Information Network Security, Ministry of Public Security un-
der Grant No.C11606, the National Natural Science Foundation of China under
Grant No. 61170263.

References

1. Saito, J., Imamoto, K., Sakurai, K.: Reassignment scheme of an RFID tags key for
owner transfer. In: Enokido, T., Yan, L., Xiao, B., Kim, D.Y., Dai, Y.-S., Yang, L.T.
(eds.) EUC Workshops 2005. LNCS, vol. 3823, pp. 13031312. Springer, Heidelberg
(2005)
2. Kapoor, G., Piramuthu, S.: Single RFID Tag Ownership Transfer Protocols. IEEE
Transactions on Systems, Man, and Cybernetics, 110 (2011)
3. Song, B., Mitchell, C.J.: Rfid authentication protocol for low-cost tags. In: WISEC,
pp. 140147 (2008)
4. Elkhiyaoui, K., Blass, E.-O., Molva, R.: ROTIV: RFID Ownership Transfer with
Issuer Verification. In: Juels, A., Paar, C. (eds.) RFIDSec 2011. LNCS, vol. 7055,
pp. 163182. Springer, Heidelberg (2012)
5. Sarma, S., Weis, S., Engels, D.: Radio-Frequency Identification: Security Risks and
Challenges. Cryptobytes, RSA Laboratories 6(1), 29 (2003)
6. Hopper, N.J., Blum, M.: Secure human identification protocols. In: Boyd, C. (ed.)
ASIACRYPT 2001. LNCS, vol. 2248, pp. 5266. Springer, Heidelberg (2001)
7. Juels, A., Weis, S.A.: Authenticating Pervasive Devices with Human Protocols. In:
Shoup, V. (ed.) CRYPTO 2005. LNCS, vol. 3621, pp. 293308. Springer, Heidelberg
(2005)
 An Ecient Privacy-Preserving RFID Ownership Transfer Protocol 549

8. Bringer, J., Chabanne, H., Emmanuelle, D.: HB++ : a Lightweight Authentication

Protocol Secure against Some Attacks. In: IEEE International Conference on Per-
vasive Services, Workshop on Security, Privacy and Trust in Pervasive and Ubiq-
uitous Computing SecPerU 2006, Lyon, France. IEEE Computer Society (June
2006)
9. Ohkubo, M., Suzuki, K., Kinoshita, S.: Cryptographic Approach to Privacy-
Friendly Tags. In: RFID Privacy Workshop. MIT, Massachusetts (2003)
10. Ohkubo, Suzuki, K., Kinoshita, S.: Ecient Hash-Chain Based RFID Privacy Pro-
tection Scheme. In: International Conference on Ubiquitous Computing Ubicomp,
Workshop Privacy: Current Status and Future Directions, Nottingham, England
(September 2004)
11. Avoine, G., Dysli, E., Oechslin, P.: Reducing time complexity in RFID systems. In:
Preneel, B., Tavares, S. (eds.) SAC 2005. LNCS, vol. 3897, pp. 291306. Springer,
Heidelberg (2006)
12. Berbain, C., Billet, O., Etrog, J., Gilbert, H.: An ecient forward private rfid
protocol. In: ACM Conference on Computer and Communications Security, pp.
4353 (2009)
13. Chang-She, M.: Low cost rfid authentication protocol with forward privacy. Chinese
Journal of Computers 34(8) (2011)
14. Berbain, C., Billet, O., Etrog, J., Gilbert, H.: An Ecient Forward Private RFID
Protocol. In: Al-Shaer, E., Jha, S., Keromytis, A.D. (eds.) Conference on Computer
and Communications Security ACM CCS 2009, Chicago, Illinois, USA, pp. 4353.
ACM Press (November 2009)
15. Kulseng, L., Yu, Z., Wei, Y., Guan, Y.: Lightweight mutual authentication and
ownership transfer for rfid systems. In: INFOCOM, pp. 251255 (2010)
16. Osaka, K., Takagi, T., Yamazaki, K., Takahashi, O.: An ecient and secure RFID
security method with ownership transfer. In: Wang, Y., Cheung, Y.-m., Liu, H.
(eds.) CIS 2006. LNCS (LNAI), vol. 4456, pp. 778787. Springer, Heidelberg (2007)
17. Fouladgar, S., Afifi, H.: A simple privacy protecting scheme enabling delegation
and ownership transfer for rfid tags. JCM 2(6), 613 (2007)
 Fractal Based Anomaly Detection
over Data Streams

Xueqing Gong1 , Weining Qian1 , Shouke Qin2 , and Aoying Zhou1

1
Software Engineering Institute, East China Normal University, Shanghai, China
2
Baidu Inc., Beijing, China
{xqgong,wnqian,ayzhou}@sei.ecnu.edu.cn, [email protected]

Abstract. Robust and ecient approaches are needed in real-time mon-

itoring of data streams. In this paper, we focus on anomaly detection on
data streams. Existing methods on anomaly detection suer three prob-
lems. 1) A large volume of false positive results are generated. 2) The
training data are needed, and the time window of appropriate size along
with corresponding threshold has to be determined empirically. 3) Both
time and space overhead is usually very high. We propose a novel self-
similarity-based anomaly detection algorithm based on piecewise fractal
model. This algorithm consumes only limited amount of memory and
does not require training process. Theoretical analysis of the algorithm
are presented. The experimental results on the real data sets indicate
that, compared with existing anomaly detection methods, our algorithm
can achieve higher precision with reduced space and time complexity.

Keywords: Data Streams, Anomaly Detection, Fractal.

1 Introduction
Real-time monitoring over data streams have been attracting much attention.
Anomaly detection aims to detect the points which are significantly dierent
from others, or those which cause dramatic change in the distribution of under-
lying data stream. It is a great challenge to accurately monitor and detect the
anomalies in real time because of the uncertainty of properties of the anomaly.
In general, existing anomaly detection methods detect abnormality based on
a model derived from normal historical behavior of a data stream. They try to
reveal the dierences over short-term or long-term behaviors that are inconsis-
tent with the derived model. However, the detected results largely depend on the
model defined, as well as the data distribution and time granularity of the under-
lying data stream. For the data stream whose distribution changes constantly,
the existing methods will result in a large volume of false positives. On the other
hand, due to the application-dependent definition of abnormal behavior, each
existing method can only detect certain types of abnormal behaviors. In real
situation, it is meaningful to define abnormal behavior in a more general sense,
and is highly desirable to have an ecient, accurate, and scalable mechanism for
detecting abnormal behaviors over dynamic data streams.

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 550562, 2013.
c Springer-Verlag Berlin Heidelberg 2013

 Fractal Based Anomaly Detection over Data Streams 551

Self-similarity is ubiquitous both in natural phenomena and social events. It

represents the intrinsic nature of the data. Abnormal behavior defined based on
the change of self-similarity naturally captures the essence of the event. Fractals,
which are built upon the statistical self-similarity, have been found with wide
variety of applications [14]. In this paper, we focus on modeling data using
fractals based on recurrent iterated function systems (RIFS).
Most existing data stream algorithms are proposed for specific monitoring
tasks [10,16,19,5,15].
In [20], the authors use a Shifted Wavelet Tree (SWT) to detect bursts, which
is an important kind of anomalies, on multiple windows. A more ecient aggre-
gate monitoring method is put forward in [4]. However, they can only monitor
monotonic aggregates, and both the maximum window size and the number of
monitored windows are limited. The moving average, or Kalman-like filter [9] are
also important measures to be monitored for anomaly detection. An adaptive
change detecting method for such measurements, based on Inverted Histogram,
is proposed in [19]. However, no any overall bound for the error is given.
Monitoring abnormal behaviors in the network trac systems has been stud-
ied in [7,11,10]. They can monitor the changes of the behavior on only one gran-
ularity. The detection of trends and surprise in time series database is studied
in [17]. In [6], the exception of trend is defined and detected.
Fractal geometry has been known to be a powerful tool to describe complex
shapes in nature [12]. It has been successfully applied to modeling and com-
pressing the data sets [18,13,14]. However, it is dicult to apply them directly
in the data stream scenario, since the time complexity of constructing the fractal
model is polynomial, while multi-pass scan of the whole data is required.
In this paper, we propose a fractal-based novel approach to detect anomaly
in data streams. Our contributions are summarized below.

We propose a general-purpose and adaptable definition for abnormal behav-

ior based on the change of self-similarity. It captures the intrinsic property
of the data streams.
We introduce the fractal analysis based on self-similarity into anomaly detec-
tion of the data stream. The piecewise fractal model is presented to process
data streams in guaranteed space cost and error.
Both the theoretical analyses and experimental results illustrate that the
proposed method can monitor data streams accurately and eciently with
limited resource consumption.

2 Preliminaries of Fractals

If a quantitative property q is measured on a time scale s, then the value of q

depends on s according to the q = psr scaling relationship [12]. This is called
power law scaling relationship of q with respect to s. p is a factor of proportion-
ality and r is the scaling exponent. The value of r can be determined by the
slope of the linear least squares fitting to the graph of data pairs (log s, log q):
552 X. Gong et al.

log q = log p + r log s. For discrete time series data, the power-law scaling rela-
tionship can be transformed to q(st) = sr q(t), where t is the basic time unit and
s(s R) is the scaling measurement [3]. r can be determined by r = log(q(st)/q(t))
log s .
The power-law scaling relationship is also retained when q is a first order
aggregate function (e.g. sum, count) or some second order statistic variables
(e.g. variance).
Intuitively, objects satisfying self-similarity property can be mapped (by a se-
ries of maps) to part of itself. This mapping process can be conducted iteratively
until a fixed point called attractor, is found. Given the attractor. by iteratively
applying the inverse of the mapping, we can approximate the original objects.
The mapping process introduced above is mathematically called the iterated
function system (IFS).
More precisely, an IFS consists of a finite collection of contraction maps Mi
(i = 1, 2, 3, ..., m). Each Mi can map a compact metric space onto itself, provided
that the metric space is linear.
G H AGpoint (x,
HGy) can
H Gbei mapped/transformed
H by
x ai11 ai12 x b1
Mi [x, y] with the form Mi = i i + i .
y a21 a22 y b2
(m
Assume A is the attractor, then A = i=1 Mi (A). By modeling the data with
an IFS, we can use the attractor of the IFS to approximate the data. When ai12
is zero, the mapping is called shear transformation [14]. Then, ai22 < 1 is called
the contraction factor for map Mi .
IFS can only be used to manipulate linear fractals, which have a fixed contrac-
tion factor for all objects in all scales. However, the natural objects may be more
complex. Thus recurrent iterated function systems (RIFS) are introduced. RIFS
can be generalized for arbitrary metric spaces [2]. Intuitively, dierent from IFS,
in RIFS each map can be applied on part of the objects. The ranges of objects
the maps can be applied on may overlap. Thus, dierent objects may be mapped
to dierent parts and with dierent contraction scale. (
In RIFS, given the contraction maps Mi , the attractor A Rd is N i=1 Ai
where Ai Rd (i = 1,( ..., N ) are possibly overlapping partitions. Further-
more, Mj (A) = Aj = <i,j>G Mj (Ai ). Here G is a directed graph, and
A = (A1 , ..., AN ) (Rd )N . Each edge < i, j > G indicates that the map-
ping composition Mj Mi is allowed. The attractor can also be represented by
A = M(A), in which M : (Rd )N (Rd )N is defined as (M1 , M2 , ...MN ). Thus,
given a set vector D = (D1 , D2 , ..., DN ) with Di Rd , A = limi Mi (D).
Similar to IFS, attractor A can be used to approximate the data. When the
graph G is complete, the RIFS degrades to an IFS [8].
IFS and RIFS are often used to summarize and compress fractal data. The
problem lies in the determination of the attractor. Since the attractor can be
computed given the mapping, the computation of the parameters of the mapping
is usually called the inverse problem of IFS or RIFS [8].
Given the mapping M, D = M(D), we consider the condition that D D
in this paper. Under this condition, the open set property holds, so that the
Collage Theorem can be applied to guarantee the precision of approximating
the original data using the attractor [14,2]. This theorem provides a way to
 Fractal Based Anomaly Detection over Data Streams 553

measure the goodness of fit of the attractor associated with an RIFS and a given
function.
Collage Theorem: Let (X, d) be a complete metric space where d is a distance
measure. Let D be a given function and > 0 be given. Choose(an RIFS, with
N
contractility factor s = max{si : i = 1, 2, 3, ..., N } so that d(D, i=1 Mi (D))

. Then d(D, A) 1s where A is the attractor of the RIFS.
Since no assumption that the data streams to be monitored are linear fractals
can be made, we consider RIFS in the rest of this paper.

3 Problem Statement
A data stream X is considered as a sequence of points x1 , ..., xn . Each element
in the stream can be a nonnegative value, which is denoted as a (timestamp :
i, value : yi ) pair or i for simplicity. Function F being monitored can be not only
the monotonic aggregates with respect to the window size, for example sum
and count, but also the non-monotonic ones such as average and variance. One
common property of the above functions is that they all retain the power-law
scaling relationship [16].
Monitoring anomalies in data stream is to detect the change of power-law
scaling relationship, i.e., self-similarity, on the sequences with continuous incom-
ing new value xi . Under such a data stream model, the problem about anomaly
detection can be described as follows.
PROBLEM STATEMENT : A data stream X is represented as a sequence of
points x1 , ..., xn . An anomaly is detected if the self-similarity of X changes (i.e.,
the historic one is violated) when new value xn comes.
Our basic idea is to use the attractor to represent the data stream, and moni-
tor events that change the attractor, which mean anomalies. Thus, the problem
lies in How to eciently estimate the attractor in a streaming data environ-
ment?, How good is it to use estimated attractor to approximate the original
data stream?, and How to modeling the anomaly based on the attractor?.

4 Piecewise Fractal Model for Data Streams

Constructing a piecewise fractal model is equivalent to solving a inverse problem
in data stream scenario. We investigate two approaches. One is L2 error optimal,
whose complexity is O(n2 ) in time and O(n) in space. The other is an approx-
imate method with at most (1 + ) times the error for the optimal solution. It
uses O(n) time and O(m) space, where m is the number of pieces needed, and
m < log n. For ease of understanding, we list the notations used in Figure 1.

4.1 Inverse Problem

RIFS maps a partition Di of the whole metric space D to another partition Di
under two constraints. The first constraint is: Di = Mi (Di ), Di Di .
554 X. Gong et al.

xi i-th data point in stream X

m number of pieces/contraction maps
Ai distance
Bi distance
error constraint of piecewise fractal model
Mi i-th contraction map
Di i-th partition of whole metric space D
Di the partition mapped from Di , Di Di
Pi i-th piece of fractal model
Pi the piece mapped from Pi by Mi , Pi Pi

Fig. 1. List of notations Fig. 2. Contraction mapping on

stream X in our piecewise frac-
tal model

The second constraint which needs to be maintained in RIFS is that every

contraction mapping maps the two endpoints (x(i,0) , y(i,0) ) and (x(i,1) , y(i,1) )
of Di to two endpoints (x(i,0) , y(i,0)

) and (x(i,1) , y(i,1)

) of Di (Figure 2). Mi
transforms a larger interval Di of stream X to a smaller interval Di of the same
stream. Thus
I the endpoints
J G of Di are
I also Jthe endpoints of the attractor of Mi ,
H G H
x x(i,0) x x(i,1)
that is, Mi (i,0)
= , Mi (i,1)
= .
y(i,0) y(i,0) y(i,1) y(i,1)
Once the contraction factor ai22 for each mapping is determined, the remaining
parameters can be obtained using the endpoint constraint (4.1) as follows.

x(i,1) x(i,0) x(i,1) x(i,0) x(i,0) x(i,1)

ai11 = , bi1 = (1)
x(i,1) x(i,0) x(i,1) x(i,0)

y(i,1) y(i,0) y(i,1) y(i,0)
ai21 = ai22 (2)
x(i,1) x(i,0) x(i,1) x(i,0)

x(i,1) y(i,0) x(i,0) y(i,1) x(i,1) y(i,0) x(i,0) y(i,1)
bi2 = ai22 (3)
x(i,1) x(i,0) x(i,1) x(i,0)

The contraction mapping Mi denotes the similarity between Di and Di . Pro-

vided that Di is a subset of Di , the Mi denotes the self-similarity between Di
and Di . With the self-similarity/Mi , we can reconstruct Di from Di . More im-
portantly, we alarm anomaly when the self-similarity/Mi of the latest partition
Di is violated by newly arrived data point. Piecewise fractal model is used to
maintain such contraction mappings (self-similarity) over data streams.
Now the problem of constructing the piecewise fractal model for data streams
is reduced to determining the appropriate value of contraction factor ai22 .

4.2 Piecewise Fractal Model

In a piecewise fractal model, the goal of the global optimal solution is to minimize
the sum of error of each piece.
 Fractal Based Anomaly Detection over Data Streams 555

Optimal Model. Let the start point of a piece P of stream X be (s, ys ), and
the end point of P be (e, ye ), e > s. A contraction mapping M can be defined
as M (i, yi ) = (int(i a11 + b1 ), i a21 + yi a22 + b2 ). The data points in a larger
piece of a data stream are mapped by M from P {(s, ys ), (e, ye )} to a smaller
piece of the same stream P {(s , ys ), (e , ye )}, where P P and e s > e s .
Every two adjacent P s meet only at the endpoints and do not overlap. For
all pieces( of the (whole data stream, however, there might be some overlaps,
that is, Pi = Pi = X. Suppose the L2 error between the data points in
2
mapping M (P ) and those in the original piece #e P is E(M ) = (M (P ) P )2 ,
then the error of a mapping M is E(M ) = i=s ((i a21 + yi a22 + b2 ) yj ) ,
where j = i a11 + b1 . Given P and P , the optimal mapping Mopt is the one
such that the minimum value of E(M ) is reached.
For data stream X, it is ensured that each contraction mapping M maps the
start point and end point of piece P to corresponding ones of piece P , namely,
(s , ys ) = M (s, ys ) and (e , ye ) = M (e, ye ). a21#and b2 can be obtained by
e
equations 2 and 3. Therefore, we have E(M ) = i=s (Ai a22 Bi )2 , in which
ei is ei is
Ai = yi ( es ys + es ye ), and Bi = yj ( es ys + es ye ).
The coecient a22 of the optimal mapping Mopt can be computed using
E(M) #e #e #e 2
a22 = 2 i=s (Ai a22 Bi )Ai = 0, then a22 = i=s Ai Bi / i=s Ai . Conse-
quently, the optimal contraction mapping Mopt can be obtained by maintaining
e e
i=s Ai Bi and i=s A2i over the data streams. Algorithm 1 illustrates the method
for constructing the optimal piecewise fractal model for a data stream.

Theorem 1. Assume the length of a data stream is n. Algorithm 1 maintains

optimal piecewise fractal model with O(n + m) space in O(n2 ) time.

Approximate Model. It is impossible to store the whole data stream in mem-

e
ory. Thus we cannot maintain the value of i=s Ai Bi in the data stream scenario,
as Algorithm 1 requires. To address this problem, we construct a approximate
e
piecewise fractal model to compute the approximate value of i=s Ai Bi .

i
Denote the approximate value of a22 to be a22 . We have 1) Ci = yi ( ee s

ys +
# #
" e 2
e e Bi
is 2 es 2 2
e s ye ) , 2) i=s Bi = e s i=s Ci , and 3) (a22 ) = . Hence we only
"ei=s
A2
i=s i

e 2 e 2
need to maintain i=s Ci and i=s Ai incrementally for current pair of P and P
556 X. Gong et al.

on the data stream. In this way, the a22 and corresponding approximate results
Mapp can be computed.
The approximate piecewise fractal model consists of m pieces which can be
used to reconstruct the original data stream. The error of reconstruction is
bounded by Collage Theorem. Each piece Pi corresponds to a contraction map-
ping Mi and a pair of partition P and P . One value needed to be stored for Pi
is the sux sum F (n si ) starting from (si , ysi ), where a sux sum of data
n
stream is defined as F (wi ) = j=nwi+1 xj . The other value needed to be stored
for Pi is the number of points included in this sux sum. We first try to add
each data point of the evolving stream into the current piece. If failed, then a
new piece is created for this point. Because the end point of piece Pi is the start

point of piece Pi+1 , it is not necessary to store the end point. The method of con-
structing an approximate piecewise fractal model for a data stream is described
in Algorithm 2.
Theorem 2. Algorithm 2 can maintain the approximate piecewise fractal model
for a data stream with O(m) space in O(n) time.

4.3 Error Bound

Theorem 3. Assume Algorithm 2 produces an approximate solution Mapp which

maps the partition P to P with error E(Mapp ), and the optimal method maps
the partition P to P with error E(Mopt ), then E(Mapp ) 2E(Mopt ).

The error bound of Algorithm 2 to the optimal result is shown in Theorem 3.

The proof is omitted due to space limitation. For some applications in which
the bounded absolute error is wanted, we provide a parameter to control the
error. Then the error of mapping Mapp on partitions P and P can be guaranteed:

E(Mapp ) E(PLSF ), where E(PLSF ) is the Least Linear Fitting error on

partition P . Therefore, the error produced by approximate piecewise fractal
model on partition P is less than the smaller value of E(PLSF

) and 2E(Mopt ),

which is E(Mapp ) min{E(PLSF ), 2E(Mopt )}.
To control the parameter , we only need to check whether E(Mapp )

E(PLSF ) holds or not when a new xn comes. If E(Mapp ) E(PLSF ) holds,
xn is put into the current partition; otherwise, a new partition is created for xn .
The rest is the same as that of Algorithm 2.

5 Anomaly Detection Mechanism

5.1 History-Based Model

The history-based model trains a model incrementally, and compares the new
coming data against the model. The data that deviate from the model are re-
ported as anomalies. Existing burst detection [20,4,16] and change detection [10]
methods conform to this model. For instance, the threshold of bursty behaviors
of aggregate function F on i-th window is set as follows. At first, the moving
 Fractal Based Anomaly Detection over Data Streams 557

F (wi ) is computed over some training data which could be the foremost 1% of
the data set. The training data forms another time series data set, say, Y . Then
the threshold of anomaly is set to be T (wi ) = y + y , where y and y are
the mean and standard deviation, respectively. The threshold can be tuned by
varying the coecient of standard deviation. These methods then look for the
significant dierence in short-term or long-term behaviors which are inconsistent
with the model.
To maintain the model, piecewise fractal model is used. For each piece, the
line connecting end points of a piece is used to approximate the data within this
piece. The error is denoted by E(M ) and is guaranteed. The data generated by
all pieces are summed up to approximate the original data stream. Thus, the
piecewise fractal model is used as a synopsis data structure.

5.2 Self-similarity-Based Model

It is natural for us to investigate the another mechanism for detecting abnormal
behaviors in the data streams. The behaviors observed by the mechanism on the
data streams can be depicted with the equation: Y = X + e, where Y is either an
aggregate function(e.g., sum, count, average), measurement of trend and deviant
(say, variance), or distance of two distributions on the data stream. X is the
result of Y in normal condition. e, the eect caused by abnormal behaviors on
Y , can be seen as a noise. We assume that X and e have their own distributions
and the uncertainty noise e is white Gaussian [9].
In the scenario of discrete data points, Gaussian white noise can only be
modelled and processed in time series data. The constraints imposed on the
data stream processing include limited memory consumption, sequentially lin-
ear scanning, and on-line accurate result reporting over evolving data streams.
Therefore, we build piecewise fractal model on the data streams to model the
change of self-similarity induced by white Gaussian noise.
The occurrence of abnormal behaviors is rooted in the change of self-similarity
in the data stream. The piecewise fractal model can find the pieces with self-
similarity in a data stream. That a new piece is created means there is a change
in current self-similarity. When a new piece is created for the incoming xn , an
alarm is reported on the anomaly of xn . If the bounded absolute error is wanted,
a parameter can be used in the algorithm to control the error.
With this model, we can evaluate both the signal and noise in the data stream.
Since the anomaly detection only focuses on the prominent noise which causes
the change of self-similarity, the evaluation of anomaly on e is not aected by
the evolving of normal signal X.

6 Performance Evaluation
6.1 Experimental Setup
The experiments are conducted on a platform with a 2.4GHz CPU, 512MB main
memory on Windows 2000. We applied our algorithms to a variety of data sets.
558 X. Gong et al.

a) Varying (D1) b) Varying (D3) (a) Varying (D1) (b) Varying (D3)

Fig. 3. The accuracy when varying Fig. 4. The accuracy when varying

Due to the space limitation, only results on three real-life datasets are reported:
1) Stocks Exchange (D1): This data set contains one year tick-by-tick trading
records in 1998 from a corporation on Hong Kong main board. 2) Network Trac
(D2): This data set BC-Oct89Ext4 is a network trac tracing data set obtained
from the Internet Trac Archive [1,7]. 3) Web Site Requests (D3): It consists of
all requests made to the 1998 World Cup Web site between April 26, 1998 and
July 26, 1998 [1].
Two accuracy metrics, recall and precision, are considered. Recall is the ratio
of true alarms raised to the total true alarms which should be raised; precision
is the ratio of true alarms raised to the total alarms raised.

6.2 Experimental Results

Detecting Anomalies in History-Based Model. The piecewise fractal model
can be applied to reconstruct data and detect history-based anomalies defined by
existing methods [20,4,16]. Here we present the results of comparison tests with
these methods.
To set the threshold for F (wi ) on i-th window wi , we compute moving F (wi )
over some training data. The training data is the foremost 1% of each data set.
It forms another time series data set, called Y . The absolute threshold is set to
be T (wi ) = y + y . The length of windows is 4 N W time units, where N W
is the number of windows ranging from 50 to 1100. Similar testing results can
be got from all the data sets.
Given N W = 70, and m = 8, being 7 (D1) or 9 (D3), Figure 3 shows
that under any setting of the precision and recall of our method are always
above 97%. Larger means larger threshold. For our method can model such
significant dierences more accurate with self-similar pieces, it can be concluded
that with the increasing value of , our method can guarantee better accuracy
for detecting history-based anomalies.
While fixing the threshold parameter of piecewise fractal model (7 on D1
and 5 on D3), Figure 4 shows that when the value of increases, the precision
and recall also increase in certain domain. When exceeds certain value, both
precision and recall decrease. cannot be too small, because smaller means a
piece can only store less data points, hence the length of the overlapping pieces
for the piecewise fractal models becomes shorter when the number of pieces m is
fixed. Thus the estimated aggregate value of larger windows is more error prone,
 Fractal Based Anomaly Detection over Data Streams 559

(a) Space cost (b) Accuracy (a) D1 (b) D3

Fig. 5. The eciency of anomaly detection Fig. 6. Comparing number of anomalies,

on D3 when varying number of windows with N W = 100, = 7 for SWT (history-
based method)

(a) Anomalies on D1 (b) Anomalies on D3

Fig. 7. Anomalies detected on D1 and D3

which leads to the increase of the error of anomaly detection. On the other hand,
it is easy to understand that larger results in longer pieces, therefore larger
reconstruction error.
Given = 5 and = 9, with the increase of number of monitored windows, the
number of used pieces also increases. But the number of consumed pieces can still
be guaranteed by m < log2 4 N W . It can be seen from Figure 5 that with only
a few pieces the precision and recall are still high. The accuracy becomes better
when N W increases. This is because the average size of monitored windows is
larger with the increase of N W . The fractal approximation is more accurate over
large time scales. So the error decreases with the increasing value of N W . This
illustrate our algorithms ability to monitor large amounts of windows with high
accuracy over streaming data.

Anomaly Detection in Self-similarity-Based Model. In Figure 7 (m =

1, = ) (a) and (b), the solid red points are anomalies identified by our
algorithm on two pieces of data streams D1 and D3. Both the precision and
recall are 100%. Such significant dierence depicts the abnormal behavior in
stock exchanging and web site requesting.
In this experiment, we also compare the self-similarity-based method with
existing history-based anomaly detection methods [20,4,16]. Since they have the
same definition on anomalies, we choose [20] to represent the rest. When de-
tecting anomaly on data streams, existing methods will return large number of
false positive results. In the experiment, we set N W = 100. It can be seen from
560 X. Gong et al.

(a) D2 (b) D3 (a) D2 (b) D3

Fig. 8. Time consumption comparison on Fig. 9. Space consumption comparison on
dierent data sets by varying the number dierent data sets by varying the number
of monitored windows of monitored windows

Figure 6 that SWT returns large number of false positive results in every case.
For example, in Figure 6.(b), 108 alarms are returned, which is definitely not
informative to the users. Moreover, it is dicult to set the threshold for anoma-
lies. A slightly smaller or larger can either boost false alarms or false positive,
respectively. Therefore, our self-similarity-based method is more adaptable for
providing accurate alarms in anomaly detection.
Space and Time Eciency. To show the time and space eciency of our
algorithm, we compare our method with Stardust [4] and the query-based method
[16]. It can be observed from Figures 8 and 9 that our method uses much less
memory, runs more than tens of times faster than other methods. We use the
code provided by the authors of [4] and [16] to do the comparing test. The rest
of setting is the same as before, except for the special setting for Stardust with
box capacity c = 2.
Figure 8 shows the time eciency testing results. With the increasing of the
number of monitored windows, the processing time saved by using self-similarity-
based algorithm is getting larger and larger. Our method can process 400 107
tuples in 10 seconds. This means that it is capable of processing trac rates on
100Mbs links, and with some work than 1Gbps or even higher are within reach.
Figure 9 shows the space eciency testing results. The space cost of our piece-
wise fractal model is only aected by the number of pieces maintained. However,
Stardust has to maintain all the monitored data points and the index structure at
the same time. The query-based method has to maintain an inverted histogram
(IH) in memory. It can be seen that the space saved by the piecewise fractal
model is considerable. Thus, our method is more suitable to detect anomalies
over streaming data.

7 Concluding Remarks
In this paper, we incorporate the fractal analysis technique into anomaly de-
tection. Based on the piecewise fractal model, we propose a novel method for
detecting anomalies over a data stream. Fractal analysis is employed to model
the self-similar pieces in the original data stream. We show that this fractal-
based method can be used to detect not only the bursts defined before, but
 Fractal Based Anomaly Detection over Data Streams 561

more general anomalies that cause the fractal characteristics change. Piecewise
fractal model is proved to provide accurate monitoring results, while consuming
only limited storage space and computation time. Both theoretical and empirical
results show the high accuracy and eciency of the proposed method. Further-
more, this approach can also be used to reconstruct the original data stream
with the error guaranteed.

Acknowledgments. This work is partially supported by National Science Foun-

dation of China under grant numbers 60925008 and 61170086.

References
1. Internet trac archive, http://ita.ee.lbl.gov/
2. Barnsley, M., Elton, J., Hardin, D.: Recurrent iterated function systems. Construc-
tive Approximation 5(1), 331 (1989)
3. Borgnat, P., Flandrin, P., Amblard, P.: Stochastic discrete scale invariance. IEEE
Signal Processing Letters 9(6), 181184 (2002)
4. Bulut, A., Singh, A.: A unified framework for monitoring data streams in real time.
In: Proceedings of the 21st International Conference on Data Engineering, ICDE
2005, pp. 4455. IEEE (2005)
5. Chandola, V., Banerjee, A., Kumar, V.: Anomaly detection: A survey. ACM Com-
puting Surveys (CSUR) 41(3), 15 (2009)
6. Chen, Y., Dong, G., Han, J., Wah, B., Wang, J.: Multi-dimensional regression anal-
ysis of time-series data streams. In: Proceedings of the 28th International Confer-
ence on Very Large Data Bases, pp. 323334. VLDB Endowment (2002)
7. Cormode, G., Muthukrishnan, S.: Whats new: Finding significant dierences in
network data streams. In: Twenty-third Annual Joint Conference of the IEEE
Computer and Communications Societies, INFOCOM 2004, vol. 3, pp. 15341545.
IEEE (2004)
8. Hart, J.: Fractal image compression and recurrent iterated function systems. IEEE
Computer Graphics and Applications 16(4), 2533 (1996)
9. Jain, A., Chang, E., Wang, Y.: Adaptive stream resource management using kalman
filters. In: Proceedings of the 2004 ACM SIGMOD International Conference on
Management of Data, pp. 1122. ACM (2004)
10. Kifer, D., Ben-David, S., Gehrke, J.: Detecting change in data streams. In: Proceed-
ings of the Thirtieth International Conference on Very Large Data Bases, vol. 30,
pp. 180191. VLDB Endowment (2004)
11. Krishnamurthy, B., Sen, S., Zhang, Y., Chen, Y.: Sketch-based change detection:
methods, evaluation, and applications. In: Proceedings of the 3rd ACM SIGCOMM
Conference on Internet Measurement, pp. 234247. ACM (2003)
12. Mandelbrot, B.: The fractal geometry of nature. Times Books (1982)
13. Mazel, D., Hayes, M.: Using iterated function systems to model discrete sequences.
IEEE Transactions on Signal Processing 40(7), 17241734 (1992)
14. Michael, F.: Fractals everywhere. Academic Press, San Diego (1988)
15. Patcha, A., Park, J.: An overview of anomaly detection techniques: Existing solu-
tions and latest technological trends. Computer Networks 51(12), 34483470 (2007)
16. Qin, S., Qian, W., Zhou, A.: Approximately processing multi-granularity aggregate
queries over data streams. In: Proceedings of the 22nd International Conference on
Data Engineering, ICDE 2006, p. 67. IEEE (2006)
562 X. Gong et al.

17. Shahabi, C., Tian, X., Zhao, W.: Tsa-tree: A wavelet-based approach to improve
the eciency of multi-level surprise and trend queries on time-series data. In: Pro-
ceedings of the 12th International Conference on Scientific and Statistical Database
Management, pp. 5568. IEEE (2000)
18. Wu, X., Barbar a, D.: Using fractals to compress real data sets: Is it feasible. In:
Proc. of SIGKDD (2003)
19. Zhou, A., Qin, S., Qian, W.: Adaptively detecting aggregation bursts in data
streams. In: Zhou, L., Ooi, B.-C., Meng, X. (eds.) DASFAA 2005. LNCS, vol. 3453,
pp. 435446. Springer, Heidelberg (2005)
20. Zhu, Y., Shasha, D.: Ecient elastic burst detection in data streams. In: Proceed-
ings of the Ninth ACM SIGKDD International Conference on Knowledge Discovery
and Data Mining, pp. 336345. ACM (2003)
 Preservation of Proximity Privacy
in Publishing Categorical Sensitive Data

Yujia Li1, Xianmang He2,*, Wei Wang1, Huahui Chen2, and Zhihui Wang1
1
School of Computer Science and Technology, Fudan University,
No.220, Handan Road, Shanghai, 200433, P.R. China
{071021056,weiwang1,zhhwang}@fudan.edu.cn
2
School of Information Science and Technology, NingBo University
No.818, Fenghua Road, Ning Bo, 315122, P.R. China
{hexianmang,chenhuahui}@nbu.edu.cn

Abstract. In this paper, we address the problem of proximity privacy in pub-

lishing categorical sensitive data. When using traditional approach to anonym-
ize the data and generalize the QI-groups, some sensitive attribute values with
semantical proximity may exist in the same QI-group, and thus lead to privacy
leakage. To solve this issue, the concept of m-Color constraint is introduced and
a method based on the m-Color constraint is proposed to prevent this kind of
privacy leakage. The properties of m-Color constraint and related generalization
algorithm are given, which reduce the loss of information greatly. The experi-
ment results are provided to explain practicality and efficiency of the algorithm
proposed in this paper.

Keywords: privacy preservation, k-anonymity, proximity privacy, m-Color.

1 Introduction

The information age has witnessed a tremendous growth of personal data that can be
collected and analyzed, including a great deal of released microdata (i.e. data in raw,
non-aggregated format) [12].These released microdata offer significant advantages in
terms of information availability, which is particularly suitable for ad hoc analysis in
a variety of domains, such as medical research.
However, the publication of microdata leads to the concerns of individual private
information. For example, the medical institute and its cooperative organization need
a large amount of private data on patients for disease research. Table 1 is such an
example. One straightforward approach to achieve this is to exclude unique identifier
attributes, such as Name from Table 1, which is not sufficient for protecting privacy
leakage under linking-attack [1-3]. Suppose the attacker has known about Andy with
attribute values <age=5, sex=M, zip=12k>, then according to Table 1, s/he can deduce
Andys disease, which is pneumonia. Such combination of Age, Zipcode and Sex can

*
Corresponding author.

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 563570, 2013.
Springer-Verlag Berlin Heidelberg 2013
564 Y. Li et al.

be potentially used to identify an individual in Table 1 and is called a quasi-identifier

(QI for short) in literatures.
To avoid linking attack, numerous approaches have been proposed. Generalization
is applied on the quasi-identifiers and replaces QI-values with the less-specific but
semantically consistent values Achieving k-anonymity privacy protection using the
method of L. Sweeney [1]. Consequently, more records will have the same set of QI-
values, in which are indistinguishable from each other. For example, Table 2 is 3-
anonymized by generalizing QI-values in Table 1. Age=5, Zipcode=12k and Sex =M
of Andy have been replaced with intervals [5-9], [12k-18k] and M respectively. Clear-
ly, even if an adversary has the exact QI-values of Andy, s/he can not infer exactly
which tuple in the first group is Andy. In general, we refer to the set with the same
projection on the quasi-identifier as QI-Group.

Table 1. Microdata Table 2. Generalization Table

Name Age Sex Zip D iseas e ID Age Sex Zip Disease
A ndy 5 M 12 k Pneum on ia [5-9] M [12k-18k] Pneum onia
Bill 9 M 14 K Pulm on ary em bo lism 1 [5-9] M [12k-18k] Pulm onary em bolism
Ken 6 M 18 K Flu [5-9] M [12k-18k] Flu
Nash 8 M 19 K Bro nchit is [8-12] M [19K-22k] Bronchitis
J oe 12 M 22 K C olo tis 2 [8-12] M [19K-22k] Colotis
Sam 11 M 20 K C olo n can cer [8-12] M [19K-22k] Colon cancer
Lin da 21 F 58 k Sto m ach cancer [21-56] F [33k-58k] Stom ach cancer
Jam e 26 F 36 k Gast ric ul cer 3 [21-56] F [33k-58k] Gastric ulcer
Sarah 28 F 37 k G as tritis [21-56] F [33k-58k] Gastritis
Mary 56 F 33 k Pu lmo nary edem a [21-56] F [33k-58k] Pulm onary edem a

1.1 Motivation
Recently, there are many works to address the problem of privacy preserving data
publishing. However, most of them ignore the case that there may exist semantically
proximity for sensitive categorical data. For example, an adversary possessing the QI-
values of Andy is able to find out that Andys record is in the first QI-group of Table
2. Without further information, assume that each tuple in the group has an equal
chance of being owned by an adversary, s/he can conclude that Andy suffers from
some respiratory infection disease with 2/3 probability.
The existing anonymization principles fall short of solving such proximity privacy
leakage in our setting, which is a privacy threat specific to categorical sensitive
attribute (i.e. pneumonia in Table 2). To address the issue, this paper proposes a new
anonymization principle termed as m-Color constraint.
The rest of the paper is organized as follows. The previously related work is re-
viewed in Section 2. Section 3 clarifies the basic concepts and problem definition.
Section 4 discusses the properties of m-Color constraint, and Section 5 elaborates the
generalization algorithm. In Section 6, we experimentally evaluate the efficiency
and effectiveness of our techniques. Finally, the paper is concluded in Section 7.

2 Related Work
In this section, previous related work will be surveyed. All the privacy-preserving
transformation of the microdata is referred to as recoding, which can be classified into
 Preservation of Proximity Privacy in Publishing Categorical Sensitive Data 565

two classes of models: global recoding and local recoding[11,16]. In global recoding,
a particular detailed value must be mapped to the same generalized value in all
records. Local recoding allows the same detailed values being mapped to different
generalized values of different QI-groups. Obviously, global recoding is a more spe-
cial case of local recoding. Efficient greedy solutions following certain heuristics
have been proposed [7],[9],[11],[15] to obtain a near optimal solution. Incognito [8]
provides a practical framework for implementing full-domain generalization, borrow-
ing ideas from frequent item set mining, while Mondrian [9] takes a partitioning ap-
proach reminiscent of kd-trees. To achieve k-anonymity, Ghinita [16] presents a
framework mapping the multi-dimensional quasi-identifiers to 1-Dimensional(1-D)
space. It is discovered that k-anonymity a data set is strikingly similar to building a
spatial index over the data set, so that classical spatial indexing techniques can be
used for anonymization [17]. The idea of non-homogeneous generalization was first
introduced in k-anonymization revisited [18], which studies techniques with a guaran-
tee that an adversary cannot associate a generalized tuple to less than K individuals,
but suffer additional types of attack. Authors of paper [19] proposed a randomization
method that prevents such type of attack and showed that k-anonymity is not com-
promised by the attack, but its partitioning algorithm is only a special of the top-down
algorithm presented in [11]. However, in our algorithm, we mainly discuss privacy
preservation of categorical sensitive attributes over m-Color constraint.

3 Problem Statement

Let T be a microdata table that contains the private information of a set of individuals.
T has d QI-attributes A1,....,Ad, and a sensitive attribute (SA)S. A partition P consists of
several subsets Gi (1 i k) of T, such that each tuple in T belongs to exactly one sub-
set and T = ik Gi . We refer to each subset Gi as a QI-group or a bucket.

Definition 1. (Normalized Certainty Penalty (NCP) [11]). Suppose a table T is ano-

nymized to T*. In the domain of each attribute in T, suppose there exists a global
order on all possible values in the domain. If a tuple t in T* has range [xi ,yi] on
y i xi
attribute Ai(1 i d), then the normalized certainty penalty in t on Ai is NCPA (t ) = ,
i

Ai

where Ai is the domain of the attribute Ai. For tuple t, the normalized certainty
d
penalty in t is NCP (t ) = i
NCPA (t )
i
.The normalized certainty penalty in T is tT
NCP (t ) .
In reality, an attacker with sufficient background knowledge [10] may be able to
deduce a concrete QI-group. Assume that adversary deduce the sensitive attribute
values of victims with the same probability. For example, after realizing that Andys
record appears in the first QI-group of Table 2, an adversary derives P[X = s] = 1/3,
where X models the disease of Andy, and s can be any disease (i.e., Pneumonia, Pul-
monary embolism, Flu) in that group.
566 Y. Li et al.

3.1 Problem Definition

In our anonymization framework, we present a new anonymization method, which

eliminates proximity privacy leakage in publishing categorical sensitive attributes. In
general, we classify each tuple in table T based on a given taxonomy tree. For in-
stance, as shown Fig 1, there is a taxonomy tree of respiratory and digestive system
diseases. In the tree, ten types of diseases are described and belong to different sub-
trees. We simply divide these diseases into four classes, and color each class with red,
blue, green and black in Fig 2 respectively. According to the Fig 2, then we can get
the Table 3, where these tuples are labeled by four class-tags, red, blue, green and
black.

Table 3. Color Table Table 4. 3-Color Constraint

Name Age Sex Zip Disease Color ID N m ae A ge Sex Zip D i s ea s e C o lo r

Andy 5 M 12k Pneumonia red A n dy 5 M 12k P n e u m o n ia re d

Bill 9 M 14K Plu em bolism blue 1 B ill 9 M 14k P u l e m b o lis m b lu e
Ken 6 M 18K Flu red L in d a 21 F 58k S to m a c h c a n c e r g r e en
Nash 8 M 19K Bronchitis red Ken 6 M 18k F lu re d
Joe 12 M 22K Colotis black 2 Jo e 12 M 22k C o lo t i s b la ck
Sam 11 M 20K Colon cancer black Ja m e 26 F 36k G a s t r ic u lc e r g r e en
Linda 21 F 58k Stomach cancer green N a sh 8 M 19k B r o n ch i tis re d
Jam e 26 F 36k Gastric ulcer green S a r ah 28 F 37k G a s t ri t is g r e en
3
Sarah 28 F 37k Gastritis green M ary 56 F 33k P u l ed e m a b lu e

Mary 56 F 33k Pul edema blue Sam 11 M 20k C o lo n ca n ce r b la ck

While generalizing the table T, we do not change the sensitive attribute values of
disease but take colors as the sensitive values. From the Table 3, we can observe that
the generalization algorithm will not cause any information loss in sensitive attribute
values while the personality privacy [5] was not. We may label the tuples whose sen-
sitive values are similar with the same color class-tags, and assign the tuples with
the same color class-tags into different QI-groups to prevent the problem of proximity
privacy leakage. Hence, m-Color constraint is introduced.
Definition 2. A QI-group satisfies m-Color constraint, if the QI-group has at least m
tuples, and each tuple of this QI-group owns different color class-tags. Then a table
fulfills m-Color constraint, if in this table, each QI-group satisfies m-Color constraint.
Without considering the information loss, we can provide a possible partition of
Table 1 again (see the Table 4). Based on the definition 2, the partition fulfills 3-
Color constraint.
In our paper, m-Color constraint is used to solve the categorical attribute proximity
privacy problem, which limits proximity privacy leakage effectively.

4 The Properties of m-Color

Given table T and m, after each tuple was labeled different color class-tags, we won-
der the following question: Given a table T and an integer m, can we find a genera-
lized T* that fulfills m-Color constraint? This question is equivalent to the following
definition 5. Hence, the results are given as follows.
 Preservation of Proximity Privacy in Publishing Categorical Sensitive Data 567

Definition 3. Given table T and m, if there exists a partition that each QI-group has at
least m tuples, and all the sensitive attribute values of tuples in each QI-group are
unique. We refer to such table as m-eligible.
Theorem 1. T is a m-eligible, if and only if the number of tuples with the same sensi-
tive values is no more than 1/m the total number of tuples in the table T.
Respiratory & Digestive Respiratory & Digestive
system diseases system diseases

Vascular lung Digestive system

Vascular lung Digestive system
diseases diseases
diseases diseases

Respiratory Vascular lung Stomach Colon

Respiratory Vascular lung Stomach Colon infection dieseases disease disease
infection dieseases disease disease

Pulmonary Pulmonary Stomach Gastric Colon

Flu Prenmonia Bronchitis
Flu Prenmonia Bronchitis
Pulmonary Pulmonary Stomach Gastric Colon edema embolism disease ulcer Gastritis Colitis cancer
edema embolism disease ulcer Gastritis Colitis cancer

Fig. 1. Taxonomy Tree Fig. 2. Color Taxonomy Tree

5 Generalization Algorithm

The principle of generalization algorithm is to partition the microdata into QI-groups

and to reduce the information loss of table as much as possible. Here, the whole algo-
rithm consists of three steps: partition, assign and split. The first step, according to
top-down algorithm [11], we can reduce the scale of questions continuously. Second-
ly, we partition each sub-table, which comes from the first step, into buckets. The last
step, the buckets were partitioned into QI-groups.

5.1 Top-Down Partitioning Algorithm

The basic idea of the algorithm is to partition the table iteratively. A set of tuples is
partitioned into subsets if each subset is more local and satisfies m-Color constraint.
The algorithm framework is shown in ALG1. To keep the algorithm simple, we con-
sider that in each round, we partition a set of tuples into two subsets, and to check
whether the both subsets satisfy the m-Color constraint. Otherwise, after trying k
times in the first step 1-5, we use the Assign algorithm to create the buckets.
ALG1: Partition Algorithm
Input dyed T, m, K
Output the set of sub-tables S
Method:
Let S1 S2 initial value empty set and K partition times.
FOR I=1 TO K
1. S1 ={t1} S2={t2}
2. FOR each w T-{t1,t2},
Compute 1=NCP(w S1)-NCP(S1), 2=NCP(w S2)-NCP(S2);
3. IF ( 1> 2), S2 =S2 {w};
ELES S1 =S1 {w};
568 Y. Li et al.

4. Find the max value a1,b1;

5. Check if S1 and S2 are m-eligible
a1 m1 (a1 + a2 + + a ) b1 m1 (b1 + b2 + + b ) ;
6. Random Shuffler the set T;

Step1: We can employ the following strategy to select t1, t2 : select a point u ran-
domly, then to choose t1 to minimize NCP(t1, u) (see the definition 3). We can scan all
tuples once to find tuple t2 that maximizes NCP(t1, t2).
Step 2: For each tuple w in table T, we compute the 1= NCP(w S1)-NCP(S1), and
2=NCP(w S2)-NCP(S2), and distribute w to the group that leads to a lower infor-
mation loss.
Step 3: If 1= 2, then w was put to the group with less cardinality.

Step 4-5 is to check whether S1, S2 satisfy m-eligible. If both are generalizable, we

remove T from S, and add S1, S2 to S; otherwise T is retained in S. The attempts to

partition T are tried K times and tuples of G are randomly shuffled for each time.
In the second step, we use assign algorithm [3] to create the set of buckets.

5.2 Split Algorithm

In this section, we use buckets passed from the assign algorithm to generate QI-
groups. Let the buckets signature be {v1,v2, , v } ( m) and each color tag corres-
ponds to tuples. If 1, the bucket is a QI-group; if >1, we divide the bucket
into two disjoint buk1 and buk2, and each bucket has the same signature. The process
to split the bucket until each bucket contains exactly tuples, each color tag corres-
ponds to one tuple.
The idea underlying the split algorithm is given as follows: sort all the tuples in
the bucket according to attribute A1. Then put the first /2 tuples from the bucket
into buk1 and the rest tuples into buk2. In this way, we can obtain a partition of the
bucket. For the attribute A2, A3,., we can get another d-1 choices, where d is the
dimension of QI-attributes. Among all the d partitions, we pick the one that minimizes
the sum of NCP(buk1) and NCP(buk2) as the final partition.

6 Experiments

In this section, we experimentally evaluate the effectiveness and efficiency of the

proposed algorithm. In our paper, the experiments run on AMD Dual Core CPU
1.9GHZ, RAM 1GB, Window 2003 Server machine. Our experimentation deploys a
real database INCOME from http://ipums.org. We treat the first five columns as QI-
attributes, and Income as the sensitive attribute. We compare our algorithm (denoted
by m-Color) with perturbation technique [21] (perturbation in short).
 Preservation of Proximity Privacy in Publishing Categorical Sensitive Data 569

We measure utility by answering aggregate queries on published anonymous data,

since such queries are the basic operations for many database tasks. Specifically, we
use queries of the following form: select count(*) from INC where A b1 And
A b2 And A bv
In the experiment, we compare our algorithm with perturbation technique [21]. We
fixed s as 0.1 and the dimensionality of queries is two. The technique of perturbation
from the experiment results in the first set is relatively poor. With the increase of m,
from about 40% to about 280%, the results of our paper are not to exceed 10%, seeing
the Fig 3.
Fig 4 illustrates the cost of our anonymous algorithm, while the perturbation tech-
nique is finished in less than 1 second (omitted from the Fig 4). The executing time of
generalization algorithm probably decreases from 350s to 110s when m from 4 to 30.
3 av erage 400
2.5 error 350 m-co lor
m-co lor
300
2 m-u niq ue
250
1.5
200
time(s

1 150
0.5 100
50
0
0
4 8 10 6 20 25 30
m 4 8 10 m 15 20 25 30

Fig. 3. Query Accuracy vs. m Fig. 4. Running Time

7 Conclusion

The weakness of k-anonymity motivates us to develop a novel anonymization prin-

ciple called categorical sensitive attributes m-Color constraint, which provides better
privacy preservation in order to prevent the same categorical sensitive attributes prox-
imity privacy leakage. We systematically analyze the theoretic properties of this new
technique and propose a corresponding algorithm. As verified by extensive experi-
ments, our algorithm supports more effective query and data analysis, and simulta-
neously providing enough privacy preservation.

Acknowledgement. This work was supported in part by the National Natural Science
Foundation of China (NO. 61033010, 61170006, 61202007), the Project of Education
Department of Zhejiang Province(Y201224678).

References
[1] Sweeney, L.: k-anonymity: A model for protecting privacy. Journal on Uncertainty,
Fuzziness, and Knowlege-Based Systems 10(5), 557570 (2002)
[2] Machanavajjhala, A., Gehrke, J., Kifer, D.: l-diversity: Privacy beyond k-anonymity. In:
ICDE (2006)
[3] Xiao, X., Tao, Y.: m-invariance: Towards privacy preserving re-publication of dynamic
datasets. In: SIGMOD 2007, pp. 689700 (2007)
570 Y. Li et al.

[4] Li, N., Li, T.: t-closeness: Privacy beyond k-anonymity and l-diversity. In: ICDE (2007)
[5] Xiao, X., Tao, Y.: Personalized privacy preservation. In: SIGMOD 2006, pp. 229240
(2006)
[6] Bayardo, R., Agrawal, R.: Data privacy through optimal k-anonymization. In: ICDE
2005, pp. 217228 (2005)
[7] Fung, B.C.M., Wang, K., Yu, P.S.: Top-down specialization for information and priva-
cy preservation. In: ICDE 2005, pp. 205216 (2005)
[8] LeFevre, K., DeWitt, D.J., Ramakrishnan, R.: Incognito: Efficient full-domain k-
anonymity. In: Sigmod 2005, pp. 4960 (2005)
[9] LeFevre, K., DeWitt, D.J., et al.: Mondrian multidimensional k-anonymity. In: ICDE,
pp. 277286 (2006)
[10] Li, J., Tao, Y., Xiao, X.: Preservation of Proximity Privacy in Publishing Numerical
Sensitive Data. In: Sigmod 2008 (2008)
[11] Xu, J., Wang, W., Pei, J., et al.: Utility-based anonymization using local recoding. In:
KDD 2006, pp. 785790 (2006)
[12] Willenborg, L., de Waal, T.: Elements of Statistical Disclosure Control. Lecture Notes
in Statistics. Springer (2000)
[13] Samarati, P., Sweeney, L.: Protecting Privacy when Disclosing Information: k-
Anonymity and Its Enforcement through Generalization and Suppression (1998)
[14] Samarati, P.: Protecting RespondentsIdentities in Microdata Release. IEEE Trans.on
Konwl. and Data Eng. 13(6), 10101027 (2001); Control. Lecture Notes in Statistics.
Springer (2000)
[15] Wong, W.K., Mamoulis, N., Cheung, D.W.L.: Non-homogeneous generalization in pri-
vacy preserving data publishing. In: SIGMOD 2010: Proceedings of the 2010 Interna-
tional Conference on Management of Data, pp. 747758. ACM, New York (2010)
[16] Ghinita, G., Karras, P., Kalnis, P., Mamoulis, N.: Fast data anonymization with low in-
formation loss. In: VLDB 2007: Proceedings of the 33rd International Conference on
Very Large Data Bases, pp. 758769. VLDB Endowment (2007)
[17] Iwuchukwu, T., Naughton, J.F.: K-anonymization as spatial indexing: toward scalable
and incremental anonymization. In: VLDB 2007: Proceedings of the 33rd International
Conference on Very Large Data Bases, pp. 746757. VLDB Endowment (2007)
[18] Gionis, A., Mazza, A., Tassa, T.: k-anonymization revisited. In: ICDE 2008: Proceed-
ings of the 2008 IEEE 24th International Conference on Data Engineering, pp. 744753.
IEEE Computer Society, Washington, DC (2008)
[19] Wong, W.K., Mamoulis, N., Cheung, D.W.L.: Non-homogeneous generalization in pri-
vacy preserving data publishing. In: SIGMOD 2010: Proceedings of the 2010 Interna-
tional Conference on Management of Data, pp. 747758. ACM, New York (2010)
[20] Zhang, Q., Koudas, N., Srivastava, D., Yu, T.: Aggregate query answering on anony-
mized tables. In: ICDE 2007: Proceedings of the 23nd International International Confe-
rence on Data Engineering, vol. 1, pp. 116125 (2007)
[21] Agrawal, R., Srikant, R.: Privacy-preserving data mining. SIGMOD Rec. 29(2), 439
450 (2000)
S2MART: Smart Sql to Map-Reduce Translators

Narayan Gowraj, Prasanna Venkatesh Ravi, Mouniga V, and M.R. Sumalatha

Department of Information Technology, Anna University,

Chennai, India
{ngowraj,prasanvenkyr,meek.mounik,
sumalatha.ramachandran}@gmail.com

Abstract. With the advent of rapid increase in the size of data in large
cluster systems and the transformation of data into big data in major
data intensive organizations and applications, it is very necessary to build
ecient and flexible Sql to Map-Reduce translators that would make Tera
to Peta bytes of data easy to access and retrieve because conventional Sql-
based data processing has limited scalability in these cases. In this paper
we propose a Smart Sql to Map-Reduce Translator (S2MART), which
transforms the Sql queries into Map-Reduce jobs with the inclusion of
intra-query correlation for minimizing redundant operation, sub-query
generation and spiral modeled database for reducing data transfer cost
and network transfer cost. S2MART also applies the concept of views
in database to perform parallelization of big data easy and streamlined.
This paper gives a comprehensive study about the various features of
S2MART and we compare the performance and correctness of our system
with two widely used Sql to Map-Reduce translators hive and pig.

Keywords: Sql to Map-Reduce Translators, big data, intra-query rela-

tionship and sub-query generation.

1 Introduction

According to a recent presentation from the vice-president of Facebook Je

Rothschild at the UC San Diego, Facebook has 30,000 servers supporting its
operations which were 10,000 servers on April 2008. This drastic increase in the
amount of data in large online organizations had led to the development of ex-
tensive data processing applications like the Sql to Map-Reduce translators [1]
[2] [3]. Due to the ever increasing size of data in these large clusters, traditional
relational databases can be expensive; hence many companies have maneuvered
their focus towards NoSql database framework such as BigTable, HadoopDB,
Simple DB, Microsofts SDS cloud database, MongoDB, Apache CouchDB etc
[4] [5] [6] [7]. The importance and uniqueness of cloud infrastructure is closely
coupled with the development of the Map-Reduce paradigm [8] [9]. In summary
the major contributions of this paper are: 1) Support intra-query correlations
by building a Sql relationship tree to minimize redundant operations and com-
putations.2) Build a spiral modeled database to store and retrieve the recently

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 571582, 2013.
c Springer-Verlag Berlin Heidelberg 2013

572 N. Gowraj et al.

used query results to minimize the data transfer cost and network cost and in-
crease the time eciency.3) The spiral modeled database is adaptive to dynamic
workload patterns.4) Sub-query generation to increase the throughput of the
system.5) Use the concept of Views in database to store intermediate results
which reduces the amount of records to be checked for every computation.6)
Experiment our system with legacy systems like Hive, pig etc.

2 Literature Survey
The popularity of a hybrid system transforming Sql queries to Map-Reduce jobs
has increased due to the emergence of big data in many data centric organiza-
tions. S2MART uses nuoDB [10] [11] as its cloud storage as nuoDB provides full
indexing support, easy and flexible Sql querying with DB replication, provides
drivers for ODBC, JDBC and Hibernate, it can be deployed anywhere in the
cloud, provides scalability upto 50 nodes and the most important thing is that it
provides support for Hadoop HDFS. In this paper we focus on three important
legacy Sql to Map-Reduce translators namely Pig, Hive and Scope. Companies
and organizations which deal with big data in the form of click streams, logs and
web crawls have focused their attention towards developing an open source high-
level dataflow system. This has led yahoo develop an open source incarnation
called Pig to support parallelization of data retrieval in the cloud environment
[12]. Pig supports a simple language called Pig Latin [13] to support query formu-
lations and data manipulation. Pig Latin is written using java class libraries and
provides support for simple operations like filtering, grouping, projection and
selection. The major issue concerned with Pig is that it provides limited flexibil-
ity towards join operations and nested operators. Hive developed by Facebook
adds Sql like functions to Map-Reduce but follows syntax dierent from the
conventional Sql and it is called as HiveQl [14] [15]. HiveQl transforms the Sql
query into Map-Reduce jobs that are executed in Hadoop (an open source in-
carnation of Map-Reduce). Hive provides the functionalities of a metastore that
supports query compilations and optimizations [16]. Even though Hive follows
a structured approach towards database concepts, it provides limited function-
ality towards join predicate. The Sql to Map-Reduce translator developed by
Microsoft was called as Scope [17]. The compiler and optimizer of Scope gen-
erates a new scripting language for managing and retrieving data from large
data repositories. The compiler and optimizer are responsible for automatic job
generation and parallelization of job execution. However, Scope provides no sup-
port towards user defined types and metastore which reduces the performance
of the system. Hence Scope doesnt provide a rich data model when compared
to Hive and pig as it doesnt include partitioning of tables. The other translators
developed by Microsoft are Cosmos [18] and Dyrad [19].

3 Proposed System
The proposed system consists of two linchpin modules which are:
 S2MART: Smart Sql to Map-Reduce Translators 573

3.1 System Architecture

The full-fledged system architecture is shown in fig.1 which is divided into three
major layers.The architecture of S2MART uses the classic master/worker pattern
of the Map-Reduce framework where the master node is referred as the query
distributor and the worker nodes are referred as query analyzers. The input
to the system is the Sql query from the user. This Sql query is given to the
Sql query Parser which in turn is given to the IQ-RT (Intra-query relationship
tree) parser which constructs the tree based on the input query. IQ-RT parser
generates the tree structure which we call as the relationship tree for the input
query from the Sql query parser. The tree is then given as an input to the IQ-RI
(Intra-query relationship identifier). After identifying the existing relationships,
the operations similar to each other are consolidated so that the number of scans
of a particular database reduces drastically. The resultant IQ-RT is subjected
to the Sub-Query generator engine which disintegrates the dierent operations
in the IQ-RT into sub-queries rather than a single large query. Now each query
are distinctly separated and every query is compared with the queries stored in
the Spiral database which uses the FQR (Frequent Query Retrieval) algorithm
to check whether the input query matches any one of the entries in the Spiral
database. Now there exists two possibilities, the first scenario is the case where
the input query and the results of the input query are cached from the Spiral
database and the other scenario is that the query and the results are not found
in the Spiral database. In the first case the results of the individual sub-query
are cached from the Spiral database [20] and are given as an input to the query
optimizer which optimizes the query results and gives the query results back to
the user. In the other case, each individual query is given to the Map-Reduce
framework which is built inside a Hadoop infrastructure [21][22]. Here in this case

Fig. 1. System Architecture

574 N. Gowraj et al.

a single node alone is not subjected to retrieve data from the database nodes
hence avoiding a bottleneck situation and no single node does the entire mapping
job of a single query hence increasing the eciency of data retrieval from the
database nodes without changing the underlying structure of the Map-Reduce
framework. Data transfer cost is reduced in this method by the use of spiral DB.
The frequently used queries are stored, thus reducing the need to transfer data
unnecessarily. Network transfer cost is reduced by the use of buers after the
map-reduce step.

3.2 Methodology

The entire working of the S2MART basically depends on the following modules:

3.2.1 SQL Query Parser

The Sql Query Parser is a module which accepts Sql queries as input from the
user, converts them into a format which is accepted by the S2MART and outputs
the query to the IQ-RT module to generate the relationship tree for the input
query.

3.2.2 Intra-query Relationship Tree

The relationship tree is constructed based on a standard prototype where every
operation (selection, aggregation, ordering and sorting) is represented within an
oval and every database node is represented with a square. Let us consider an
example where a large input Sql query shown in fig.2 gets converted into the re-
lationship tree. The query is used to retrieve the department name (dept name),
department id (dept id) and employee name (empname) where the employee is
of Indian origin with a salary greater than 25000 who belongs to the IT de-
partment and lives in the state Mumbai (MUM). Now given the input query,
we construct the relationship tree based on the IQ-RT algorithm shown in fig.6.
Fig.3 represents the relationship tree for the given input query where D, E, C
represents the tuples which represent the Department, Employee and Country
tables respectively. Here the total number of scans are 7 and requires 7 selection
operations and 4 join operation to get the required query results.Each node in
the fig.3 represents the individual database.

3.2.3 Intra-query Relationship Identifier

The first step in IQ-RI module is to find out the intra-query relationship existing
between the various operations performed by the queries. The intra-query rela-
tionships are based on a set of procedures which are: Procedure 1: The selection
operations on the same database can be grouped so as to eliminate redundant
scans on the same table. Procedure 2: Two join operations involving the same
partition key are combined into one join operation either using OR or AND.
Procedure 3: Aggregation operations involving sum, max, min, avg etc on the
same database can be grouped into one. Procedure 4: Combine individual outer
join and inner join into a single query. Based on these procedures certain oper-
ations are grouped together to avoid redundant scans on the same table and to
 S2MART: Smart Sql to Map-Reduce Translators 575

JOIN

Select d.dept_name,d.dept_id (Select e.empname from JOIN JOIN JOIN

emp e where (d.empno=e.empno)) from (dept d, emp e)

where e.nationality=indian and e.salary>25000 and SEL SEL SEL SEL SEL SEL SEL

d.dept_name=IT and c.state_name=MUM and

c.empno=e.empno; D D E E E C C

Fig. 2. Input Query to SQL Parser Fig. 3. Relationship Tree for Input Query

JOIN
Select d.dept_name, d.dept_id, e.empname from
dept d, emp e, country c where d.empno=e.empno JOIN JOIN

and e.empno=c.empno and e.nationality=indian

and e.salary>25000 and d.dept_name=IT and SEL SEL SEL SEL

c.state_name=MUM; C E D D

Fig. 4. Resultant Query after IQ-RI mod- Fig. 5. Resultant Tree after IQ-RI mod-
ule ule

optimize the query eciency. Now the relationship tree is generated again based
on all the groupings. The query which has been subjected to the IQ-RI module
is shown in fig.4. Now, the IQ-RT is again generated for the query obtained
which is shown in fig.5. The IQ-RT obtained represents 4 selection operations
and 3 join operations totally, hence shows less number of operations compared
to the IQ-RT obtained in fig.3. Hence the number of table scans and operations
that has to be performed against the database has decreased which increases
the throughput and reduces execution time. The algorithm used to find the
intra-query relations and combine similar operations are shown in fig.7.

3.2.4 Sub-query Generator Engine

The input to the sub-query generator engine is the relationship tree which has
been subjected to the IQ-RI module. Now the consolidated query obtained from
the relationship tree is given a sequence number which is used as a reference.
This query is checked with the queries present in the Spiral Database to find

IQRI Generator
IQRT Generator
Input: IQRT, a tree
Input : query, a string create empty tree IQRT_new;
if no_of_table_ids is greater than six then
create empty tree IQRT; if join.sel.table_ids are repeating then
while query parsing till the end do group selection operations on same table_id;
make Join node; if join.partitionkey are same then
use OR or AND operation into one join;
assign Join as root node of IQRT;
if aggregrate operations nodes on same table_id then
if select operation is found in query then group aggregate operations on same table_id;
make select node; combine individual outer join and inner join into a -
assign the input table_id to select node; single query, IQ;
IQRT_new=IQRT(IQ);
do point node to Join root node of query;
GenQl(IQRT_new);
return IQRT; else
GenQl(IQRT)

Fig. 6. IQ-RT Algorithm Fig. 7. IQ-RI Algorithm

576 N. Gowraj et al.

any match. If there is any match then the results of the query is directly given
to the buer or else the query is divided into a number of sub-queries which is
done based on the GenQL algorithm. In order to bring concurrency between the
queries to get the proper results we use the concept of Views in database. After
the sub-queries are generated, each sub-query represents a unique independent
operation performed by the relationship tree. The number of views that has to
be created is equal to the number of sub-queries generated for the same. The
query which has been subjected to the IQ-RI module shown in the fig.4 can be
divided into 4 sub-queries based on the GenQL algorithm shown in fig. 12. Since
the number of sub-queries is 4 the number of views required is also 4. Every
individual query is subjected to a view creation which reduces the amount of
records or data that has to be accessed or referred for further operations on the
table and the view creation can group operations that scans the same tables,
minimizing the number of operations. The sub-queries generated for the query
shown in fig.4 are shown in fig 8, 9, 10, 11.

Select d.dept_name, e.empno, e.empname Select v.empno, v.dept_name, v.empname from

from dept d, emp e, country c where view1 v, emp e where .empno=e.empno and
d.empno=e.empno and e.empno=c.empno; e.salary>25000 and e.nationality=indian;

Fig. 8. Sub-Query 1 Fig. 9. Sub-Query 2

Select v.empno, v.dept_name, v.empname Select v.empno, v.dept_name, v.empname

from view2 v, dept d where from view3 v, country c where
v.empno=d.empno and d.dept_name=IT; v.empno=c.empno and c.state=MUM;

Fig. 10. Sub-Query 3 Fig. 11. Sub-Query 4

3.2.5 Spiral Database

The spiral database acts as the cache storage for the S2MART system which
is used to directly retrieve the most frequently used Sql query and its results.
Eventually, this reduces the data and network transfer cost for the overall system.
The name Spiral database comes from the fact that the Sql queries are stored
dynamically in the database based on the frequency count. In this case we store
the six frequently referred queries along with its results so as to have a less space
complexity for the system. If the number of queries is more than six then we
use the FQR (Frequent Query Retrieval) to perform deletion of unused queries.
Thus we perform an operation which involves dynamic inclusion and exclusion
of queries and its results and hence the name Spiral Database. The input to the
Spiral database is from the buer which stores the results of the query executed.
The results along with the sequence number of the query are stored in the
database for caching purposes. The FQR (Frequent Query Retrieval) algorithm
is implemented within the spiral database to dynamically store the dierent
queries based on the frequency counts. In the FQR algorithm implemented, we
have assumed six as the maximum number of queries that can be stored in the
spiral DB. The number of queries that are stored in the spiral DB can be changed
 S2MART: Smart Sql to Map-Reduce Translators 577

depending on our needs. The work flow of the Spiral Database has already been
discussed in Section 3. The complete work procedure of the FQR algorithm is
depicted in fig. 13.

Frequency Query Retrieval

GenQl Sub_Query Generator
Input: sub_query, a string
Input : IQRT, a tree
no_of_queries = get_count() from spiral_db;
while IQRT parsed do while there are queries in spiral_db and no_of_queries less than six do
if sub_query does not match any of the sub_queries in
identify objects and table fields of same; spiral_db then
if group the table fields with relational operations assign a sequence_no_val pair(autoincrement key, identifier )
then to sub_query;
do query insertion in the spiral_db;
create view for each table id; no_of_queries=getcount() from spiral_db;
do make a select statement with views; if no_of_queries exceeds six then
lru_val=least_recently_used(sequence_no_val);
else do query removal on lru_val in spiral_db;
assign the entire query to query distributor;
else
return sub_query; increment key value of sequence_no_val pair;
retrieve results;

Fig. 12. GenQl Algorithm Fig. 13. FQR Algorithm

3.2.6 Query Distributor and Analyzer

The query distributor and the query analyzer are set up inside the Hadoop
framework where the query distributor represents the master node of the Map-
Reduce framework and the query analyzers represent the worker nodes. The
master node sends the individual sub-query to the query analyzers which are
connected to the various multifaceted database nodes. The general format for
the execution of every individual query is shown in fig.14. The final records are
sent to the intermediate buer once all the sub-queries have been executed, to
undergo the reduce function. The point of using the intermediate buers is to
increase the fault tolerance of the entire system in situations where the buer
is subjected to single point of failure. In case of a bottleneck situation on the
buer, the query results can be retrieved back from the intermediate buer
which eliminates the requirement of performing the entire process once again.
This increases the robustness of the system. A copy of the results are sent back
to the Spiral Database to associate with the sequence number of the query.

3.2.7 Query Optimizer

The query results from the buer in the Map-Reduce framework or results from
the Spiral database are the two possibilities of query input to the query optimizer.
The main function of the query optimizer is to check whether all the sub-queries
which are generated from the relationship tree are executed successfully and
the query results are optimized and are ordered in a presentable manner to be
presented to the user.

3.2.8 Query Results

The final step in the methodology of the S2MART is the query results which
provide the user interface where the query results are presented back to the user.
578 N. Gowraj et al.

4 Experimental Results

We have conducted various experiments to verify the sanity of S2MART. We have

conducted our experiments by comparing our novel system with our benchmarks
Hive and Pig [23]. A cluster with 40 nodes where each node uses a 4GB ram, 250
GB hard disk with 2.1GHz speed is used for implementation of various queries
[24] [25].The size of the database used is 1GB with 7 attributes, 3 databases, and
10,000 tuples. Comprehensive evaluation of our novel system is done by using
various paradigms of queries such as [26] [27] [28]:

4.1 Lightweight Query without Intra-query Relations and

Sub-query Generation
In this case we consider a simple query which cannot be subjected to any of the
modules of the S2MART. The query shown in fig.15 is used to find the employee
name (emp name) from the Employee table where employee id (emp id) is 101.

create view <view name> as ( < Individual Select empname from employee where
sub-query is given here > ); emp_id=101;

Fig. 14. General Format for sub-queries Fig. 15. Simple Lightweight query

The comparison between the various execution time of the query in S2MART,
Hive and Pig is shown in fig.19 where x-axis represents the number of nodes and
y-axis represents the execution time in seconds. The graph shows that there is
no drastic decrease in the execution time for simple light weighted queries.

4.2 Lightweight Query with Intra-query Relationships But No

Sub-query Generation
In this case we consider a query that can be subjected to the IQ-RT and the
IQ-RI modules but it cannot be divided into sub-queries as it is atomic. The
query shown in fig.16 is used to retrieve the Department name (dept name),
Department id (dept id), Employee name (empno) from the tables Employee,
Department and Country where the join operation is performed between the ta-
bles using the primary key of each table. Now the query which has been subjected
to the IQ-RT and IQ-RI modules is shown in fig.17.
Fig.20 shows the comparison between S2MART and the benchmarks. The
graph where x-axis represents the number of nodes and y-axis represents the
execution time in seconds shows that there is gradual decrease in the execution
time of the query when compared to the other systems after 25 nodes.
 S2MART: Smart Sql to Map-Reduce Translators 579

Select d.dept_name,d.dept_id (Select e.empname Select d.dept_name,d.dept_id, e.empname from

from emp e where (d.empno=e.empno)) from employee e , department d and country c where
(dept d, emp e) where e.empno=c.empno; d.empno=e.empno and e.empno=c.empno;

Fig. 16. Not Subjected to IQ-RT & IQ-RI Fig. 17. Subjected to IQ-RT & IQ-RI

4.3 Lightweight Query with Sub-query Generation and No

Intra-query Relations
In this case we consider a consolidated query which can be divided into a num-
ber of Sub-queries but there are no intra-query relationships existing within
the queries. The query shown in fig.18 is used to find the names of the Indian
employees whose Department id (dept id) is 101 from the Employee and the De-
partment table. This query is divided into two sub-queries which are executed
in two dierent views.

5
Select e.empname from employee e, department d
and country c where c.country=indian and 4
Execution Time in Seconds

d.dept_id=101;
3 Hive
Sub-Query 1: Create view view1 as (Select
e.empname from employee e, country c where 2 Pig
c.country=indian); S2MART
1
Sub-Query 2: Create view view2 as (Select 0
v.empname from view1 v, department d where
1 5 10 15 20 25 30 35 40
d.dept_id=101); Number of Nodes

Fig. 18. Subject to Sub-queryGeneration Fig. 19. Comparison of systems - case 4.1

Fig.21 shows the experimental comparison between the systems. The x-axis of
the graph represents the amount of data in thousands and y-axis represents the
execution time of the query. The graph shows that the execution time drastically
decreases after 20,000 records and thus it can be inferred that S2MART can be
very useful when we deal with large amount of data in the data repositories.

4.4 Heavyweight Query with Both Intra-query Relationships and

Sub-query Generation
In this case we consider the same query which is shown in fig.4. The query is
subjected to all the modules of the S2MART which includes IQ-RT, IQ-RI and
Sub-Query Generation. Fig.22 shows the experimental comparison between the
various systems where the x-axis represents the number of nodes and y-axis
represents the execution time in seconds. The graph shows that the execution
time gradually decreases as the number of nodes or the amount of data increases.
We can infer that S2MART can be more ecient than its benchmarks when we
have a complex heavyweight Sql query as the input.
580 N. Gowraj et al.

12 20

10
15

Execution Time in Seconds

Execution Time in Seconds

8
Hive Hive
6 10
Pig Pig
4
S2MART 5 S2MART
2

0 0
1 5 10 15 20 25 30 35 40 1 5 10 15 20 25 30 35 40
Number of Nodes Number of Nodes

Fig. 20. Comparison of systems - case 4.2 Fig. 21. Comparison of systems - case 4.3

25
Execution Time
20
14
Execution Time in Seconds

15 12
Hive
10
10 Pig 8
6 Execution Time
S2MART
5 4
2
0 0
1 5 10 15 20 25 30 35 40 Hive S2MART
Number of Nodes

Fig. 22. Comparison of systems - case 4.4 Fig. 23. Comparing systems of cached data

4.5 Query Results Cached from the Spiral Database

In this case we consider the result of the query to be cached from the Spiral
Database. In this case there is no query execution time spent for the query
because the FQR algorithm finds whether a match is present before the execution
of the query. Fig.23 shows the comparison between the execution time of the
query for Hive and S2MART. We dont consider Pig for the comparison as it
doesnt provide the functionalities of a metastore or caching of data. The graph
shows that the execution time in hive[29] is very much greater when compared
to S2MART.

5 Conclusion
Execution of complex queries involving join operations between multiple tables,
aggregation operations are very much necessary for data intensive applications,
thus we developed a system that could support the concept of big data in data
intensive applications and ease data retrieval to provide useful information. We
implement many novel modules such as Intra-Query relationships (IQ-RT and
IQ-RI), Sub-Query generation and Spiral Database in the S2MART system that
makes the complete system a novel one. We verified the correctness of S2MART
system by performing various experimental results using dierent paradigms of
queries and we achieved eciency in all the experimental results. The intended
applications for S2MART are all data intensive applications which use big data
 S2MART: Smart Sql to Map-Reduce Translators 581

analytics. Hence in the future we can build the system with more ecient query
optimization for a particular application rather than being generic and enable
secure cloud storage [30] for confidentiality and integrity of data.

References
1. Jiang, D., Tung, A.K.H., Chen, G.: MAP-JOIN-REDUCE: Toward Scalable and
Ecient Data Analysis on Large Clusters. IEEE Transactions on Knowledge and
Data Engineering 23(9) (September 2011)
2. Warneke, D., Kao, O.: Exploiting Dynamic Resource Allocation for Ecient Par-
allel Data Processing in the Cloud. IEEE Transactions on Parallel and Distributed
Systems 22(6) (June 2011)
3. Jiang, W., Agrawal, G.: MATE-CG: A MapReduce-Like Framework for Accelerat-
ing Data-Intensive Computations on Heterogeneous Clusters. In: 2012 IEEE 26th
International Parallel and Distributed Processing Symposium (2012)
4. Sakr, S., Liu, A., Batista, D.M., Alomari, M.: A Survey of Large Scale Data Man-
agement Approaches in Cloud Environments. IEEE Communications Surveys and
Tutorials 13(3) (Third Quarter 2011)
5. Han, J., Song, M., Song, J.: A Novel Solution of Distributed Memory NoSQL
Database for Cloud Computing. In: 10th IEEE/ACIS International Conference on
Computer and Information Science (2011)
6. Bisdikian, C.: Challenges for Mobile Data Management in the Era of Cloud and
Social Computing. In: 2011 12th IEEE International Conference on Mobile Data
Management (2011)
7. Nicolae, B., Moise, D., Antoniu, G., Bouge, L., Dorier, M.: BlobSeer: Bringing High
Throughput under Heavy Concurrency to Hadoop Map-Reduce Applications. In:
International IEEE Conference (2010)
8. Zhang, J., Wu, X.: A 2-Tier Clustering Algorithm with Map-Reduce. In: The Fifth
Annual ChinaGrid Conference (2010)
9. Pallickara, S., Ekanayake, J., Fox, G.: Granules: A Lightweight, Streaming Run-
time for Cloud Computing With Support for Map-Reduce. In: IEEE International
Conference (2009)
10. Starkey, J.: Presentation of nuoDB,
http://www.siia.net/presentations/software/AATC2012/NextGen_NuoDB.pdf
11. Starkey, J.: Presentation of nuoDB,
http://www.cs.brown.edu/courses/cs227/slides/dtxn/nuodb.pdf
12. Zhang, Z., Cherkasova, L., Vermam, A., Loo, B.T.: Optimizing Completion Time
and Resource Provisioning of Pig Programs. In: 2012 12th IEEE/ACM Interna-
tional Symposium on Cluster, Cloud and Grid Computing (2012)
13. Olsten, C., Reed, B., Srivastava, U., Kumar, R., Tomkins, A.: Christopher Olsten,
Benjamin Reed, Utkarsh Srivastava, Ravi Kumar, Andrew Tomkins. Presented by
Welch, D.
14. Thusoo, A., Sen Sarma, J., Jain, N., Shao, Z., Chakka, P., Zhang, N., Antony, S.,
Liu, H., Murthy, R.: Hive A Petabyte Scale Data Warehouse Using Hadoop. In:
IEEE International Conference (2010)
15. Michel, S., Theobald, M.: HadoopDB: An Architectural Hybrid of MapReduce and
DBMS Technologies for Analytical Workloads. Presented by Raber, F.
16. Thusoo, A., Sen Sarma, J., Jain, N., Shao, Z., Chakka, P., Anthony, S., Liu, H.,
Wycko, P., Murthy, R.: Hive A Warehousing Solution Over a MapReduce Frame-
work. In: VLDB 2009, Lyon, France (August 2009)
582 N. Gowraj et al.

17. Chaiken, R., Jenkins, B., Larson, P.-A., Ramsey, B., Shakib, D., Weaver, S., Zhou,
J.: SCOPE: Easy and Ecient Parallel Processing of Massive Data Sets. In: VLDB
2008, Auckland, New Zealand, August 24-30 (2008)
18. Yang, C.: Osprey: Implementing MapReduce-Style Fault Tolerance in a Shared-
Nothing Distributed Database. In: ICDE Conference (2010)
19. Isard, Y.Y.M.: DryadLINQ: A System for General-Purpose Distributed Data-
Parallel Computing Using a High-Level Language. Microsoft research labs
20. He, Y., Lee, R.: RCFile: A Fast and Space-ecient Data Placement Structure in
MapReduce-based Warehouse Systems. In: ICDE Conference (2011)
21. Foley, M.: High Availability HDFS,
http://storageconference.org/2012/Presentations/M07.Foley.pdf
22. Chandar, J.: Join Algorithms using Map/Reduce. University of Edinburgh (2010)
23. Chen, T., Taura, K.: ParaLite: Supporting Collective Queries in Database System
to Parallelize User-Defined Executable. In: 2012 12th IEEE/ACM International
Symposium on Cluster, Cloud and Grid Computing (2012)
24. Leu, J.-S., Yee, Y.-S., Chen, W.-L.: Comparison of Map-Reduce and SQL on Large-
scale Data Processing. In: International Symposium on Parallel and Distributed
Processing with Applications (2010)
25. Hsieh, M.-J.: SQLMR: A Scalable Database Management System for Cloud Com-
puting. In: 2011 International Conference on Parallel Processing (2011)
26. Zhu, M., Risch, T.: Querying Combined Cloud-Based and Relational Databases.
In: 2011 International Conference on Cloud and Service Computing (2011)
27. Husain, M.F.: Scalable Complex Query Processing Over Large Semantic Web Data
Using Cloud. In: 2011 IEEE 4th International Conference on Cloud Computing
(2011)
28. Hu, W.: A Hybrid Join Algorithm on Top of Map Reduce. In: 2011 Seventh Inter-
national Conference on Semantics, Knowledge and Grids (2011)
29. Introduction to Hive by Cloudera 2009,
http://www.cloudera.com/wp-content/uploads/2010/01/6-IntroToHive.pdf
30. Wallom, D.: myTrustedCloud: Trusted Cloud Infrastructure for Security-critical
Computation and Data Management. In: 2011 Third IEEE International Confer-
ence on Cloud Computing Technology and Science (2011)
 MisDis: An Efficent Misbehavior Discovering Method
Based on Accountability and State Machine in VANET

Tao Yang1,2,3, Wei Xin1,2,3,*, Liangwen Yu1,2,3, Yong Yang4,

Jianbin Hu1,2,3, and Zhong Chen1,2,3
1
MoE Key Lab of High Confidence Software Technologies, Peking University,
100871 Beijing, China
2
MoE Key Lab of Computer Networks and Information Security, Peking University,
100871 Beijing, China
3
School of Electronics Engineering and Computer Science, Peking University,
100871 Beijing, China
4
School of Software Engineering, Beijing Jiaotong University,
100044 Beijing, China
{ytao,xinwei,yulw,hujb,chen}@infosec.pku.edu.cn,
[email protected]

Abstract. VANET has some prominent features such as channel-opening,

large-scale, fast, dynamic-changing, so there is some misbehavior vehicles. Due
to the large number of vehicles, even if there is only 1% percent of misbehavior
vehicles, there can be thousands of misbehavior vehicles. How to identify these
vehicles quickly, and then use appropriate cancellation and isolation measures
to reduce system losses to its minimum and controllable range. In this paper, we
propose MisDis, a method that can identify misbehavior vehicle by implement-
ing state automata and supervision, as well as a special security log to record
the behavioral characteristics of the target vehicle. If the host vehicle violates
the established security policy, then the system need to record the behavioral
and try to dump to a safer roadside unit or department in charge.

Keywords: VANET, secure log, secure record, accountability, misbehavior

detection.

1 Introduction

VANET (Vehicular ad hoc network) are gaining significant prominence from

both academia and industry in recent years. VANET aims to enhance the safety and
efficiency of road traffic, and have attractive prospects for development. VANET can
improve Peoples livelihood and has a good prospect. We can get better driving expe-
rience, ease the traffic pressure and reduce traffic accidents through the establishment
of VANET with which we constituent network infrastructure environment of "smart
car". Due to the large number of vehicles, even if there is only 1% percent of

*
Corresponding author.

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 583594, 2013.
Springer-Verlag Berlin Heidelberg 2013
584 T. Yang et al.

misbehavior vehicles, there can be thousands of misbehavior vehicles. It also faces

serious security threats because of its huge scale of vehicle number, its variable node
velocity, and its openness. VANET has been strictly constrained by security because
users would not accept or participate it with fear for their safety or personal privacy
leakage. As a result, the VANET security issue becomes a key, urgent, fundamental
and challenging problem, and the research about that is becoming a hot spot.
Vehicles play a basic role in the VANET because they are the original location of
messages generation and transmission. Some types of messages are exchanged
among vehicles such as traffic information, emergency incident notifications, and
road conditions. When something goes wrong with vehicles, such as message leakage
or misbehaviors, the whole VANET will be problematic, and its correctness and ef-
fectiveness cannot be guaranteed. Malicious vehicles can start malfunctioning due to
some internal failures or propagate false identity, position or time information to solve
their selfish motives. For example, an attacker node may create an illusion of traffic
congestion by pretending multiple vehicles simultaneously and launch DoS attack by
impairing the normal data dissemination operation.
There are number of security threats[1-3], which may degrade the performance of
VANET and put the lives of passengers in danger. Misbehaving vehicles perform in
several ways and have different objectives such as attackers eavesdrop the communi-
cation, drop, and change or inject packets into the network. Therefore, one solution is
that vehicles should cooperate together to enhance the security performance of a net-
work. Therefore, security mechanisms, facilities, and protocols are needed to diminish
and to eliminate the attackers effect. In this paper, the terms attacker, illegitimate
behavior, malicious vehicles are synonymously used. However, current research
mainly focuses on the integrity and confidentiality of the communication in the
VANET and seldom cares about the possibility of OBUs abnormal behaviors.
In this paper, we propose MisDis, a method focus on the misbehavior vehicles de-
tection. In general, the contributions of MisDis are:

1. MisDis can introduce accountability to the vehicles behavior by record evidence,

and confirm the property both inside and outside the vehicle.
2. MisDis can identify misbehavior vehicle by implementing state automata and su-
pervision.
3. MisDis designs a special security log to record the behavioral characteristics of the
target vehicle. And if the host vehicle violates the established security policy, then
the system need to record the behavioral and try to dump to a safer roadside unit or
department in charge.

Organization: The rest of this paper is organized as follows: in Section 2, we present

the related work. Section 3, introduces the preliminaries such as system model and
security objectives. Section 4 describes the design rationale and details of our scheme.
Then, the security analysis result is discussed in Section 5. Finally, we will give con-
clusion in Section 6.
 MisDis: An Efficent Misbehavior Discovering Method 585

2 Related Work

There are only few misbehavior detection schemes suggested for finding attacker.
Golle et al. [4] propose a scheme to detect malicious data by finding outlying data.
If there is data that does not fit to the current view of the neighborhood, it will be
marked as malicious. The current view, in turn, is established cooperatively. Vehicles
share sensor data with each other. An adversary model helps to find explanations for
inconsistencies. The main goal is to detect multiple adversaries or a single entity car-
rying out a sybil attack. This work, however, does not provide concrete means to
detect misbehavior but only mentions a sensor-driven detection.
Raya et al. [5] introduced a solution of immediate revocation of certificates of a
misbehaving vehicle and formulated a detection system for misbehavior. It is as-
sumed that the PKI is not omnipresent and hence the need for an infrastructure as-
sisted solution. One component of their work is a local misbehavior detection system
running at each node. This system relies on gathering information from other nodes,
and then comparing the behavior of a vehicle with some model of average behavior of
the other vehicles built on the fly. It is based on the deviation of attacker node from
normal behavior and used timestamp, assigned messages and trusted components
(hardware and software). Honest and illegitimate behavior is differentiated by using
clustering techniques. The basic idea behind the autonomous solution is to evaluate
deviation from normal behavior of vehicles, while they always assume an honest ma-
jority. Once, misbehavior is detected, a revocation of a certificate is indicated over a
base station, a vehicle connects to. This revocation is then distributed to other vehicles
and the Certification Authority itself. But, the assumption of an honest majority of
vehicles does not hold as identities may be generated arbitrarily.
Ghosh et al [6,7] presented and evaluated a misbehavior detection scheme (MDS)
for post crash notification (PCN) application. Vehicles can determine the presence of
bogus safety message by analyzing the response of driver at the occurrence of an
event. The PCN alert raised will be received by multiple vehicles in the vicinity, and
more than one of these receiving vehicles may use an MDS in their OBUs to detect
misbehavior. The basic cause-tree approach is illustrated and used effectively to joint-
ly achieve misbehavior detection as well as identification of its root-cause.
Schmidt et al [8] introduced a framework (VEBAS) based on reputation models.
By combining the output of multiple behavior analysis modules, a vehicle calculates
the trustworthiness value in other vehicles in its near vicinity from the various sensor
values obtained. The value is then used to identify a misbehaving vehicle and build up
reputation. VEBAS requires continuous exchange of information between the ve-
hicles that is used by a vehicle to analyze the behavior of nearby vehicles. Based on
this information, vehicles are classified into trustworthy, untrustworthy or neutral.
Applications may then take this trust rating into consideration in order to react appro-
priately on incoming information. However, the paper only describes the scheme, no
evaluation for any specific application is presented.
A message filtering model is introduced by Kim et al [9] that uses multiple com-
plementary sources of information. According to this model, drivers are sent alert
messages after the existence of certain event is proved. The model leverages multiple
586 T. Yang et al.

complementary sources of information to construct a multi-source detection model

such that drivers are only alerted after some fraction of sources agree. The model is
based on two main components: a threshold curve and a Certainty of Event (CoE)
curve. A threshold curve implies the importance of an event to a driver according to
the relative position, and a CoE curve represents the confidence level of the received
messages. An alert is triggered when the event certainty surpasses a threshold.
PeerReview[10], presented by Haeberlen et al., is a system that provides accounta-
bility in distributed systems. Its useful in accounting for the behavior of every peer
and identifying those peers that break the protocol. It ensures that Byzantine faults
whose effects are observed by a correct node are eventually detected and irrefutably
linked to a faulty node. It also ensures that a correct node can always defend itself
against false accusations. These guarantees are particularly important for systems that
span multiple administrative domains, which may not trust each other. PeerReviews
key idea is maintaining a per-node secure log which records the messages a node has
sent and received, and the inputs and outputs of the application. Any node i (as a giv-
en reference implementation) can request the log of another node j and independently
determine whether j has deviated from its expected behavior. PeerReview is widely
applicable: it only requires that a correct nodes actions are deterministic, that nodes
can sign messages, and that each node is periodically checked by a correct node. To
do this, i replays js log using a reference implementation that defines js expected
behavior. By comparing the results of the replayed execution with those recorded in
the log, PeerReview can detect Byzantine faults without requiring a formal specifica-
tion of the system.

3 Preliminaries

3.1 Definition

First, we define two concepts:

Malicious vehicles: it means that vehicles drop or duplicate packets in network in

order to mislead other vehicles or abolish important messages based on their per-
sonal aims.
Honest (normal) vehicles: A vehicle that has a normal behavior, where a vehicle
with normal behavior forwards messages correctly or generates right messages.

3.2 Tamper Resistant Hardware

A basic approach that aims to prevent attacking over V2V and V2I is the usage of
tamper-resistant hardware. Tamper resistant in general means that something is resis-
tant against tampering, independent of the type of access a person might have to the
system. With respect to hardware, it means the device in question is difficult to be
manipulated or exchanged by another device, without the system would take notice.
The implementation of tamper resistance ranges from complicating access to device
internals up to self-destruction of the device upon the detection of tampering attempts.
 MisDis: An Efficent Misbehavior Discovering Method 587

Closely related are the following two terms, tamper proof and tamper evident. Tam-
per-proof is a more strict definition of tamper resistant, which claims to be 100%
secure against tampering, whereas tamper evident means that one is able to detect that
a device has been tampered with. Tamper resistance means the vehicles devices have
to be secured, and communication buses have to be protected. Furthermore, tamper
resistant devices can provide secure storage for keys and certificates, and maybe even
for a history of recent messages sent over the external communication system, like an
event data recorder (EDR), which could be used for legal purposes. When properly
applied, tamper-proof and tamper-resistant hardware enable secured communication
and will prevent most attacks on the active safety communication system from inside
of the vehicle.

3.3 Time Verification

Time is the most important element in a secure log. Timing information based
verification correlates the time data fields in secure log against the vehicles internal
clock (synchronized and updated using information provided by the GPS system).
Accurate time service can support the upper modules, like Detector and State
Machine. Verifications with the secure log in order to detect misbehavior are possible
with regard to the following aspects. A first step consists in comparing the dump log
time to the creation log time stamp. Time verification with several logs originating
from a single node can provide additional insight on the nodes behavior (or on the fact
another node is trying to impersonate this node). In combination with position
information, time based plausibility checks for single nodes also leads towards
vehicle speed related plausibility checking.

3.4 Secure Log

Security log is the basic data set for vehicle misbehavior detection. Each OBU main-
tains a security log with the following security features: it can only add normal per-
mission, only entities with high privileges, such as RSU and DoT, can dump or
delete the security log.
The secure log entry is defined as follows:
len = ( n, n, n, n)

where:
n: a strictly increasing sequence number;
n: either SEND or RECV which is the category of the log entry;
n: some type-specific content;
n: a recursively defined hash value which is defined as n = H( n-1 || n|| n ||H( n));
|| stands for concatenation and the base hash n-1 is a predefined value. The hash
function H(.) is pre-image resistant, second pre-image resistant, and collision
resistant.
588 T. Yang et al.

The secure log is an append-only list and is secure controlled by DoT. If required,
the log can be dump to the DoT by some security channels. In this paper, we suppose
the secure log about a node vehicle can take enough behavior evidence for judging.

3.5 OBU Model

The OBU model architecture is as shown in Fig. 1.

Fig. 1. OBU components

In general, the OBU i is modeled as a deterministic state machine Si, a detector

module Di, and an application Ai. We use the state machine to denote all the
functionality that should be checked, and use the application to denote other functions
which are unnecessary to check (e.g., sending messages). The detector module Di,
which observes all inputs and outputs of Si, implements MisDis, and it can
communicate with other vehicles detector modules. A normal vehicle i implements Si
and Di as specified; on the other hand, a misbehavior vehicle may behave arbitrarily.
Di sends the failure indications about other vehicles to its Ai. Informally, exposed(j) is
set when vehicle i has obtained enough proof of vehicle js misbehavior; suspected(j)
indicates that i suspects that j does not send a message that it should send; trusted(j) is
raised otherwise. The model have two important database in the different level. One is
the security policy which can directly instruct the OBUs misbehavior checking task,
and another is the secure log which keep the behavior evidence of the host vehicle.
In the core level, there is the OBUs key management module which include the
secret and public key of the OBU, the pseudonyms (related to the vehicles electronic
license plate), the digital certificates, the check token given by RSU and the
Revocation List (RL for short) issued by DoT.
 MisDis: An Efficent Misbehavior Discovering Method 589

In the bottom level, the 802.11p communication module guarantees the interactive
with the other OBUs and RSUs.

3.6 Secure Objectives

We define the security requirements and will show the fulfillment of these require-
ments after presenting the design details.
Accountability: The requirement states that the misbehavior can be detectable by
the evidence kept by the system. We consider the use of accountability to detect
and expose node faults in distributed systems [11]. An accountable system main-
tains a tamper evident record that provides nonrepudiable evidence of all nodes
actions. Non-repudiation also requires that violators or misbehaving users cannot
deny the fact that they have misbehaved. Based on this record, a faulty node whose
observable behavior deviates from that of a correct node can be detected eventual-
ly. At the same time, a correct node can defend itself against any false accusations.
Identificability: An accountable system requires strong node identities. Otherwise,
an exposed faulty node could escape responsibility through assuming a different
identity. In this paper, we assume that each node is in possession of a cryptograph-
ic key pair that can be linked to a unique node identifier. And we also assume that
correct actions are deterministic and that communication between correct nodes
eventually succeeds.

4 Design of MisDis
We expire by Timeweave[12] and employ some ideas of the scheme in PeerRe-
view[11]. The basic idea of MisDis is to set a security log for each vehicle to record
its behavioral state determined behavioral state machines strictly, and introduce a
moderate size of monitoring OBUs which have check-token authorized by DoT. Then
we use this collection to monitor the security log of the target vehicle real-time and
on-site. If we find any exceptions, we should report to DoT through RSU for future
investigation and isolation.
First, Table 1 shows the notation table in our scheme. Next, we elaborate on the
technical design of the proposed scheme as following.

Table 1. Notation Table

Notation Description
Vi ith Vehicle
RSUj jth RSU
RIDi , PIDi Vis true and pseudo ID
s, Ppub DoTs master and public key
+
K RSU j
, K RSU j RSUjs secret and public key
+
K ,K
i i Vis secret and public key
590 T. Yang et al.

Table 1. (Continued)
Notation Description
H(.) Hash function H :{0,1}* *
q

Sign(.) Signature function

Enc(.) symmetric encryption function
T The total database in DoT
T1 The mid-database in RSU1
T2 The mid-database in RSU2

4.1 System Initialization

In the system initialization, the DoT initializes the system parameters in offline man-
ner, and registers vehicles and RSUs.

1) Setup: Given the security parameter k, the bilinear parameters ( q , G1 , G2 , e, P , Q )

are first chosen, where | q |= k . Then DoT randomly selects a master secret
s R Z q ,and computes the corresponding public key Ppub = sP . In addition, the
*

DoT chooses a secure symmetric encryption algorithm Enc(), a cryprographic hash

function H :{0,1}* *q , and a secure signature algorithm sign(). In the end, DoT sets
the public system parameters as ( q , G1 , G2 , e, P , Ppub , H , HMAC , Enc , Sign ) .

2) RSU-KeyGen: DoT randomly select K RSU R Z q* for every RSU, and com-
+
putes the corresponding public key K RSU = K RSU P . DoT sends the pair
+
( K RSU , K RSU ) to the target RSU by secure channel.
3) OBU-KeyGen: Suppose every vehicle has a unique real identity, denote RID.

Vehicle Vi first randomly chooses K i R Z q* as its secret key and computes the
corresponding public key K i+ = K i P . Then Vi submits a four-tuple
( K i+ , RIDi , ai , bi ) to DoT, where ai = H (ti P || RIDi ), bi = (ti K i ai ), ti R
Z
*

DoT verifies the tuples validity by judge a i

= H ((b P + ( K ) a ) || RIDi ) . If pass,
i
+

i
i

+
DoT assigns a PID for Vi, issues a anonymous certification binding K i to Vi.
4) OBU-TokenGen: According to the DoTs history record and the choice policy,
DoT authorizes RSU to issue Check-Token to some OBUs. Each token has a limit
valid time.

4.2 Behavior Logging

Suppose a vehicle i wants to send a message msg to a vehicle j, the process of mes-
sage transmission is as follows:
 MisDis: An Efficent Misbehavior Discovering Method 591

1) i creates a log entry ( n , SEND, j, msg ) , attaches n 1 , n and i ( n || n ) to

msg, and sends the result to j.
2) j has enough information to calculate n and to extract uni : If the signature in
uni is not valid, j discards msg; otherwise, j creates its own log entry
( l , RECV , i, n , msg ) and returns an ack with l 1 , l , and j ( l || l ) to i.

3) i extracts and verifies ulj : If the signature in ulj is valid, transmission

completes; otherwise, i sends a challenge to j's monitors.
From the process above, MisDis ensures that the sender (or the receiver) of each
message msg obtains verifiable evidence that the receiver (or the sender) of msg has
logged the transmission. We focus on two categories of vehicles (i.e., misbehavior
and correct vehicles) and their situations. For misbehavior vehicles, we detect and
identify them; for correct vehicles, we prevent them from false accusations.

4.3 Secure Log Checking

Fig. 2 shows the secure log checking procedure (suppose Vj/Vk have check-token and
meet Vi on the road).

T1 T

Vj

Vi
Vk T2

Fig. 2. OBUs and their monitor OBUs

In MisDis, an authenticator is defined as uni = i ( n , n ) , a signed statement by

vehicle i that its log entry len has a hash value n ; i () represents that the argument
is signed with vehicle i's private key. Through sending uni to vehicle j, a vehicle i
commits to having logged entry len and to the contents of its log before len .
592 T. Yang et al.

If vehicle i subsequently cannot produce a prefix of its log that matches the hash value
in uni , then vehicle j has verifiable evidence that i has tampered with its log and is
faulty. In addition, uni can be used by vehicle j as the verifiable evidence to persuade
other vehicles or RSUs an entry len exists in vehicle i's log. Any OBU or RSU can
also use uni to inspect len and the entries before it in i's log.

The secure log checking includes the following 3 procedures:

1) Hello: OBUi broadcasts Hello-Message periodically to show its existence for the
near vehicle. Hello-Message is [ K i+ ] .
2) Check-Req: After received the Hello-Message from Vi, Vj sends back a Check-
+
Req-Message which is [ K i || check token || Log || T ] s encryption version by a
session key between Vi and Vj. The Check-Req Algorithm is shown as following.

Table 2. Check-Req Algorithm

Algorithm 1: Check-Req Algorithm
1. Vj randomly selects a secret r R Z q* ;
Vj computes a session key = r ( K i ) ;
+
2.
3. Vj computes a hint = rP ;
+
4. Let M = [ K || check token || Log || T ] where Log denotes
i
Log check part, T denotes Timestamp;
5. Let M ' = [ , Enc ( M )] , and send it to Vi.

3) Check-Confirm: After received the Check-Req from Vj, Vi sends a message

include the target log back.

4.4 Log Dumping

The token Vj can dump the secure log part into the RSU1s T1 database through the
same way of Check-Req Algorithm. Furthermore, RSU1 can forward the T1 to DoTs T
database for the thorough misbehavior analysis by secure channel such as SSL.

5 Security Analysis

Accountability: The OBUs secure log keeps the evidence of the vehicle node,
and can be further dump onto the DoTs total database. OBUs core components
are tamper-resistant hardware. Based on this outsourcing record, a misbehavior car
whose observable behavior deviates from that of a correct node can be detected
eventually.
 MisDis: An Efficent Misbehavior Discovering Method 593

Identificability: we assume that each vehicle is in possession of a cryptographic

key pair that can be linked to a unique vehicle identifier. And more, the correct ve-
hicle actions are deterministic and that communication between correct nodes
eventually succeeds.

6 Conclusion

In VANET, each car on the road is equipped with one OBU. In this paper, we intro-
duce MisDis, a method that can identify misbehavior vehicle by state automata and
supervision, as well as a special security log to record the behavioral characteristics of
the target vehicle. If the host vehicle violates the established security policy, then the
system need to record the behavioral and try to dump to a safer roadside unit or de-
partment in charge.

Acknowledgments. I would like to thank Yonggang Wang from Peking University

and Le Cai from UIBE for their helpful advice and kind encouragement. This work
was supported in part by the NSFC under grant No. 61170263 and No. 61003230.

References
1. Aijaz, A., Bochow, B., Dtzer, F., Festag, A., Gerlach, M., Kroh, R., Leinmller, T.: Attacks
on Inter Vehicle Communication Systems - an Analysis. In: Proceedings of WIT, pp. 189
194 (2006)
2. Raya, M., Hubaux, J.P.: Securing Vehicular Ad Hoc Networks. Journal of Computer Secu-
rity, Special Issue on Security of Ad Hoc and Sensor Networks 15(1), 3968 (2007)
3. Grover, J., Prajapati, N.K., Laxmi, V., Gaur, M.S.: Machine Learning Approach for Mul-
tiple Misbehavior Detection in VANET. In: Abraham, A., Mauri, J.L., Buford, J.F., Suzu-
ki, J., Thampi, S.M. (eds.) ACC 2011, Part III. CCIS, vol. 192, pp. 644653. Springer,
Heidelberg (2011)
4. Golle, P., Greene, D., Staddon, J.: Detecting and Correcting Malicious Data in VANETs.
In: Proceedings of the First ACM Workshop on Vehicular Ad Hoc Networks (VANET),
pp. 139151. ACM Press, Philadelphia (2004)
5. Raya, M., Papadimitratos, P., Gligor, V.D., Hubaux, J.P.: On data centric trust establish-
ment in ephemeral ad hoc networks. In: 27th IEEE Conference on Computer Communica-
tions (INFOCOM 2008), pp. 12381246 (2008)
6. Raya, M., Papadimitratos, P., Aad, I., Jungels, D., Hubaux, J.P.: Eviction of Misbehaving
and Faulty Nodes in Vehicular Networks. IEEE Journal on Selected Areas in Communica-
tions, Special Issue on Vehicular Networks 25(8), 15571568 (2007)
7. Ghosh, M., Varghese, A., Kherani, A.A., Gupta, A.: Distributed Misbehavior Detection in
VANETs. In: 2009 IEEE Conference on Wireless Communications and Networking Con-
ference (WCNC 2009), pp. 29092914. IEEE Press, Piscataway (2009)
8. Ghosh, M., Varghese, A., Kherani, A.A., Gupta, A., Muthaiah, S.N.: Detecting Misbeha-
viors in VANET with Integrated Root-cause Analysis. In: Ad Hoc Networks, vol. 8, pp.
778790. Elsevier Press, Amsterdam (2010)
594 T. Yang et al.

9. Schmidt, R.K., Leinmuller, T., Schoch, E., Held, A., Schafer, G.: Vehicle Behavior Analy-
sis to Enhance Security in VANETs. In: 4th Workshop on Vehicle to Vehicle Communica-
tions (V2VCOM 2008), pp. 168176 (2008)
10. Kim, T.H., Studer, H., Dubey, R., Zhang, X., Perrig, A., Bai, F., Bellur, B., Iyer, A.:
VANET Alert Endorsement Using Multi-Source Filters. In: Proceedings of the Seventh
ACM International Workshop on Vehicular Internetworking, pp. 5160. ACM Press, New
York (2010)
11. Haeberlen, A., Kouznetsov, P., Druschel, P.: PeerReview: practical accountability for dis-
tributed systems. In: 21st ACM Symposium on Operating Systems Principles (SOSP
2007), pp. 175188. ACM Press, New York (2007)
12. Maniatis, P., Baker, M.: Secure History Preservation Through Timeline Entanglement, pp.
297312. USENIX Association, Berkeley (2002)
13. NCTUns 5.0, Network Simulator and Emulator,
http://NSL.csie.nctu.edu.tw/nctuns.html
 A Scalable Approach
for LRT Computation in GPGPU Environments

Linsey Xiaolin Pang1,2, Sanjay Chawla1 , Bernhard Scholz1 , and Georgina Wilcox1
1
School of Information Technologies, University of Sydney, Australia
2
NICTA, Sydney, Australia
{qlinsey,scholz}@it.usyd.edu.au,
{sanjay.chawla,georgina.wilcox}@sydney.edu.au

Abstract. In this paper we propose new algorithmic techniques for massively

data parallel computation of the Likelihood Ratio Test (LRT) on a large spatial
data grid. LRT is the state-of-the-art method for identifying hotspots or anoma-
lous regions in spatially referenced data. LRT is highly adaptable permitting the
use of a large class of statistical distributions to model the data. However, standard
sequential implementations of LRT may take several days on modern machines
to identify anomalous regions even for moderately sized spatial grids.
This work claims three novel contributions. First, we devise a dynamic pro-
gram with a pre-processing step of O(n2 ) that allows us to compute the statistic
for any given region in O(1), where n is the length of the grid. Second, we pro-
pose a scheme to accelerate the likelihood computation of a complement region
using a bounding technique. Third, we provide a parallelization strategy for the
LRT computation on GPGPUs. In concert all three contributions result in a speed
up of nearly four hundred times reducing the LRT computation time of large spa-
tial grids from several days to minutes.

Keywords: Spatial outlier, GPGPUs, LRT, 1EXP, upper-bounding.

1 Introduction

With the widespread availability of GPS-equipped smartphones and mobile sensors,

there has been an urgent need to perform large scale spatial data analysis. For example,
by carrying out a geographic projection of Twitter feeds, researchers are able to narrow
down hotspot regions where a particular type of activity is attracting a disproportion-
ate amount of attention. In neuroscience, high resolution MRIs facilitate the precise
detection and localization of regions of the brain which may indicate mental disorder.
The statistical method of choice for identifying hotspots or anomalous regions is the
Likelihood Ratio Test (LRT) statistic. Informally, the LRT of a spatial region compares
the likelihood of the given spatial region with its complement, and hence can be used
to identify hotspot regions. In [10], it was known that the LRT value always follows a
2 distribution, independent of the distribution of the underlying data.
For a n n spatial grid, the worst case execution time for identifying the most
anomalous region is O(cn4 ), where c is the execution time of computing the LRT over
a single region. As noted by Wu et al. [8], a naive implementation of LRT for a moderate

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 595608, 2013.
c Springer-Verlag Berlin Heidelberg 2013
596 L.X. Pang et al.

64 64 spatial grid may take nearly six hundred days.1 Wu et al. [8] proposed a method
which reduces the computation time to eleven days. However as noted previously in [9],
this approach will not scale for larger data sets and the biggest spatial grid reported in [8]
was 64 64.
The nature of LRT permits the computation of regions independently of each other,
which facilitates parallelization to some degree. However, new algorithmic techniques
and parallelization strategies are required in order to fully harvest the computational
power of GPGPUs. We have identified the following challenges that need to be ad-
dressed for achieving a speed up of several orders of magnitude with GPGPUs:

Spatial Grid Exact

DM-GPU BF-GPU LRT
1EXP Family

Spatial Grid
Approx.
Generalised AX-GPU
LRT
Distribution

Fig. 1. Workflow: We distinguish between distributions which belong to the one-parameter ex-
ponential family (1EXP) and distributions outside of 1EXP. We then design two algorithms for
1EXP which are used to efficiently and exactly estimate the LRT. For general data distributions,
we design the AX-GPGPU algorithm which uses the exact algorithms as subroutines.

Computing the LRT of a given region R requires the computation of the likelihood
i.e. the complement of R. While R has a rectangular shape, R
of R, is irregularly
shaped making the parallelization more difficult [11,5,6]. The experiments of [9]
showed that enumerating the complement regions is a bottleneck for efficient com-
putation.
The LRT computation on sequential computers is intractable for large spatial grids.
GPGPUs can perform massively data parallel computations, which is well-suited to
LRT computation. However, the LRT computation must be adapted to the GPGPU
programming model that exhibits a complex memory hierarchy consisting of a grid,
blocks and threads. To minimize the communication between CPU and GPGPU
memory, it is advantageous to perform the majority of the computation on GPG-
PUs using appropriate schemes for dividing work between threads and allocating
threads to blocks.

Our strategy to overcome the challenges noted above is shown in Figure 1. First, in
our framework, we distinguish between data distributions following the one-parameter
exponential family (1EXP) and those that are outside of 1EXP. When the data belongs

1
The experiment results were reported in 2009.
 A Scalable Approach for LRT Computation in GPGPU Environments 597

to a 1EXP distribution, we can estimate the LRT of a given region R by omitting the
computation of R (c.f. [3]). When the data is assumed to be generated by an arbitrary
statistical distribution, we provide a solution to eliminate the bottleneck associated with
the computation of R: we use an upper-bounding technique to replace the computa-

tion of R with that of its constituent regions [8]. Second, we provide three GPGPU
algorithms named as: Brute-Force GPGPU (BF-GPGPU), De Morgan GPGPU (DM-
GPGPU) and Approximate GPGPU (AX-GPGPU). The BF-GPGPU and DM-GPGPU
implementations provide exact solutions for 1EXP. The AX-GPGPU algorithm provides
approximate solution for any underlying data distribution and it uses the BF-GPGPU
or DM-GPGPU as a sub-routine. We designed our blocking scheme for partitioning
the work by dividing the spatial grid into overlapping sub-grids and mapping these re-
gions onto blocks of threads [4]. The majority of the computation is performed on the
GPGPUs and we utilize shared memory for each block by pre-loading the data that will
be used by multiple threads.
The rest of the paper is structured as follows. In Section 2, we provide background
materials on LRT computation and upper-bounding technique. Related work is given
in Section 3. In Section 4, we explain how we use De Morgans law and dynamic pro-
gramming to speed up the enumeration and processing of regions measurements. The
description and details of three designed algorithms (i.e. BF-GPGPU, DM-GPGPU and
AX-GPGPU) are presented in Section 5. In Section 6, we evaluate the algorithms on
both synthetic and real data sets consisting of Magnetic Resonance Imaging (MRI)
scans from patients suffering from dementia. We give our conclusions in Section 7.

2 Background
2.1 The Likelihood Ratio Test (LRT)
We provide a brief but self-contained introduction for using LRT to find anomalous
regions in a spatial setting. The regions are mapped onto a spatial grid. Given a data
set X, an assumed model distribution f (X, ), a null hypothesis H0 : 0 and an
alternate hypothesis H1 : 0 , LRT is the ratio

sup0 {L(|X)|H0 }
=
sup {L(|X)|H1 }
where L() is the likelihood function and is a set of parameters for the distribution [8].
In a spatial setting, the null hypothesis is that the data in a region R (that is currently
being tested) and its complement (denoted as R) are governed by the same parameters.
Thus if a region R is anomalous then the alternate hypothesis will most likely be a
better fit and the denominator of will have a higher value for the maximum likelihood
estimator of . A remarkable fact about is that under mild regularity conditions, the
asymptotic distribution of 2 log follows a 2k distribution with k degrees of
freedom, where k is the number of free parameters2. Thus regions whose value falls
in the tail of the 2 distribution are likely to be anomalous [8].
2
If the 2 distribution is not applicable then Monte Carlo simulation can be used to ascertain
the p-value.
598 L.X. Pang et al.

For example, if we assume the counts m(R) in a region R follow a Poisson distri-
bution with baseline b and intensity , then a random variable x P oisson() is
a member of 1EXP with T (x) = x/, = 1/, a() = , = log(), Be =
m(R) and m(R)
exp(), ge (x) = log(x). For any regions R and R, are independently

Poisson distributed with mean {exp(R )b(R)} and {exp(R )b(R)} respectively. Then
b(R) m(R)
bR = b(R)+b( and mR = m(R)+m(R)
R) . The log-likelihood ratio is calculated by:
c(mR log( m 1mR
bR ) + (1 mR )log( 1bR )) (c.f. [3]). The closed-form formula for LRT
R

for Poisson distribution generalizes to the 1EXP family of distributions [3].

2.2 The Upper-Bounding Technique

The upper-bounding technique for LRT was introduced by Wu et al. [8] to reuse the
likelihood computation, i.e., the likelihood of a region could be upper-bounded in terms
of its sub-regions. The basic observation is that the likelihood value of any given region
R under the complete parameter space is not greater than the product of the likelihood
value of all its non-overlapping sub-regions under the null parameter model. For in-
stance, if a region R is composed of two non-overlapping sub-regions R1 and R2 , then
L(R |XR ) L( R1 |XR1 ) L( R2 |XR2 ). This is equivalent to logL(R |XR )
logL( R1 |XR1 ) + logL( R2 |XR2 ). Here R , R1 and R2 are the maximum likeli-
hood estimators (MLEs). The R is computed under the complete parameter space, and

R1 and R are computed under the null parameter space.
The upper-bounding technique can be used to prune non-outliers as follows: if we
replace the likelihood of a region R by the product of the likelihoods of its sub-regions
and the new LRT is below the anomalous threshold (i.e. confidence level ), then R
cannot be anomalous. As noted earlier, in this paper we use the upper-bounding tech-
nique to speed up the computation of the likelihood of the complement of a region R
when the data is assumed to follow a general statistical distribution.

3 Related Work
Previous attempts to parallelize LRT computation have only achieved limited success.
For example, the Spatial Scan Statistic (SSS), which is a special case of LRT for Poisson
data, is available as a program under the name SatScan [2]. It has been parallelized for
multi-core CPU environments and its extension for a GPGPU hardware [7] has achieved
improved speed up of two over the multi-core baseline. The GPGPU implementation
in [7] has proposed loading parts of the data into shared memory but has achieved only
a modest speed up. The other attempt of [14] applied their own implementation of a
spatial scan statistic program on the GPU to the epidemic disease dataset. The number
of blocks was decided by both the number of the exploring villages and the number of
diseases to take full advantage of stream multiprocessors on GPGPUs. This solution is
only applicable to its special disease scenario. In each of these cases, we believe there is
further room for optimising the algorithms for the GPGPU by carefully exploiting the
thread and block architecture and further utilising shared memory.
Furthermore, all the existing parallel solutions perform simplified Poisson LRT tests
in a circular or cylindrical way, not in a grid-based scenario. Our parallel GPGPU
 A Scalable Approach for LRT Computation in GPGPU Environments 599

solution is different and provides a fully paralleled template for a general LRT com-
putation in a grid, based on the work of [8].

4 The De Morgan Model

For the 1EXP family, we can estimate the LRT of a region R by just computing the
statistic of the region. We do not need to separately compute the statistic of R. For a
more general statistical family, we can upper-bound the likelihood of R by estimating
the likelihood of the four sub-regions that are anchored at the corners of the grid.
We now present a novel and unique approach, based on De Morgans law and dy-
namic programming to accumulate the statistic of a region R. We will build a table in
time O(n2 ) and then we will be able to compute the statistic of any region R in constant
(O(1)) time.
Consider Figure 2(a) that shows a rectangular area R embedded in a grid G. Instead
of counting the number of elements in R directly, we express set R as set intersections
of two sets A and B as shown in Figure 2(b). Set A is a rectangular region that starts
in the upper left corner of the grid and ends at the lower right corner of R. Set B is a
rectangular region that starts at the upper left corner of R and ends at the lower right
corner of the grid. Hence, R = AB. We denote the region from the lower left corner of
G to the lower left corner of R by X and the region from the upper right corner of R to
the upper right corner of G by Y . By applying De Morgans law and inclusion/exclusion
principle we obtain following relationship:

|R| = |A B| (1)
= |G| (|A| + |B| |A B|) (2)
= |G| (2|G| |A| |B| (|X| + |Y |)) (3)
= |A| + |B| + |X| + |Y | |G| (4)

To obtain a query time of O(1), we need to pre-compute sets A, B, X, and Y for

all possible regions in G. Since one of the corner is fixed we can pre-compute the
cardinalities of these sets in tables of size O(n2 ). Hence, the query of counting the

G G
Y
A

R R

X B

(a) Region R in Grid G (b) R= A B , A B = X Y

Fig. 2. Set Relations for Region Problem

600 L.X. Pang et al.

number of elements of region R(x1 , y1 , x2 , y2 ) where (x1 , y1 ) is the upper left corner
and (x2 , y2 ) is the lower right corner is expressed by:
|R(x1 , y1 , x2 , y2 )| = |A(x2 , y2 )| + |B(x1 , y1 )| + |X(x1 , y2 )|
(5)
+ |Y (x2 , y1 )| |G|
To obtain the tables for A, B, X, and Y , we employ dynamic programming. For ex-
ample the table for A can be computed using the following recurrence relationship:

|A(i, j)| = |A(i, j 1)| + |A(i 1, j)| |A(i 1, j 1)|

(6)
+ |R(i, j, i, j)|
where |R(i, j, i, j)| counts whether there is an element in the cell location (i, j). The
first element and the first column and row need to be populated (initialized) so that all
cardinalities of A can be computed.

5 Three GPGPU Algorithms

We present three GPGPU-based algorithms and provide a complete parallel solution
for LRT computation. The Brute-Force GPGPU (BF-GPGPU) and De Morgan GPGPU
(DM-GPGPU) algorithms provide an exact solution of LRT when data is assumed to
be generated from 1EXP. For general data distributions the Approximate GPGPU (AX-
GPGPU) uses the BF-GPGPU and DM-GPGPU as a subroutine to upper bound the
likelihood of the complement region R.
In the GPGPU memory hierarchy architecture, global memory is large and shared
memory is small. Access to shared memory is nearly two orders of magnitude faster
compared to global memory. A common strategy to exploit this mismatch is to partition
the data into subsets called tiles which can fit into shared memory [13,12]. We now
describe our block and thread schemes which underpins the three algorithms.

5.1 Exact Solution for 1EXP

5.1.1 BF-GPGPU
Smart Block and Thread Scheme: We assume that spatial grid3 G (n n) is small
enough so that it can be loaded into GPGPU global memory but too big to be loaded
into shared memory directly. The entire spatial grid G is partitioned into tiles which
can be separately read into shared memory. Assume that the size of cell data is f bytes
and shared memory accommodates s bytes. There are s/f cells or equivalently w h
sub-grids that can be stored into shared memory, where w h = s/f . Each such
rectangular region will be assigned to a block of threads. If the number of threads in
a block is tw th, which is less than the maximum number of threads allowed per
block, then each thread can process (w/tw) (h/th) cells. We use an example with
the Poisson data model to illustrate how the computation is done, but our algorithm also
works with more general distributions.
3
To differentiate our grid in outlier detection with the grid in C UDA, we call the grid in our
C UDA program C UDA grid and the grid in the outlier detection context grid the spatial
grid. We assume the grid size is power of 2.
 A Scalable Approach for LRT Computation in GPGPU Environments 601

(a) (b)
Fig. 3. Block and thread schemes: (a) is applicable for exact computation and pre-computation
of R in AX-GPGPU. Each (x, y) point corresponds to a block coordinate. Each thread is re-
sponsible for processing (w/tw, h/th) regions and each thread processes (2, 2) regions in our
implementation; Figure (b) is applicable to DM-GPGPU and pre-computation of R. The grid is
divided into equal disjoints parts. Each part is associated with a block.

In a Poisson model, each cell has a population count and a success count, which takes
8 bytes. If the size of shared memory is 16KB, a maximum of 2048 cells can be loaded
into the shared memory. This resembles a rectangular region of 45 x 45 cells. To use
threads efficiently, the rectangular region is set to 44 44 cells and associated with a
block. A block with coordinates (x,y) in the C UDA grid is associated to the region with
(x, y)(x+43, y+43) in the spatial grid. In our implementation, the maximum number
of threads per block is limited by the hardware to 512. Therefore, 22 22 threads can
be assigned to process 44 44 cells for maximizing the usage of shared memory. So
each thread is responsible to process four sub-regions: (x1 , y1 , x2 , y2 ), (x1 , y1 , x2 +
1, y2 ), (x1 , y1 , x2 , y2 + 1), (x1 , y1 , x2 + 1, y2 + 1), where (x1 , y1 ) is the upper left cell
and (x2 , y2 ) is the lower right cell. See Figure 3a and algorithm 1.
A two-pass scheme is used to obtain the likelihood ratio of each region:
(a) In the first pass, all rectangular regions of fixed sizes up to 44 44 are enumerated
in the grid. The data for each block is copied to its shared memory. After the block
data is stored into shared memory, each thread computes the score function of its
four rectangular regions and stores it back into global memory.
(b) The second pass enumerates all rectangular regions that do not fit into a single block
on the GPGPU and uses the intermediate results from the first pass to calculate the
top outlier over all regions.

5.1.2 DM-GPGPU
We now illustrate the GPGPU pre-computation step based on De Morgans Law. Rest
of the computation is identical as BF-GPGPU. In our De Morgan model, four datasets
are pre-computed for all possible rectangular regions. From above, we already know
that the pre-computed datasets come from four groups of regions, each group of which
shares one corner with the grid G. The computation of a n n grid requires the pre-
computation of 4 n n regions.
602 L.X. Pang et al.

Algorithm 1 . Exact GPGPUs solution (all regions with size up to (w, h) )

Input: grid G (grid width, grid height), block size (w, h) ,thread size (tw, th)
Output: subset of scores, m R, b R

1: //Step 1: Load sub-grid into shared memory of each block

2: x1 blockIdx.x , y1 blockIdx.y
3: for i 0 to tw do
4: for j 0 to th do
5: x2 blockIdx.x+twthreadIdx.x+i
y2 blockIdx.y+ththreadIdx.y+j
6: if (x2<grid width and y2<grid height) then
7: //Copy region (x1,y1, x2,y2) data from grid G to shared array (SM)
8: SM[(tw threadIdx.y+j) th blockDim.x + th threadIdx.x+i] grid[y
grid width+ x2]
9: synchronize threads();
10: //Step 2: process all the regions within each block
11: for i 0 to tw do
12: for j 0 to th do
13: //Retrieve region (x1,y1, x2,y2) data from SM
14: x2 blockIdx.x +twthreadIdx.x+i
y2 blockIdx.y+ththreadIdx.y+j
15: for tmpx 0 to x2-x1 do
16: for tmpy 0 to y2-y1 do
17: m R accumulate mr(); b R accumulate br();
18: score lrt(m R, b R)
19: save(score); save(m R); save(b R);

To parallelize the pre-computation, we split the whole spatial grid G into equal dis-
joint parts. For a grid of size nn, if the shared memory can load region with maximum
size of w h, we create (n/w) (n/h) non-overlapping parts. Each such disjoint part
is associated to a block. If the thread number in a block is tw th, each thread can
be assigned to process (w/tw) (h/th) regions. Each thread is only responsible for
the regions that share one of the four corners of its correspondent disjoint sub-grid.
The approach for pre-computing R in the next section will be along similar lines. See
Figure 3b.
An additional second pass is carried out to merge the results from each sub-grid to
get the full pre-computed set. In our implementation, each block holds 32 16 cells to
fully utilize the threads capability.

5.2 AX-GPGPU

5.2.1 Block and Thread Scheme

When we cannot assume that data follows 1EXP, the likelihood of R cannot be esti-

mated from the statistic of R. Since R has an irregular shape, a naive implementation
will not be able to leverage the strengths of a GPGPU environment. Experiments in
 A Scalable Approach for LRT Computation in GPGPU Environments 603

a CPU environment [7] have identified that the LRT computational bottleneck comes
from the enumeration of the complement region for each given region.
However we can leverage the strategy proposed by Wu et al. [8] for pruning non-
outliers, to provide an efficient parallelization of R. The likelihood of R is upper-
bounded by the sum of the likelihood of four regions. These four regions share one
of the corner of the grid separately and they constitute R. We use an approach similar

as DM-GPGPU for R computation and BF-GPGPU for R computation.
Each point (x,y) in the spatial grid is the coordinate of a block. To satisfy the tight-
bound criteria and fully utilize the thread capability, our block size is defined as 3216.
Therefore, 32 16 regions are stored into shared memory for processing R. There are
8568 sub-regions to be enumerated in each block. So each thread is responsible for 2
sub-regions. Figure 3b shows the block scheme in pre-computing R. It is different from

Figure 3a. The pre-computation of R only involves the group of regions that share one
of the four corners [8]. Each point (x, y) does not need to be associated to a block, the
entire sub grid is divided into several equal disjoint parts and each part corresponds to a
block. We choose 32 32 regions to be block size for pre-computation of R and there
are (n/32) (n/32) blocks.

5.2.2 Concurrent Execution of Different Kernels

For the general distribution model we have to first compute an bounds for R, R and
then use them to get an upper bound for the LRT. We can break the code up into three
distinct kernels: pre compute r , pre compute r and search. Since the two kernels
pre compute r, pre compute r are independent of each other, we can use streaming
operations to run them concurrently.
After each kernel finishes execution, the results are transferred back to the CPU
where the merging is carried out for bigger regions. Instead of loading the cell values,
kernel search loads the scores of pre-computed sets from the previous two kernels into
shared memory. In our implementation, we can only load a sub-grid of maximum size
32 8 into shared memory. After the final score of each region is obtained, one more
kernel for parallel sorting is executed and the top outlier is obtained.

5.3 Processing Larger Data Grid

All our algorithms assume that the entire spatial grid can be loaded into GPGPU global
memory. This may not be a realistic scenario. For example, in our environment, the
maximum size of the spatial grid that can be processed by a single GPGPU is 128128.
We provide several solutions to overcome this constraint.
Split the Grid into Small Sub-grids. We can split the spatial grid into several sub-
grids and load each of them in a sequential fashion onto a GPGPU. The block and
thread scheme is applied to each sub-grid and the final computation is merged on the
CPU. This may involve multiple iterations of the same kernel function.
Multiple GPGPUs Approach. We combine multiple GPGPU cards into a computa-
tional node. For example, in our case, a single GPGPU implementation can handle a
maximum grid size of 128 128 grid due to the 4GB global memory limitation. For
a 256 256 grid, we can split it into 4 equal 128 128 sub grids and send the grids
604 L.X. Pang et al.

to the two GPGPUs , separately. Merging is carried out on the cpu after the results are
obtained from the two GPGPUs.
Multi-tasked Streaming. C UDA offers APIs for asynchronous memory transfer and
streaming. With these capabilities it is possible to design algorithm that allows the com-
putation to proceed on both CPU and GPGPUs, while memory transfer is in progress.
In our implementation, we use streaming operation to support asynchronous memory
transfer while the kernel is being executed.

(a) (b)

Fig. 4. Figure (a) The running time of CPU implementation of De Morgan vs. brute-force search-
ing. Figure (b) pre-computation time for De Morgan: CPU vs. GPGPUs implementation.

6 Experiments and Analysis

We have designed and implemented a set of experiments to answer the following ques-
tions.

What are the performance gains of the brute-force versus the De Morgan technique
on a sequential CPU?
What are the performance gains of sequential versus parallel outlier detection?
What are the performance gains of sequential versus parallel pruning using our
upper-bounding technique?
How does our technique perform on real-world data sets?
The experiments were conducted on an 8-core E5520 Intel server that is equipped with
two GPGPU T esla C1060 cards supporting CUDA 4.0. Each GPGPU card has 4GB
global memory, 16KB shared memory, 240 cores and 30 multiprocessors. The experi-
ments are performed on a Poisson distribution model and a randomly generated anoma-
lous region was planted for verification in synthetic data sets. To ensure correctness of
the parallel approaches, the results were verified by implementing a sequential version
of all the algorithms which ran on a single CPU machine. For the real data, we used
MRI images of people suffering from dementia and used MRIs of normal subjects as
the baseline.
 A Scalable Approach for LRT Computation in GPGPU Environments 605

(a) (b)
Fig. 5. (a) The running time of exact outlier detection algorithm on a CPU, and single GPGPU.
It shows that BF-GPGPU is faster than the BF-CPU. BF-GPGPU becomes nearly 40 times faster
than BF-CPU on the grid of 128128. (b) The running time of exact searching on single GPGPU
and two GPGPUs.

6.1 Performance on Synthetic Data

6.1.1 De Morgan vs. Brute-Force Processing on 1EXP Family
Figure 4a illustrates the performance of the brute-force method versus the De Morgan
method executed on a single CPU core. The x-axis represents the size of the input
instances and the y-axis the runtime in milliseconds. It is clear that the De Morgan
approach consistently out performs the brute-force and for the largest grid size, the
difference is almost 90 times more.
We compare the overheads of De Morgan pre-computation in a CPU vs. a GPGPU
environment. The results are shown in Figure 4b. For smaller grid sizes, the CPU imple-
mentation has a smaller overhead compared to the GPGPU implementation. However,
as the size of the spatial grid increases, the GPGPU implementation is three times faster
than its CPU counterpart.

6.1.2 Sequential vs. Parallel Outlier Detection

To compare the performance of the brute-force approach on a CPU and a GPGPU
environment for outlier detection, we implemented the BF-CPU and BF-GPGPU al-
gorithms. Experiments were conducted on a on single-core CPU, one GPGPU, and
two GPGPUs. We have applied parallel sorting provided by the C UDA Thrust API.
Figure 5a illustrates the results. The BF-CPU performs well on small data sets but is
clearly outperformed by parallel GPGPU searching for larger dataset. BF-GPGPU is
six times faster than BF-CPU on the grid of size 64 64 and around forty times faster
than on a grid of size 128 128. With the current memory limitation of 4GB, sin-
gle GPGPU implementation cannot process very large data sets. For larger data sets,
multiple GPGPUs become a viable option. Figure 5b depicts the running time of exact
outlier detection algorithm on a single GPGPU versus two GPGPUs without perform-
ing parallel sorting. The gains obtained from using two GPGPUs is around a factor of
three compared to a single GPGPU.
606 L.X. Pang et al.

(a) (b)

Fig. 6. (a) The running time of AX-CPU is nearly 10 times faster than BF-CPU. (b) The running
time of AX-GPGPU is 10 times faster than AX-CPU and almost 400 times than BF-CPU.

(a) Normal subject (b) Dementia subject (c) Baseline subject

Fig. 7. MRI Images: the LRT value of each region in (a) and (b) is compared to a corresponding
region in the baseline image (c). The region in the yellow box in (b) has a LRT value significantly
higher than the baseline (7562), so it is a potential outlier. The corresponding region in (a) has a
low LRT value (2134) so it is normal.

6.1.3 Sequential Pruning vs. Parallel Pruning

Figure 6a and Figure 6b depict the performance gains obtained by using exact pruning
versus approximate pruning on a GPGPU and CPU implementation respectively. As
shown in the figures, the pruning approach accelerates the execution. Furthermore, the
parallelized pruning approach shows a vast improvement with a running time of at least
20 times faster than AX-CPU. Moreover, it is almost 400 times faster than BF-CPU.
The generalized GPGPU pruning approach provides us with a suitable template for fast
computing of the likelihood ratio under any statistical model.

6.2 Outlier Detection on MRI Image

To illustrate a real-world application, the GPGPU implementation was tested on a MRI

data set (OASIS) made available by Washington Universitys Alzheimers Disease Re-
search Center [1]. It consists of a cross-section collection of subjects including indi-
viduals with early-stage Alzheimers Disease (AD). MRIs of a subject suffering from
dementia and two normal subjects were selected as shown in Figure 7. The resolution
 A Scalable Approach for LRT Computation in GPGPU Environments 607

of the MRI data is 176 176 208. Figure 7 shows one slice of each subjects data,
which is 176 208. A large number of voxels are black on the first and last 20 rows,
and the first 40 and the last 40 columns. The area tested has dimensions of 128 128.
The application of LRT using both AX-GPGPU and BF-GPGPU identifies regions
where the subjects suffering from dementia have a high likelihood ratio value for the af-
fected regions in the brain. The same areas of normal subjects have low likelihood ratio
values. The AX-GPGPU found the anomalous region in 24s compared to 1min6s by
BF-GPGPU. The BF-CPU on the other hand required 17min55s to find the anomalous
region.

7 Conclusion

The Likelihood Ratio Test Statistic (LRT) is the state-of-the-art method for identifying
hotspots or anomalous regions in large spatial settings. To speed up the LRT computa-
tion, this paper proposed three novel contributions: (i) a fast dynamic program to enu-
merate all the possible regions in a spatial grid. Compared to a brute force approach, the
dynamic programming method is nearly a hundred times faster. (ii) a novel way to use
an upper bounding technique, initially designed for pruning non-outliers, to accelerate
the likelihood computation of a complement region. (iii) a systematic way of mapping
the LRT computation onto the GPGPU programming model. In concert the three con-
tributions yield a speed up of nearly four hundred times compared to their sequential
counterpart. Until now, existing GPGPU implementations for specialized versions of
LRT such as Spatial Scan Statistic had given only modest gains. Moving the computa-
tion of the LRT statistics to the GPGPU enables the use of this sophisticated method of
outlier detection for larger spatial grids than ever before.

Acknowledgment. We thank Dr. Jinman Kim at the University of Sydney for providing
us with the MRI image resources and helping us analyze the data.

References

1. http://www.oasis-brains.org/
2. SatScan, http://www.SatScan.org
3. Agarwal, D., Phillips, J.M., Venkatasubramanian, S.: The hunting of the bump: On maximiz-
ing statistical discrepancy. In: SODA, pp. 11371146 (2006)
4. Beutel, A., Mlhave, T., Agarwal, P.K.: Natural neighbor interpolation based grid dem con-
struction using a gpu. In: GIS 2010, pp. 172181. ACM, New York (2010)
5. Gregerson, A.: Implementing fast mri gridding on gpus via. cuda. Nvidia Tech. Report on
Medical Imaging using CUDA (2008)
6. Hong, S., Kim, S.K., Oguntebiv, T., Olukotun, K.: Efficient parallel graph exploration on
multi-core cpu and gpu. In: Proceedings of the 16th ACM Symposium on Principles and
Practice of Parallel Programming, PPoPP 2011 (2011)
7. Larew, S.G., Maciejewski, R., Woo, I., Ebert, D.S.: Spatial scan statistics on the gpgpu.
In: Proceedings of the Visual Analytics in Healthcare Workshop at the IEEE Visualization
Conference (2010)
608 L.X. Pang et al.

8. Wu, M., Song, X., Jermaine, C., Ranka, S., Gums, J.: A LRT Framework for Fast Spatial
Anomaly Detection. In: Proceedings of the 15th ACM SIGKDD International Conference
on Knowledge Discovery and Data Mining (KDD 2009), pp. 887896 (2009)
9. Pang, L.X., Chawla, S., Liu, W., Zheng, Y.: On mining anomalous patterns in road traf-
fic streams. In: Tang, J., King, I., Chen, L., Wang, J. (eds.) ADMA 2011, Part II. LNCS,
vol. 7121, pp. 237251. Springer, Heidelberg (2011)
10. Wilks, S.S.: The large sample distribution of the likelihood ratio for testing composite hy-
potheses. Annals of Mathematical Statistics (9), 6062 (1938)
11. Vuduc, R., Chandramowlishwaranv, A., Choi, J., Guney, M., Shringarpure, A.: On the limits
of gpu acceleration. In: HotPar 2010 Proceedings of the 2nd USENIX Conference on Hot
Topics in Parallelism, pp. 237251 (2010)
12. Wu, R., Zhang, B., Hsu, M.: Gpu-accelerated large scale analytics. In: IACM UCHPC 2009:
Second Workshop on UnConventional High Performance Computing (2009)
13. Wu, R., Zhang, B., Hsu, M.C.: Clustering billions of data points using gpus. In: IACM
UCHPC 2009: Second Workshop on UnConventional High Performance Computing (2009)
14. Zhao, S.S., Zhou, C.: Accelerating spatial clustering detection of epidemic disease with
graphics processing unit. In: Proceedings of Geoinformatics, pp. 16 (2010)
ASAWA: An Automatic Partition Key Selection Strategy

Xiaoyan Wang1,2,3, Jinchuan Chen3, and Xiaoyong Du1,3,*

1
School of Information, Renmin University of China
2
School of Information and Electrical Engineering, Ludong University
3
Key Laboratory of Data Engineering and Knowledge Engineering, MOE
{wxy,jcchen,duyong}@ruc.edu.cn

Abstract. With the rapid increase of data volume, more and more applications
have to be implemented in a distributed environment. In order to obtain high
performance, we need to carefully divide the whole dataset into multiple parti-
tions and put them into distributed data nodes. During this process, the selection
of partition key would greatly affect the overall performance. Nevertheless,
there are few works addressing this topic. Most previous projects on data parti-
tioning either utilize a simple strategy, or rely on a commercial database sys-
tem, to choose partition keys. In this work, we present an automatic partition
key selection strategy called ASAWA. It chooses partition keys according to
the analysis on both dataset and workload schemas. In this way, intimate tuples,
i.e. co-appearing in queries frequently, would be probably put into the same
partition. Hence the cross-node joins could be greatly reduced and the system
performance could be improved. We conduct a series of experiments over the
TPC-H datasets to illustrate the effectiveness of the ASAWA strategy.

Keywords: partition key selection, data partitioning.

1 Introduction

Nowadays, more and more applications need to manage huge volume of data and
serve large amount of requests concurrently. As it is too hard for a single machine to
support the whole system operations, these applications are mostly installed on clus-
ters containing multiple data nodes. Thus, the whole data in an application needs to be
partitioned and then stored in different machines. The goal of a partitioning algorithm
is to achieve low response time and high throughput by minimizing inter-node com-
munications.
During data partitioning, partition key selection, which would select attributes for
tables as the baseline in partitioning, is the most important decision made regarding
the scalability of database [1]. Data in different partitions are serviced by different
processing nodes, and concurrent Read and Write operations over multiple partitions
will be a bottleneck for applications. Poor partition key selection can then greatly
affect query performance [2]. For example, an application, which has partitioned data

*
Corresponding author.

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 609620, 2013.
Springer-Verlag Berlin Heidelberg 2013
610 X. Wang, J. Chen, and X. Du

by the Location attribute but accesses data with queries containing many joins on the
Date column, would undoubtedly has an expensive communication cost.
Previously, there have been many efforts on the problem of data partitioning [1, 2,
5-8]. However, few of them address the problem of partition key selection. Instead,
most of them utilize some simple strategy, while some others rely on a commercial
database system for this selection.
In this work, we discuss how to select partition keys automatically for tables in ap-
plications and propose ASAWA (Automatic Selection based on Attribute Weight
Analysis) strategy. Its general idea is based on the observation that if an attribute is
found to be accessed in many queries, it would probably act as keys in join opera-
tions. Thus it would be better to partition datasets according to the attributes in search
condition defined in query statements. The basic process of ASAWA would first
compute an aggregated score for each attribute by identifying the data dependency
among tables and finding out associated attributes, and then select partition keys with
top scores evaluated by grouping features in workload schema. Combined both data
and workload information predefined, it is expected to reduce communication cost
among data nodes and improve system performance.
The contributions of this paper are summarized as follows:
1. We propose an automatic partition key selection strategy.
2. We implement the strategy within a novel algorithm for choosing partition
keys based on data and workload information.
3. We conduct a series of experiments on TPC-H [3] to show both the efficiency
and effectiveness of our proposed approaches.
The remainder of this paper is organized as follows. Section 2 presents related works
in data partitioning with particular focus on partition key selection. Next, Section 3
illustrates the automatic partition key selection method ASAWA. Using three kinds of
partition key selection strategies, Section 4 describes the performance comparison to
show the efficiency of ASAWA. Finally, we have briefly summarized the contribu-
tions in this work and discussed works in the future in Section 5.

2 Related Works

Data partitioning [4] is known as one of the most famous technologies for its efficien-
cy in large-scale application scalability. Traditionally, there are 3 kinds of partitioning
strategy: horizontal partitioning [5], vertical partitioning [6] and hybrid approaches
[7], any of which needs to select a partition key for each relation as the basis of parti-
tioning algorithms. In recent years, there have emerged some new researches, which
do not need to select partition keys. Schism [8] uses Metis [9] to divide mapping
graph of relations in data into fine-grained partitioning results. In [2], it has proposed
a method to have a skew-aware partitioning solution based on large-scale neighbor-
hood search. It would like to select the most frequently accessed column in the work-
load as the partition key for each table originally, and change it when a better one
 ASAWA: An Automatic Partition Key Selection Strategy 611

founded during the process of relaxation. As followed by the partitioning strategies,

when there are too many columns in a large amount of tables, the search space would
be fairly huge. Designed in a bottom-up view, both of them need to be recalculated
and partitioned if there is incremental data. Consequently, this kind of solutions is
much more suitable for the applications with fixed data and workloads.
The ideal solution is supposed to partition data without manual assignment, while
the performance is as well as the best-known. In most research works, they assume
[10] that the proper partitioning keys have been chosen, and concentrate on partition-
ing algorithms with less aware on partition key selection. Although many traditional
commercial database systems (e.g. Oracle [11], DB2 [12], SQL Server [13], etc.) have
mentioned the importance of partition key selection and have their own mechanisms
for recommendation, most of them are not open sources and combined deeply with
underlying database optimizers. Therefore, they could not be applied on others direct-
ly. Based on some guidelines [14, 15], it is mostly operated manually in cases that the
data management architecture is new or open source.
Basically, there are four kinds of partition key selection strategies:

Primary Key Selection: This is the most widely used partition key selection
strategy as default in commercial database systems. One of its variations is to
use foreign key or the first column in a table as its partition key.
Random Selection: This occurs when partition keys are specified without suffi-
cient database management experiences or by systems without any constrains.
Exhaustive Selection: This would like to try all possible partition key groups to
have the one with best performance. It needs to use a sample data and workload
set to simulate as in the real world, which is hard to predict the cost correctly.
Analytical Selection: This is based on data and workload schema analysis,
which is usually applied manually, or recommended by optimizing advisors in-
tegrated within expensive commercial database systems.

As data amount increasing rapidly and workload requests variously, this work would
concentrate on automatically partition key selection from a top-down view that suits
the requirements from data and workload. It prompts ASAWA method for automatic
partition key selection, which not only takes the data and workload schemas into con-
sideration, but also combined with design and execution information.

3 Partition Key Selection

To illustrate the problem of partition key selection, consider the following motivating
scenario: Given a data center with many query workloads, it needs to build a data
management configuration that consists of many storage nodes and to partition data
to suitable nodes so that the workloads could be executed in parallel. As the first step
in data partitioning, a critical problem is how to select partition keys for relationships
in the data center to minimize communication costs as much as possible.
612 X. Wang, J. Chen, and X. Du

3.1 Problem Description

For a well-designed application, the basic information of the system, like data sche-
ma, workload pattern and their utility features, is given before data partitioning. In
this work, we focus on applications with well defined data schema and static work-
load pattern which could be obtained with access frequencies for queries in it.
The partition key selection problem can be generally formulated as follows:
Definition. Partition Key Selection: Given a database D = {R1,R2,,Rd}, a query
workload W = {Q1,Q2,,Qw}, and a storage bound S = {B1,B2,,Bs}, select an
attribute PKSx (1 x ) for each table as the partition key. Here,

Rx (1 x ) = <Ax1, Ax2,, Axt> is a predefined table schema with a scale factor

coefficient Cx in database D. Axi (1 i ) is one of the attributes in Rx.
Qy (1 y ) is a predefined query pattern with an access frequency AFy in work-
load W,
Bz (1 z ) is the storage bound of node z.

The total execution time Ttotal consists of two parts: data loading time Tload and the
whole workload execution time Texec. In this work, we assume all the data nodes are
the same in physical features and use the widely used Hash function h1:S [0,1] to
distribute datasets into data nodes uniformly. The best partition key selection solution
should partition data to nodes and complete queries within the least response time.

3.2 ASAWA Method

The basic process of ASAWA would first compute an aggregated score for each
attribute based on identifying the dependency among tables and finding out the cor-
responding associated attributes, and then select partition keys with top scores opti-
mized by grouping features in workload schema.

Database Analysis
Data Schema
As mentioned in Section 3.1, we assume that the data schema is well defined to guar-
antee data integrity for the database and all of the basic information (including
attribute name, data type, constraints in the table, etc.) could be obtained.
To guarantee the efficiency of ASAWA, we would like to select the attributes with
data integrity declarations as partition key candidates, including Primary Key, For-
eign Key and columns declared Unique or Not Null. The weight of affect in dimension
of data schema for an attribute could then be assigned equally as the values of
wDS(PK), wDS(FK), wDS(U) and wDS(NN) respectively for each of declarations men-
tioned above. The values of weights could be the same or different, which depend on
specific applications. The existences of the 4 kinds of statements are present as
eDS(PK), eDS (FK), eDS (U) and eDS (NN) respectively, whose value is 1 or 0 based on
the integrity statement on the relative attribute.
 ASAWA: An Automatic Partition Key Selection Strategy 613

For a given attribute in the data schema, its weight in dimension of data schema is
shown as wDS(GA) in Formula (1), which is the product of the existences that have
defined integrity statements on it and the values of weights in a relative sequence:
wDS(GA) = (wDS(PK) eDS(GAPK) + wDS(FK) eDS (GAFK)
+ wDS(U) eDS(GAU) + wDS(NN)) eDS (GANN)) (1)

Scale Factor.
A scale factor is the number used as a multiplier in scaling. This is mainly used in
data generation. In ASAWA, the weight of affect for a given attribute in dimension of
scale factor is as in Formula (2):

wSF(GA) = 1 + log ( SFx / SFMIN ) (2)

in which SFMIN = min{ SF1, SF2 ,, SFd }.Obviously, min(wSF(GA))=1.

In some applications, each tables volume might increase in the same growth rate,
which means that the scale factor for each of them is 1. Thus, wSF of each table is 1.In
most of real world applications, tables volumes are not increased in the same pace.
For example, in TPC-H, the scale factor coefficient of table Lineitem is 6000000,
while Orderss is 15000000 and Suppliers is 10000. According to Formula (2),
wSF(AMIN) = wSF(ASUPPLIER) = 1, wSF(ALINEITEM) 3.78, and wSF(AORDER) 4.18.
An extreme case is that, there are tables that too tiny to be worth partitioning. For
example, in an OLAP application with Star Schema, the volume of fact tables in-
creases fast while dimension tables always stay the same. These tiny tables are consi-
dered do not have scale factor. Instead of partitioning, they should be replicated for
considerations on system performance.
Thus, the total weight of a given attribute in database is present in Formula (3):
wDATABASE(GA) = wDS(GA) wSF(GA) (3)

Workload Analysis
Query Pattern
Query pattern is a piece of code that will make a reasonably normal query expression
compile. As mentioned in Section 3.1, we assume that each query would be executed
using the same structure with variable parameters. This ensures that the attributes
touched in each query is the same in every execution of it. As mentioned in [10], The
most relevant operations that present the attributes touched in SQL statements include
equi-joins, Group By operations, duplicate elimination, and selections. The weights of
them in dimension of query pattern are as wQP(Join), wQP(GB), wQP(DR),
wQP(SelectConstant) and wQP(SelectHostVariable) relatively, while the existences as
eQP(Join), eQP (GB), eQP (DR), eQP(SelectConstant) and eQP(SelectHostVariable)
respectively, which are the counts of their appearance in the query.
614 X. Wang, J. Chen, and X. Du

For a given attribute in a given query, the weight of it in dimension of data schema
is shown as wQP (GAGQ) in Formula (4):
wQP(GAGQ) = (wQP(Join) eQP(Join) + wQP(GB) eQP (GB)

+ wQP(DR) eQP (DR) + w QP(SelectConstant) eQP(SelectConstant)

+ wQP(SelectHostVariable) eQP(SelectHostVariable)) (4)

Access Frequency
Access frequency represents how many times an application runs in a given period.
Its measurement is decided by the designer, e.g. an hour, a day, et al. In ASAWA, we
assume the access frequency is steady for queries in the workload and define that the
weight of affect in dimension of access frequency is as in Formula (5):
wAF(Qy) = 1 + log ( AFy / AFMIN ) (5)

in which AFMIN = min{ AF1, AF2 ,, AFw }.Obviously, min(wAF(Qy))=1.

For example, we assume that the access frequency of Q1 and Q11 is 7000 and 150
respectively during an hour in TPC-H, while the minimized access frequency is 3 on
Q21. According to Formula (6), wAF(QMIN) = wAF(Q21) = 1, wAF(Q1) 3.37 and
wAF(Q11) 2.70.
Thus, the total weight of a given attribute in workload is present in Formula (6):
w
wWORKLOAD (GA) = ( wQP (GAQy ) w AF (Q y )) (6)
y =1

According to the above analysis, the general weight of a given attribute w(a) could be
calculated as presented in Formula (7):

w(a) = wDATABASE(a) wWORKLOAD(a)

w
= wDS (a) wSF (a) ( w QP (aQy ) w AF (Q y )) (7)
y =1

ASAWA Algorithm
Although there have been many partition key selection strategies for manual operation
tips, how to combine these tips for automatic partition key selection is a problem.
Here we propose ASAWA (Automatic Selection based on Attribute Weight Analysis)
method to solve this issue.
The whole ASAWA algorithm is listed in Fig.1. Combining both data and work-
load information predefined, it takes attributes with data integrity constrains into con-
sideration to reduce the complexity in practice and is expected to reduce communica-
tion cost among data nodes and benefit for system performance.
 ASAWA: An Automatic Partition Key Selection Strategy 615

4 Experimental Study

In this section, we have shown results of our experiments using different partition key
selection strategies and compared with the results with analysis.

Input:
1. Data schema of the designed database
2. Scale Factors of tables defined in the data schema
3. Query Pattern of the designed workload
4. Access Frequency of each query in the workload
Output:
Solution PKSD = {PKS1, PKS2,, PKSx} (1 x ), in which
PKSj (1 j x) is an attribute in table Tx D.
Procedure ASAWA:
1. Eliminate attributes without impact and obtain par-
tition key candidates with its weight w(a) for each.
a. Extract the attributes with data integrity con-
strains from the data schema.
b. Assign the weight of each attribute in dimen-
sion of data schema, scale factor, Query Pattern
and access frequency as w(a) using Formulate (8).
2. Exhaustively consider all combinations of
attributes that are not eliminated:
a. Call optimizer with candidate set and grouping
to estimate execution times of queries.
b. Estimate aggregate Response time of all que-
ries, and update {best partition key, relation
grouping} to the current one to choose the best.
3. Output PKSD = { PKS1, PKS2, , PKSx} (1xd)
xd) over D
as the partition key for the given data schema.
Fig. 1. The ASAWA Algorithm

4.1 Environment and Configuration

In the experiment, we have setup 3 data nodes, each of which is equipped with
2.66GHz * 8 cores, more than 300G disk, and CentOS 6.1 x86-64 with Open SSH 5.2
inside. As the scenario mentioned in Section 3, we select Greenplum Database
(GPDB for short) 4.2.0.0 [26] as the platform. The dataset and queries used are gener-
ated from TPC-H 2.14.2, which has well-defined data schema and query patters to
illustrate a system that examine large volumes of data, execute queries with a high
degree of complexity, and give answers to critical business questions. The node with
32G main memories is chosen as the master node, while the other two nodes have
16G main memory and are set as slave nodes. Each of them has 8 segments.
616 X. Wang, J. Chen, and X. Du

4.2 Experimental Procedure

A complete experimental procedure is shown in Fig. 2. It would be executed in every
round of each strategy on a generated dataset with a specific scale factor. To be fair,
this procedure should be repeated in each round of every case.
Input:
1. A specific scale of datasets generated from TPC-H
2. Queries defined in the benchmark
Output:
1. Data partitioning time
2. Executing time of each query
3. Total time of the whole process
Procedure:
1. Analysis the partition key for each tables based on
applied strategy within the round.
2. Generate a new TPC-H dataset with the specific
scale factor
3. Partition data to data nodes based on keys se-
lected, use Hash function h1:S[0,1]
[0,1] to distribute it
to nodes and record this time period as data loading
time Tload.
4. Execute specific query set for this round and
record this time period as query execution time Texec.
5. Record the whole process executing time as Ttotal
when the last query finishes its execution.
Fig. 2. One Round of Experimental Procedure in Each Case

It needs to select partition keys based on different strategies in the first step. In this
work, we implement three partition key selection strategies for comparison, which are
based on Primary Key, Random Assigned and ASAWA respectively. The detail de-
scriptions of the first two strategies could be found in Section 2. For ASAWA, we set
the weights in dimension of query pattern as: wQP(Join) = 1.0, wQP(GB) = 0.1,
wQP(DR) = 0.08, wQP(SelectConstant) = -0.05 and wQP(SelectHostVariable) = 0.05 as
in [10]. In order to present the underlying structure information in the schema, we
assign wDS(PK) = wDS(FK) = 1 and wDS(U) = wDS(NN) = 0.5 respectively for
wDS(GA) in this work. Then we could run Procedure ASAWA shown in Fig. 1 for
partition key selection for case ASAWA with the specific scale factor in this round.
In the TPC-H Schema, there are two tables named Nation and Region respectively
that are not affected by scale factors. For the size for both of them are fixed and quite
limited, instead of partitioned, both of them are copied to each data node in every
selected strategy case for performance consideration.

4.3 Results and Analysis

As a matter of fact that the servers may have many different networks and application
status, the running result is most probably different from each other, more or less.
 ASAWA: An Automatic Partition Key Selection Strategy 617

In order to get a steady result and have a much more reliable result analysis, each
strategy with relative dataset and workload in a specific scale should be executed
multi round in the same physical environment. Due to the limitation of current hard-
ware environment, the scale factor of dataset is set to be 1, 10 and 100 respectively.
We run query set 5 times within the newly partitioned dataset in every round.

Overall Results
As a decision support oriented benchmark, it is hard to predict the access frequency of
queries in TPC-H. We first all of the 22 queries in the data set of scale factor =1, 10
and 100 respectively. The results are shown in Figure 3.

(a) scale factor = 1 (b) scale factor = 10 (c) scale factor = 100
Fig. 3. Executing Time of 22 Queries

From the series figures shown in Fig. 3, we could see that in each scale factor
round, Case 2: Random Strategy always runs the longest time and shows the worst
performance, Case 1: Primary Key Strategy is faster than Case 2 and shows fine per-
formance, while Case 3: ASAWA is even better than Case 1 and always shows the
best performance in the 3 scales of datasets. Besides the results shown in Figure 3, we
have also recorded data loading time in each round of every case. In these records,
Case 2 shows obviously longer data loading time than the others.
The result verifies again that the performance in queries is affected by the key se-
lection switch. Also, we found that compared with Case 1, the advantages of Case 3
in Figure 3(c) are not so distinct than it in Figure 3(a) and 3(b). This is because GPDB
has implemented a mechanism named Motion to do live migration and eliminate par-
tition differences during running time, which means the long time running, the less
distance. But this would add extra workload to data nodes and are not common im-
plemented in every product.

Specific Query Results

The series of figures shown in Fig. 4 indicate the average executing time of each
query in the data set of scale factor =1, 10 and 100 respectively. From these figures,
we could see that for most TPC-H queries in each scale factor round, Case 2 runs the
618 X. Wang, J. Chen, and X. Du

longest time and shows the worst performance, Case 1 shows fine performance than
Case 2 especially in large scales, while Case 3 is shows the best. This indicates that
the partition strategy should take the actual workloads of applications into considera-
tion to get much better performance. Due to identify tuples specifically but having not
considered the business semantical information, Primary Keys strategy always obtains
reasonable but limited performance. For improper partition keys selection, Random
strategy leads to bad system performance. Having taken data and workload features
into consideration, ASAWA shows the best performance in all of these 3 strategies.

(a) scale factor = 1

(b) scale factor = 10

(c) scale factor = 100

Fig. 4. Executing Time of Each Query
 ASAWA: An Automatic Partition Key Selection Strategy 619

Besides, we have intentionally selected Q1, Q5 and Q13 from the 22 queries of
TPC-H, in which Q1 is represented for queries on a single large table, Q5 for queries
on multi tables and Q13 for queries only on small tables, to take a continuous running
for 7 rounds respectively on the dataset with the sale factor set as 100. The result
shown in Fig.5 indicates that, for the queries that data distributed originally and mem-
ory could not cached, the executing time does not change too much, like Q1 and Q5,
while for the queries that memory cacheable as Q13, the executing time is the most at
the first time, less in the second time, and lest and maintaining in the following 5
times. It also shows the downtrend slightly in Q1 and Q5. Due to the automatic mo-
tion operation (like live migration) in GPDB, Round 3-7 show steady results. Ob-
viously, if similar queries could be scheduled closer, which our strategy want to
achieve, executing performance would be improved a lot.

Fig. 5. Continuous Running of Single Query (scale factor = 100)

5 Conclusion
As more and more applications need to manage data in huge volume and response
large amount of requests concurrently, the whole dataset for an application need to be
partitioned into multiple parts and put into different machines. During this process,
the selection of partition key would greatly affect the partitioning results. Neverthe-
less, there are few works addressing this topic. In this work, we present an automatic
partition key selection strategy, called ASAWA. It would first compute an aggregated
score for each attribute based on identifying the dependency among tables and finding
out the corresponding associated attributes, and then select partition keys with top
scores optimized by grouping features in workload schema. In this way, close tuples,
i.e. co-appearing in queries frequently, would be probably put into the same partition.
Hence the inter-node joins could be greatly reduced and the system performance
could be improved. We then conduct a series of experiments over TPC-H queries with
datasets generated in different scale factors. The experiment results illustrate the ef-
fectiveness of ASAWA method.
For most applications with huge data volume and great concurrent requests, parti-
tioning strategy should be well planed before data storage and would be difficult to
change once it adopted. If it could be adjusted automatically according to the actual
applications, this would be undoubtedly more meaningful and useful for the widely
use of database system products and release the burden of database administrators or
developers. Partition key selection is the first step in data partitioning, and we only
620 X. Wang, J. Chen, and X. Du

use empirical weight values assigned in ASAWA in the experiments. In the future
work, we would like to adjust the values by learning methods to get a more practical
weight set. Of course, there must be some differences among application features and
visions. We would like to explore an automatic partitioning method combined with
both ASAWA and proper partitioning algorithms in further research.

Acknowledgements. This work is supported by State Key Laboratory of Software

Development Environment Open Fund under Grant No.SKLSDE-2012KF-09, the
National Science Foundation of China under Grant No. 61003086, and the Graduate
Student Scientific Research Foundation of Renmin University of China under Grant
No. 42306176. The authors would like to thank the anonymous reviewers for their
helpful comments and valuable suggestions.

References
1. Zilio, D.C.: Physical Database Design Decision Algorithms and Concurrent Reorganiza-
tion for Parallel Database Systems. PhD Thesis, Department of Computer Science, Univer-
sity of Toronto (1998)
2. Pavlo, A., Curino, C., Zdonik, S.: Skew-aware Automatic Database Partitioning in Shared-
Nothing Parallel OLTP Systems. In: Proc. of the ACM SIGMOD, pp. 6172 (2012)
3. TPC BenchmarkTM H, http://www.tpc.org/tpch/
4. Stonebraker, M., Cattell, R.: 10 Rules for Scalable Performance in Simple Operation Da-
tastores. Communications of the ACM 54, 7280 (2011)
5. Ceri, S., Negri, M., Pelagatti, G.: Horizontal Data Partitioning in Database Design. In:
Proc. of the ACM SIGMOD, pp. 128136 (1982)
6. Navathe, S., Ceri, G., Wiederhold, G., Dou, J.: Vertical Partitioning Algorithms for Data-
base Systems. ACM Transactions on Database Systems 9(4), 680710 (1984)
7. Agrawal, S., Narasayya, V., Yang, B.: Integrating Vertical and Horizontal Partitioning into
Automated Physical Database Design. In: Proc. of the ACM SIGMOD, pp. 359370 (2004)
8. Curino, C., Jones, E., Zhang, Y., Madden, S.: Schism: a Workload-Driven Approach to
Database Replication and Partitioning. Proc. of the VLDB Endowment 3, 4857 (2010)
9. Metis, http://glaros.dtc.umn.edu/gkhome/views/metis/index.html
10. Zilio, D.C., Jhingran, A., Padmanabhan, S.: Partition Key Selection for a Shared-nothing
Parallel Database System. Technical Report RC 19820(87739) 11/10/94, IBM T. J. Wat-
son Research Center (1994)
11. Eadon, G., Chong, E.I., Shankar, S., Raghavan, A., Srinivasan, J., Das, S.: Supporting Table
Partitioning by Reference in Oracle. In: Proc. of the ACM SIGMOD, pp. 11111122 (2008)
12. Zilio, D.C., Rao, J., Lightstone, S., Lohman, G., et al.: DB2 Design Advisor: Integrated
Automatic Physical Database Design. In: Proceedings of the VLDB, pp. 10871097
(2004)
13. Nehme, R., Bruno, N.: Automated Partitioning Design in Parallel Database Systems. In:
Proc. of the ACM SIGMOD, pp. 11371148 (2011)
14. zsu, M.T., Valduriez, P.: Principles of Distributed Database Systems, 3rd edn. Springer,
New York (2011)
15. Rahimi, S., Haug, F.S.: Distributed Database Management Systems: A Practical Approach.
IEEE Computer Society, Hoboken (2010)
16. Greenplum Database,
http://www.greenplum.com/products/greenplum-database
 An Active Service Reselection Triggering
Mechanism

Ying Yin, Tiancheng Zhang, Bin Zhang, Gang Sheng, Yuhai Zhao, and Ming Li

College of Information Science and Engineering, Northeastern University

Shengyang, 110004, China
[email protected]

Abstract. Actively identifying service faults and actively trigger service

reselection are important management methods for eciently promot-
ing system reliability. Composite Web services, however, are often long-
running, loosely coupled and cross-organizational applications. They
always run in a highly dynamic and changing environment(The Web)
which imposes many uncertainties, such as server unavailable or network
interruption or temporarily interrupt and so on. Under these uncontrol-
lable circumstances, it is impractical to monitor the changes in Quality
of Service parameters for each and every service in order to timely trigger
service reselection, due to high computational costs associated with the
process. In order to overcome the above problem, this paper proposes an
ecient reselection mechanism by mining early patterns in advance. The
system will trigger service reselection once potential execution failure is
detected by matching these early patterns with the current situation. The
process is real time and low-cost, and as such, the proposed mechanism
will improve system reliability.

Keywords: Early Pattern, Prediction, Service Reliability.

1 Introduction

The advantage of composite Web services is that it realizes a complex appli-

cation by connecting multiple component services seamlessly. However, in real
applications, Web service lives in a highly dynamic environment, both the net-
work condition and the operational status of each of the constituent Web Ser-
vices(WSs) may change during the life-time of a business process itself. The
instability brought on by various uncertain factors often makes the composite
services to be interrupted or interrupted temporally. Therefore, it is very impor-
tant to ensure the normal execution of the composite service applications and
to provide a reliable software system[11].
Reselection mechanisms for failed services are important technologies which
ensures the reliability of composite services. On the other hand, reselection of
services triggered by interruptions may not be suitable as the pre-computing
of the Quality of Service (QoS) is costly both in time and financial cost. The
system can identify the unreliable service by predicting potential interrupting

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 621628, 2013.
c Springer-Verlag Berlin Heidelberg 2013

622 Y. Yin et al.

factors in advance and then avoids the exceptional service interruption. That is,
prediction has gradually become a valid method to trigger service reselection [9].
Prediction is realized by some algorithms which monitor and evaluate the quality
of composite services. There are several prediction methods, such as ML-based
methods known in literature[2], QoS-aware based methods[3,4] and collaborative
filtering-based methods[5,9] and so on.
Despite the existence of previous work, prediction methods still three major
challenges remain, as defined below. First of all, most of the traditional prediction
models, such as Support Vector Machines (SVM) and Artificial Neural Networks
(ANN), can only take input data as vectors of features [12]. However, there are
no explicitly features in sequential data. Secondly, prediction models based on
QoS monitoring, such as Naive Bayes and Markov model[10], which assume that
sequences in a class are generated by an underlying model M and the probability
distributions are described by a set of parameters. These parameters however,
are obtained by predicting the QoS during the whole lifecycle of the services and
will therefore lead to high overhead costs. Another method is prediction based on
sequence distance, such as collaborative filtering[9], this type of methods require
the definition of to define a distance function to measure the similarity between
a pair of sequences. However, how to select an optimal similarity function is far
from trivial, as it will introduce numerous parameters and measures for distances
can be subjective.
In order to maintain the reliability of composite services, an important thing
for us is actively identify exception patterns and trigger service reselection algo-
rithm quickly. In this paper, we proposes an ecient service reselection mech-
anism by mining previous patterns in advance. The system will trigger service
reselection once the currently executing sequence matches any mined historical
sequence indicating imminent failure.
The remainder of this paper is organized as follows: Section 2 discusses the
system architecture. In Section 3, we present the process of extracting execution
instances from execution logs while in Section 4 the basic related definitions
are given, and an ecient method for identifying early patterns based on an
interest measure is proposed, and methods for implementation are presented. We
summarize our research and discuss some future work directions in Section 5.

2 System Architecture
Below we describe the system architecture of our service reselection trigger-
ing mechanism. After accepting a users query specification, our BPEL Engine
should be able to provide a composite service execution sequence. Once some
latent exception is identified during execution, our system will trigger the rese-
lection process.
There are four phases in the trigger process: (1) recording the composite
service execution log information to generate the execution service sequence
repository; (2) cleaning collected data and converting service log into service
execution sequences; (3) mining early prediction patterns from converted execu-
tion sequences and deducing predictive rules; (4) triggering the service reselection
 An Active Service Reselection Triggering Mechanism 623

process as soon as an executing service sequence matches one of the early predic-
tion patterns. Figure 1 shows the whole reselection triggering process and below
each of the steps will be briefly discussed.

Busine ss Pruning
Proce ss Early Rules Rules

Pruning
Measure Function
Selection/ Y/N?
Matching Early Factor
Reselection
Conciseness Factor

Execution
Engine Monitor S equence

Preparing
Web services
Execution Log

Fig. 1. The framework

Recording of composite service execution log. During this phase, the system needs
to collect large amounts of failed composite service execution information, aiming
to mine useful knowledge and to analyze the causes of failure reason so as to
trigger service reselection as quickly as possible. The generated composite service
execution log repository forms the basis for early prediction.
Extracting of the service execution sequence. the original service log includes a
large variety of information, such as Web Service Description Language (WSDL)
information, keywords, IP address, QoS and so on. The data needs to be cleaned
and QoS information needs to be transferred into sequences so as to mine early
pattern eectively. For the functionally similar Web services, even for each Web
service itself, perceptive QoS values may dier at dierent time points or dierent
network environments thus presumably indicating the dierence in quality. That
means that the quality of Web service might dierent in dierent situation.
Therefore, extracting descriptive information from the service log and converting
it into forms that can readily be mined are the key part of our methods. The
details of this step will be given in Section 3.
Mining of early prediction pattern. Various behavioral patterns in dierent envi-
ronment can be lead to composite service execution failure. However, both the
time when the fault occurs and the length of the behavioral pattern influence
when the service reselection will be triggered. Under the various environments,
mining triggering patterns with minimal cost and triggering the reselection as
earlier as possible are the important issues in our work. The contents of this part
will be described in Section 4.
624 Y. Yin et al.

Triggering of Reselection. These mined execution sequence patterns can be used

to timely detect whether the composite service deviates from normal execution.
This kinds of knowledge should be applied to improve the reliability of the
composite service. Based on these mined rules, how to construct an ecient
search tree aiming to trigger the reselection process becomes an more important
issues in promoting the reliability. The details of this part will be described in
the further work.

3 Preliminary
Once a service-oriented application or a composite service is deployed in a run-
time execution environment, the application can be executed in many execu-
tion instances. Each execution instance is uniquely identified with an identifier
(i.e., id). In each execution instance, a set of service sequences can be triggered.
Various uncertain factors may make a execution instance fail. Reason for this
phenomenon is that there exists a large number of potential exception status
information under dierent uncertain factors. We record triggered events in a
log which contains dierent types of information. We are interested in failed
QoS information. It is helpful for service quality management by extracting the
execution status information from execution log to predict the service reliability.

Table 1. Composite QoS status Information

A CompositeQoS Description
S0 < av 0 , exe0 > Server unavailable, runtime delay
S1 < av 0.5 , exe0 > Server available intermittently, runtime delay
S2 < av 1 , exe0 > Server available, runtime delay
< av 0 , exe1 > Server unavailable, normal execution
S3 < av 0.5 , exe1 > Server available intermittently, normal execution
S4 < av 1 , exe1 > Server available, normal execution

In order to depict various exception status information clearly, for each com-
ponent service, we only consider two QoS attributes as an example, such as the
availability attribute (AV ) and the execution time attribute(EXE). There are
three possible states for AV , namely inaccessible, intermittently accessible and
accessible, which are denoted by av 0 , av 0.5 and av 1 respectively. EXE on the
other hand can only signify two states, namely exe0 and exe1 representing de-
layed execution and normal execution respectively. Under these situations, we
can obtain five possible groups of service execution statuses information which
are shown in Table 1: < av 0 , exe0 >, denoted by S 0 , represents server unavail-
able but runtime delay, < av 0.5 , exe0 >, denoted by S 1 , represents server avail-
able intermittently but runtime delay, < av 1 , exe0 >, denoted by S 2 , represents
server available but runtime delay, < av 0.5 , exe1 >, denoted by S 3 , represents
server available intermittently but normal execution and < av 1 , exe1 >, denoted
by S 4 , represents server available but normal execution. However, for the status
 An Active Service Reselection Triggering Mechanism 625

information < av 0 , exe1 >, denoted by S 0 , represents server unavailable but nor-
mal execution. We do not consider this status, as it cannot occur in
practice.

Table 2. An example of service execution instances

ID ExecuteLogEntry Count Class

1 S10 3 f ailed
2 S1 , S31
1
3 f ailed
3 S1 , S22 , S32
2
2 f ailed
4 S11 , S21 , S31 3 f ailed
5 S1 , S22 , S32 , S42
4
2 f ailed
6 S14 , S22 , S32 2 f ailed

Based on the description assumption above large collections of composite

service execution records D which are stored as sequential data as shown in
Table 2. Let S={S1 , S2 ,...,Sn }, then each record consists of one or more com-
ponent services from S. Note that every component service will have dierent
QoS expression values, which can evaluate describe the execution status of that
component service. For simplify, we let the capital letters with subscript, such
as S1 , S2 , S3 , S4 , denotes dierent abstract service class and the superscript,
denotes dierent QoS statuses of the component service. Table 2 only shows a
service execution data with 15 failed users(not including successful users) and
four abstract service class. Bellow are some definitions related.

4 Early Prediction Pattern Mining

In order to classify the each service sequence status into dierent fault types,
we need to extract features from execution service log. Due to the unsupervised
nature of the service log data, the first step is to mine those patterns that are
eective for predicting services to be failed.

4.1 Basic Definition

In this section, we introduce some of the basic concepts for mining statuses and
give the problem definition.

Definition 1. Pattern. Let D be an execute service log with service set S,

S={S1 ,S2 ,...,Sn }. Let P = {S1i S2j ...Slk } S(l=1,2,...,n) be a component service
or a subset of the execution sequence with dierent status information, skl denotes
the k th status information of the ith service, where {i, j, k} [0, 1, 2, 3, 4]. A
component service with status information or a set of component pairs is called
a Pattern. .

Given a pattern p=si1 sj2 , ...pkm , we say pattern p appears in a sequence s if there
exists an index 1i0 n m + 1 such that ai0 =si1 , ai0 +1 =sj2 ,..., ai0 +m1 =skm .
626 Y. Yin et al.

For example, p = {acd} appears in sequence s = {aacdf gacde}. However, a

pattern p may appear multiple times in a sequence. Below, we give the minimal
prefix length definition.

Definition 2. Earliest Matching Pattern. Given Pattern p and sequence s,

where p s. The Earliest Matching Pattern(EMP) is the distance from initial
position of s to the first occurrence of p in s, and is denoted by EM P (P, S).

Definition 3. Earliest Support. Let p be a pattern. The Earliest Support(ES)

of pattern p is the ratio of the sum of the reciprocals of the earliest position of p
to the number of instances in the data set, i.e.:
# 1
pS EMP (p,S)
ES(p) =
S

Problem Statement. For a given failed service log entry, a Pattern p should
satisfy: (1) Pattern p should frequently be present in failed type; (2) Pattern p
should be as short as possible and should have low prediction cost; All patterns
satisfy the above conditions called Early Patterns. the optimal rules are deduced
called Early and Concise (EC) Rule Sets.

4.2 Early Pattern Mining Algorithm

Previous existing pattern mining methods require the definition of various thresh-
olds as quality measures such as support, confidence , such that only those
relatively frequent rules are selected for classifier construction. However, these
mining methods are ineective in practice and the set of mined rules are still
very large. It is very hard for a user to seek out appropriate rules. In our case,
we consider the earliness and conciseness properties to design a utility measure
and only preserve those early, concise patterns passing the measure from the
training database.
The mining process is conducted on a prefix tree. Limited by space, we omit
the description of constructing such a prefix tree structure. A complete pseudo-
code for mining optimal EP sets is presented in Algorithm 1.
Finding interesting contrast sequence rules by checking all subsequence along
the paths is still a tough task. Various pruning rules are introduced below to
prevent unnecessary rule generation and only preserve interesting rules. Reduc-
ing the search space are extremely important for eciency of EP. In algorithm
1, we use two ecient pruning rules to improve eciency.

Pruning Rule 1. Pruning by ES.

Given ES(Definition 3), Pattern pa denotes pattern p and all its possible proper
supersets. If 0! EP (p)! , then pattern p and its supersets will not be an
optimal rule.
 An Active Service Reselection Triggering Mechanism 627

Algorithm 1. Mining the EP

Input: data set D
Output: Pattern sets EC
1: Set EC =
2: Count numbers of 1-patterns
3: Generate 1-pattern set
4: Calculate ES of 1-pattern
5: Select 1-pattern respectively and add them to EC
6: new pattern set Generate (2-pattern set)
7: while new pattern set is not empty do
8: Calculate ES(p) of candidates in new pattern set
9: For each pattern p in the (l+1)-pattern set
10: Apply pruning rule 1: IF EP (p)<
11: remove pattern EP;
12: Else if there is a superpattern pa in l-pattern set
13: Apply pruning rule 2: that EP (pa)= EP (p)
14: Then remove pattern P;
15: Select EP to EC;
16: ENDIF
17: end while
18: new pattern set Generate(next level pattern sets)
19: Return R;

Note: The presentation of pruning rule 1 describes the early support of a pat-
tern, not the general support. Because in a exception execution log, pattern p
may emerge several times. We only record the first position so as to trigger ser-
vice reselection procedures as soon as possible. This is dierence with traditional
association rules.

Pruning Rule 2. Pruning by concise.

If pattern p satisfy EP (p)= EP (pa), where pa denotes its proper superset, then
pattern pa and all its possible proper supersets will not be useful for EP.

Limited by space, we omit the proof. The above pruning rules are very ecient
since it only generates a subset of frequent patterns with maximal interestingness
instead of all ones.

5 Conclusion
In this paper, we discuss the mining of early prediction patterns which are im-
portant for predicting service reliability, and propose two interest measures in
order to decide whether the rules are concise and as short as possible in length for
quickly identifying exception patterns. Based on the proposed interest measures,
we propose a new algorithm with ecient pruning rules to mine all optimal EP
rules.
628 Y. Yin et al.

Acknowledgments. This work was supported by a grant from the National

Natural Science Foundation of China under grants (No.61100028, 61272182,
61272180, 61073062); the New Century Excellent Talents in University Award
(NCET-11-0085); the Ph.D. Programs Foundation of Ministry of Education of
China(young teacher)(No.20110042120034); the Fundamental Research Funds
for the Central Universities under grants (No. N110404017, N110404005); China
Postdoctoral Science Foundation (No. 20100481204, 2011M500568, 2102T50263);

References
1. Yu, Q., Liu, X.M., Bouguettaya, A., Medjahed, B.: Deploying and managing Web
services: issues, solutions, and directions. The VLDB Journal 17, 537572 (2008)
2. Wang, F., Liu, L., Dou, C.: Stock Market Volatility Prediction: A Service-Oriented
Multi-kernel Learning Approach. In: IEEE SCC, pp. 4956 (2012)
3. Goldman, A., Ngoko, Y.: On Graph Reduction for QoS Prediction of Very Large
Web Service Compositions. In: IEEE SCC, pp. 258265 (2012)
4. Tao, Q., Chang, H., Gu, C., Yi, Y.: A novel prediction approach for trustworthy
QoS of web services. Expert Syst. Appl. (ESWA) 39(3), 36763681 (2012)
5. Lo, W., Yin, J., Deng, S., Li, Y., Wu, Z.: Collaborative Web Service QoS Prediction
with Location-Based Regularization. In: ICWS 2012, pp. 464471 (2012)
6. Yu, T., Zhang, Y., Lin, K.J.: Ecient Algorithms for Web Services Selection with
End-to-End QoS Constraints. ACM Trans. Web 1(1), 125 (2007)
7. Leitner, P., Michlmayr, A., Rosenberg, F., Dustdar, S.: Monitoring, Prediction and
Prevention of SLA Violations in Composite Services. In: ICWS, pp. 369376 (2010)
8. Zheng, Z., Ma, H., Lyu, M.R., King, I.: QoS-Aware Web Service Recommendation
by Collaborative Filtering. IEEE Transactions on Service Computing (TSC) 4(2),
140152 (2011)
9. Chen, L., Feng, Y., Wu, J., Zheng, Z.: An Enhanced QoS Prediction Approach for
Service Selction. In: IEEE Conference on Web Service, ICWS (2011)
10. Park, J.S., Yu, H.-C., Chung, K.-S., Lee, E.: Markov Chain Based Monitoring
Service for Fault Tolerance in Mobile Cloud Computing. In: AINA Workshops, pp.
520525 (2011)
11. Zhang, L., Zhang, J., Hong, C.: Services Computing. TSingHua University Press,
Beijing (2007)
12. Han, J., Kamber, M.: Data Mining: Concepts and Techniques, 2nd edn. The Mor-
gan Kaufmann Series in Data Management Systems. Morgan Kaufmann Publishers
(March 2006)
 Linked Data Informativeness

Rouzbeh Meymandpour and Joseph G. Davis

School of Information Technologies

The University of Sydney
{rouzbeh,jdavis}@it.usyd.edu.au

Abstract. By leveraging Semantic Web technologies, Linked Open Data pro-

vides an extensive amount of structured information in a wide variety of
domains. Principles of Liked Data facilitate access and re-use of semantic in-
formation, both for human and machine consumption. However, information
overload due to the availability of a large amount of semantic data, as well as
the need for automatic interpretation and analysis of the Web of Data require
systematic approaches to manage the quality of the published information. Iden-
tifying the value of information provided by Linked Data enables a general un-
derstanding about the significance of semantic resources. This can lead to better
information filtering functionalities in Semantic Web-based applications. The
aim of this paper is to propose a novel approach, derived from information
theory, to measure the informativeness in the context of the Web of Data. We
experiment with the metric and present its applications in a variety of areas, in-
cluding Linked Data quality analysis, faceted browsing, and ranking.

Keywords: Semantic Web, Linked Data, the Web of Data, DBpedia, Semantic
Network, Partitioned Information Content, Information Filtering.

1 Introduction

The Semantic Web is a major evolution to the current World Wide Web (WWW) that
provides machine-readable and meaningful representation of Web content. Formal
knowledge representation languages such as RDF (Resource Description Framework)
[1] and OWL (Ontology Web Language) [2], and query languages such as SPARQL
[3] are key enablers of semantic applications. These technologies enable the Web of
Data, a giant graph of interconnected information resources, known as Linked Data.
Based on the principles of the Linking Open Data (LOD) project, an extensive
amount of structured data is now published and available for public use [4]. This in-
cludes interlinked datasets in a wide variety of domains that range from encyclopedic
knowledge bases such as DBpedia1, which is the Semantic Web-based representation of
Wikipedia2 [5], to domain-specific datasets such as Geo Names3 and Bio2RDF4.

1
http://dbpedia.org
2
http://wikipedia.org
3
http://www.geonames.org
4
http://bio2rdf.org

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 629637, 2013.
Springer-Verlag Berlin Heidelberg 2013
630 R. Meymandpour and J.G. Davis

According to the Linking Open Data cloud5, more than 295 datasets in various domains,
including Geography, Publications, Life Sciences, and so on are now interconnected
together. This valuable source of semantic knowledge can be exploited in practical and
innovative applications. It has also introduced several research opportunities.
Due to the fact that most of Linked Datasets have been created by automatic or
semi-automatic extraction of structured data from Web pages, assessing the quality of
the Web of Data has become a critical challenge for semantic data consumption.
Moreover, not all available Linked Data is valuable for semantic software agents or
human understanding. For example, it is likely that among all relations available in
the Web of Data relating to a particular movie, its director and actors are more inter-
esting to most users than its language or runtime. Likewise, for a software agent that
compares two movies, having a same director conveys more information than sharing
a same language. Understanding the informativeness of resources and structured data
associated with them facilitates browsing as well as automatic consumption and inter-
pretation of Linked Data. This paper, introduces two novel metrics to evaluate the
amount of information that a resource and its relations carry. After proposing the
metrics, we present a set of experiments conducted in various domains and applica-
tions in order to demonstrate their characteristics.

2 Measuring Informativeness in Linked Data

In this section, we first introduce the concept of informativeness measurement accord-
ing to its mathematical theory. Then, by formalizing the definition of Linked Data, we
propose metrics to assess the value of information in the Web of Data.

2.1 Measuring the Information Content

Information Theory describes the mathematical foundation of communication for
transmitting an information source over a communication channel by means of cod-
ding schemes [6, 7]. Shannon [8] proposed entropy as a measure of information con-
tained in a process. He showed that entropy, also known as Information Content (IC)
or Self-Information, is the amount of binary symbols or bits that the receiver needs in
order to recreate the original transmitted process. Derived from the probability theory
paradigm, Information Content is defined as the engendered amount of surprise [9]:
(1)
where is the probability of presence of term or concept in a given corpora. The
logarithm is often calculated to the base 2. Thus, the unit of information is bits. For
the base 10 and natural logarithms, the unit is bans and nats, respectively.
By looking closely at this metric, we can describe its important characteristics. The
amount of surprise evoked by occurrence of a particular term has a negative correla-
tion with its probability. More popular terms, with a higher chance of presence, are
more probable and therefore, there is less surprise in their occurrence. In other words,
less frequent entities are more specific and informative than common terms.

5
http://lod-cloud.net
 Linked Data Informativeness 631

2.2 Formal Description of Linked Data

In order to be able to present a systematic approach for Linked Data analysis, we

mathematically explain and formalize the Web of Data based on a combination of
graph and set theories. Linked Data is a labeled directed graph consists of resources as
nodes and relations as edges that are identifiable by their URIs (Uniform Resource
Identifiers). Thus, nodes, edges, and the directions of edges have to be considered for
a proper representation of Linked Data and its resources:
Definition 1 (Linked Data Graph): Linked Data is a labeled directed graph LD, de-
fined as , where is a set of resources, and
is a set of links, in which is defined as , where is the link that connects
resource to resource .
Therefore, incoming and outgoing edges can be used to describe a resource:
Definition 2 (A Resource in Linked Data): A resource in Linked Data LD is
represented as its set of features , which is defined as:
(2)

2.3 Informativeness of Resources in Linked Data

By deriving from the formal definition of Linked Data and the principles of informa-
tion theory, we propose a metric to measure the value of information associated with
resources in the Web of Data. According to the Definition 2, each resource is
represented as its set of characteristics, which is a collection of its incoming and out-
going edges. Based on the basis of the probability theory, we define the surprise value
associated with a resource in Linked Data:
Definition 3 (Probability of a Resource in Linked Data): For a resource such as
, represented as a set of features , the probability of is

(3)

Thus, we have

(4)

Based on this, we define the Partitioned Information Content:

Definition 4 (Partitioned Information Content for Linked Data): Information Content
of a resource is defined as the sum of the Information Content values of its features:

(5)

where IC of each feature is calculated using the Equation (1):

632 R. Meymandpour and J.G. Davis

(6)

such that is the frequency of the feature and is the frequency of the most
common feature in a given Linked Dataset.
In this definition, the amount of information conveyed by a given resource is calcu-
lated according to the probability of its features. Popular features carry less informa-
tion while specific ones are more informative. This metric also allows us to study the
Web of Data in terms of the value of information contained in each resource and do-
main. It highlights the most informative and influential information sources among a
large number of available resources in a particular domain.

2.4 Informativeness of Relations in Linked Data

Another key component of Linked Data is relations, also known as properties. Under-
standing the importance and informativeness of relations allows better automatic in-
terpretation and enhanced filtering of Linked Data. From the users perspective, it
supports information browsing and visualization by identifying the most significant
properties in a particular domain. The informativeness of a given property is indicated
by measuring the value of information produced by the relation in a given context:
Definition 5 (Generated Information Content of Relations): Let and be
the set of all features and all links in the domain , respectively, and let be
the subset of all features that have as their link:

(7)

The domain could be the whole Linked Data graph or a subset of it that is filtered by a
search query. This metric enables us to indicate the most important relations that con-
vey the highest amount of information about resources inside a particular domain.

3 Experiments

3.1 Experimental Context

The main Linked Dataset used in our experiments was DBpedia. We downloaded the
main DBpedia Ontology as well as five fundamental datasets in English, including
Ontology Infobox Properties, Ontology Infobox Types, Article Categories, SKOS
Categories, and YAGO Links. Finally, by simultaneous executing of a number of
SPARQL queries, we extracted instances and their features from the DBpedia dataset
in three domains: Music, Actors, and Films (see Table 1).
 Linked Data Informativeness 633

At the time of writing this paper (October 2012), the latest version of DBpedia was
3.8, released on August 2012. In our final DBpedia dataset, the parameter in the
Equation (6) was set as the number of (rdf:type, owl:thing) relations (2,350,906).

Table 1. Different domains and their size characteristics

Films Music Actors
rdf:type a dbo:Filmb dbo:MusicalArtist dbo:Band yago:Actor c
Resources 71,715 34,246 27,061 31,482
Unique Features 237,429 640,157 236,184
Total Triples 1,729,718 2,302,008 1,002,901
a
rdf: is the prefix for the namespace http://www.w3.org/1999/02/22-rdf-syntax-ns#
b
dbo: is the prefix for the namespace http://dbpedia.org/ontology/
c
The actual property value is http://dbpedia.org/class/yago/Actor109765278

3.2 Distribution of Information in Different Domains

The first experiment was performed using the Partitioned Information Content-based
metric in order to compare information distribution across several domains in DBpe-
dia. It is observable that information in the Music domain is more distributed than the
others, with standard deviation of 590.2 in comparison to 148.2 for Films and 437.6
for Actors (see Fig. 1 (a) and Fig. 2). Moreover, there is more distinctive information
available for musical artists and actors than films (Fig. 1 (b)). This may be explained
by the fact that each artist has an increasing number of outputs during their career.

(a) (b)

Fig. 1. (a) Distribution of IC and (b) Cumulative distribution of IC among DBpedia resources

Further experiments were performed in order to compare the informativeness of

semantic data provided by various Linked Datasets. LinkedMDB6 (Linked Movie
Database) is a specialized open Linked Data for movies [10]. It provides structured
information for 85,600 movies described using more than 6 million triples7. The expe-
riments illustrate a wider distribution of information in LinkedMDB in comparison to
6
http://www.linkedmdb.org
7
According to the latest downloadable version of LinkedMDB, released on February 2012.
634 R. Meymandpour and J.G. Davis

the Films dataset extracted from DBpedia (see Fig. 2). Comparing these figures un-
derline the difference between the qualities of structured information provided by
DBpedia, which is a general purpose Linked Data, against LinkedMDB.

600
500
400
300
200
100
0
Films Music Actors LinkedMDB
Average (bits) 282.5 485 369.4 157.8
Standard Deviation (bits) 148.2 590.2 437.6 198.9

Fig. 2. Information distribution among instances in various domains

3.3 Applications to Faceted Browsing and Information Filtering

Information filtering techniques help users to overcome data overload by facilitating

browsing and exploring the knowledge base. Facets improve semantic data browsers
by providing suggestions according to the currently viewed item. We applied the
proposed metrics to discover the most relevant resources to a given instance. Table 2
lists the top 10 recommended artists and bands when the user browses a specific
domain.

Table 2. Applications of PIC for faceted browsinga

rdf:type dbo:MusicalArtist dbo:genre db:Country_musicb dc:subject db:RockandRoll_HFcde
Artist Name PIC (bits) Artist Name PIC (bits) Artist Name PIC (bits)
David Bowie 19,008.9 Bob Dylan 14,901.7 Bob Dylan 14,901.7
Prince 16,220.9 Van Morrison 11,652.0 Van Morrison 11,652.0
Bob Dylan 14,901.7 Elvis Presley 11,096.9 Elvis Presley 11,096.9
The Beatles 12,337.7 Dolly Parton 8,536.0 Johnny Cash 8,091.1
Kanye West 12,176.2 Johnny Cash 8,091.1 Chet Atkins 6,346.4
Paul McCartney 11,920.7 Tony Brown 7,986.3 Ray Charles 4,959.8
Ilaiyaraaja 11,887.7 Reba McEntire 7,888.2 Bob Seger 3,936.0
Van Morrison 11,652.0 George Strait 7,858.5 Jackson Browne 3,421.4
Elvis Presley 11,096.9 Willie Nelson 7,758.0 Tom Petty 3,370.6
Elton John 11,005.2 Rick Rubin 7,186.3 Neil Diamond 3,245.9
a
For a clear presentation, artist names are shown instead of DBpedia URIs.
b
The applied filters also include rdf:type dbo:MusicalArtist
c
dc: is the prefix for the namespace http://purl.org/dc/terms/subject
d
The actual applied filter is dc:subject db:Rock_and_Roll_Hall_of_Fame_inductees
e
The applied filters also include rdf:type dbo:MusicalArtist and dbo:genre db:Country_music

Another application of the metrics is to identify the most important relations in

each domain. Table 3 shows the top 10 list of properties with the highest amount of
 Linked Data Informativeness 635

Generated Information Content. These properties hold the most valuable and distinc-
tive information about instances. It is useful especially for Linked Data browsers and
search engines to allow users easily navigate through a large collection of results.

Table 3. Applications of GIC for faceted browsing

Films Music Actors
Relation GIC (bits) Relation GIC (bits) Relation GIC (bits)
dbo:starring 1,175,262.8 dbo:artist 2,166,102.0 dbo:starring 991,276.0
owl:sameAsa 937,578.8 dbo:producer 1,389,607.9 owl:sameAs 666,309.8
dbo:writer 643,112.6 dbo:associatedBand 1,230,862.1 dbo:occupation 461,851.4
dbo:producer 467,498.6 dbo:associatedMusicalArtist 1,230,862.1 dbo:director 418,187.0
dbo:director 404,065.1 owl:sameAs 998,533.4 dbo:writer 335,585.8
dc:subject 287,922.5 dbo:musicalBand 846,536.5 dc:subject 309,823.7
dbo:musicComposer 217,413.4 dbo:musicalArtist 846,536.5 dbo:producer 286,722.6
dbo:editing 150,631.0 foaf:homepage 626,596.5 rdf:type 210,529.5
dbo:cinematography 142,578.4 dbo:bandMember 603,403.8 dbo:spouse 143,141.6
rdf:type 123,186.2 dbo:writer 558,319.8 dbo:birthPlace 137,186.4
a
owl: is the prefix for the namespace http://www.w3.org/2002/07/owl#

4 Related Work

Since being introduced by Resnik [11], Information Content-based metrics have been
widely exploited for semantic similarity measurement between concepts in lexical
taxonomies [12-15]. These metrics demonstrated better performance than convention-
al edge-counting and feature-based methods. However, the notion of informativeness
has been rarely used in Linked Data. Cheng et al. [16] embed the mutual information
of resources into a random surfer model and present an approach for centrality-based
summarization of entities in Linked Data.
A number of metrics have been developed to analyze the Web of Data from vari-
ous aspects. Ell et al. [17] propose metrics to evaluate labels in Linked Data in terms
of completeness, unambiguity, and multilinguality. Guret et al. [18] apply network
analysis techniques to measure and improve the robustness of Linked Data graph.
Another series of research has been carried out for entity ranking in Linked Data.
Although traditional graph-based ranking methods, such as PageRank [19], SimRank
[20], and HITS [21], can also be applied to the link structure of the Web of Data, its
semantics, represented using various relation types, will not be taken into account.
Numerous studies have attempted to extend the existing algorithms to consider link
types in the rating methodology [22-27]. ObjectRank [22] adds link weights to Page-
Rank for ranking in a directed labeled graph. Bamba and Mukherjea [23] extend
PageRank to arrange the results of Semantic Web queries. TripleRank [26] is a gene-
ralization of the authority (importance) and connectivity (hub) scores of HITS algo-
rithm for Linked Data, with applications in faceted browsing. These methods rate the
importance of resources according to the link structure of the Web of Data graph
while our approach is based on the amount of useful information that each individual
resource and relation conveys. Nevertheless, further research needs to be done to
assess the performance of different algorithms in terms of ranking functionality.
636 R. Meymandpour and J.G. Davis

5 Conclusion and Future Work

In this paper, we addressed the problem of Linked Data quality assessment, which has
a significant impact on the performance of Semantic Web applications. By relying on
the principles of information theory, we have developed two metrics, namely Parti-
tioned Information Content (PIC) and Generated Information Content (GIC), for mea-
suring the informativeness of resources and relations in Linked Data. We analyzed the
Web of Data using the metrics in several domains. As it was demonstrated in the ex-
periments, assessing the quality of information contained in Linked Data offers the
potential for its deployment in a variety of applications including information filter-
ing, multi-faceted browsing, search engines, and entity ranking. Moreover, it can also
be exploited for Linked Data clustering and visualization to highlight the most impor-
tant information sources and relations. Currently, we are verifying and evaluating the
proposed metrics in a range of application areas. We are also devising them in order
to enable further analysis and comparison of resources in the Web of Data.

References
1. Lassila, O.: Resource Description Framework (RDF) model and syntax specification.
World Wide Web Consortium, W3C (1999)
2. McGuinness, D.L., Harmelen, F.V.O.: Web Ontology Language: Overview. World Wide
Web Consortium, W3C (2004)
3. Prudhommeaux, E., Seaborne, A.: SPARQL Query Language for RDF. World Wide Web
Consortium, W3C (2008)
4. Bizer, C., Heath, T., Berners-Lee, T.: Linked Data - The Story So Far. International Jour-
nal on Semantic Web and Information Systems 5(3), 122 (2009)
5. Auer, S., Bizer, C., Kobilarov, G., Lehmann, J., Cyganiak, R., Ives, Z.: DBpedia: A Nuc-
leus for a Web of Open Data. Springer, Heidelberg (2007)
6. Gray, R.M.: Entropy and Information Theory. Springer, New York (2009)
7. Edwards, S.: Elements of information theory, 2nd edition. Information Processing & Man-
agement 44(1), 400401 (2008)
8. Shannon, C.E.: A Mathematical Theory of Communication. Bell System Technical Jour-
nal 27, 379423, 623656 (1948)
9. Ross, S.M.: A First Course in Probability. Prentice Hall (2002)
10. Hassanzadeh, O., Consens, M.P.: Linked Movie Data Base (2009)
11. Resnik, P.: Using information content to evaluate semantic similarity in a taxonomy. In:
Proceedings of the Fourteenth International Joint Conference on Artificial Intelligence, pp.
448453 (1995)
12. Jiang, J.J., Conrath, D.W.: Semantic Similarity Based on Corpus Statistics and Lexical
Taxonomy. In: Proceedings of the International Conference Research on Computational
Linguistics (ROCLING X), Taiwan (1997)
13. Lin, D.: An information-theoretic definition of similarity. Morgan Kaufmann, San Francis-
co (1998)
14. Resnik, P.: Semantic Similarity in a Taxonomy: An Information-Based Measure and its
Application to Problems of Ambiguity in Natural Language. Journal of Artificial Intelli-
gence Research 11, 95130 (1999)
 Linked Data Informativeness 637

15. Zili, Z., Yanna, W., Junzhong, G.: A New Model of Information Content for Semantic Si-
milarity in WordNet (2008)
16. Cheng, G., Tran, T., Qu, Y.: RELIN: Relatedness and Informativeness-Based Centrality
for Entity Summarization. Springer, Heidelberg (2011)
17. Ell, B., Vrande i , D., Simperl, E.: Labels in the Web of Data. Springer, Heidelberg
(2011)
18. Guret, C., Groth, P., van Harmelen, F., Schlobach, S.: Finding the Achilles Heel of the
Web of Data: Using Network Analysis for Link-Recommendation. Springer, Heidelberg
(2010)
19. Brin, S., Page, L.: The anatomy of a large-scale hypertextual Web search engine. In: Pro-
ceedings of the Seventh International Conference on World Wide Web 7, Brisbane, Aus-
tralia. Elsevier Science Publishers B. V. (1998)
20. Jeh, G., Widom, J.: SimRank: a measure of structural-context similarity. In: Proceedings
of the Eighth ACM SIGKDD International Conference on Knowledge Discovery and Data
Mining, Edmonton, Alberta, Canada. ACM (2002)
21. Kleinberg, J.M.: Authoritative sources in a hyperlinked environment. J. ACM 46(5), 604
632 (1999)
22. Balmin, A., Hristidis, V., Papakonstantinou, Y.: Objectrank: authority-based keyword
search in databases. In: Proceedings of the Proceedings of the Thirtieth International Con-
ference on Very Large Data Bases - Volume 30, Toronto, Canada. VLDB Endowment
(2004)
23. Bamba, B., Mukherjea, S.: Utilizing Resource Importance for Ranking Semantic Web
Query Results. Springer, Heidelberg (2005)
24. Ding, L., Pan, R., Finin, T., Joshi, A., Peng, Y., Kolari, P.: Finding and Ranking Know-
ledge on the Semantic Web. Springer, Heidelberg (2005)
25. Hogan, A., Harth, A., Decker, S.: ReConRank: A Scalable Ranking Method for Semantic
Web Data with Context. In: Proceedings of Second International Workshop on Scalable
Semantic Web Knowledge Base Systems (SSWS 2006), in conjunction with International
Semantic Web Conference (ISWC 2006) (2006)
26. Franz, T., Schultz, A., Sizov, S., Staab, S.: TripleRank: Ranking Semantic Web Data by
Tensor Decomposition. Springer, Heidelberg (2009)
27. Delbru, R., Toupikov, N., Catasta, M., Tummarello, G., Decker, S.: Hierarchical Link
Analysis for Ranking Web Data. Springer, Heidelberg (2010)
 Harnessing the Wisdom of Crowds
for Corpus Annotation through CAPTCHA*

Yini Cao and Xuan Zhou

DEKE Lab, Renmin University of China

Beijing, China 100872
{novgddhy,xzhou}@ruc.edu.cn

Abstract. CAPTCHA is a type of widely applied security measures to prevent

abuses of online services. Our research is focused on designing a new type of
CAPTCHA called AnnoCAPTCHA, which can harness the intellectual power
of Web users to perform corpus annotation an expensive step for computa-
tional linguistic analysis. We analyzed the applicability and the efficiency of
AnnoCAPTCHA. We also conducted user study to demonstrate its accessibility.

Keywords: CAPTCHA, corpus annotation, human computation.

1 Introduction

The term CAPTCHA (Completely Automated Public Turing test to tell Computers
and Humans Apart) was created by Luis von Ahn in 2000 [1], to name a simple Tur-
ing test which requires users to type letters from distorted images. Since its creation,
CAPTCHA has become a widely applied security measure on the Web, to prevent
large scale abuses of online services. According to Ahns estimates [2], humans
around the world type more than 100 million CAPTCHAs every day. Suppose it takes
10 seconds to enter a CAPTCHA. In accordance with the use of 100 million times a
day, it will be huge labor resources of 150,000 hours. It is desirable if we can turn
them into useful services. In recent years, there have been several proposals for better
utilization the labor resources on CAPTCHA. A typical example is ReCAPTCHA, a
special type of CAPTCHA that can be used to digitize printed material.
In this work, we explore the possibility of utilizing the labor resources on
CAPTCHA to annotate textual corpus. We propose a new type of CAPTCHA, named
AnnoCAPTCHA, which not only provides the functionality of CAPTCHA but also
works a platform to perform corpus annotation. In each Turing test issued by Anno-
CAPTCHA, a user is presented with a set of short textual items and is asked to select
the right annotation for each item. Among the textual items, some items annotations
are already known, so they can be used to verify if the user is a human being. The
other items annotations are unknown, such that the users selections become new

*
This work is supported by the Basic Research fund in Renmin University of China from the
central government-No.12XNLJ01.

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 638645, 2013.
Springer-Verlag Berlin Heidelberg 2013
 Harnessing the Wisdom of Crowds for Corpus Annotation through CAPTCHA 639

annotations to the target corpus. If AnnoCAPTCHA can be applied broadly, it can

help us construct large scale annotated corpus and there after enhance the perfor-
mance of NLP. In the paper, we conduct analytical study to evaluate its effectiveness
and efficiency. In addition, we report a user study to demonstrate its accessibility to
human users.
The rest of this paper is organized as follows. We review some related work in
Section 2. In Section 3, we introduce our system and analyze its effectiveness and
efficiency. In Section 4, we present a user study for evaluating the accessibility of
AnnoCAPTCHA. Finally, Section 5 provides a conclusion.

2 Related Work

2.1 Human Computation

Our work can be regarded as an application of human computation to the field of
NLP. The concept of human computation was first coined by Luis Von Ahn. Guided
by its philosophy, Ahn invented ReCAPTCHA [2] and the well known ESP Game
[3]. Game is the most common area to implement human computation. There have
already been a number of online games [3,4] that can exploit players intelligence to
label the contents of images. CAPTCHA is another typical area to apply human com-
putation. Except ReCAPTCHA, there are a number of proposals for generating values
out of CAPTCHA. For instance, ASIRRA [5] has shown that CAPTCHA can be used
for humanitarian purpose. TagCaptcha [6] uses CAPTCHA to perform image annota-
tion on the Web. Besides, SolveMedia1, captcha.tw2 and other ads applications use
CAPTCHAs to display ads. To the best of our knowledge, we are the first to use
CAPTCHA to perform corpus annotation.

2.2 Corpus Annotation

Corpus annotation is an expensive preparing step in natural language processing and

text mining. According to our study, the existing methods for large-scale corpus anno-
tation usually fall in the following three categories:
1. Traditional method. The traditional method mainly counts on domain experts to
perform manual annotation. The well known Brown Corpus was generated by this
method. Using this method, large-scale annotation tasks will require considerable
human efforts and thus are very costly.
2. WWW-based non-experts method. The main idea is to call up non-expert volun-
teers on the Web to annotate corpuses. The most famous example is the Open
Mind Commonsense project3. The efficiency of this method is also questionable, as
there is a lack of incentive for Web users to perform annotation for free.

1
For more information, go to http://www.solvemedia.com/
2
For more information, go to http://captcha.tw/
3
For more information, go to http://openmind.media.mit.edu/
640 Y. Cao and X. Zhou

3. Game-based method. This method applies human computation. It exploits human

intelligence in playing games to perform corpus annotation. For instance, by using
competitive games, the Learner project [7] creates 1.3 million labels over a large
image repository. However, the existing methods only deal with multimedia data
such as images. There is no game designed to annotate textual corpus.

In summary, corpus annotation is a labor intensive task. There is a dearth of human

computation mechanisms that exploit the free labor on the Web to perform textual
corpus annotation. Our work in this paper aims to provide such a mechanism.

3 The AnnoCAPTCHA System

To illustrate the functionality of AnnoCAPTCHA, we use sentiment annotation as an

example. Sentimentally annotated corpus is the base for sentiment analysis, a hot
topic in the areas of NLP and Web mining. As the creation of such a corpus is very
expensive, it is an ideal application scenario for AnnoCAPTCHA.

3.1 User Interface

The user interface of AnnoCAPTCHA is illustrated in Fig.1. In each test issued by

AnnoCAPTCHA, a user is presented with four context-adjective phrase. The user is
asked to judge their sentiment polarity (either positive or negative). Among the four
phrase items, twos sentiment polarities are already known to the system and the other
two are waiting to be annotated. As the four context-adjective phrases are randomly
placed, the user cannot distinguish between the known ones and the unknown ones.
The items with known sentiment polarities are used to verify if the user is a human. If
the verification succeeds, i.e. the users judgment on the known items is correct, we
regard that the users judgment on the unknown items is very likely to be correct too.
Then, the user judgment on each unknown item is recorded and counted as a vote for
that items sentiment polarity. As more and more users use AnnoCAPTCHA, the
number of polarity votes for each unknown item increases. When the overall judg-
ment on an unknown item reaches a certain statistical significance, such that we are
confident about its real sentiment polarity (e.g. 9 out of 10 vote for positive polarity),
we label the items sentiment polarity and turn it into a known item. Then, the new
known item becomes a member of the target corpus. It can subsequently be used to
perform user verification.

Fig. 1. The User interface

 Harnessing the Wisdom of Crowds for Corpus Annotation through CAPTCHA 641

3.2 Effectiveness as CAPTCHA

AnnoCAPTCHA can provide the functionality of CAPTCHA. It performs human

verification using the two context-adjective phrases whose sentiment polarities are
already known. Usually, a humans judgment on sentiment polarity is much more
accurate than a machines. Suppose a humans accuracy can achieve at least 95%.
Then, a human users chance of passing the test is around 90% (=0.952). The test is
much harder to a computer. Suppose a computer can only perform random guess. Its
chance of passing the test is thus only about 25% (=0.52). Even though a user may fail
the test occasionally, the user can try more tests. Eventually, the verification will suc-
ceed. In contrast, to perform attacks, such as email spamming, an adversary has to
rely on a machine to obtain a large number of accounts in a short period of time.
Thus, it has to be tested by AnnoCAPTCHA repeatedly. A machine may pass some
tests by chance. According to the theory of probability, as the number of tests increas-
es, the machines success rate will quickly approach its expected accuracy. As a con-
sequence, a machine can be easily recognized by a small number of tests, before it can
perform any effective attack. Upon detection, the website can immediately cut off the
connection with the attacker to prevent it from taking further action.

3.3 Efficiency of Annotation Generation

In the following, we analyze how fast sentiment annotations can be generated by

AnnoCAPTCHA.

The Model.

We assume that each CAPTCHA contains k+2 items, in which 2 items annotations
are known and the other k items annotations are unknown.
A user of AnnoCAPTCHA is either a human or a machine. Let s% be the accuracy
of a humans judgment. Let t% be the accuracy of a machines judgment. Obvious-
ly, s% is much greater than t%.
As mentioned previously, once a machine is detected by AnnoCAPTCHA, its con-
nection will be cut off. This effectively restricts the number of annotations by ma-
chines. Let the proportion of CAPTCHA tests finished by humans be m% and that
by machines/attackers be 1-m%. Thus, m% should be far greater than 1-m%.
After a user passes the verification based on the 2 known items, the system will
record users votes on the k unknown items in a database. The vote record of each
unknown item is in the form i : j, where i represents the number of users voting for
positive polarity and j represents the number of users voting for negative polarity.
When the difference between i and j reaches a certain statistical significance, that
is, when i (or j) is significantly larger than j (or i), we turn the unknown item into a
known item. We use error probability to measure the significance level, which is
calculated as:

(1)
642 Y. Cao and X. Zhou

where P0 denotes the probability that a vote is incorrect, that is

In principle, we regard the sentiment polarity of an unknown item positive if i > j

and negative if i < j. The error probability Pij measures the probability that we
make a wrong judgment. AnnoCAPTCHA uses a predefined threshold of the error
probability, e.g. 1%, to represent the maximum error rate it can accept. Once the Pij
of an unknown item falls below the threshold, AnnoCAPTCHA finalizes its senti-
ment polarity and turns it into a known annotation.

Analysis of Efficiency

When an item receives i positive votes and j negative votes, if the error probability Pij
falls below the threshold, the annotation process for this item will terminate. Thus, i+j
measures the cost of the annotation, i.e. the number of votes required to determine the
sentiment polarity of a context-adjective phrase. The higher the number, the longer
the duration we need to generate an annotation. We are interested in the expected
annotation cost, which we use to depict the efficiency of AnnoCAPTCHA.
Let Pt denote the threshold of error probability. Given an arbitrary pair of i and j,
let Qij denote the probability that the process does not terminate before it receives i
positive votes and j negative votes. It can be inferred that the value of Qij depends on
that of Q(i-1)j and Qi(j-1), because the vote record i : j must be evolved from either i-1 : j
or i : j-1 . Thus, Qij can be calculated recursively as follows:

(2)

Let Oij denote the probability that the process terminates when it receives i positive
votes and j negative votes. Then Oij can be computed by:

(3)

Finally, the expected annotation cost can be calculated as:

(4)

We did not manage to integrate Equations (1), (2), (3) and (4) into a single formula to
measure the expected annotation cost. However, it does not prevent us from evaluat-
ing the relationship between the expect cost and the various factors in our model. As
suggested by our analysis, the main influence factors of the expected annotation cost
 Harnessing the Wisdom of Crowds for Corpus Annotation through CAPTCHA 643

include the accuracies of the annotations, i.e. s% and t%, and the proportion of attack-
ers votes, i.e. 1-m%.
We created a Java program to compute the expect annotation cost based on Equa-
tions (1)~(4). To evaluate the influence of humans accuracy (s%) on the expected
cost, we set m% to 80% and t% to 50% (machines only make random guess), and
vary s% from 55% to 100%. Then, the relationship between the annotation cost and
humans accuracy (s%) is plotted in Fig. 2. Obviously, the higher the human accura-
cy, the smaller of the expected cost. As we can see, to guarantee the efficiency of the
annotation process, we need to ensure that the accuracy of humans be significantly
higher than that of computers. For example, to ensure that the average cost is below
10 votes, the human accuracy must be above 75%. Fortunately, this condition can be
satisfied for most corpus annotation tasks.
To evaluate the influence of attacks, we vary m % from 50% to 100%, and fix s%
to 90% and t% to 50%, and. Then, the relationship between the annotation cost and
the proportion of humans votes (m %) is plotted in Fig.3. As we can see, the higher
the proportion of humans votes, the smaller the average cost. The proportion of hu-
mans votes depends on how effectively a website detects and blocks attackers. When
the proportion is more than 70%, the average cost can be controlled below 5 votes.

50 12
average cost

10
average cost

40
30 8
6
20 4
10 2
0 0
0.55 0.75 0.95 0.50 0.70 0.90
human accuracy human annotations

Fig. 2. Annotation cost vs human accuracy Fig. 3. Annotation cost vs proportion of

(s %) human annotations (m %)

4 Evaluation of Accessibility

We carried out a user study to evaluate AnnoCAPTCHAs accessibility. In the study,

we asked 9 volunteers to do the same 10 CAPTCHAs. The volunteers are all universi-
ty students. Every CAPTCHA has 4 items, which are all Chinese sentiment phrases
selected from the SemEval-2010 test set 4. Each phrase contains an adjective and a
context. We recorded the users votes and the time each user took to finish each
CAPTCHA. In addition, we let the machine to annotate the same 10 CAPTCHAs by
random guess, to study the impact of attacks to the annotation efficiency.
4
For more information, go to
http://semeval2.fbk.eu/semeval2.php?location=tasks#T.
644 Y. Cao and X. Zhou

4.1 me
User Response Tim
User response time measurres the time a human user needs to complete a CAPTCHA
test. Fig.4 shows the response times of the 9 volunteers on each CAPTCHA. As we
can see, there is a learning curve
c for this new type of CAPTCHA. When a volunteeer is
presented with the first CA APTCHA, she / he needs a certain time to understandd its
requirements and get used to t the test environment. This time usually takes aroundd 25
seconds, and can occasionaally go up to 1 minute. After the second CAPTCHA, the
volunteers become familiarr with the exercises. Most of the response times fall in the
interval between 5 seconds to 25 seconds. The average response time is below w 10
seconds. Overall, no volun nteer encountered difficulty when doing the CAPTCH HAs.
The accessibility of AnnoCAPTCHA appears acceptable to most users.

80 volunteer11
volunteer22
60 volunteer33
time(s)

40 volunteer44
volunteer55
20 volunteer66
0 volunteer77
volunteer88
0 2 4 6 8 10 volunteer99
CAPTC
CHA's group number 1-10
Fig. 4. User Respond Time

4.2 User Accuracy

The accuracy of the 9 volun nteers is shown in Fig.5. As we can see, all volunteers can
provide correct answers to more
m than 90% of the test questions. The average accurracy
is 96.67%. This also indicaates that AnnoCAPTCHA is not a difficult exercise to llite-
rate users.

100% 12
10
8
accuracy

average cost

95%
6
90% 4
2
85% 0
1 2 3 4 5 6 7 8 9 100% 80% 60% 40%
volunteerr 1-9 human annotations
Fig. 5. User Accurracy Fig. 6. Annotation Cost vs Proportionn of
Human Annotations
 Harnessing the Wisdom of Crowds for Corpus Annotation through CAPTCHA 645

4.3 Efficiency

Using the annotations collected from the user study, we simulated the process of an-
notation generation by AnnoCAPTCHA. We increased the proportions of random
annotations made by machines to evaluate the impact of attacks on the annotation
efficiency. We assume that humans accuracy is 96.67%, the exact value we measured
in the user study. And we set the error probability threshold to 1%. The average cost
for annotating each context-adjective phrase (i.e. the average number of users we
need to annotate each phrase) is shown in Fig. 6. As we can see, with the accuracy
of 96.67%, the annotation is very efficient. Without the disturbance of machines, the
annotation cost is only about 2 votes. As more and more test results of machines are
added, the annotation cost goes up. However, the overall cost still appears acceptable.

5 Conclusions

The main contribution of this paper is to provide a new type of CAPTCHA that can
utilize crowd intelligence to do corpus annotation. We used the sentiment corpus an-
notation as example and discussed the effectiveness and efficiency of our system. We
clarified the relationship between the efficiency of the system and the accuracy of
human annotation as well as the influence of attacks. Finally we conducted a user
study to prove our system is easy to access and efficient in annotation generation.

References
1. von Ahn, L., Blum, M., Hopper, N.J., Langford, J.: CAPTCHA: Telling Humans and Com-
puters Apart. In: Biham, E. (ed.) EUROCRYPT 2003. LNCS, vol. 2656, Springer, Heidel-
berg (2003)
2. von Ahn, L., Maurer, B., McMillen, C., Abraham, D., Blum, M.: reCAPTCHA:human-
Based Character Recognition via Web Security Measures,
http://www.sciencemag.org/content/321/5895/1465.full
3. von Ahn, L., Dabbish, L.: Labeling Images with a Computer Game,
http://www.cs.cmu.edu/~biglou/ESP.pdf
4. von Ahn, L., Liu, R., Blum, M.: Peekaboom: A Game for locating Objects in Images,
http://www.cs.cmu.edu/~biglou/Peekaboom.pdf
5. Elson, J., Douceur, J.R., Howell, J., Saul, J.: Asirra: A CAPTCHA that Exploits Interest-
Aligned Manual Image Categorization. In: Proceedings of the 14th ACM Conference on
Computer and Communications Security. Association for Computing Machinery, New
York (2007)
6. Morrison, D., Marchand-Maillet, S., Bruno, E.: Tagcaptcha: annotating images withcapt-
chas. In: Proceedings of the ACM SIGKDD Workshop on Human Computation, HCOMP
2009, pp. 4445. ACM, New York (2009)
7. Chklovski, T.: Improving the design of intelligent acquisition interfaces for collecting world
knowledge from web contributions. In: Proceedings of the 3rd International Conference on
Knowledge Capture, pp. 3542 (2005)
 A Framework for OLAP in Column-Store Database:
One-Pass Join and Pushing the Materialization to the End

Yuean Zhu1,2, Yansong Zhang1,2,3, Xuan Zhou1,2, and Shan Wang1,2

1
DEKE Lab, Renmin University of China, Beijing 100872, China
2
School of Information, Renmin University of China, Beijing 100872, China
3
National Survey Research Center at Renmin University of China, Beijing 100872, China
{iwillgoon,swang}@ruc.edu.cn

Abstract. In data warehouse modeled with the star schema, data are usually re-
trieved by performing a join operation between the fact table and dimension ta-
ble(s) followed by a selection and project operation, while join operator is the
most expensive operator in RDBMS. In column-store database, there are two
ways to do join. The first way is early materialization join (EM join); the other
way is late materialization join (LM join). In EM join, the columns involved in
the query are glued together firstly, then the glued rows are sent to join opera-
tor. Whereas, in LM join, only the attributes participated in the join operator are
accessed. The problem that access to inner table is out-of-order cant be ig-
nored for LM join. Otherwise, the nave LM join is usually slower than EM join
[9]. Since the late materialization is good for memory bandwidth and CPU effi-
ciency, the LM join attracts more attention in academic research community.
The state-of-art LM joins in column-store such as radix-cluster hash join [8] in
MonetDB, invisible join [10] in C-Store all try to avoid accessing table random-
ly. In this paper, we devised a framework for OLAP called CDDTA-MMDB
where a new join algorithm called CDDTA-LWMJoin (we contract it to
LWMJoin in the following) is introduced. The LWMJoin is on the basis of our
prior work: CDDTA-Join [7]. We equip the CDDTA-Join with light-weight
materialization (LWM) which is designed to cut down the memory access and
reduce production of intermediate data structure. Experiments show that
CDDTA-MMDB is efficient and can be 2x faster than MonetDB and 4x faster
than invisible join in the context of data warehouse modeled with star schema.

Keywords: OLAP, in-memory column-store database, join, materialization.

1 Introduction
Data warehousing (DW) and On-line analytical processing (OLAP) consist of essen-
tial elements of decision support, increasingly becoming a focus of the database in-
dustry. Deviating from the queries to transactional databases, the nature of the queries
to DWs is more complex, longer lasting and read-intensive, but less attributes to
touch. Since column-store database only retrieves attributes required in the query, it
can optimize cache performance. Thus columnar storage is more suitable for DW in
this sense, and now the Microsoft also adds columnar storage in SQL server 2012 to

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 646653, 2013.
Springer-Verlag Berlin Heidelberg 2013
 A Framework for OLAP in Column-Store Database 647

improve performance on data warehousing queries [5]. With the increment of RAM
density and reduction of its price there are a number of in-memory column-oriented
database systems such as MonetDB [1, 2] and C-Store [3] emerging for better per-
formance. These systems offer order-of-magnitude gains against traditional, row-
oriented databases on certain workloads, particularly on read-intensive analytical
processing workloads, such as those encountered in DWs. In this paper, we propose a
framework for OLAP called CDDTA-MMDB (Columnar Direct Dimensional Tuple
Access-Main Memory Database) to efficiently process the query on DWs modeled
with star scheme. This work is on the basis of our prior work to develop one-pass join
algorithm [7] which is variant of invisible join [10]. The following is our focus inno-
vations:

The Storage Model. In CDDTA-MMDB we employ the decomposed storage

model (DSM) [4] which vertically partitions the tables and maintain a collection of
triplets <surrogate, key, value> tables. In addition to make use of CPU cache line
more efficiently, this representation is one of crucial measures for our goal to gain
light-weight materialization and LWMJoin algorithm.
Materialization Strategy in Column-Store Database. We propose a new materia-
lization strategy called light-weight materialization (LWM) which adopts notion of
late materialization (LM) to push the materialization as late as possible. Compared
to the LM, LWM tries to avoid producing intermediate data structure such as bit-
map and reduce random memory access during the whole query plan. Experiments
show that CDDTA-Join [7] equipping with LWM (CDDTA-LWMJoin, in the fol-
lowing we contract it to LWMJoin) outperforms the state-of-art join algorithm
such as radix-cluster join [8], invisible join [10] in the context of DWs modeled
with star schema. To the best of our knowledge, this is first time to propose such
materialization for column-store database.

The rest of this paper is organized as follows: In section 2, prior work on columnar
storage, materialization and join algorithms is described. Section 3 gives the light-
weight materialization techniques. In Section 4, system overview is presented. Sec-
tion 5 shows representative experimental results. Finally, conclusion is drawn in
section 6.

2 Related Works

Storage. The decomposition storage Model (DSM) [4] reduces each n-attribute rela-
tion to n separate relations. DSM organizes values from same attribute close in physi-
cal. Supposed the record reconstruction cost is low, cache performance of main-
memory database systems can be improved by DSM. As mentioned to the detail of
representation of DSM, MonetDB proposed the Binary Units (or BUNs) [2] to con-
struct the decomposed table in the form of <OID, value > pairs.
Materialization and Join Algorithm in Column-Store. Column-store modifies the
physical data organization in database. In logical, the applications treat column-store
the same as row-store. At some point in a query plan, a column-oriented DBMS must
648 Y. Zhu et al.

glue multiple attributes from the same logical tuple into a physical tuple. The strate-
gies of materializing are divided into two categories: early materialization (EM) and
late materialization (LM). A recent work on materialization in column-store is in
[14]. It used a dynamic and cache-friendly data structure, called cracker maps which
provide direct mapping between the pairs of attributes in the query, to reduce the ran-
dom data access. In [6], trade-offs between different strategies are systematically
explored and advices are given for choosing a strategy for a particular query.
The nave join algorithm using LM strategy (nave LM join) in column-store data-
base is about 2x slower than the join algorithm using EM strategy (EM join) [9]. The
main reason is that in nave LM join the access to inner table is out-of-order, causing
the high cache miss. The state-of-art LM join algorithms in column store such as radix
hash join [8], MCJoin[13] and invisible join [10] all try to avoid the random memory
access to inner table. Radix-clustered hash join [8] partitions the input relation into H
clusters. By controlling the number of clustering bits each pass, the H clusters gener-
ated can be kept under the count of Translation Lookaside Bu er (TLB) entries and
cache lines available on that CPU. The random access only happens in storage block.
Invisible join [10] uses bitmap to discard tuples that dont satisfy the prediction and
can reduce the random memory accesses. LWMJoin tries to avoid producing the in-
termediate data structure and further reduces memory access compared to invisible
join [10] and CDDTA-Join [7]. Experiments show that LWMJoin runs about 4x faster
than invisible join and 2x faster than radix-cluster join on average.

3 Light-Weight Materialization

Column-store is physical modification to database. In terms of logical and view level

it is identical to row-store. The applications involved with database, row-oriented or
column-oriented, treat the interface as row-oriented. So a column-oriented DBMS
must glue multiple attributes, at some time, from the same logical tuple into a physi-
cal tuple, which is called materialization. There are two ways to materialize: early
materialization (EM) and late materialization (LM). At each point when a column is
needed, it is read into CPU and added to intermediate tuple representation. This policy
to reconstruct the tuple is called early materialization. It is not surprising that EM is
not always the best strategy for a column-store [9], because some CPU works are
useless. Consider this scenario: a column-store DBMS place all columns of a table in
separate files and sort them in the same order; a query has two selection operators 1
and 2 over two columns (R.a, R.b), a GROUP BY operator on R.a, where 1 has
higher selective than 2 . Under the EM policy, the operations are as follows: read in
a block of R.a and a block of R.b; glue them together to form the rows and apply 1
and 2 on the rows in turn. Then, send the tuples for aggregation. A more efficient
way for this process is that: keep two intermediate position lists; scan the block of R.a
and R.b, meanwhile apply 1 and 2 on them respectively; use the lists to record
 A Framework for OLAP in Column-Store Database 649

whether attributes in the corresponding position in column satisfy the prediction; then
use position-wise AND to intersect the two position lists. According to intersected
result, the block of R.a is re-accessed for aggregation. The latter strategy is called late
materialization. Light-weight materialization (LWM) adopts notion of LM to push
the materialization as far as possible in query plan, but try to reduce the memory
access and avoid producing intermediate data structure. Instead of re-accessing the
block of R.a to get record value for GROUP BY, LWM just lets the key of triplet pro-
duced by scanning (R.a, R.b) go through whole query plan. Furthermore, LWM
doesnt need to be assisted by the intermediate data structure except triplets (in some
case such as the attributes that are not used any more but just serve as filters, the trip-
lets can be replaced by bitmap to save space). Considering the above scenario again,
we dont need to keep two intermediate position lists and use position-wise AND to
intersect the two positions to produce intersected result, but just utilize key in triplet
and let it go through query plan. So, the LWM can accelerate the system speed, and it
greatly facilitates aggregation operation.

4 System Design

The sub-queries on fact table and dimension tables are handled separately in CDDTA-
MMDB. When a query is received, it is broken off. Then, sub-queries on dimension
table(s) are rewritten in order to produce triplets, while the sub-query on fact table is
parsed and information is stored for later use. The Query Parser is responsible for
those tasks. Based on the triplets and information from parsing sub-query on fact
table, the result set can be obtained by scanning fact table one pass. Figure 1 illu-
strates the overview of the system.

Fig. 1. System overview

650 Y. Zhu et al.

4.1 Details of CDDTA-LWMJoin Algorithm

CDDTA-LWMJoin is derived from invisible join [10]. Invisible join is also divided
into three phases. First, apply the prediction on the dimension table and use key of
rows that pass the prediction to build hash table. Second, scan the fact table; mean-
while fetch the foreign key to probe the hash table. The bitmaps on dimension tables
are built up in this step. Then intersect the bitmaps to product final bitmap. At last,
according to final bitmap scan the fact table again to construct result relation. If the i-
th value is set to true in the final bitmap, the i-th record in the fact table is retrieved.
From above analysis, the time t used by invisible join can be expressed as t0 + 2t1 + t2,
where t0 is spent on producing hash table; t1 on scanning the fact table; t2 on produc-
ing the final bitmap. Since size of the final bitmap is equal to size of fact table, we
can assume that t1 is equal t2. So, the total time t spent by invisible join is
t0 + 3 t1.
Compared with invisible join, LWMJoin algorithm just utilizes the key information
in triplet which was produced in first phase and thus avoids scanning fact table twice.
The following is the algorithm procedure. First, decomposed queries involved with
dimension table are rewritten; then we get the result sets from dimension tables. The
last step of this phase is to produce the <surrogate, key, value> triplets. For those sub-
queries such as RQ1, bitmap is enough to represent them; for the other case, the <sur-
rogate, key, value> triplet is to be produced. This representation forms the base of
LWMJoin. Second, Fact table is scanned sequentially; foreign keys are used to probe
the key vectors producing in first phase, deciding whether the row satisfies the predic-
tion. If keys equal to zero is greater than one, this row should be discarded. Other-
wise, a physical tuple is constructed. Now, the operators such as GROUP BY can be
applied on rows passing qualification.The benefits brought by LWM are obvious:
dynamically allocating the memory to build the temporary position lists is no need; it
accelerate the aggregation especially when the fact table is big and result sets are big
because the records don need to be fetched again for aggregation. In general, the key
of triplet serves as grouper in star schema and in such case we dont need to re-access
attributes in dimension table for aggregation. Therefore, random memory access can
be further reduced. Figure 2 is the illustration of CDDTA-MMDB handling Q4.1.
First apply the rewritten SQL RQ1, RQ2, RQ3, and RQ4 on corresponding dimension
table to produce triplets or bitmaps. Then scan the fact table and use foreign key to
probe key/OID arrays. In this example, there are four tuples passing the prediction,
and result set contains two groups. Key/OID vectors with broken line are used in
GROUP BY (if it is end of query plan, the program will fetch value by key to con-
struct tuple directly). The time t used by LWMJoin is expressed as: t0 + t1, where the
time t0 is used to produce the triplets and t1 is used to scan the fact table. So, the time
spent by invisible join is almost three times more than LWMJoin. The following ex-
periments will show that the consumed time by invisible join is more than three times
compared with time used by LWMJoin, which accords with above analysis.
 A Framework for OLAP in Column-Store Database 651

2 1 228490 1 1 1014839

2396604 3083649

Fig. 2. Demonstration of CDDTA-LWMJoin

5 Experiments

In this section, we will evaluate CDDTA-MMDB using SSBM. We generated 100G

data set of SSBM using data generator with SF = 100. Only attributes of fact table
participating in experiment are loaded into memory, and all dimension tables are resi-
dent in memory. The experiment environment is as following: 48GB memory, two
Intel xeon processors (each runs at 2.4GHz and has six superscalar processor
cores that support simultaneous multi-threading with two hardware threads per core),
1TB disk space and 64-bit Ubuntu 10.10 OS (linux kernel version 3.0.0-19, gcc ver-
sion 4.6.1). We focus the benefits LWM brings, the performance of CDDTA-MMDB
and the scalability of the system.

5.1 Aggregation Time Comparison

Query 1.x doesnt have GROUP BY, so we omit them in the experiment. We simulate
state-of-art LM join algorithms such as in [7, 10] that after join phase the data is re-
trieved for aggregation (LMJoin). Then we compare aggregation time used by
LMJoin with that used by LWMJoin. To our surprise, as is shown in Figure 3 time
consumed by aggregation in LMJoin even exceeds the time by join phase on Q3.1. On
Q3.1 the time took up by aggregation is up to 61% of total time, followed by 35% on
Q4.1, 21% on Q2.1 and 17% in Q4.1. Whereas, in LWMJoin the time took by aggre-
gation is reduced to 28%, 13%, 8% and 5% respectively. The difference of aggrega-
tion time between LMJoin and LWMJoin lies in the random memory access. After
join phase, LMJoin fetches the tuple for aggregation. This out-of-order memory
access will incur cache miss. So the aggregation time in LMJoin is much more than
that in LWMJoin, especially when the selectivity is high.
652 Y. Zhu et al.

Fig. 3. Comparison of aggregation time by CDDTA-LWMJoin and by CDDTA-LMJoin

5.2 CDDTA-MMDB vs. MonetDB and Invisible Join

CDDTA-MMDB and invisible join implemented by us can automatically detect
parameters of the host CPU, and then horizontally partitions the fact table evenly
according to CPU cores. Figure 4 gives the performance comparison of CDDTA-
MMDB, MonetDB (v11.7.9) and invisible join. We implement invisible join using the
C++ Boost Library 1.48. All systems run on the thirteen queries several times such
that the experiment is done on hot data set. CDDTA-MMDB has best performance
among them. The time consumed by invisible join is almost four times more than the
time by CDDTA-MMDB, which accords with the above analysis. Our system has
worse performance than MonetDB on Q.2.2 and Q2.3, but it outperforms MonetDB
greatly on Q1.x and Q4.x. On Q3.x, our system is slightly better than MonetDB.
Overall, the total time to execute all queries by our system is about 22532ms, and
total time by MonetDB is about 50311ms.

Fig. 4. Time used by CDDTA-MMDB vs. MonetDB vs. invisible join

6 Conclusion

The main contribution of this paper is that we propose a framework which introduces
LWMJoin to efficiently deal with the queries on DW modeled with star scheme.
 A Framework for OLAP in Column-Store Database 653

LWMJoin is the combination of CDDTA-Join [7] and LWM. Compared to the state-
of-art materialization in columnar database, LWM reduces production of intermediate
data structure and can cut down memory access greatly. Experiments show that
CDDTA-MMDB has good scalability and can be 2x faster than MonetDB and 4x
faster than invisible join. In future, we will add triplet representation and LWM strat-
egy into open source columnar database such as MonetDB.

Acknowledgement. This work is supported by the Important National Science &

Technology Specific Projects of China ("HGJ" Projects, Grant No.2010ZX01042-
001-002), the National Natural Science Foundation of China (Grant No.61070054).

References
1. Boncz, P.A., et al.: MonetDB/X100: Hyper-pipelining query execution. In: CIDR, pp.
225237 (2005)
2. Boncz, P.A., et al.: MIL primitives for querying a fragmented world. VLDB Journal 8(2),
101119 (1999)
3. Stonebraker, M., et al.: C-Store: A Column-Oriented DBMS. In: VLDB, pp. 553564
(2005)
4. Copeland, G.P., et al.: a decomposition storage model. In: ACM SIGMOD, pp. 268279
(1985)
5. Larson, P.-A., et al.: Columnar Storage in SQL Server 2012. In: ICDE, pp. 1520 (2012)
6. Abadi, D.J., et al.: Integrating compression and execution in column-oriented database sys-
tems. In: ACM SIGMOD, pp. 671682 (2006)
7. Yan-Song, Z., et al.: One-size-fits-all OLAP technique for big data analysis. Chinese Jour-
nal of Computers 34(10), 19361946 (2011)
8. Boncz, P.A., et al.: Breaking the Memory Wall in MonetDB. Communications of the
ACM 51(21), 7785 (2008)
9. Abadi, D.J., et al.: Materialization strategies in a column-oriented DBMS. In: ICDE, pp.
466475 (2007)
10. Abadi, D.J., et al.: Column-Stores vs. Row-Stores: How Different Are They Really? In:
ACM SIGMOD, pp. 967980 (2008)
11. ONei, P.E., et al.: The Star Schema Benchmark (SSB),
http://www.cs.umb.edu/~poneil/StarSchemaB.PDF
12. ONeil, P.E., et al.: Adjoined Dimension Column Index (ADC Index) to Improve Star
Schema Query Performance. In: ICDE, pp. 14091411 (2008)
13. Begley, S., et al.: MCJoin: A Memory-Constrained Join for Column-Store Main-Memory
Databases. In: ACM SIGMOD, pp. 121132 (2012)
14. Idreos, S., et al.: Self-organizing Tuple Reconstruction in Column-stores. In: ACM
SIGMOD, pp. 112 (2009)
 A Self-healing Framework for QoS-Aware Web
Service Composition via Case-Based Reasoning

Guoqiang Li1,2 , Lejian Liao1 , Dandan Song1, , Jingang Wang1 , Fuzhen Sun1 ,
and Guangcheng Liang1
1
Beijing Engineering Research Centre of High Volume Language Information
Processing & Cloud Computing Applications, Beijing Key Laboratory of Intelligent
Information Technology, School of Computer Science, Beijing Institute of Technology,
Beijing, China
2
School of informatics, Linyi University, Linyi, China
{lgqsj,sdd,liaolj,bitwjg,10907023,195812}@bit.edu.cn

Abstract. The self-healing ability is very important for service-oriented

systems in a highly dynamic environment. Case-based reasoning is
adopted to cope with the component services failing to meet with the
functional and nonfunctional requirements in this paper. We take previ-
ous failure instances as cases which are stored in a case base. When a
new fault occurs, its symptoms are extracted and matched against the
case base to look for the most similar case. A case representation and
a similarity function are proposed. Meanwhile, a novel reuse approach
is designed to find the solutions satisfying the symptoms. Moreover we
present a self-healing framework and conduct experiments with real QoS
dataset. Experimental results show that case-based reasoning improves
the self-healing ability encouragingly.

Keywords: Quality of Service,case-based reasoning, self-healing frame-

work, web service.

1 Introduction

Service-oriented architecture has become a mainstream technology for business

process integration. However, in complex network environments, composite ser-
vices may encounter unexpected risks, e.g. network breaking down, or a service
unavailability. Therefore, service-oriented systems should be equipped with self-
healing ability in order to run without interruption.
The main idea of our approach is that each failure instance can be taken as
a failure case. We engage in using previous experiences to understand and solve
new problems, which is the core ideology of the case-based reasoning (CBR).
Its advantages include: (1) proposing solutions to problems quickly, (2) allowing
us to make assumptions and predictions based on the past, (3) remembering
previous experiences to avoid repeating past mistakes. (4) adapting to changes

Corresponding author.

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 654661, 2013.
c Springer-Verlag Berlin Heidelberg 2013

 A Self-healing Framework for QoS-Aware Web Service Composition 655

of the environment [1]. To the best of our knowledge, our work is the first one
to use CBR in the service adaptive domain.
The remainder of this paper is structed as follows. In Section 2 we present
the related work. The self-healing framework is described in Section 3. Section
4 sets forth a case representation, a retrieval algorithm and a reuse approach in
detail. In Section 5 we present experiments of our work. At last, in Section 6 we
discuss conclusion and future work.

2 Related Work
Although WSBPEL1 specifies business processes behavior as a standard, its
fault-handling mechanism is limited as it is dicult for users to consider the dy-
namic information of services and their running environment in advance. Many
works extended BPEL engine to enhance their fault-handling capabilities. A
multi-layered control loop is proposed in [2], which focuses on the framework
structure. VieDAME [3] monitors BPEL process according to Quality of Service
(QoS) and replaces existing partner services based on various strategies. In [4],
AOP (aspect-oriented programming) is used to extend the engine ActiveBPEL
by using ECA (event-condition-action) rule paradigm. In [5], Web service Re-
covery Language is defined inspired by the ECA rule paradigm with emphasis
on the process recovery. Similarly, in [6], ECA rules are designed to change the
components states according to the fault events. In addition, vital and non-
vital components are distinguished to express the importance of activities in the
workflow. In Dynamo [7] and DISC [8], JBOSS rule and EC (event calculus)
model are used respectively. But in the above mentioned works, it is dicult
for designers to design the complete set of rules and event models before the
process execution. Our approach can automatically generate the failure cases on
the basis of the fault information. With this advantage, our approach is easier
to be accepted by the users.
When QoS constraints are violated during the execution, the problem of select-
ing supplementary service is formalized as a mixed integer linear programming
problem in [9], where negotiation techniques are exploited to find a feasible so-
lution. In [10], a staged approach generates multiple workflows for one template
(abstract workflow), and some workflows are selected as an IP optimization
problem by considering QoS parameters. In [11], a reconfigurable middleware is
proposed, covering the whole cycle of adaptation management including moni-
toring, QoS analysis and substitution-based reconfiguration. In [12], failures are
classified into several categories from the perspective of a service composition.
Then process reorganization, substitution and retrial/data mediation are pro-
posed to react to the failures. In [13], a region-based reconfiguration algorithm
is designed to minimize the reconfiguration cost. Supplementary services are se-
lected as a backup source for recovery by using the linear programming. In [14],
DSOL re-planning technique is leveraged by Q-DSOL to optimize the orches-
tration supporting QoS in presence of faults. Dierent from previous works, our
1
http://docs.oasis-open.org/wsbpel/2.0/wsbpel-v2.0.html
656 G. Li et al.

chief consideration is on the selection of adaptive action according to the typical

fault information considering QoS constraints. We carry on the candidate chain
search based on the similarity using previous experiences as much as possible.

3 Self-healing Framework
There are three parts in our framework. The first part is the business process
which achieves users requirements. The framework is shown in Figure 1. The
second part extractor gets the fault information (FIN), services functional in-
formation (FI) and non-functional information (NFI) which are passed to the
third part case-based reasoner. After reasoning, a case with solution (CWS)
is returned to the business process and the satisfied solution keeps the process
running forward rightly. A case without solution (CWOS) is viewed as a new
fault. We assume that fault information has been caught using some method(e.g.
AOP). The reasoning pseudo code is presented in Algorithm 1.

s31
extractor s1 s21 s51 s 61
business
Business
s41
process

throw fault
return solution

representation retrieve reuse verify case-based

reasoner

Fig. 1. Self-healing framework

Algorithm 1. Implementation of CBR

1: symbols: fault information data base(FIDB), case base(CB)
2: input: FIN, NFI
3: output: solution (sn)
4: Begin:
5: interacting with FIDB , CWOS= generate(FIN, NFI)
6: searching CWS in CB with the similarity function, CWS = retrieve(CWOS)
7: extracting the solution of CWS, denoted as sn
8: if sn is replace
9: verify(sn) goto 12.
10: else sn is repeat
11: return sn. goto 15.
12: if sn satisfies QoSC
13: return sn. goto 15.
14: else goto 6.
15: updating CB with sn and CWOS.
16: end
 A Self-healing Framework for QoS-Aware Web Service Composition 657

4 Detailed Implementation

Suppose there is a case base CB = (C1 , C2 , C3 ) and a services composition SC =

{f (S1 , S2 , S3 ,. . . ), QoSC}, where f (S1 ,S2 ,S3 ,. . . ) is the composition connecting
all services by using sequence, condition and loop structures. QoSC denotes the
users constraints about QoS. So, our task is to find a case Ci whose solution
can deal with a failed service Sj and satisfy QoSC simultaneously.

4.1 Case Representation

In our approach, FI and NFI information can be easily extracted from WSDL
files. FIN is represented by using the HTTP status codes as in [15]. The case
symptoms (CS) can be described as a triple: CS = (FI, NFI, FIN), where FI =
(input, output), NFI = (availability, cost, successability, response time), FIN =
(fri, nfri). Their definitions refer to WSBPEL and [15]. Here we only give one
part as follows:

1. functional information(FI)
(a) input: abstract message format for the solicited request.
(b) output: abstract message format for the solicited response.
2. nonfunctional information(NFI)
(a) availability: the percentage of time is operating during a time interval
(b) throughput: number of successful invocations/total invocations
3. fault information(FIN)
(a) fri: function related information
(b) nfri: non-function related information

We classify adaptive actions into two categories repeat and replace which are
commonly used in previous works, e.g.[12]. For example, the code 404 denotes
that service is not found, so replace should be chosen. There are 20 standard
faults defined in WSBPEL specification where they are distinguished according
to whether they are related to input data. For example, the fault invalidEx-
pressionValue is data related, and repeat can be used. Analysis shows only
missingReply and missingRequest can be thought as data unrelated.
We construct a case representation with solution and solution templates shown
in Table 1 and 2 respectively. Codes with the prefix d can be defined by the
users in the first column in Table 2. Codes with prefix b come from WSBPEL
and the prefix is added by us. The remaining codes come from [15]. The under-
lined tasks in template column will be replaced with a concrete service in Table
2. A new problem is a case without the solutions part of Table 1.

4.2 Retrieval Algorithm

In light of our initial study, the case base is stored in a linear list. In real word,
we can index the cases according to attributes of the service to improve the
retrieval eciency. Due to the fact that attributes of QoS vary in units, they
658 G. Li et al.

Table 1. Case in the case base

Problem(symptoms) Solutions
(1)problem: fault description
(2)service location:
(3)service input:
(4)service output: (1)Diagnosis: the used data is faulty
(5)availability:89 (2)Repair: repeat invoking the service with
(6)throughput:7.1 correct data
(7)successability:90
(8)response time:302.75

Table 2. Solution templates

code explanation templates

403 forbidden Diagnosis: the network doesnt work well.
405 method not allowed Repair: replace this service with another one

b901 ambiguousReceive Diagnosis: the service doesnt work well with incorrect
data.
411 Length required Repair: repeat this service with correct input data.

b906 MissingReply Diagnosis: the service violates the QoS constraints

d801 services availability Repair: replace this service with another one, reset the
is lower weight factors of the QoS.

should be normalized to be [0,1]. A QoS attribute matrix is shown in Eq. (1)

same as the one in [16].

q11 q12 . . . q1n
q21 q22 . . . q2n
A=
. . . . . . . . . . . . . . . . (1)
qm1 qm2 . . . qmn
where qij is in the range [0,1]. So, we use the Euclidean distance to compute two
cases similarity: @
AN
A"
d =B
2
ij (C C )
ik jk k (2)
k=1

The factor k is the weight denoting the users preference for an attribute of
QoS and sum of these factors is 1. Cik comes from the matrix A of Ci .

4.3 Reuse Approach

After retrieving top N cases, we use the solutions to achieve adaption. Two
conditions are considered.
 A Self-healing Framework for QoS-Aware Web Service Composition 659

Condition 1: if the adaptive action is repeat, the failed service can be re-
invoked directly.
Condition 2: if the adaptive action is replace, the solution is verified to check
whether QoS attributes deviate from users constraints.

Our approach automatically generates two alternatives for one new case. Assume
S1 , S2 and S3 occur in the case base corresponding to C1 , C2 and C3 . C2 is similar
to C1 and its solution is: replace S2 with S3 . So, C3 is also a candidate of C1 .
Even though their similarity may be bigger than the similarity threshold, we
think that our method has a great success rate as long as the threshold is as
small as possible.
Finally, we transform the validation of solution against QoS requirements into
a constraint satisfaction problem. In our paper, it is assumed that the candidate
service of the failed service can be found in the case base. The reason is that the
business process will return to a correct state by using the compensation handler
in WSBPEL. How to implement this compensation mechanism with CBR is a
future research topic.

5 Experiment

The QWS dataset2 is used in our experiments including 2,507 Web services.
We use five attributes: response time, availability, throughput, successability,
reliability. We assume all services are functionally equivalent and the services
composition uses sequential structure including 20 services. We record the com-
parison times(CPT) about the failed services which are generated randomly.
We build the case base only using the first 2400 services. In the first experi-
ment, we do not limit the search step (SS). The results in Figure 2(a) show that
CPT is not proportional to the number of failed services. We can see that the
CPT is very high at point 8 and 10, where the CPT of point 10 is 938.2 because
there are 30% of found services coming from the remaining 107 services. So, it
is very important to build a case base as completely as possible.
In the second experiment, two conditions are compared: step-limited and step-
unlimited. The latter is our reuse approach and the value of SS is equal to 5.
In Figure 2(b), we can see that the CPT of step-unlimited is lower than that
of step-limited. In particular, when there are 9 failed services, the comparison
times reduce 26 times. So, our method takes less time to achieve the adaption.
At last, we make a further study on the influence on CPT by dierent search
steps . In Figure 2(c), the results show that the CPT can be reduced significantly
with the increase of search steps. But, it does not mean that the bigger steps are
better. For example, when SS is equal to 4 and 5, the comparison times are very
close. So, in practice, we should determine a proper value of the search steps.
2
http://www.uoguelph.ca/ qmahmoud/qws/index.html#Quality of Web Service
(QWS) This dataset has more QoS attributes than that on the web site
http://www.zibinzheng.com/
660 G. Li et al.

1000 40 90
stepunlimited step3
900 80
35 steplimited step4
step5
800
70
30
700
Number of Comparisons

60

Comparisons Times

Comparisons Times
25
600
50
500 20
40
400
15
30
300
10
20
200
5 10
100

0 0 0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
Number of the Failed Services Number of the Failed Services Number of the Failed Services

(a) services influence (b) reuse approach (c) dierent steps

Fig. 2. Comparison times under dierent conditions

6 Conclusion and Future Work

We presented a framework by means of the case-based reasoning. This framework
can dynamically adapt to the failure of composite services eciently. According
to the principle of the case-based reasoning, we focus on the developments of the
case representation, the retrieval algorithm and the reuse approach. Experiments
are conducted to show the eectiveness of our approach with real dataset.
In practice, the case base may become very large. So, how to maintain the
base will be a challenge to use this method in the future. An interesting research
direction is to determine how the eciency of our approach is aected by complex
services composition structures.

Acknowledgment. This work is funded by the National Program on Key Ba-

sic Research Project(973 Program, Grant No.2013CB329605), Natural Science
Foundation of China (NSFC, Grant Nos. 60873237 and 61003168).
We are very grateful to Qusay H. Mahmoud for providing the QWS Dataset.

References
1. Kolodner, J.L.: An introduction to case-based reasoning. Artif. Intell. Rev. 6(1),
334 (1992)
2. Guinea, S., Kecskemeti, G., Marconi, A., Wetzstein, B.: Multi-layered monitoring
and adaptation. In: Kappel, G., Maamar, Z., Motahari-Nezhad, H.R. (eds.) ICSOC
2011. LNCS, vol. 7084, pp. 359373. Springer, Heidelberg (2011)
3. Moser, O., Rosenberg, F., Dustdar, S.: Non-intrusive monitoring and service adap-
tation for ws-bpel. In: Proceedings of the 17th International Conference on World
Wide Web, WWW 2008, pp. 815824. ACM, New York (2008)
4. Liu, A., Li, Q., Huang, L., Xiao, M.: Facts: A framework for fault-tolerant composi-
tion of transactional web services. IEEE Transactions on Services Computing 3(1),
4659 (2010)
5. Baresi, L., Guinea, S.: A dynamic and reactive approach to the supervision of bpel
processes. In: ISEC 2008, pp. 3948 (2008)
6. Ali, M.S., Rei-Marganiec, S.: Autonomous failure-handling mechanism for wf long
running transactions. In: Proceedings of the 2012 IEEE Ninth International Con-
ference on Services Computing, SCC 2012, pp. 562569. IEEE Computer Society,
Washington, DC (2012)
 A Self-healing Framework for QoS-Aware Web Service Composition 661

7. Baresi, L., Guinea, S., Pasquale, L.: Self-healing bpel processes with dynamo and
the jboss rule engine. In: International Workshop on Engineering of Software Ser-
vices for Pervasive Environments: in Conjunction with the 6th ESEC/FSE Joint
Meeting, ESSPE 2007, pp. 1120. ACM, New York (2007)
8. Zahoor, E., Perrin, O., Godart, C.: Disc: A declarative framework for self-healing
web services composition. In: IEEE International Conference on Web Services,
ICWS 2010, Miami, Florida, USA, July 5-10, pp. 2533. IEEE Computer Society
(2010)
9. Ardagna, D., Pernici, B.: Adaptive service composition in flexible processes. IEEE
Transactions on Software Engineering 33(6), 369384 (2007)
10. Chafle, G., Dasgupta, K., Kumar, A., Mittal, S., Srivastava, B.: Adaptation inweb
service composition and execution. In: International Conference on Web Services,
ICWS 2006, pp. 549557 (September 2006)
11. Ben Halima, R., Drira, K., Jmaiel, M.: A qos-oriented reconfigurable middleware
for self-healing web services. In: IEEE International Conference on Web Services,
ICWS 2008, pp. 104111 (September 2008)
12. Subramanian, S., Thiran, P., Narendra, N.C., Mostefaoui, G.K., Maamar, Z.: On
the enhancement of bpel engines for self-healing composite web services. In: Pro-
ceedings of the 2008 International Symposium on Applications and the Internet,
SAINT 2008, pp. 3339. IEEE Computer Society, Washington, DC (2008)
13. Li, J., Ma, D., Mei, X., Sun, H., Zheng, Z.: Adaptive qos-aware service process
reconfiguration. In: Proceedings of the 2011 IEEE International Conference on
Services Computing, SCC 2011, pp. 282289. IEEE Computer Society, Washington,
DC (2011)
14. Cugola, G., Pinto, L., Tamburrelli, G.: Qos-aware adaptive service orchestrations.
In: IEEE 19th International Conference on Web Services (ICWS), pp. 440447
(June 2012)
15. Al-Masri, E., Mahmoud, Q.H.: Investigating web services on the world wide web.
In: Proceedings of the 17th International Conference on World Wide Web, WWW
2008, pp. 795804. ACM, New York (2008)
16. Al-Masri, E., Mahmoud, Q.: Qos-based discovery and ranking of web services. In:
Proceedings of 16th International Conference on Computer Communications and
Networks, ICCCN 2007, pp. 529534 (August 2007)
 Workload-Aware Cache for Social Media Data

Jinxian Wei, Fan Xia, Chaofeng Sha, Chen Xu,

Xiaofeng He, and Aoying Zhou

Institute of Massive Computing, Software Engineering Institute,

East China Normal University
{jinxianwei,fanxia,chenxu}@ecnu.cn, [email protected],
{xfhe,ayzhou}@sei.ecnu.edu.cn

Abstract. The success of social network services has brought up many

interesting Web 2.0 applications, while posed great challenges for human-
real-time data management for huge volume of data with unstructured
nature. Timeline query is a specific type of queries that are widely used in
social network services and analysis. A workload-aware cache scheme for
ecient evaluation of home timeline queries in human-real-time manner
is proposed in this paper. It utilizes the communities within followship
network and considers the high skew of access frequency across users
to generate cache units. Thus, timeline queries are transformed to a
process of merging cache units. Empirical studies show the superiority
of overlapping cache strategy over other three existing strategies.

1 Introduction

The success of social network services (SNS), such as Facebook and Twitter, has
brought up many interesting Web 2.0 applications, while posed great challenges
for human-real-time data management for huge volume of data with unstruc-
tured nature.
Many SNS tasks, including those for both SNS providers and end-users, can
be represented by a specific kind of queries, i.e. timeline query (TQ). Intuitively,
a timeline query is a top-k or windowed query for records satisfying specific
conditions, in reverse order of timestamps associated with those records. Home
timeline query (HTQ), as a kind of typical timeline queries, is to retrieve a
timeline with records created by a given users followees.
Fig. 1, for example, shows a global timeline within the period between times-
tamps t0 and tj of six users, i.e. u1 , u2 , u3 and v1 , v2 , v3 , whose following rela-
tionships are also given. Here, ui,j means the jth message posted by user ui .
The top-5 results of home timeline queries for u1 , u2 , u3 at timestamp tj are also
listed. Here, the query condition is that the message queried should belong to
a querying user himself or his followees. Thus, v3,1 is included in the results of
HT Q(u1, tj , 5) and HT Q(u2 , tj , 5), but not of HT Q(u3, tj , 5), since user u3 is
not following v3 .
However, evaluation of home timeline queries in a human-real-time manner is
non-trivial. On the one side, the data should be cached in memory with orders or

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 662673, 2013.
c Springer-Verlag Berlin Heidelberg 2013

 Workload-Aware Cache for Social Media Data 663

Fig. 1. An example of home timeline queries over a global timeline of six users

indices on timestamps, so that the top-k or time-window part of the query can
be eciently evaluated. On the other side, a specific query may only retrieve a
very small portion of the data distributed in the whole timeline, while dierent
queries may extract dierent set of records with overlaps. The sparsity of the
results makes the ordering of the global timeline in the system meaningless.
The report in [1] shows three caching schemes to facilitate the evalation of
home timeline queries over social media data. The pull -base method caches the
messages of each user separately by putting any item to the cache unit of its
corresponding user simply. This strategy requires to merge the messages to con-
struct home timeline so that it suers the problem of high-cost in cache merge
leading to high latency. The push-base approach directly caches home timeline
query result by delivering user generated message to his followers eagerly. Clearly,
it consumes too much memory for duplicated items and imposes heavy load to
CPU for vast deliveries. To strike a balance between pull-base and push-base
approach, a hybird strategy distinguishes popular users, such as celebrities who
have a large volume of followers, from the ordinary ones. The messages gener-
ated by popular users are cached following pull-based method, while messages
from the other users are processed as push-based approach. Consequently, ac-
tive users are most likely to have worse user experience in the view of retrieval
latency since they are always following a certain number of popular ones.
A workload-aware caching scheme named overlapping caching is proposed in
this paper with the observation that users of SNS tend to form communities
with homogenization and access frequency has high skew among users in SNS.
It distinguishes itself from previous work in the following aspects.

A cache unit is shared by a group of users with close relationships in the

followship network, while user groups may have overlap with each other.
The evaluation of timeline queries is transformed to an ecient process of
merging cache units.
Users access frequency is taken into consideration ensuring that our method
outperforms other native caching schemes in general case.
Empirical studies upon data sets simulated on the basis of real-life social
media data show the superiority of overlapping cache strategy over other
three existing strategies in the view of both system cost and performance.
664 J. Wei et al.

Table 1. Symbols used in this paper

Notation Description Notation Description

u, ux , ui User identifiers ty , tj Timestamps
U A set of user identifiers TG Global timeline
i Update frequency of ui i Request frequency of ui
Fi Set of users followed by ui S Set of all shared timelines
T sl a shared timeline Cl Set of users covered by T sl
Pl Set of users who share T sl Xi Set of shared timelines owned by ui

The rest of the paper is organized as follows. In Section 2, we state the prob-
lem of home timeline queries formally. The workload-sensitive caching scheme
is illustrated in Section 3 and the empirical study results are shown in Section
4. Section 5 lists the related work followed by the conclusion of this paper in
Section 6.

2 Problem Statement
A timeline is a sequence of items (e.g., messages) that are ordered chronologi-
cally. Each item is a triple < ux , ty , mx,y >, in which ux is the user identifier,
ty denotes the timestamp, and mx,y is the id of the item created by ux at ty . A
followship network is a directed graph G : (V, E), in which V is the set of users,
and E is a subset of V V . (u, ui ) E means that user u is following user ui .
In many real-life SNS, every user in default subscribes his own messages, i.e.
u, (u, u) E. We denote all ui s followees as Fi , i.e. Fi = {u|(ui , u) E}.
The raw data that to be queried is a global timeline, denoted as TG , in which
items from dierent users are interlaced. A timeline query T Q(U, tj , k) is a query
at tj to retrieve last k items < ux , ty , mx,y > in TG satisfying both ty tj and
ux U , where U is a set of user identifiers.
Home timeline queries, as a kind of typical timeline queries, are notoriously
common in SNS applications. Every time a user logs in or refreshes his/her web
page, home timeline query is triggered. It retrieves ui s home timeline, denoted
as Ti , which is the timeline with all items from this users followees. Thus, the
home timeline query of user ui at time tj can be denoted as HT Q(ui , tj , k) =
T Q(Fi , tj , k).
The main problem in this work is how to achieve low latency and high through-
put with limited system resources. Here, latency is the response time of home
timeline query, whereas throughput means the operations consisting of creating
items and retrieving home timelines per time unit (e.g., seconds). In addition,
we take RAM and CPU as system resources into consideration, since most of
the data is stored in RAM for performance reasons in many real applications.
Formally, let us define CostRAM () as the total consumption of memory and
CostCP U () as the aggregate CPU utilization. Then, the home timeline query
problem is to minimize both CostRAM () and CostCP U (), while ensuring the
 Workload-Aware Cache for Social Media Data 665

     

       






 













 


(a) pull (b) push

             



 






 



 









    





(c) hybrid (d) overlap

Fig. 2. An overview of caching for timeline query processing

average query latency is less than M time units (e.g., milliseconds) and the
average throughput is higher than N operations per time unit in the premise
that home timelines are provided correctly.

3 Workload-Sensitive Caching

3.1 Overview of Four Caching Strategies

Two trivial solutions for home timeline query processing are pull -based method
and push-based method, which are illustrated in Fig. 2(a) and Fig. 2(b) respec-
tively. Under pull strategy, any message item < ux , ty , mx,y > is cached only once
in ux s user timeline where all messages posted by u are organized chronologi-
cally. Thus, while processing a home timeline query HT Q(ui , tj , k), the problem
is to retrieve user timelines of ui s followees, i.e, Fi . Dierently, the push-based
method assumes that all home timeline queries are known in advance and re-
sults of them are pre-computed and cached. While the query is triggered, the
system just retrieve the result from cache and deliver it to the user. Clearly, to
support push strategy, each item < ux , ty , mx,y > is cached as many times as the
number of ux s followers. However, neither pull-based nor push-based method
is ideal for home timeline query processing. The former suers the problem of
high-cost in cache merge leading to high latency, whereas the later consumes too
much memory for duplicated items and imposes heavy load to CPU.
666 J. Wei et al.

As a trade-o between the two strategies aforementioned, a hybrid approach

is reported that combines pull-based approach with push-based one [1], which is
shown in Fig. 2(c). In this method, messages from popular users, i.e., the users
with many followers, are cached once in their own user timelines as pull-based
method, while messages from the other users are delivered to home timelines of
their followers as push-based approach. Generally, the hybrid strategy decreases
CPU cost and requires less memory space in comparison with the push one,
and reduces the latency in cache merge in contrary to pull one. However, it is
oblivious to the active user who have higher request rates and tend to have a
higher possibility to follow more popular users. Unfortunately, those active users,
who play crucial roles in applications, will have worse user experience in the view
of home timeline query latency under hybrid strategy.
Those drawbacks above drive us to study the properties of workload from
an alternative aspect rather than simply distinguishing the popular users from
the other ones. It is observed that users of SNS tend to form communities with
homogenization and users within the same community are most likely to have
similar followships such as many common followers. Intuitively, the messages
from a group of users can be shared by their common followers so that it is
unnecessary to cache these messages in duplicate.
Shared timeline, denoted as T sl , is a timeline generated by a corresponding
shared group of users Cl V . In other words, each item < ux , ty , mx,y >
in T sl satisfies that ux Cl . The overlapping caching strategy we propose is
illustrated in Fig. 2(d), where each message < ux , ty , mx,y > is delivered to any
T sl satisfying ux Cl and each T sl is used to answer home timeline query
HT Q(ui , tj , k) if Cl Fi .
Besides exploiting the community characteristic in SNS as above, this workload-
sensitive caching method also takes users access frequency into consideration due
to the fact that both read (query a home timeline) frequency and write (post a
message) frequency have high skew across users. On one hand, write frequencies
of users in one shared group are relevant to the size of the corresponding shared
timeline, which influences system cost including RAM usage and CPU utiliza-
tion. On the other hand, decreasing latency for users with high read frequency
results in low aggregated read latency, in which way, system throughput is im-
proved. A natural question is how to determine shared groups while taking both
social networks and user access frequency into account followed by answering
home timeline queries via shared timelines based on those shared groups.

3.2 Shared Groups Determination

The result shared groups are expected to resolve the problem stated in Section
2. That is how to achieve low latency and high throughput with limited system
resources. Thus, the determination of shared groups is forced to facilitate the
evaluation of home timeline queries to consume RAM and utilize CPU in an
economical manner while satisfactory performance is guaranteed.
 Workload-Aware Cache for Social Media Data 667

We first formulate the shared groups determination problem, then present

our greedy solution to this problem followed by a linear community detection
algorithm that making use of user profiles, which serves as the basis of our greedy
solution.

Problem Formulation. To formulate the shared groups determination prob-

lem, we additionally introduce some notations. Let Pl denote the set of users
who share T sl covering updates from users in Cl . Clearly, followees of users in
Pl are supersets of Cl . We use Xi to denote the set of shared timelines that user
ui participating to share. Given the update (write) and request (read) frequency
of user ui , denoted as i and i respectively, we now define the shared groups
determination problem, with a setting threshold , as follows:
" "
min i
T sl S ui Cl

s.t. K
Cl = Fi , i (1)
ui Pl

|Xi | /i , i (2)
Constraint 1 ensures that one users home timeline can be generated by merging
all his/her shared timelines. Constraint 2 limits the number of shared timelines
each user owns, which is bound to request frequencies of users home timelines.
The problem of determination of the shared timelines is non-trivial. We sim-
plify the original problem by setting all i = 1 and formally define the simplified
problem as Volume Bounded Set Basis Problem, which can be reduced to the
classic Set Basis problem [2].
Theorem 1. The Volume Bounded Set Basis Problem is NP-hard.
Details of the definition of Volume Bounded Set Basis Problem and its complexity
analysis along with the proof of Theorem 1, are attached in Appendix.

Community-Based Greedy Identification of Groups. Although a greedy

algorithm upon the entire network gives a good approximation to the problem
of determining shared groups, it is expensive for large social networks like Sina
weibo who has a user base of hundreds of millions. The main idea of our solution
is to divide the network into overlapped communities, and adopt greedy algo-
rithm for selecting shared groups in communities for the purpose of eciency.
The calculated shared groups, on the other side, can be reused by users across
communities.
Suppose that we already divide the network into communities(to be discussed
in the following subsection).We now present our greedy algorithm to determine
shared groups as Algorithm 1, where the gain of Ct is calculated as:
"
gain(Ct ) = u
uCt ,Ct Fu
668 J. Wei et al.

Algorithm 1. Greedy algorithm for determination of shared groups

Input: The followship network: G(V, E); community: C
Output: A set of user groups: R = {Ri |Ri V }
1: for all vi V do
2: Compute the number of shared timeline wi for vi ; wi wi ;
3: Fi {u|(v, u) E}; Fi Fi ;
4: end for
5: while there are vi s that cannot be satisfied by T sj s for Rj R do
6: for all vi whose wi = 1 do +
7: Create a new Rk Fi ; R R {Rk }; wi 0;
8: end for
9: for all vi with wi > 0 in ascendant order of wi do
10: Create a new Rk arg+maxgain(Ct ) Ct : {u|u Fi , (u, t) C, (vi , t) C};
11: Fi Fi Rk ; R R {Rk }; wi wi 1;
12: for all u with (u, t) C and Fu Rk = do
13: Fu Fu Rk ; wu wu 1;
14: end for
15: end for
16: end while

Profile-Based Community Detection. Intuitively, users of one community

tend to share a considerable portion of common followers and followees. Derived
from that, we propose a community detection algorithm to filter users in the
determination of shared groups in a more ecient manner. Graph cluster al-
gorithms can be used to find out the communities, but their high complexity
make it impractical to apply them to a real life social network[3]. We here ex-
plore the profiles of users and assign initial communities to each user based on
their profiles. Then Label Propagation Algorithm(LPA)[4] is used to decide the
final communities for them. In LPA, the communties of users are voted by their
followees. First, LPA will propagate the communities of a user to each of its
followers. Then, it chooses topK voted communities for each user. We provide a
map reduce implementation of this algorithm. In this implementation, one job
is used to propagate the communities to the followers. The other job aggregates
the votes and decides the top communities for each user. Clearly, labels of a user
are sent to all his followers, so the network trac of these two jobs is O(| E | C),
where | E | stands for the number of edges and C is an constant constraining
the maximum number of communities one user belongs to.

3.3 Timeline Query Processing in Overlapping Caching

Once shared groups are determined, each shared timeline, which organizes mes-
sages from users in corresponding shared group, can be generated. Under over-
lapping caching strategy, home timeline query can be answered by merging all
related shared timelines. The merging phrase is to remove duplicates and return
top items among all the related shared timelines, whose algorithm can be found
in Algorithm 2.
 Workload-Aware Cache for Social Media Data 669

Algorithm 2. Merge Algorithm

Input: all Tsi for u, return count l;
Output: top l items stored in set Q,
1: create priority queue P
2: enqueue all Tsi into P according to the first element of Tsi
3: while P.size > 0 and Q.size < l do
4: tl P.poll
+
5: Q Q {tl[0]}
6: tl tl {tl[0]}
7: if tl = then
8: enqueue tl into P
9: end if
10: end while
11: return Q;

4 Empirical Study
We now examine our experimental results. The four strategies(i.e., pull, push,
hybrid and overlap) described in Section 3 are implemented to handle the mes-
sage load obtained from a real-life social media(Sina Weibo). We show a com-
parison of the above four caching strategies from the aspects of system cost and
performance.
Our system cost metric is aggregate memory usage and CPU utilization across
all servers. We put a limit on memory capacity in order to evaluate performance of
the strategies when some cache misses happen, which is the common case in many
real-life applications. CPU utilization is measured by Unix sar command. Perfor-
mance is evaluated in terms of latency and throughput. We measure the latency
as response time observed from the client, for retrieving one users home timeline.
Throughput is measured as the number of operations performed per second. Here,
an operation means posting a message or retrieving a home timeline.
In summary, our results show: The overlapping caching strategy strikes best
balance between system load and performance: it utilizes least CPU, consumes
second least memories, and can achieve lowest query latency and highest through-
put at most time under the memory constraint in our experiment.

4.1 Experimental Data and Setup

We construct our data set on the basis of a real-life data set crawled from Sina
Weibo, the Chinese edition of Twitter. The real data set consisting of 1,289,986
users and all the messages they post from in 2011. We got the number of their
followers and the users they are following. Unfortunately, the social network
we crawled is incomplete and the retrieval log of home timeline is unavailable.
Therefore, as a substitute we reconstruct their followships among the 1,289,986
users following the distribution of the number of their followers, which is about
100 on average, and generate a retrieval log in which users has bigger post
frequencies tend to retrieve more often. Also, we rank the users by the number of
670 J. Wei et al.

1 14
push push
pull pull
hybrid 12 hybrid
0.8 overlap overlap

Utilization of CPU(%)
Usage of Memory
10
0.6
8

0.4 6

4
0.2
2
0
0 100 200 300 400 500 600 700 800 0
Operations(x104)
0 100 200 300 400 500 600 700 800
Time Series

(a) RAM usage (b) CPU utilization

Fig. 3. A comparison of system cost for the four strategies

their followers and regard the top 7,509 users as popular ones, who has more than
1,000 followers. All the popular users are treated dierently in hybrid strategy.
To make a fair comparison of system cost and performance among the four
caching strategies, each of them is run against the same workload. That is, the
same sequence of messages posted in real-life application and the same retrieval
log we simulate as mentioned are used in our experiments. The ratio of posts and
retrievals is 1:4. We also launch the same number of threads for each strategy.
We run our experiments on server-class machines(8 core [email protected] Intel
Xeon CPU, 16GB of RAM, 1TB of disk and gigabit ethernet). In particular,
we use twelve Redis storage servers(six for meta data) and one node to launch
and process requests. Memory capacity for each the Redis server is limited to
100MB, under which a portion of cache misses encounter under all the strategies
except the pull one.

4.2 System Cost

We start by examining the memory usage and CPU utilization of the six servers
under four strategies. Fig. 3 indicates: The push strategy reaches the maximum
capacity of memory first and imposes heaviest load to system CPU with increas-
ing number of posted messages; The pull based strategy uses the memory most
economically; Our proposal utilizes least CPU and consumes the second least
memory, confirming our hypothesis.

4.3 Performance
Next, we evaluate the performance of the four strategies. As for the latency and
throughput reported in Fig. 4, we use average values. Our experiment results
show that: The push based strategy brings lowest query latency and highest
throughput before reaching the maximum memory capacity and results in a
sharp drop in performance since it consumes most memory and meets the limit
of memory capacity first; Performance of the pull method is quite stable yet
trivial; The overlap based strategy achieves lowest query latency and highest
throughput at most time.
 Workload-Aware Cache for Social Media Data 671

Latency of Home Timeline Query(ms)

8 7000
push push
pull pull
7 hybrid 6000 hybrid
overlap overlap
6

Throughput(ops/s)
5000
5
4000
4
3000
3
2000
2

1 1000

0 0
0 100 200 300 400 500 600 700 800 0 100 200 300 400 500 600 700 800
Operations(x104) Operations(x104)

(a) latency (b) throughput

Fig. 4. A comparison of performance for the four strategies

5 Related Work
Both industry and research community have paid great eort to develop ap-
proaches to support home timeline query, which is a key feature in SNS. There
are three strategies(i.e pull,push and hybrid) that were proposed previously to
generate the home timelines for users . Facebook is one of the most giant SNS
providers in the world. According to its report[5], pull based strategy was im-
plemented to support its service and push based one was constructed on the fly.
On the other hand,Twitter[1] reported recently a hybrid strategy: push messages
from the majority of ordinary users and pull messages from popular ones.
A.Silberstein.et.al[6] have introduced a formal solution to hybrid strategy from
the aspect of materialization. They proposed a cost based heuristic method to
decide whether the tweets should be materialized for each pair of followees and
followers. Besides, some arguments can be tuned to make a tradeo between
the latency and throughput. Some other researchers[7][8] try to approach this
problem from the distributed management of graph data. Each user in the social
network is represented by a node in the graph and home timeline is generated by
issuing read requests to the neighbors. They try their best eort to minimize the
communications between sites. In order to sustain the local semantic of query,
[7]replicates all the nodes that need to be accessed on remote sites to local.
However, this approach causes a great of storage redundancy and may result in
great eorts to maintain the replications. [8] introduces a more elaborate ap-
proach. The nodes are partitioned randomly and clustered into multiple clusters
in each partition.The replication decision is made for each pair of clusters and
partitions in according to their access behaviors.The experiments illustrate that
their method is better than [7]. Besides, there are several decentralized social
network systems[9][10][11] proposed. [10] incorporates a concept similar to hy-
brid strategy. Users in its system are divided into social users and media users.
Media users typically have to broadcast messages to a lot of users. As a result,
messages of media users are disseminated by gossiping.
672 J. Wei et al.

Our work focuses on the microbloging service and proposes a new strategy.
[4][3] all aim to detecting overlapping communities in the social network. [4]
provides an algorithm that is linear to the number of edges in the network. We
provide a MapReduce implementation to preprocess the dataset and help us
understand the data.

6 Conclusion

In this paper, to minimize memory usage and CPU utilization while ensuring
high throughput and low latency, we propose a novel workload-aware cache
scheme, i.e., overlap-base strategy, to evaluate the home timeline query. This
strategy makes full use of workload information including social communities
and user access frequency in SNS. Our experimental studies show overlap-based
approach outperforms all other three strategies including pull-based, push-based
and hybrid strategies in both system cost and performance.

Acknowledgement. This work is partially supported by National Major Projects

on Science and Technology under grant number 2010ZX01042-002-001-01.

References

1. Asemanfar, A.: Timelines @ Twitter. In: International Software Development Con-

ference (QCon London) (2012)
2. Stockmeyer, L.: Set Basis Problem is NP-Complete. Technical report, Report RC-
5431, IBM T.J. Watson Research Center, Yorktown Heights, NY (1976)
3. Wei, F., Qian, W., Wang, C., Zhou, A.: Detecting overlapping community struc-
tures in networks. World Wide Web 12(2), 235261 (2009)
4. Xie, J., Szymanski, B.K.: Towards linear time overlapping community detection in
social networks. In: PAKDD (2), pp. 2536 (2012)
5. Kumar, S.: Social networking at scale. In: HPCA. IEEE (2012)
6. Silberstein, A., Terrace, J., Cooper, B.F., Ramakrishnan, R.: Feeding frenzy: se-
lectively materializing users event feeds. In: SIGMOD Conference, pp. 831842
(2010)
7. Pujol, J.M., Erramilli, V., Siganos, G., Yang, X., Laoutaris, N., Chhabra, P., Ro-
driguez, P.: The little engine(s) that could: scaling online social networks. In: SIG-
COMM, pp. 375386 (2010)
8. Mondal, J., Deshpande, A.: Managing large dynamic graphs eciently. In: SIG-
MOD Conference, pp. 145156 (2012)
9. Xu, T., Chen, Y., Jiao, L., Zhao, B.Y., Hui, P., Fu, X.: Scaling microblogging
services with divergent trac demands. In: Kon, F., Kermarrec, A.-M. (eds.) Mid-
dleware 2011. LNCS, vol. 7049, pp. 2040. Springer, Heidelberg (2011)
10. Sandler, D., Wallach, D.S.: Birds of a fethr: open, decentralized micropublishing.
In: IPTPS, p. 1 (2009)
11. Buchegger, S., Schioberg, D., Vu, L.H., Datta, A.: Peerson: P2p social networking:
early experiences and insights. In: SNS, pp. 4652 (2009)
 Workload-Aware Cache for Social Media Data 673

Appendix: Volume Bounded Set Basis Problem

Definition 1. Volume Bounded Set Basis Problem

Instance: Collection F of subsets of a finite set U , and positive integers m, W
such that m W . #
Question: Is there a collection B of subsets of U and bB |b| W such that,
for each f F there is a subcollection of B with size not greater than m whose
union is exactly f ?

To analyze the complexity of Volume Bounded Set Basis Problem, we introduced

the classic Set Basis problem [2] first.

Definition 2. Set Basis Problem [2]

Instance: Collection C of subsets of a finite set S, postive interger K |C|.
Question: Is there a collection B of subsets of S with |B| = K such that, for
each c C, there is a subcollection of B whose union is exactly c?

Lemma 1. The Set Basis Problem is NP-hard. [2]

We prove Theorem 1 that the Volume Bounded Set Basis Problem is NP-hard by
transforming the set basis problem to the Volume Bounded Set Basis Problem.
Proof. Let collection C of subsets of a finite set S, postive interger K <= |C|
be an instance of the set basis problem. We construct the instance of Volume
Bounded Set Basis Problem as follows: collection
# F = C of subsets of a finite
set U = S, postive interger m = K, W = cC |c|.
Let B be subsets of S with |B| = K such that, for each c C, there is a
subcollection of B whose union is exactly c. Therefore for each f F = C, there
is a subcollection of B whose union is exactly c which has size not greater than
|B| = k = m.
# #And the volume of all basis sets in B is upper bouned by W , i.e.
bB |b| cC |c| = W , which is also due to the fact that each c C is an
union of a subcollection of B. Then B is a solution to the instance of Volume
Bounded Set Basis Problem.
The another direction is oblivious.
 Shortening the Tour-Length
of a Mobile Data Collector in the WSN
by the Method of Linear Shortcut

Md. Shaifur Rahman and Mahmuda Naznin

Bangladesh University of Engineering & Technology, Dhaka, Bangladesh

[email protected], [email protected]

Abstract. In this work, we present a path-planning method for the

Mobile Data Collector (MDC) in a wireless sensor network (WSN). In
our method, a tour for the M DC is generated such that the latency
of delivering data to the sink is reduced. We show that the TSP tour
covers all the nodes of the WSN. However, the latency for the TSP tour
may be prohibitively high for delay-sensitive real-time WSNs. We reduce
the latency by shortening the given TSP tour using a method called
Linear Shortcut. We observe that the M DC need not visit the exact
location of the node. It only need to be in the proximity of the node
as required by the transmission radius. We present an algorithm that
iteratively shortens tour-length and hence, reduces the latency. We term
the resulting tour as Tight Label Covering (TLC) tour. Experimental
results show that TLC tour reduces latency by a significant margin.

Keywords: Wireless Sensor Network, Path-planning, Mobile Data Col-

lector, TSP tour.

1 Introduction
Wireless Sensor Network (WSN) is widely used for tracking, monitoring and
other purposes. The problem of collecting data packets from the sensor nodes
and depositing those to the sink node is known as Data Gathering Problem [1,2].
Using mobile elements for data gathering in the WSN is a recent trend [3]. It
has many advantages. For example, it increases connectivity, reduces cost of
deployment of a dense WSN and increases the lifetime of the WSN. However,
it has the disadvantage of high latency as these mobile elements have limited
speed (compared to the high speed data gathering by routing). The dedicated
mobile element in the WSN which collects data packets and brings those to
the static sink is called Mobile Data Collector (M DC). A convenient way to
control the latency in the case of data collection by the M DC is to carefully
plan its path so that the path is as short as possible. In this work, we present a
path-planning method that shortens the path of the M DC iteratively. We test
our path-planning method in a simulation which is run on a realistic testbed.
Experimental results show that the shortening of the path indeed translates into
decreased latency and other improvements.

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 674685, 2013.
c Springer-Verlag Berlin Heidelberg 2013

 Energy-Aware Path of M DC for WSN 675

Rest of the sections are organized as follows. Related works are discussed in
Section 2. The problem is formulated in Section 3. Our method is presented
in Section 4. Experimental results are presented in Section 5. The prospect of
future work is presented in Section 6.

2 Related Works
A complete survey on using mobile elements for data collection in the WSN is
presented in [3]. Earlier works on mobile elements can be found in [4], [5], [6] and
[7]. However, random motion of the mobile elements in these works is not suitable
for optimization. Mobile sinks have been considered in [8] and [9]. Mobile relay
based approaches for opportunistic networks have been surveyed in [10]. How-
ever, these methods are not suitable for the WSN because of its dierence with
such networks. In [11], an energy-ecient data gathering mechanism for large-
scale multi-hop network has been proposed. The inter-cluster tour proposed in
this work is N P -hard. Latency issue is not addressed in it. One of the heuristics
used in this work produces edges which are not connected with any nodes of the
WSN. Latency is considered while planning path for the mobile collector in [12]
and [13]. Methods presented in these works produce a shorter tour termed Label
Covering tour from a TSP-tour. However, the transmission range of the sensor
nodes is ignored in the shortening process. In [14], authors propose an approxi-
mation algorithm which is based on the TSP-tour of the M DC. Although, the
computation time (O(n)) is impressive, the solution is applicable to only certain
kind of TSP-tour (tours for which the centroid of the tour-polygon lies within
that polygon). If a condition regarding the concavity of the given TSP-tour is
not met, the problem of finding the optimal solution becomes N P -hard. In [15],
authors address the problem of planning paths of multiple robots so as to collect
the data from all sensors in the least amount of time. The method presented
here exploits earlier work on T SP -tour with neighborhood problem. However,
this work does not utilize the available location information of the sensor nodes
to the fullest as it allows the traversal of the full boundary of the transmission
region of a node. In sparse network, one or more sensor nodes have no neighbors
at all. As a result, the traversal of the boundary those nodes is futile and adds
up to the tour-length.

3 Preliminaries
We represent a WSN with n nodes by a complete graph Kn where the graph-
nodes represent the sensors and the sink. The edges in this graph represent the
Euclidian distances between two nodes. We adopt the disk model of the given
transmission range T XR. A circle with radius T XR centered at a node represent
the area of radio transmission of that node. We assume that there are one sink
and one M DC in the WSN and that both the sink and the sensor nodes are
static. A tour or cycle for the M DC is a closed path in the graph Kn which
starts and ends at the sink node.
676 Md.S. Rahman and M. Naznin

Definition 1. A TSP-tour is a tour in which the M DC visits the exact location

of all the nodes in the WSN exactly once. The min-cost T SP tour1 is a TSP-tour
in which the M DC covers the minimum Euclidian distance.
Definition 2. A tour T by the M DC is complete if each sensor node of the
WSN can send data packets to the visiting M DC directly or via a neighboring
which forwards packets. Otherwise, the tour is incomplete.
By definition, a TSP-tour is complete. However, it is not a good choice for a
delay-sensitive WSN as explained in the following section.
Definition 3. Data Delivery Latency (DDL) of a packet is the time-dierence
between the generation and delivery of the packet.
Let a packet i be generated at tg time after the M DC has set out from the sink
node. The M DC completes the current tour in tT time according to some tour
plan T . DDL for a packet i is given by:

tl (i) = tT tg (i) (1)

If n packets are collected in the current tour by the M DC, average packet
delivery latency tavg is given by:
#n
[tT tg (i)]
tavg = i=1
#nn
tg (i)
= tT i=1 (2)
n
"n
t (i)
The term i=1n g in Equation 2 is the average packet generation time. This
parameter is not controllable as it depends on the sampling rate of the sensor
nodes and the event frequency. However, we may improve both per packet DDL
(tl ) and average DDL (tavg ) by decreasing the tour-time tT (See Equation 1 and
2). The tour-time of the TSP-tour i.e. tT SP has two components: the fraction of
tour-time th when the M DC halts and collects data from the nearby nodes and
the fraction of tour-time tm when the M DC travels between the node positions.

tT SP = th + tm (3)

When the number of nodes is very high and/or the network is sparse, th << tm
and thus tm dominates tour-time tT SP . This assumption is logical for practical
scenario where the speed of a commercially available robotic car used as M DC
usually is 5 ms1 whereas packet transfer from a sensor node to the M DC
happens in the order of miliseconds [16]. Thus, decreasing the motion time tm
contributes to improving latency. If the speed of the M DC is vMDC , and if we
assume that it accelerates to this speed instantly and stops instantly, then
|tT SP |
tm = (4)
vMDC
1
We use TSP-tour to denote the minimum cost TSP-tour in this work.
 Energy-Aware Path of M DC for WSN 677

where |tT SP | is the path-length of the TSP-tour. Given a particular M DC, vMDC
is fixed. The only way to decrease tour-time is decreasing the length of the tour
i.e. |tT SP | (See Equation 4). However, decreasing the tour-length arbitrarily has
the risk of making the resulting tour incomplete. Therefore, we address the issue
carefully so that, the resulting tour is complete and shorter than the TSP-tour.

Problem Statement
Given a TSP-tour of the M DC in a WSN, find a tour Td that is complete and
shorter than the TSP-tour.

4 Improving Latency by Means of Linear Shortcut

4.1 Linear Shortcut of a Tour
Definition 4. A linear shortcut of given a tour is derived by choosing 0, 1 or
2 points (called Anchor Points) from each tour-edge according to some strategy
and connecting those points by straight lines in the order of visiting those edges.
It is called linear as only new straight lines are introduced in the resulting tour
instead of any curves.

Original Tour Path

Linear Shortcut Tour Path
Anchor Point
n2
n4
p3

p2 p4

n1 n3

p5

p1

n0

Fig. 1. Example of Linear Shortcut of a tour

An example of linear shortcut of a tour is shown in Figure 1. Five Anchor Points

p1 , p2 , p3 , p4 and p5 are chosen from five tour-edges. Those are connected in the
order of their visiting in the given tour to produce the linear shortcut. The
first and the last points are also connected to make the path a cycle. The tour
< p1 , p2 , p3 , p4 , p5 , p1 > is a linear shortcut but < p1 , p3 , p2 , p4 , p5 , p1 > is not.
Using principle of triangle inequality, Lemma 1 can be easily proved.
Lemma 1. If at least one anchor point is not coincident with the endpoint of
the tour-edge, the linear shortcut is shorter than the given tour. (Proved using
Triangle Inequality)
678 Md.S. Rahman and M. Naznin

4.2 Linear Shortcut of the TSP-tour

In [12] and [13], a tour known as Label Covering or LC tour is derived from
the TSP-tour. As shown in Figure 2, the tour-edges between nodes 1 and 2 are
also within the range of Node 3 and 5. Therefore, the label set of this edge is
{1, 2, 3, 5}. Similarly, the sets of labels of all other edges are determined. the
minimum length label covering tour is determined by making shortcut of zero
or more edges of the given TSP-tour. For the graph shown in Figure 2, the
TSP-tour is < 1, 2, 3, 4, 5, 1 > and the LC-tour derived from it is < 1, 2, 3, 1 >.

1,2,3,4,5
3
1,2,3,5 2,3,4,5
3 4
4
1,2,3
1,2,3,4,5
2 1,4,5
2
1,2,3,5 1,2,4,5
5 5
1 1,2,4,5
1
1,2,3,4,5
(a) Complete Graph derived from the Connectivity Graph (b) Label Covering Tour (tour-edges makred in double-line)

Fig. 2. Label Covering tour in a cluster with five nodes

4.3 Linear Shortcut of the LC-tour

Definition 5. The segment of the tour-edge which is within the circle represent-
ing the transmission area of a node is called the Contact Interval (CI) of that
node on that edge.

As shown in Figure 3, four nodes are covered by an edge connecting Node ni

and nj . Each of their CIs is represented by two points on the edge i.e the l
(which is encountered first by the M DC on this edge) and the r points. For
example, the CI of Node ni+1 is given by (lni+1 , rni+1 ). If the intersection of

ni+3 ni+5
ni+1
ni nj

ni+4
ni+2

lni+1 rni+1
lni+2 rni+2
lni+3 rni+3
lni+4 rni+4
lni+5 rni+5

Fig. 3. Contact Intervals of an edge

 Energy-Aware Path of M DC for WSN 679

the circle and the straight line is beyond the edge, the end-point of the CI is the
nearest end-point of the edge. For example, in Figure 3, r-point of Node ni+5 is
the location of Node nj .
Definition 6. Given a list of contact intervals CIe of a tour-edge e, Critical
Contact Interval or CCI is the interval of the minimum length which has at least
one point from each contact interval.
Critical Contact Interval or CCI of i-th edge is represented by two points- lcci
and rcci on that edge. These points can be determined as follows:

lcci r-point closest to the first end-point along the tour-edge

rcci l-point closest to the last end-point along the tour-edge

lni+1 rni+1
lni+2 rni+2
lni+3 rni+3
lni+4 rni+4
lni+5 rni+5
lcci = rni+2 rcci = lni+5

Fig. 4. Critical Contact Interval for a given list of intervals

For example, the CCI of the edge shown in Figure 3 is determined in Figure 4.
After the CCIs have been computed, we can connect r point of the CCI of
an edge with the l point of the CCI of the next edge. However, the nodes visited
in the given tour may be missed as shown in Figure 5(a). Therefore, to cover
those nodes, we apply the following method for a Node ni :
1. If both the edges have non-Null CCIs, i.e. there are intermediate nodes on
both the edges, then we add the r point of the incoming edge with the l
point of the outgoing edge. We call this line segment r-l line segment.
(a) If r-l line segment intersects circle of Node ni , then CCIs of both of the
adjacent edges are kept unchanged (See Figure 5(b)).
(b) If r-l line segment does not intersect the circle of Node ni , then we draw
a straight line that is parallel to the r-l line segment and tangent to that
circle. Let this line intersects the incoming and outgoing edges at points
pi and po respectively. (see Figure 5(b) for Node n11 ). We update the
r point of the incoming edge and the l point of the outgoing edge as pi
and po respectively.
2. If the incoming edge does not have any intermediate node with overlapping
circle or (in case it has) its r point is farther from point ni by at least T XR,
then we compute the point pi as the intersection between the incoming edge
and the circle centered at ni . If the incoming edge has non-null CCI, then we
update its r point as pi . Otherwise, we set the incoming edges r and l point
as pi . This case applies for Nodes n1 , n2 and n13 as shown in Figure 5(b).
680 Md.S. Rahman and M. Naznin

n5 n7 n5 n7
n3 n3
rcc2 rcc2
n8 n8
n2 lcc2 n2 lcc2
n6 n6
n4
lcc1 = rcc1 n4

n10 n10

lcc3 = rcc3 lcc3

lcc4 = rcc4
rcc3
n1 n1
lcc5 = rcc5

rcc4 lcc4
n11 n11
n13 n12 n13 n12

(a) Connecting the CCIs of successive (b) Updated l and r points to cover vis-
tour-edges ited nodes

lcc2 = rcc2
n5 n7
n5 n7 n3
n3 lcc3 rcc3
rcc2 n8
n8 n2
n2 lcc2 n6
n6
lcc4
n4
lcc1 = rcc1 n4 rcc1
rcc4
n10 n10

lcc6 = rcc6
lcc8 = rcc8
lcc3 lcc1

lcc7 = rcc7
rcc3
n1 n1
lcc5 = rcc5
lcc4
rcc4 n11 n11
n13 n12 n13 n12

(c) T LC-tour derived in Iteration 1 (d) Updating the l and r point after Iter-
ation 1
n5 n7 n5 n7
n3 n3
rcc3 lcc4 = rcc4
n8 lcc3
n8
n2 n2
n6 n6
n4 lcc5 = rcc5
lcc1 = rcc1 n4

n10 n10
lcc6
lcc7 = rcc7

lcc8 = rcc8

lcc9 = rcc9
rcc6
n1 n1



n11 n11
n13 n12 n13 n12

(e) T LC-tour derived in Iteration 2 (f) Updating the l and r point after Iter-
ation 2
lcc3 = rcc3
lcc2 = rcc2

n5 n7 n5 n7
n3 n3
lcc4 rcc4
n8 n8
n2 n2
n6 n6

n4 lcc = rcc n4
1 1

lcc6 = rcc6
n10 n10
lcc10 = rcc10

lcc9 = rcc9

n1 n1

lcc7 = rcc7

n11 n11
n13 n12 n13 n12

(g) T LC-tour derived in Iteration 3 (h) Updating the l and r point after Iter-
ation 3
Fig. 5. Generating TLC-tour using Linear Shortcut
 Energy-Aware Path of M DC for WSN 681

Now, we join the r point of the previous edge with the l point of the next edge.
The final edges are shown as bold straight lines in Figure 5(c). We term this
shortening as tightening of the given tour by the linear shortcut method. We
term the shorter tour derived from the LC-tour as Tight Label Covering tour or
T LC-tour.

4.4 Iterative Improvement of the TLC -tour

The path found in Figure 5(c) can be further shortened using the linear shortcut
method. We divide each iteration of improvement into 2 steps:
1. Connect the r point rcci of i-th edge with l point lccj of the next edge (j-th
edge such that j > i) with non-Null CCI and include the edge connecting
lccj and rccj in the edge set.
2. Re-associate the intermediate circles with the resulting edges and recompute
the CCIs for each edge.
Now, we use the following policy to re-associate the circles when existing tour-
edges break into shorter ones and new edges are added :
1. For each circle adjacent with two edges, the CIs for the both the edges are
calculated. Then the circle is associated with the edge on which the circle
has longer CI.
2. If the CIs for both the edges are equal, the circle is associated with the
incoming edge.
As shown in Figure 5(d), there are eight edges. Circle of Node n1 has overlaps
with both the outgoing and the incoming edges. However, unlike the incoming
edge, the circle has a CI of non-zero length with the outgoing edge. Therefore,
we associate Node n1 with this edge. Node n2 has CIs of zero length with
both the incoming and the outgoing eges. Therefore, we associate it with the
incoming edge.The CCI of the edge connecting n1 and n2 is updated after this
re-association. In similar ways, we determine the l and r points of the CCIs of
the remaining edges. After this round of re-associating of circles and computation
of the CCIs of respective edges, we join the r point of an edge with the l point
of the next edge. Thus, the tour as shown in Figure 5(e) is derived. According
to the Lemma 1, it is shorter than the tour derived in the previous iteration.
We can continue this way in to more iterations to tighten the given tour. The
steps are illustrated in Figure 5(f), 5(g) and 5(h).
The given LC-tour and the T LC-tour derived after Iteration 4 has been im-
posed on each other for comparison in Figure 6. The derived T LC-tour is a
significantly shorter than the LC-tour.

Condition for the Termination of Iterations: We can define path gain

gi (tT LC ) for a T LC-tour derived in iteration i as follows:
|tT LC |i1 |tT LC |i
gi (tT LC ) = (5)
|tT LC |i
682 Md.S. Rahman and M. Naznin

n5 n7
n3
n8
n2
n6
n4

n10

n1

n11
n13 n12

Fig. 6. Comparison between input LC-tour (doted path) and TLC-tour (solid path)
derived in Iteration 4

Here |tT LC |i is the length of the T LC-tour derived in iteration i. We stop the
iterations for path-shortening as soon as the path gain is below a threshold like
1%, 5% etc.

Computation Complexity: Our method of generating TLC-tour is formally

presented in Algorithm 1. First, we generate the CIs for all the circles in O(n)
time. Then, we sort the CIs in the non-decreasing order of the distance from the
first endpoint of the edge in O(n log n) time. Therefore, we generate the CCI for
an edge in O(n + n log n) = O(n) time. For a given tour, there are O(n) edges.
Therefore, we generate CCIs of all the edges in O(n2 ) time. Then, we connect
successive r and l Points in O(n) time. For each edge, we test the circles for
re-association. There may be O(n) such circles associated with each edge and a
total of O(n) edges. Therefore, updating the CCIs takes O(n2 ) time.

Algorithm 1. Generating TLC-tour from LC-tour

Input: LC-tour TLC and a path-gain threshold gt
1: Compute the CIs of all the nodes
2: Compute the CCIs of all the edges
3: while path-gain > tg do
4: derive tour TT LC by connecting the r point of an edge with the l point of the
next edge
5: for each edge e do
6: for each Circle c (of Node n) associated with Edge e do
7: re-associate Circle c (if necessary)
8: end for
9: update CCI of Edge e
10: end for
11: determine path-gain
12: end while
Output: TLC-tour TT LC
 Energy-Aware Path of M DC for WSN 683

T SP -tour generator O(T SP )

LC-tour generator O(n3 )

O(n2 )
T LC-tour generator

for m iteration: O(mn2 )

O(T SP ) + O(n3 ) + O(mn2 )

Fig. 7. Stages of computation along with the time complexity

If there are m iterations, then generating TLC-tour from LC-tour takes

O(mn2 ) time. We illustrate the stages of computation along with the time com-
plexity in Figure 7.

5 Experimental Results
We use Castelia[17] framework of OMENT++ simulator to distribute sensor
nodes randomly. The sensor nodes also generated packet randomly. We use Con-
cord TSP-Solver [18] to compute the optimal TSP-tour. We derive the LC-tour
from the TSP-tour and the T LC-tour from the LC-tour using path-gain thresh-
old of 5%. The M DC tours continuously in TSP-tour, LC-tour and TLC-tour.
During its travel, it collects the packets from the sensor nodes and deposits
those to the sink node. We vary the value of T XR from 2m to 32m. Low value
of T XR indicates lower degree of connectivity and hence, sparse network. Sim-
ilarly, higher value of T XR indicates dense network.
As shown in Figure 8(a), average packet delivary latency is always the lowest
for TLC-tour and the highest for TSP-tour. The value is comparatively better
in the case of the sparse WSN. In Figure 8(b), throughput for the entire run
is shown. The value is always the highest for T LC-tour and the lowest for the
TSP-tour.

(a) Average Packet Delivery Latency (b) Throughput

Fig. 8. Comparison among TSP-tour, LC-tour and T LC-tour
684 Md.S. Rahman and M. Naznin

6 Conclusion
We have given a framework for shortening a given tour of the M DC. The re-
sulting tour decreases packet delivery latency and increases throughput. The
TLC-tour derived by Linear Shortcut of the LC-tour is highly suitable for real-
time WSN in which high latency is undesirable.
In our future work, we shall consider more objectives besides minimizing la-
tency, for example- facilitating multi-hop forwarding among the sensor nodes,
load-balancing of the network trac etc. We shall also extend the path-planning
for a WSN with multiple M DCs and multiple sinks.

Acknowledgements. The authors would like to thank the department of Com-

puter Science & Engineering, Bangladesh University of Engineering & Technol-
ogy (BUET) for its generous support, laboratory, library and online resource
facilities.

References
1. Akyildiz, I.F., Su, W., Sankarasubramaniam, Y., Cayirci, E.: Wireless sensor net-
works: a survey. Comput. Netw. 38(4), 393422 (2002)
2. Rajagopalan, R., Varshney, P.K.: Data-aggregation techniques in sensor networks:
a survey. IEEE Communications Surveys Tutorials 8(4), 4863 (2006)
3. Di Francesco, M., Das, S.K., Anastasi, G.: Data collection in wireless sensor net-
works with mobile elements: A survey. ACM Transaction on Sensor Networks 8,
7:17:31 (2011)
4. Zhao, W., Ammar, M.H.: Message ferrying: Proactive routing in highly-partitioned
wireless ad hoc networks. In: Proceedings of the The Ninth IEEE Workshop on
Future Trends of Distributed Computing Systems, FTDCS 2003, pp. 308314.
IEEE Computer Society, Washington, DC (2003)
5. Shah, R.C., Roy, S., Jain, S., Brunette, W.: Data mules: Modeling a three-tier ar-
chitecture for sparse sensor networks. In: IEEE SNPA Workshop, pp. 3041 (2003)
6. Zhao, W., Ammar, M., Zegura, E.: A message ferrying approach for data delivery
in sparse mobile ad hoc networks. In: Proceedings of the 5th ACM International
Symposium on Mobile ad Hoc Networking and Computing, MobiHoc 2004, pp.
187198. ACM, New York (2004)
7. Jun, H., Zhao, W., Ammar, M.H., Zegura, E.W., Lee, C.: Trading latency for en-
ergy in wireless ad hoc networks using message ferrying. In: Proceedings of the
Third IEEE International Conference on Pervasive Computing and Communica-
tions Workshops, PERCOMW 2005, pp. 220225. IEEE Computer Society, Wash-
ington, DC (2005)
8. Wang, Z.M., Basagni, S., Melachrinoudis, E., Petrioli, C.: Exploiting sink mobility
for maximizing sensor networks lifetime. In: Proceedings of the 38th Annual Hawaii
International Conference on System Sciences - Volume 09, HICSS 2005, p. 287.1.
IEEE Computer Society, Washington, DC (2005)
9. Rao, J., Wu, T., Biswas, S.: Network-assisted sink navigation protocols for data
harvesting in sensor networks. In: WCNC 2008, IEEE Wireless Communications
& Networking Conference, Las Vegas, Nevada, USA, March 31-April 3. Conference
Proceedings, pp. 28872892. IEEE (2008)
 Energy-Aware Path of M DC for WSN 685

10. Conti, M., Pelusi, L., Passarella, A., Anastasi, G.: Mobile-relay Forwarding in Op-
portunistic Network. In: Adaptation and Cross Layer Design in Wireless Networks.
CRC Press (2008)
11. Ma, M., Yang, Y.: Sencar: An energy-ecient data gathering mechanism for large-
scale multihop sensor networks. IEEE Transaction on Parallel and Distributed
System 18, 14761488 (2007)
12. Sugihara, R., Gupta, R.K.: Improving the data delivery latency in sensor networks
with controlled mobility. In: 4th IEEE International Conference on Distributed
Computing in Sensor Systems, pp. 386399. Springer, Heidelberg (2008)
13. Sugihara, R., Gupta, R.K.: Path planning of data mules in sensor networks. ACM
Transaction on Senson Networks 8(1), 1:11:27 (2011)
14. Yuan, Y., Peng, Y.: Racetrack: an approximation algorithm for the mobile sink
routing problem. In: Nikolaidis, I., Wu, K. (eds.) ADHOC-NOW 2010. LNCS,
vol. 6288, pp. 135148. Springer, Heidelberg (2010)
15. Bhadauria, D., Tekdas, O., Isler, V.: Robotic data mules for collecting data over
sparse sensor fields. J. Field Robot. 28(3), 388404 (2011)
16. Huang, P., Xiao, L., Soltani, S., Mutka, M., Xi, N.: The evolution of mac proto-
cols in wireless sensor networks: A survey. IEEE Communications Surveys Tutori-
als PP(99), 120 (2012)
17. Castalia wsn simulator framwork for omnet++, http://www.castelia.org
18. Concorde tsp solver, http://www.tsp.gatech.edu/concorde.html
 Towards Fault-Tolerant Chord P2P System:
Analysis of Some Replication Strategies

Rafa Kapelko

Institute of Mathematics and Computer Science

Faculty of Fundamental Problems of Technology,
Wrocaw University of Technology, Poland
[email protected]

Abstract. In Peer-to-peer systems the process of unexpected nodes failures be-

comes critical. When some nodes depart from the network without uploading
gathered information into the system, some data which are stored on the departed
nodes will not be available. In our paper we address this issue for Chord-based
P2P systems. We investigate two replication schemes: successor list and multi-
ple Chord rings. We propose a novel analytical model to describe the process of
unexpected nodes failures for investigated replication scheme. Let > 0. Let d
be replication degree in Chord with successor-list replication. We prove that for
1
large n after failure of n1 d nodes from Chord with successor list replication
no document is lost with high probability. We also prove that for k multiple chord
1 1
n
rings after unexpected departures of 1
k
nodes, the system stays in safe config-
n 2k
uration with high probability. As a practical illustration of our detailed theoretical
analysis of replication schemes we discuss the recovery mechanism of partially
lost information. We show how the system fault tolerance changes depending of
different replica degree.

Keywords: peer-to-peer network, Chord, replication of documents, reliability.

1 Introduction
In recent years, Peer-to-Peer (P2P) networks (see e.g. [21], [17], [7] [15]) have been
powerful environments for successful sharing of certain resources. Despite the advan-
tages of the P2P systems, there exists a significant problem with fault tolerance. When
some nodes depart from the system without uploading gathered information, some data
which are stored on the departured nodes may be lost.
Among the structured P2P systems, Chord protocol introduced in [21] and developed
in [14] is one of the most popular and we focus on it in this paper. In the Chord P2P
system each of such unexpected node failures effects in loosing some data stored in the
node (see [2], [11]).
The replication techniques are used for storing documents at many nodes in the
system. The benefits of replication are that it can improve availability of documents

Supported by grant nr 2012/S20027 of the Institute of Mathematics and Computer Science,

Wrocaw University of Technology, Poland.

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 686696, 2013.
c Springer-Verlag Berlin Heidelberg 2013
 Towards Fault-Tolerant Chord P2P System 687

in case of failure [22]. Different strategies for replication in Chord P2P system such
as successor-list, symmetric replication or multiple hash functions are discussed in
many publications (see [5], [18], [10], [16], [3], [20], [19]). In the replication schemes
successor-list replication is very popular and often applied in ring-based networks.
Our main contribution include:
We propose a novel analytical model describing the process of unexpected nodes
failures. Let > 0. Let d be replication degree in Chord with successor-list repli-
1
cation. We prove that for large n after failure of n1 d nodes from Chord with
successor list replication no document is lost with high probability (see Theorem
3). Our consideration dedicated to Chord P2P system is also applicable for other
overlays based on ring such as Pastry [7] (with some modification), RP2P [1], Cas-
sandra [12], CRP-overlay [13] and others.
We also investigate the fault tolerance of overlay system called k multiple chord
1 1
rings. We prove that after simultaneous unexpected departures of n 1 k nodes from
n 2k
k multiple chord rings no information disappears with high probability.
As a practical application of our theoretical estimation we discuss the recovery
mechanism of partially lost information in both replication schemes. We show how
fault tolerance of Chord system changes depending of different replica degree. We
also explain how to choose the proper number of replication to optimize those two
replication strategies.
The remainder of this article is organised as follows. In Section 2 we describe the
successor-list replication scheme in Chord P2P system and present analytical formulas
describing the resistance to loss of document. Section 3 presents the overlay k Chord.
Then, in Section 4 we discuss the recovery mechanism of partially lost information.
Finally, Section 5 summaries and concludes our work.

2 Successor-List Replication

In this section, we briefly describe the successor-list replication scheme in Chord P2P
system [5]. For a more detailed description of Chord P2P system the reader is referred
to [21]. Then, we analyze the process of unexpected departure.
In Chord P2P system nodes are arranged in a logical ring. The positions of nodes
and documents are created by a hash operation, which results in random placement into
the Chord ring. Let n denotes the number of nodes in Chord. Then, the identifier space
of the nodes is defined as a set of integers {0, 1, . . . , n 1}.
Let (0, 1, . . . , n 1) denotes Chord ring. The first successor of a node with identi-
fier p is the first node found going in clockwise direction on the ring starting at p. The
predecessor of a node with identifier p is the first node found going anti-clokwise direc-
tion on the ring starting at p. Each node p is responsible for storing documents between
p s predecessor and p. To ensure the connectivity of the chord ring, each node knows the
688 R. Kapelko



U

U

U





Fig. 1. Successor-list replication of degree 4

first s nodes which identifier is greater than the nodes. These nodes are the nodes suc-
cessors. Additionally, each node maintains a table of fingers of size O(log(n)), which
point to other nodes in the ring. Hence the size of routing table is O(log(n) + s).
Let d < s. For a replication degree of d in successor-list replication a document is
stored on the node p which is responsible for storing document and d 1 immediate
successor of p. Let the node p leaves the system in an unexpected way, i.e without
uploading its documents into the system. After the departure there are still (d 1)
replicas of documents of the node p.
For p = 0, 1, . . . , n 1 and d > 1 we define family of sets by the formula

Up,d = {p mod n, (p + 1) mod n, . . . (p + d 1) mod n}.

Definition 1. Let n denotes the number of nodes in Chord. Let A = {n1 , . . . , nl }

{0, . . . , n 1} be a random subset of nodes from Chord with successor-list replication
of degree d. We say that, the set A is safe, if

p{0,1,...,n1} Up,d A = Up,d .

We say that, the set A is unsafe, if A is not safe.

Remark 1. Notice that, if the set A is safe, then no information disappears from Chord
with successor-list replication of degree d after simultaneous unexpected departure of
all nodes from A.

Theorem 2. Let d 2. Let n be the number of nodes in Chord with successor-list

1
n1 d
replication of degree d and let A be the set of nodes. If |A| 1 then Pr[A is safe]
n 2d
1 1 .
n

Proof. Firstly observe that for |A| < d we have Pr[A is safe] = 1. Therefore, we may
assume that |A| d. Then
 Towards Fault-Tolerant Chord P2P System 689

L M
Pr[A is unsafe] = Pr p{0,1,...,n1} Up,d A = Up,d =
In1 J n1
N "
Pr Up,d A = Up,d Pr [Up,d A = Up,d ] =
p=0 p=0
7 nd ;
|A|d n|A|d 1 1 1
n 7n; 7 .
|A|
(n (d 1))d n 1 d1 ;d n
n

=

Theorem 3. Let > 0. Let d 2. Let n be the number of nodes in Chord with
1
successor-list replication of degree d and let A be the set of nodes. If |A| n1 d
then
lim Pr[A is unsafe] = 0
n

Proof. As in the proof of Theorem 2, observe that for |A| < d we have Pr[A is safe] =
1. Therefore, we may assume that |A| d. Then
O 7 nd ; P
11/d d
lim Pr [A is unsafe] lim n n7 n
; = 0.
n n
n11/d

Hence the proof is finished. =

3 Multiple Chord Rings

The overlay system k Chord was introduced and developed in [8], [9]. It the system
of multiple chord rings. From the formal point of view k Chord is the structure
k
K
k Chord = Ri,n ,
i=1

where Ri,n is the ith Chord ring. The nodes of the first Chord ring are expressed in
the form R1,n = (0, 1, 2, ..., n1) and on the ith ring (i = 2, . . . , k) are generated by
independent random permutation i Sn i.e., Ri,n = (i (0), i (1), . . . , i (n 1)).
Each document has a unique position and is mapped into the same location on different
Chord rings.
In k Chord hybrid system each node maintains a k dimensional finger table and a
successor list of a size s (s > k) on each Chord ring. Hence the routing table is of size
O(k(log(n) + s)). With k being constant and small (e.g, k 5.), the size of routing
table is O(log(n) + s) which is the same like in Chord. Let us consider the event, when
the node p leaves the system in an unexpected way without uploading storing documents
into the system. In the classical Chord such event effects in loosing information stored
on the node p. The situation changes in hybrid k Chord. There are (k 1) replicas of
documents stored on the nodes i (p) on the ith ring (i = 2, . . . , k). The probability
1
of loosing information stored on the node p equals (n1) k1 .
690 R. Kapelko

 +GRFB 

+GRFB
 
 
+GRFB

+GRFB
   
+GRFB

Fig. 2. An example of 2Chord where 2th ring is generated by permutation (3,5,4,2,1)

Definition 4. Let A = {n1 , . . . , nl } {0, . . . , n 1} be a random subset of nodes

from the structure k Chord with n nodes. We denote by KA,1 the unions of intervals
controlled by nodes from A in the first circle and we denote by KA,i the unions of
intervals controlled by nodes from A in the ith circle (i = 2, . . . , k). We say that the
)k
set A is safe, if i=1 KA,i = . We say that the set A is unsafe if A is not safe.

Let us remark that, if the set A is safe then no information disappears from the hybrid
system after simultaneous unexpected departure of all nodes from A.

Theorem 2. Let k 2. Let n be the number of nodes in the structure k Chord. let
1
n1 k
A be a random subset of nodes from the structure k Chord. If |A| 1 then
n 2k
Pr[A is safe] 1 1 .
n

Notice that, the bounds in Theorem 2 do not depend on the number of documents put
into the system.

Proof. Let A = {n1 , . . . , nl } {0, . . . , n 1} be a random subset of nodes from the

(l
structure k Chord with n nodes. Hence |A| = l. Let us recall that KA,1 = s=1 As
are the unions of intervals controlled by nodes from A in the first circle, and KA,i are
the unions of intervals controlled by nodes from A in the ith circle (i = 2, . . . , k).
Notice that, the positions of nodes in the ith copy of Chord are chosen indepen-
dently from the position of nodes from the first copy. Let i (0), i (1) . . . i (n 1)
denotes the positions of nodes in the ith copy of Chord, where i Sn are random
and independent (i = 2, . . . , k). Let us consider an interval As controlled by a node ns
(s = 1, . . . , l).
The reason why an interval controlled by a node i (np ) (p=1,. . . ,l) may have a
nonempty intersection with the set As is caused by the fact that (p )i (np ) = ns .
Therefore,

l
K l
Pr[As KA,i = ] = Pr[(p )(i (np ) = ns )] = Pr[ i (np ) = ns ] = .
p=1
n
 Towards Fault-Tolerant Chord P2P System 691

Hence
I k
J l
I k
J
Q " Q
Pr[A is unsafe] = Pr KA,i = Pr As KA,i =
i=1 s=1 i=2
l
I k
J l =
k & 'k1
" R " l
Pr As KA,i = = Pr[As KA,i = ] l .
s=1 i=2 s=1 i=2
n

n1 k
1 7 l ;k1
Finally notice that, for l 1 we have l 1 , so the Theorem is proved.
n 2k n n
=

4 The Recovery Process

In this section we discuss the recovery mechanism of partially lost information in the
replication schemes: successor list and k Chord. We show how fault tolerance of
Chord system changes depending of different replica degree. Notice that the recovery
mechanism based on Queueing Theory was introduced in [2] and applied for another
modification of Chord. The notations defined in this chapter are listed in Table 2. From
the Littles Law from Queueing Theory (see e.g. [4]) we get two equations
N = T, Nr = u T r , (1)

Table 1. Time for recovery of partially lost information Trecovery as the function of the number of
nodes N [104 , 106 ] for different replica degrees D = 2, 3, 4, 5 with fixed T = 1800 sec and
u = 5%
8

500
6
400

4 300

200
2
100

200 000 400 000 600 000 800 000 1 ! 106 200 000 400 000 600 000 800 000 1 ! 106
D=2 replicas D=4 replicas
140
1200
120
1000
100
80 800

60 600

40 400
20 200

200 000 400 000 600 000 800 000 1 ! 10 6

200 000 400 000 600 000 800 000 1 ! 106
D=3 replicas D=5 replicas
692 R. Kapelko

Table 2. Notations

Symbol Meaning
N the average number of nodes in the system
Nr the medium number of nodes waiting for complementing
T the average time a node spends in a system (during one session)
u the proportion of unexpected departures among all departures
Tr the average time needed for recovery of partially lost information
the average number of departures from the system in a second
m the average number of unexpected departures from the system in a second
Tc,d the time in which nodes: p, p + 1, . . . p + d 1, check the node p + d
the average number of documents in the system per one node
ds the size of the information item in the system
ts the transmition speed
Trecovery time for recovery of partially lost information for different replica degrees

so
T Nr
Tr = .
u N
For two replication schemes discussed in this paper we have obtained the following
safety bound
1
N 1 D
Nr = 1 , (2)
N 2D
where D is a replication degree (D = d for successor-list replication of degree d,
D = k for k Chord) . If we want to keep our system in a safe configuration with
high probability greater than 1 1N , then the following inequality must hold

T Nr
Tr . (3)
u N
Therefore, time for recovery of partially lost information for different replica degrees
is:
T Nr
Trecovery = . (4)
u N
Let T = 30 minutes (see [6], [23]). Suppose that u = 5% nodes leave the system in
unexpected way without uploading gathered information. In Table 1 we put four figures
describing the function Trecovery of the number of nodes N [104 , 106 ] for different
replica degrees D = 2, 3, 4, 5 with fixed T = 1800 sec. and u = 5%. The system gains
more time for recovery of partially lost information when replica degree increases.
The following table contains the time for recovery of partially lost information with
fixed parameters N = 105 , T = 1800 sec, u = 5%, for different replica degree:

d 2 3 4 5 6 7 8
Trecovery 6 sec 113 sec 8 min 18.9 min 33.7 min 50.9 min 1h 9 min
 Towards Fault-Tolerant Chord P2P System 693

Algorithm 1. Recovery process in Chord with successor-list replication of degree d

1: nodes: p + 1 , . . . p + d 1 , check the node p + d
2: if the node (p + d ) left the system in unexpected way then
3: nodes: p + 1 , . . . p + d 1 , ask their successors for a portion of documents
4: nodes: p + 1 , . . . p + d 1 , send some portion of documents to some of their successors
5: end if

4.1 Successor-List Replication

Let us consider the network with N = 105 average number of nodes in the system (T =
30 minutes, u = 5%). From equation (1) we get m = N T u 2.77, so approximately
2.77 nodes unexpectedly leave the network in one second.
Assume that nodes: p+1, . . . p+d1 in the network check the node p+d periodically
with the period of Tc,d seconds (see Algorithm 1). Assume that Tc,d < 2 sec.
During this time nodes: p + 1, . . . p + d 1, do not know that the node p + d left
the network. Then, there are Tr Tc,d seconds for information recovery (see inequality
(3)). During this time nodes: p + 1 , . . . p + d 1 , ask their successors for a portion of
documents in 0.2 sec. Then, nodes: p + 1 , . . . p + d 1 , wait for receiving necessary
information. Nodes p + 1 , . . . p + d 1 , send some portion of documents to some of
their successors. Therefore, the following inequality guarantees the system is in a safe
configuration with high probability

ds T 3
Tc,d + 0.2 + 3 N 2d . (5)
ts u

The inequality (5) can be used in a practical way. It shows how to choose the proper
number of replication degree d to optimize the recovery process. For estimated param-
eters , ds , ts , T, u, N of the network and fixed Tc,d the optimal replica degree is the
minimal natural d ( d > 1) for which the inequality (5) holds.
Let us assume, that the transmition speed ts = 100 kB/s. As in [23] we consider two
scenarios with size of documents, ds of 1.57 kB and 1 MB, respectively.
The Table 3 contains the upper bounds for the average number of documents per one
node for different replica degree, when the system stays in a safe configuration.

Table 3. Values of the upper bound of the average number of documents per one node for different
replica degree and size of documents with fixed Tc,d = 2 sec., T = 30 minute, u = 5%,
N = 105 , ts = 100 kB/s

d=2 d=3 d=4 d=5

Tc,2 = 2 Tc,3 = 2 Tc,4 = 2 Tc,5 = 2
ds = 1, 57 kB 89 2370 10145 24123
ds = 1 MB 0 3 15 37
694 R. Kapelko

Algorithm 2. Recovery process in k Chord

1: nodes: p + 1 , . . . p + k 1 , check the node p + k
2: if the node (p + k ) left the system in unexpected way then
3: nodes: p + 1 , . . . p + k 1 , localise in the remind (k 1) Chord rings nodes responsible
for storing lost documents
4: nodes: p + 1 , . . . p + k 1 , ask localised nodes in the remind (k 1) Chord rings for
receiving necessary documents
5: nodes: p + 1 , . . . p + k 1 , send received documents to some of their successors
6: end if

4.2 k Chord
In k Chord network the recovery process of partialy lost information is more com-
plicated than in successor-list replication.
Let us consider the network with N = 105 average number of nodes in the system
(T = 30 minutes, u = 5%). From equation (1) we get m = N T u 2.77, so approxi-
mately 2.77 nodes unexpectedly leave the network in one second.
Assume that nodes: p + 1, . . . p + k 1, in each chord ring check the node p + k
periodically with the period of Tc,k seconds (see Algorithm 2). Assume that Tc,k < 2
sec.
During this time nodes: p + 1, . . . p + k 1, do not know that the node p + k left
the network. Then, there are Tr Tc,k seconds for information recovery (see inequal-
ity (3)). During this time nodes: p + 1 , . . . p + k 1 , localise in the remind (k 1)
Chord rings nodes responsible for storing lost documents. The time for fulfilling this
operation equals approximately 12 lg2 N 0.2 sec. Then, nodes: p + 1 , . . . p + k 1 ,
wait for receiving necessary information and the time is approximately dtss . Nodes:
p + 1 , . . . p + k 1 , send received documents to some of their successors. Therefore,
the following inequality guarantees the system is in a safe configuration with high prob-
ability
1 ds T 3
Tc,k + lg2 N 0.2 + 2 N 2k . (6)
2 ts u
Similary, to successor-list replication inequality (6) can be used in a practical way. It
shows how to choose the proper number of k Chord rings to optimize the recovery
process. For estimated parameters , ds , ts , T, u, N of the network and fixed Tc,k , the
optimal replica degree is the minimal natural k ( k > 1) for which the inequality (6)
holds.

Table 4. Values of the upper bound of the average number of documents per one node for different
k Chord and size of documents with fixed Tc,k = 2 sec., T = 30 minute, u = 5%, N = 105 ,
ts = 100 kB/s

k=2 k=3 k=4 k=5

Tc,2 = 2 Tc,3 = 2 Tc,4 = 2 Tc,5 = 2
ds = 1, 57 kB 80 3502 15165 36132
ds = 1 MB 0 5 23 56
 Towards Fault-Tolerant Chord P2P System 695

Let us assume that the transmition speed ts = 100 kB/s. As in [23] we consider two
scenarios with size of documents, ds of 1.57 kB and 1 MB, respectively.
The Table 4 contains the upper bounds for the average number of documents per one
node for different k Chord, when the system stays in a safe configuration.

5 Conclusion

In this paper, we study two replication strategies for Chord P2P system: successor-list
and multiple Chord rings. In both cases we obtain estimation describing the ressis-
tance to loss of documents. As a practical illustration of our estimated safety bounds we
discuss the recovery mechanism of partially lost information. We show how fault toler-
ance of Chord system changes depending of different replica degree and explain how
to choose the proper number of replication to optimize these two replication strategies.
In future research we plan to experimentaly evaluate efficiency of proposed recovery
algorithms. We also plan to extend our work to symmetric replication scheme.

References

1. Chen, S., Li, Y., Rao, K., Zhao, L., Li, T., Chen, S.: Building a scalable P2P network with
small routing delay. In: Zhang, Y., Yu, G., Bertino, E., Xu, G. (eds.) APWeb 2008. LNCS,
vol. 4976, pp. 456467. Springer, Heidelberg (2008)
2. Cichon, J., Jasinski, A., Kapelko, R., Zawada, M.: How to improve the reliability of chord?
In: Meersman, R., Tari, Z., Herrero, P. (eds.) OTM-WS 2008. LNCS, vol. 5333, pp. 904913.
Springer, Heidelberg (2008)
3. Cichon, J., Kapelko, R., Marchwicki, K.: Brief announcement: A note on replication of doc-
uments. In: Dfago, X., Petit, F., Villain, V. (eds.) SSS 2011. LNCS, vol. 6976, pp. 439440.
Springer, Heidelberg (2011)
4. Cooper, R.B.: Introduction To Queueing Theory, 2nd edn. Elsevier North Holland, Inc.
(1981)
5. Dabek, F., Kaashoek, M., Karger, D., Morris, R., Stoica, I.: Wide-area cooperative storgae
with cfs. In: 18th ACM Symposium on Operating Systems Principles (SOSP), New York,
USA, pp. 202215 (2001)
6. Derek, L., Zhong, Y., Vivek, R., Loguinov, D.: On lifetime-based node failure and stochas-
tic resilience of decentralized peer-to-peer networks. IEEE/ACM Transactions on Network-
ing 5(15), 644656 (2007)
7. Rowstron, A., Druschel, P.: Pastry: Scalable, Decentralized Object Location, and Routing
for Large-Scale Peer-to-Peer Systems. In: Guerraoui, R. (ed.) Middleware 2001. LNCS,
vol. 2218, pp. 329350. Springer, Heidelberg (2001)
8. Flocchini, P., Nayak, A., Xie, M.: Hybrid-chord: A peer-to-peer system based on chord. In:
Ghosh, R.K., Mohanty, H. (eds.) ICDCIT 2004. LNCS, vol. 3347, pp. 194203. Springer,
Heidelberg (2004)
9. Flocchini, P., Nayak, A., Xie, M.: Enhancing peer-to-peer systems through redundancy. IEEE
Journal on Selected Areas in Communications 1(25), 1524 (2007)
10. Ghodsi, A., Alima, L.O., Haridi, S.: Symmetric replication for structured peer-to-peer sys-
tems. In: Moro, G., Bergamaschi, S., Joseph, S., Morin, J.-H., Ouksel, A.M. (eds.) DBISP2P
2005 and DBISP2P 2006. LNCS, vol. 4125, pp. 7485. Springer, Heidelberg (2007)
696 R. Kapelko

11. Park, G., Kim, S., Cho, Y., Kook, J., Hong, J.: Chordet: An Efficient and Transparent Repli-
cation for Improving Availability of Peer-to-Peer Networked Systems. In: 2010 ACM Sym-
posium on Applied Computing (SAC), Sierre, Switzerland, pp. 221225 (2010)
12. Lakshman, A., Malik, P.: Cassandra: a decentralized structured storage systems. ACM
SIGOPS Operating Systems Review 2(44), 3540 (2010)
13. Li, Z., Hwang, K.: Churn resilient protocol for massive data dissemination in p2p networks.
IEEE Transactions on Parallel and Distributed Systems 22, 13421349 (2011)
14. Liben-Nowell, D., Balakrishnan, H., Karger, D.: Analysis of the Evolution of Peer-to-Peer
Systems. In: ACM Conference on Principles of Distributed Computing, Monterey, Califor-
nia, USA, pp. 233242 (2002)
15. Maymounkov, P., Mazires, D.: Kademlia: A Peer-to-Peer Information System Based on the
XOR Metric. In: Druschel, P., Kaashoek, M.F., Rowstron, A. (eds.) IPTPS 2002. LNCS,
vol. 2429, pp. 5365. Springer, Heidelberg (2002)
16. Pitoura, T., Ntarmos, N., Triantafillou, P.: Replication, load balancing and efficient range
query processing in dHTs. In: Ioannidis, Y., Scholl, M.H., Schmidt, J.W., Matthes, F., Hat-
zopoulos, M., Bhm, K., Kemper, A., Grust, T., Bhm, C. (eds.) EDBT 2006. LNCS,
vol. 3896, pp. 131148. Springer, Heidelberg (2006)
17. Ratnasamy, S., Francis, P., Handley, M., Karp, R., Shenker, S.: A Scalable Content-
addressable Network. In: SIGCOMM 2001, San Diego, California, USA, pp. 161172 (2001)
18. Rieche, S., Wehrle, K., Landsiedel, O., Gtz, S., Petrak, L.: Reliability of data in structured
peer-to-peer systes. In: HOT-P2P 2004, Volendam, The Netherland, pp. 108113 (2004)
19. Kapelko, R.: Towards efficient replication of documents in chord: Case (r,s) erasure codes.
In: Liu, B., Ma, M., Chang, J. (eds.) ICICA 2012. LNCS, vol. 7473, pp. 477483. Springer,
Heidelberg (2012)
20. Shafaat, T.M., Ahmad, B., Haridi, S.: ID-replication for structured peer-to-peer systems. In:
Kaklamanis, C., Papatheodorou, T., Spirakis, P.G. (eds.) Euro-Par 2012. LNCS, vol. 7484,
pp. 364376. Springer, Heidelberg (2012)
21. Stoica, I., Morris, R., Karger, D., Kaashoek, M.F., Balakrishnan, H.: Chord: A Scalable Peer-
to-Peer Lookup Service for Internet Applications. In: SIGCOMM 2001, San Diego, Califor-
nia, USA, pp. 149160 (2001)
22. Vu, Q., Lupu, M., Ooi, B.: Peer-to-Peer Computing. Springer (2010)
23. Wang, C., Harfoush, K.: On the stability-scalability tradeoff of dht deployment. In: INFO-
COM 2007, 26th IEEE International Conference on Computer Communications, Anchorage,
Alaska, USA, pp. 22072215 (2007)
 A MapReduce-Based Method
for Learning Bayesian Network from Massive Data

Qiyu Fang1, Kun Yue1,2,*, Xiaodong Fu3, Hong Wu1, and Weiyi Liu1
1
Department of Computer Science and Engineering, School of Information Science and
Engineering, Yunnan University, 650091, Kunming, China
2
Key Laboratory of Software Engineering of Yunnan Province, 650091, Kunming, China
3
Faculty of Information Engineering and Automation, Kunming University of Science and
Technology, 650500, Kunming, China
[email protected]

Abstract. Bayesian network (BN) is the popular and important probabilistic

graphical model for representing and inferring uncertain knowledge. Learning
BN from massive data is the basis for uncertain-knowledge-centered inferences,
prediction and decision. The inherence of massive data makes BN learning be
adjusted to the large data volume and executed in parallel. In this paper, we
proposed a MapReduce-based approach for learning BN from massive data by
extending the traditional scoring & search algorithm. First, in the scoring
process, we developed map and reduce algorithms for obtaining the required
parameters in parallel. Second, in the search process, for each node we devel-
oped map and reduce algorithms for scoring all the candidate local structures in
parallel and selecting the local optimal structure with the highest score. Thus,
the local optimal structures of each node are merged to the global optimal one.
Experimental result indicates our proposed method is effective and efficient.

Keywords: Massive data, Bayesian network, Learning, Scoring and search,

MapReduce.

1 Introduction

With the rapid development of Internet and Web applications, more and more data are
generated and collected in scientific experiments, e-commerce and social network
applications [1]. In this case, computationally intensive analysis results in complex
and stringent performance demands that are not satisfied by any existing data man-
agement [2, 3]. Centered on massive data management and analysis, data intensive
computing may generate many queries, each involving access to supercomputer-class
computations on gigabytes or terabytes of data [4]. Thus, knowledge discovery from
massive data is naturally indispensable for the applications of data intensive para-
digms, such as Web 2.0 and large-scale business intelligence. However, it is quite
challenging by incorporating the specialties of massive data [1].

*
Corresponding author.

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 697708, 2013.
Springer-Verlag Berlin Heidelberg 2013
698 Q. Fang et al.

As one kind of important knowledge, probabilistic graphical models (PGMs) [5]

comprise any model that uses the language of graphs to facilitate the representation
and resolution of complex problems that use probability as representation of Statistic-
al uncertainty knowledge. The most important and popular PGM is Bayesian network
(BN) [5]. It is a directed acyclic graph (DAG) to represent a joint probability distribu-
tion of random variables and each node has a conditional probability table (CPT) that
quantitatively reflects variable dependency. BN provides a framework for compact
representation and efficient inferences about joint probability distribution of several
interdependent parameters [6], and has been widely used in real applications.
Generally for BN construction, there are two kinds of methods: constructing from
expert knowledge and learning from data [5]. The former relies on the knowledge of
experts which is of subjectivity and one-sidedness frequently. However, the latter
provides us a feasible way to obtain uncertain knowledge implied in data, frequently
used to refine the initial expert knowledge. Learning BN from data has been paid
much attention in machine learning and data mining [5, 6, 7, 8]. The inherence of
massive data demands the task of BN learning to be adjusted to the large volume of
sample data and executed in parallel. It is known that the tasks of BN learning include
DAG and CPT learning [5], in which the former is the critical and challenging step.
This makes us focus on the DAG learning from massive data in this paper.
Currently, dependency analysis and scoring & search are the methods for learning
BN from data [6]. The typical dependency analysis methods focus on the dependence
between data objects [6]. The typical scoring & search algorithms are K2 [8], CH [9],
MMHC [10] and MDL [11]. These methods choose the candidate BN structure based
on the scoring function quantitatively. In this paper, we adopt the scoring & search as
the representative and initial exploration of BN learning from massive data, while the
basic ideas can be universally suitable for the dependency analysis as well.
The crucial processes in K2 are structure scoring and structure search [8]. The
structure search function is applied for each nodes candidate structures, possible
subgraphs made up of nodes likely to connect the current node, to compute their
scores respectively, so the local optimal structure with the highest score can be ob-
tained. Then, local optimal structures are merged to achieve the global optimal one.
However, K2 will have to be extended w.r.t. the storage and retrieval of massive data
for real applications. As the data volume grows, the execution time of K2 will be
increased rapidly due to the expensive cost for computing the required parameters.
The scoring function [8] in K2 is as follows:
n qi (ri 1)! ri
g (i , i ) = N ijk ! (1)
i =1 j =1 ( N + r 1)! k =1
ij i

where the parameters Nij and Nijk can be obtained by statistics from data. Actually, the
values of Nij and Nijk of each record can be obtained in parallel, and then the values of
these two parameters of all records can be summed to obtain the final values required
in Equation (1). It can be observed that the parameters can be obtained separately
(i.e., independently), and the scoring function for local structures can be evaluated
separately as well. For each node, the local optimal structures could be chosen by
 A MapReduce-Based Method for Learning Bayesian Network from Massive Data 699

scoring all candidate ones separately. So, we can merge all local optimal structures to
obtain the global optimal one as the final result. The above observation and discussion
illustrate that both the scoring and search process can be executed in parallel.
From the parallel point of view, some parallel methods for BN learning [12, 13, 14,
15] have been achieved. As a representative, the dynamic programming algorithm
[14] is used to obtain the local and global optimal structures. However, these methods
only focus on the parallelization during algorithm execution and the allocation for
processor cooperation, while have not concerned the throughput computing to access
massive data. Thus, it is desired to explore a BN learning method that can be executed
in parallel and efficiently to access massive data.
Fortunately, MapReduce is a programming model for processing and generating
large data sets [16]. It not only offers a parallel programming model, but also can
conduct and process massive data. It has been used in data analysis and dependency-
analysis-based BN learning [17, 18, 19, 20, 21]. Although MapReduce cannot reveal
its advantages on conducting data of normal volume, it is suitable for learning BN
from massive data. Using MapReduce to extend the current centralized K2 algorithm,
we are to explore the parallel and data-intensive K2 algorithm. To make the MapRe-
duce-based K2 fit the demand of massive data definitely is the purpose of our study.
Hadoop is an open source implementation of MapReduce, and has been used exten-
sively outside of Google by a number of organizations [22]. In this paper, we pro-
posed a MapReduce-based K2 in which the scoring and search process can be
executed in parallel on Hadoop. Generally, the contributions of this paper are as
follows:

We developed the architecture of BN learning system by extending the traditional

K2 algorithm based on MapReduce.
We introduced the scoring function based on MapReduce, which can be used to
obtain all parameters in parallel. In the map process, the parameters in each record
can be obtained by statistics. In the reduce process, the parameters obtained
through the map process will be summed to obtain the requited ones.
We proposed the search function based on MapReduce. In the map process, the
candidate local structures are generate and scored in parallel. In the reduce process,
all the local optimum structures are merged to generate the global optimal one.
Thus, we obtained the BN structure from massive data by extending K2 algorithm.
We implemented our proposed algorithms for learning BN from massive data.
Experimental results show the feasibility of our method.

The rest of this paper is organized as follows: Section 2 introduces K2 algorithm and
MapReduce. Section 3 extends the K2 algorithm based on MapReduce. Section 4
gives experimental results. Section 5 concludes and discusses future work.

2 Preliminaries

2.1 Scoring and Search for BN Learning

A BN is a DAG G=(V, E), in which the following holds [5]:
700 Q. Fang et al.

(1) V is the set of random variables and makes up the nodes of the network.
(2) E is the set of directed edges connecting pairs of nodes. An arrow from node X
to node Y means that X has a direct influence on Y (X, YV and XY).
(3) Each node has a CPT that quantifies the effects that the parents have on the
node. The parents of node X are all those that have arrows pointing to it.
Let Z be a set of n discrete parameters, where a variable xi in Z has ri possible value
assignments: (vi1, ..., vij). Let D be a database with m records (i.e., rows), each of which
contains a value assignment for each variable in Z. Let BS denote a belief-network
structure containing just the parameters in Z. Each variable xi in BS has a set of par-
ents, which we represent with a list of ri parameters. Let wij denote the jth unique
instantial record of Li relative to D. Suppose there are qi such unique instantial records
of Li. Define Nijk as the number of cases in D in which variable xi has the value vik and
Li is instantiated as wij. Here let
ri
N ij = N ijk (2)
k =1

For each node i and its parents i, we use Equation (1) to compute the score of the
candidate structures of each node, and use Pred(xi) to denote the set of nodes preced-
ing xt. We consider the preceding parent of each node according to the nodes
sequence. In the scoring function g(i, i), Nij and Nijk can be obtained by statistics
counting from the given massive data in parallel, which will be discussed in Section
3. The following pseudo-code expresses the heuristic search algorithm, called K2 [8].
Algorithm 1. K2
Input:
1) A set of n nodes, a sequence of the nodes
2) An upper bound u on the number of parent nodes
3) A database D containing m records
Output: A printout of the parents for each node
Steps:
For i:=1 to n Do
i :=
Pold:=g(i, i ) //by Equation (1)
OKTo Proceed:=true
while OKToProceed and | i |<u Do
Let z be a node in Pred(xi) i maximizing g(i, i {zi})
Pnew:=g(i, i {zi})
If Pnew>Pold Then
Pold:=Pnew
i := i {z}
Else OKToProceed:=false
End while
output(`Node:', xi, `Parents of this node:', i )
End for
 A MapReduce-Based Method for Learning Bayesian Network from Massive Data 701

For each node, we get a list of its parents. Actually, we can score the candidate local
structures composed of the node and its parents in parallel. In the scoring of each
local structure, the parameters in Equation (1) can be counted from data in parallel.

2.2 MapReduce Model

The computation takes or produces a set of input or output key/value pairs. Users of
MapReduce library express the computation as two functions: Map and Reduce.
Map, written by users, takes an input pair and produces a set of intermediate
key/value pairs. The MapReduce library groups all intermediate values associated
with the same intermediate key together and passes them to the Reduce function.
Reduce, also written by users, accepts
an intermediate key and a set of values
for that key. It merges these values to
form a possibly smaller set of values. The
intermediate values are supplied to the
users reduce function via an iterator.
This allows us to handle lists of values
that are too large to fit in memory [15].
Fig. 1 (a) indicates that we send the in-
put data to the Master node, by which the
data are split into many maps sent to
Slave node, which will return the result to
Master node and reduce the map result to
the final one for output. The NameNode
JobTracker is allocated on Master node to
control the whole task and DataNodes,
and TaskTracker is allocated on Slave
node to communicate with the NameNode
JobTracker, shown in Fig.1 (b). Fig. 1. MapReduce Process

3 MapReduce-Based Algorithm for BN Learning

The architecture of the MapReduce-based K2 algorithm is shown in Fig. 2:

Preprocessing. The required parameters in Equation (1) are obtained by statistical

counting from the given massive sample data in parallel based on MapReduce.
Searching. For each node (i.e., a random variable) in a BN, the candidate local
structures are scored in parallel by the map process, and the global optimal struc-
ture is obtained by the reduce process. The parameters obtained in preprocessing
will be used for scoring each candidate structure, and the local optimal ones with
the highest score of all nodes will be merged to achieve the global optimal one.

In this architecture, massive data will be split into m map processes dynamically,
whose results will be collected to the final result by R (usually R=1) reduce processes.
702 Q. Fang et al.

Fig. 2. Architecture for MapReduce-based K2

3.1 MapReduce-Based Scoring Function

In general, ri in Equation (1) is a constant, which we do not compute in parallel. As-

sume there is a sample data set like Table 1, in which each column denotes a single
variable and each record denotes a case upon the variables x1, x2 and x3 that denote
smoking, cancer and Tuberculosis respectively. For simplicity, 0 and 1 are used to
denote Absent and Present respectively. In this example, ri is 2.

Table 1. Sample data

Variable values for each case

Case
x1 x2 x3
1 Present Absent Absent
2 Present Present Present
3 Absent Absent Present
4 Present Present Present
5 Absent Absent Absent
6 Absent Present Present
7 Present Present Present
8 Absent Absent Absent
9 Present Present Present
10 Absent Absent Absent

To obtain the parameters in Equation (1), we develop Algorithm 2 including the

map process and reduce process respectively for accessing the sample data and sum-
ming the intermediate results. In each map process (Algorithm 2.1), the parameters Nij
and Nijk of each record will be obtained as intermediate results.
Algorithm 2. MapReduce-based Scoring Function
Input:
1) m records upon n variables
2) record Lx (1xm) including values in {s1, s2, , sn},
 A MapReduce-Based Method for Learning Bayesian Network from Massive Data 703

where si (1in) is 1 or 0, denoting the ith variable is

Present or Absent respectively
Output: <key/value> pairs of Nij (or Nijk) and the value
of Nij (or Nijk) (1in; 1jqi; 1kr)
Algorithm 2.1. Map (String key, String value)
//key: key of data record, value: content of data record
Compute Nij and Nijk from Lx (1in; 1jqi; 1kri)
//by Equation (2)
If Nijk =1 Then
Generate key/value pairs <Nijk, 1>
End If
If Nij =1 Then
Generate key/value pairs <Nij, 1>
End If
If Nijk =0 Then
Generate key/value pairs <Nijk, 1>
End If
If Nij =0 Then
Generate key/value pairs <Nij, 1>
End If

For the first record in Table 1, we can obtain the <key/value> pair <N(x1=0, x2=0), 0>
by Algorithm 2.1, that is the number of records satisfying x1=0, x2=0 is 0 and the val-
ue of Nij is 0. Similarly, we can obtain for the pairs for other records.
In the reduce process (Algorithm 2.2), the intermediate results <Nij, rk> and <Nijk,
rk> that have the same key will be collected and summer to obtain parameters <Nij,
rk > and < Nijk, rk> required in Equation (1).
Algorithm 2.2. Reduce(String key, Iterator values)
//key: Nij or Nijk, values: in the Iterator and have the
same key in <key/value> pairs of <Nij, rk> or <Nijk, rk>
Result:=0
For each key/value pair Do
value:=Interator.Next()
result:=result+value
End For
Output<key, result>

For the data in Table 1, we can obtain <N(x1=0, x2=0), 4> by Algorithm 2.2.

3.2 MapReduce-Based Search Function

According to a sequence of the nodes Seq and the Pred(xi) which represents the set of
nodes that precede xi in the predefined node sequence Seq, the candidate structures of
node xi can be obtained. For example, we assume Pred(x1)={x2, x3}, Pred(x2)={x3},
Pred(x3)={} and Seq={x1, x2, x3}. For node x1, there are 3 candidate structures (x1
704 Q. Fang et al.

Fig. 3. Candidate structures Fig. 4. Local optimal structures Fig. 5. Global optimal structure
of x1 and x2 of x1 and x2

x2), (x1 x3) and (x1 x2, x1 x3), shown in Fig. 3 (a), Fig. 3 (b) and Fig. 3 (c) respec-
tively. For node x2, the candidate structure (x2 x3) is shown in Fig. 3 (d).
The candidate structures generated according to Seq and Pred(xi) are formed as
<key/value> pairs <xi, Si>, where Si is the set of l candidate structures of node xi, and
Si={Si1, Si2, , Sil}. For example, the candidate structure of x2 will be <x2, x2 x3>. In
each map process (Algorithm 3.1) of Algorithm 3, the candidate local structures will
be scored in parallel to obtain the local optimal one and split into M map processes..
Algorithm 3. MapReduce-based Search Function
Input: candidate local structures of each node
Output: the local optimal structure of each node

Algorithm 3.1. Map (String key, String value)

//key: node name xi, value: a candidate structure of xi
scoreij:=g(i, Sij); (1jl)//by Equation (1)
Generate key/value pairs <i&Sij, scoreij>
//key consists of node name xi and structure Sij, and value
is the score of Sij

The intermediate results generated by map processes are formed as <key/value> pairs
<xi&Sij, scoreij>. In each reduce process (Algorithm 3.2), the structure of node xi with
the highest score will be chosen as the local optimal structure of xi.
Algorithm 3.2. Reduce(String key, String value)
// key: node name xi & structure Sij, value: score of Sij
max:=0;
For each Sij of xi Do
If max<scoreij Then
max:=scoreij
Content:=key.Sij
Node:=key.xi
End If
Output <Node,Content>
End For
 A MapReduce-Based Method for Learning Bayesian Network from Massive Data 705

Now, we exemplify the structure search process on the data in Table 1. First, the
scores of candidate structures (x1 x2), (x1 x3), (x1 x2, x1 x3), shown in Fig. 3 (a),
Fig. 3 (b) and Fig. 3 (c), can be computed in parallel by map processes. The scores of
these three structures are 1.4410-5, 7.2210-6 and 2.2310-10 respectively. Thus,
(x1 x2) can be chosen as the optimal structure for x1 with the highest score by the
reduce process. By the same way, we can choose the only candidate structure
(x2 x3), shown in Fig. 3 (d), as the optimal structure for x2 with the highest score
7.2210-5.
Second, by merging the local optimal structures (x1 x2) and (x2 x3) shown in
Fig. 4, we can obtain the global optimal structure, shown in Fig. 5. According to the
ideas in [8], the candidate local structures generated by Algorithm 1 are based on the
sequence of Seq, which makes the merging result be the global optimal structure.
If the massive sample data have m records upon n variables. The time complexity
of traditional K2 is O(m*n2). Assume the number of Slave nodes is large enough, and
a map is executed in each Slave node, that is one Slave node exactly processes one
record of data. If there are t Slave nodes, the time complexity of Algorithm 2 will be
O(m*n2/t), whose order is much lower than O(m*n2). Comparing the time consumma-
tion of Algorithm 2 and that of Algorithm 3, the former fulfills the preprocessing and
mainly determines the efficiency of the whole method. Thus, we conclude that the
time complexity of our method is O(m*n2/t).
The BNs DAG can be obtained by Algorithm 2 and Algorithm 3. Then, we can
easily compute the CPTs for each node in the DAG by the similar MapReduce-based
ideas adopted in Algorithm 2 when accessing the massive data.

4 Experimental Results

4.1 Experiment Environment

Our experiment platform includes 4 machines. One machine is Master node and the
others are Slave nodes. The CPU of each machine is Pentium(R) Dual-Core CPU
E5700 @3.00GHz @3.01GHz with 2GB main memory. On each machine, the ver-
sions of Hadoop, Linux and Java are 0.20.2, Ubuntu 10.04, and JDK 1.6 respectively.
Chest-clinic [23] is a widely used BN in medical science, including 8 nodes, where
each node corresponds to one kind of disease or its reason, and the arrows between
nodes denote their causalities. We adopted the Chest-clinic test data and the corres-
ponding BN in the experiments. The initial data set has 1000 records. To achieve the
massive data, we generated the test data by multiple copies of the original sample set
and the volume of the generated test data is larger than the initial data but the implied
BN structure is the same. On average, every one million records of data have
15.26MB. We generated the test data from 15.62KB to 610.35MB.

4.2 Efficiency Comparison

To test the efficiency of Algorithm 2 and Algorithm 3, we considered two kinds of
execution time, called total time and algorithm time and denoted as AT and TT re-
spectively. Total time is the total execution time of the Hadoop system running our
706 Q. Fang et al.

Data Volume=1.53MB Data Volume=152.59MB

60 4000
Time(s)

Time(s)
40
2000
20
0 0
1 2 3 1 2 3

Number of Datanodes Number of Datanodes

Data Volume=15.26MB Data Volume=305.18MB

400 10000
Time(s)

Time(s)
200 5000
0 0
1 2 3 1 2 3
Number of Datanodes Number of Datanodes

Fig. 6. Comparisons between total time and algorithm time on various Slave node numbers and
data volume

algorithms, and algorithm time is the execution time of Algorithm 2 without including
the time of initialization of Hadoop and I/O. First, we recorded TT and AT on differ-
ent sized data with the increase of Slave nodes, shown in Fig. 6.
Then, we recorded the total time with the increase of data volume with 1, 2 and 3
Slave nodes respectively, shown in Fig. 7. It can be seen that the total time is in-
creased slowly with the increase of data volume under 1, 2 or 3 Slave nodes. Mean-
while, for each size of the sample data, the smaller the data size, the less the total time
will be. This means that the method for learning BN from sample data can be run
efficiently on the Hadoop-based architecture. With the increase of machines in the
cluster, BN can be constructed in less time.
Meanwhile, we recorded the algorithm time (including those of Algorithm 2 and
Algorithm 3) with the increase of data volume with 1, 2 and 3 Slave nodes respective-
ly, shown in Fig. 8. It can be seen that the algorithm time is increased with the in-
crease of data volume under 1, 2 or 3 Slave nodes. Meanwhile, for each size of the
sample data, the smaller the data size, the less the algorithm time will be. This means
that the map and reduce steps can be executed efficiently in Hadoop, which verifies
that our extension to the classical algorithm for BN learning is efficient and effective.

14000 6000

12000 1 slave node 1 slave node

5000
2 slave nodes 2 slave nodes
10000 3 slave nodes 4000
3 slave nodes
Time (s)

8000
Time (s)

3000
6000
2000
4000
1000
2000
0
0
15.26 152.59 305.18 457.76 610.35 15.26 152.59 305.18 457.76 610.35
Data size (MB) Data size (MB)

Fig. 7. Total time with the increase of data Fig. 8. Algorithm time with the increase of data
volume volume
 A MapReduce-Based Method for Learning Bayesian Network from Massive Data 707

4.3 Structure Comparison

Power Constructor [23] is one of
the current BN learning systems, by
which we applied the chest-clinic
sample data with 10000 records to
construct BN, shown as Fig. 9. Also
we used the same data to construct
BN by our method, and we obtained
the same structure as that obtained
by Power Constructor. Thus, we Fig. 9. Chest-clinic BN learned by the Power Con-
can confirm that the BN structure structor and our method
learned by our algorithm is correct.

5 Conclusion and Future Work

In this paper, as the initial exploration of BN learning from massive data, we gave a
MapReduce-based method for BN learning from massive data while focusing on the
learning of BNs DAG structure. Our method is the extension but a more efficient
version of the traditional scoring & search algorithm. The extended algorithm is suit-
able for data-intensive computing and can be adopted generically for representing and
inferring uncertainties implied in massive data.
In the future, we will improve the environment for real massive data sets. Further,
we can store the learned large-scaled BN in the distributed file system by looking
upon it as a kind of informative data. Moreover, we can extend other BN learning
algorithms by incorporating the ideas explored and proposed in this paper, so that the
knowledge model can be obtained efficiently even oriented to massive data.
Acknowledgement. This paper was supported by the National Natural Science Foun-
dation of China (61063009, 61163003, 61232002), the Ph. D Programs Foundation of
Ministry of Education of China (20105301120001), the Yunnan Provincial Founda-
tion for Leaders of Disciplines in Science and Technology (2012HB004), the Founda-
tion for Key Program of Department of Education of Yunnan Province (2011Z015),
and the Foundation of the Key Laboratory of Software Engineering of Yunnan Prov-
ince (2012SE205).

References
1. Kouzes, R., Anderson, G., Elbert, S., Gorton, L., Gracio, D.: The changing paradigm of
dataintensive computing. IEEE Computer 42(1), 2634 (2009)
2. Agrawal, D., El Abbadi, A., Antony, S., Das, S.: Data Management Challenges in Cloud
Computing Infrastructures. In: Kikuchi, S., Sachdeva, S., Bhalla, S. (eds.) DNIS 2010.
LNCS, vol. 5999, pp. 110. Springer, Heidelberg (2010)
3. Borkar, V., Carey, M., Grover, R., Onose, N., Vernica, R.: Hyracks: A flexible and extens-
ible foundation for dataintensive computing. In: Abiteboul, S., Bhm, K., Koch, C., Tan,
K. (eds.) Proc. of ICDE 2011, pp. 11511162. IEEE Computer Society, Hannover (2011)
4. Deshpande, A., Sarawagi, S.: Probabilistic graphical models and their role in database. In:
Koch, C., Gehrke, J., Garofalakis, M.N., et al. (eds.) VLDB 2007, pp. 14351436. ACM
(2007)
708 Q. Fang et al.

5. Pearl, J.: Probabilistic reasoning in intelligent systems: network of plausible inference.

Morgan Kaufmann, San Mates (1988)
6. Russel, S., Norvig, P.: Artificial intelligence-A modern approach. Pearson Education,
Prentice Hall (2002)
7. Song, W., Yu, J.X., Cheng, H., Liu, H., He, J., Du, X.: Bayesian Network Structure Learn-
ing from Attribute Uncertain Data. In: Gao, H., Lim, L., Wang, W., Li, C., Chen, L. (eds.)
WAIM 2012. LNCS, vol. 7418, pp. 314321. Springer, Heidelberg (2012)
8. Cooper, G., Herskovits, E.: A Bayesian method for the induction of probabilistic networks
from data. Machine Learning. Machine Learning 9(4), 309347 (1992)
9. Heckerman, D., Geiger, D., Chickering, D.: Learning Bayesian networks: the combination
of knowledge and statistic data. Machine Learning 20(3), 197243 (1995)
10. Tsamardinos, I., Brown, L., Aliferis, C.: The max-min hill-climbing Bayesian network
structure learning algorithm. Machine Learning 65(1), 3178 (2006)
11. Suzuki, J.: Learning Bayesian belief networks based on the MDL principle: An efficient
algorithm using the branch and bound technique. IEICE Trans. Information and Systems
E82-D(2), 356367 (1999)
12. Xiang, Y., Chu, T.: Parallel learning of belief networks in large and difficult domains. Da-
ta Mining Knowledge Discovery 3(3), 315338 (1999)
13. Yu, K., Wang, H., Wu, X.: A parallel algorithm for learning Bayesian networks. In: Zhou,
Z.-H., Li, H., Yang, Q. (eds.) PAKDD 2007. LNCS (LNAI), vol. 4426, pp. 10551063.
Springer, Heidelberg (2007)
14. Yoshinori, T., Seiya, I., Satoru, M.: Parallel Algorithm for Learning Optimal Bayesian
Network Structure. Journal of Machine Learning Research 12, 24372459 (2011)
15. Zhang, Q., Wang, S., Qin, B.: Cleaning Uncertain Streams by Parallelized Probabilistic
Graphical Models. In: Chen, L., Tang, C., Yang, J., Gao, Y. (eds.) WAIM 2010. LNCS,
vol. 6184, pp. 274279. Springer, Heidelberg (2010)
16. Dean, J., Ghemawat, S.: MapReduce: a flexible data processing tool. Communications of
the ACM 53(1), 7277 (2010)
17. Chu, C., Kim, S., Lin, Y., Yu, Y., Bradski, G., Ng, A., Olukotun, K.: Map-Reduce for Ma-
chine Learning on Multicore. In: Schlkopf, B., Platt, J.C., Hoffman, T. (eds.) NIPS 2006,
pp. 281288. MIT Press, Vancouver (2006)
18. Low, L., Bickson, D., Gonzalez, J., Kyrola, A., Guestrin, G., Hellerstein, J.: Distributed
GraphLab: a framework for machine learning and data mining in the cloud. PVLDB 5(8),
716727 (2012)
19. Chen, W., Zong, L., Huang, W., Ou, G., Wang, Y., Yang, D.: An empirical study of mas-
sively parallel Bayesian networks learning for sentiment extraction from unstructured text.
In: Du, X., Fan, W., Wang, J., Peng, Z., Sharaf, M.A. (eds.) APWeb 2011. LNCS,
vol. 6612, pp. 424435. Springer, Heidelberg (2011)
20. Bahmani, B., Kumar, R., Vassilvitskii, S.: Densest subgraph in streaming and MapReduce.
PVLDB 5(5), 454465 (2012)
21. Yuan, P., Sha, C., Wang, X., Yang, B., Zhou, A., Yang, S.: XML Structural Similarity
Search Using MapReduce. In: Chen, L., Tang, C., Yang, J., Gao, Y. (eds.) WAIM 2010.
LNCS, vol. 6184, pp. 169181. Springer, Heidelberg (2010)
22. White, T.: Hadoop: The Definitive Guide. OReilly Media, Inc. (2009)
23. Cheng, J.: PowerConstructor system,
http://webdocs.cs.ualberta.ca/~jcheng/bnpc.htm
 Practical Duplicate Bug Reports Detection
in a Large Web-Based Development Community

Liang Feng1 , Leyi Song1 , Chaofeng Sha2 , and Xueqing Gong1

1
Software Engineering Institute, East China Normal University
2
School of Computer Science, Fudan University
{liang.feng,songleyi}@student.ecnu.edu.cn,
[email protected], [email protected]

Abstract. Most of large web-based development communities require a

bug tracking system to keep track of various bug reports. However, du-
plicate bug reports tend to result in waste of resources, and may cause
potential conflicts. There have been two types of works focusing on this
problem: relevant bug report retrieval[8][11][10][13] and duplicate bug
report identification[5][12]. The former methods can achieve high accu-
racy (82%) in the top 10 results in some dataset, but they do not really
reduce the workload of developers. The latter methods still need further
improvement on the performance.
In this paper, we propose a practical duplicate bug reports detection
method, which aims to help project team to reduce their workload by
combining existing two categories of methods. We also propose some
new features extracted from comments, user profiles and query feedback,
which are useful for improving the detection performance. Experiments
on real dataset show that our method improves the accuracy rate by 23%
compared to state-of-the-art work in duplicate bug report identification,
and improves the recall rate by up to 8% in relevant bug report retrieval.

Keywords: Bug Report, Duplicate Detection, Classification.

1 Introduction

For the development of large-scale software systems, project team members do

not usually work together. They are distributed all over the world and commu-
nicate each other via various tools. In addition to e-mail, instant messaging soft-
wares, web forums and wikis have become common as a communication medium
for users. Moreover, Web-based software engineering tools are also widely used,
such as bug tracking systems. Users and their various tools constitute a web-
based development community.
A bug tracking system is designed to keep track of reported software bugs in
a development community. Users can submit bug reports to the system which
describe the software problems, and comments upon existing bug reports. Each
bug report is triaged to a suitable developer who will in charge to address the re-
lated problems. Usually, there are thousands of developers in a large development

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 709720, 2013.
c Springer-Verlag Berlin Heidelberg 2013

710 L. Feng et al.

community, and hundreds of bugs will be reported daily. Dierent QAs might
report the same problem many times respectively. This causes an issue as du-
plicate bug reports. The phenomenon of duplicate bug reports is very common.
If duplicate bug reports of the same problem are triaged to dierent developer,
lots of eorts will be wasted and potential conflicts will be made.
There have been a number of works which focus on the duplicate bug reports
detection issue in a bug tracking system. Existing methods [8][11][10][13][5][12][7]
can be divided into two dierent categories: relevant bug report retrieval and du-
plicate bug report identification. Given a new submitted bug report, the former
methods will return a sorted list of existing reports that are similar to it. How-
ever, the latter methods will identify that it is a duplicate report or not. As far
as we know, the state-of-the-art search approach[7] can guarantee that the re-
lated bug reports will be found in the top-10 search results with 82% probability
in some dataset if the new bug report is really duplicate. It is already a very
good result. However, in actual development environment, it still requires users
to spend lots of eorts to review the search results manually. From a practical
point of view, duplicate bug reports identification can really reduce the devel-
opers eorts. Yuan Tian etc.[10] reported their progress in the field of duplicate
bug reports identification. The true positive rate was 24% while keeping the
false positive rate of 9%. It still need to be improved in order to meet the needs
of practical application. In this paper, we propose PDB, a practical duplicate
bug reports detection method, which will help developers to reduce the eorts
on processing bug reports by combining the relevant bug report retrieval and
duplicate bug report identification.
For duplicate bug reports issue, any type of existing methods need to
compute the similarity between reports. Dierent methods will choose dierent
features to compute the similarity. such as textual similarity[8][11], surface fea-
tures(timestamp, severity, product and component etc.)[10], execution trace[13]
etc. Almost all of these features are extracted from the bug reports themselves.
Existing works have tried almost all possible methods in the field of informa-
tion retrieval and machine learning to address the duplicate bug reports issue.
It is hard to improve the performance only depending on these features. We
propose some new features extracted from comments and user profiles according
to the observation on bug reports. In a development community, comments are
necessary complement to bug reports. Furthermore, The reporters knowledge
and experience will aect the quality of the bug reports he submitted. Exper-
iments show that these new features can eectively help us to further improve
the duplicate bug reports detection performance.
The contributions of this paper include:
We propose a practical duplicate bug reports detection method, which adopts
three stages of classifiers to combine existing two categories of methods. It
will really help developers to reduce the eorts on processing bug reports
We propose some new features from comments, user profiles and query feed-
back to improve the eectiveness of duplicate bug report detection.
 Practical Duplicate Bug Reports Detection 711

Our extensive experiments on actual dataset shows PDB method outper-

forms the start-of-the-art approach on identification of duplicate bug re-
ports, and makes sucient improvement in duplicate bug reports retrieval
compared to existing works.
The rest of this paper is organized as follows: section 2 gives a background
knowledge on web-based bug tracking system, bug report, and duplicate bug
report detection. Section 3 presents our practical solution to identify duplicate
bug reports in a large web-based development community. The experimental
study is shown in section 4. We discuss the related works in section 5. Section 6
concludes and describes some potential future works.

2 Background
2.1 Web-Based Bug Tracking System
Nowadays web-based development communities become grown up for its remov-
ing the geographical restrict, fostering communication and collaboration in group
members. Various kinds of web-based development communities are widely used:
Web forums provide place for users to submit and discuss the program questions.
Distributed revision control system like GitHub facilitates sharing codes with
other people and code management for multiple contributors.
Web-based bug tracking system also is one important kind of web-based de-
velopment communities. It is designed to keep track of reported software bugs.
It allows remote access of resources anytime, anywhere. It engages all the roles
including QAs, developers and project managers on a single collaboration plat-
form over the web. Users could review bugs and share comments and resources
freely. All the knowledge are stored and centralized. Currently many web-based
bug tracking systems such as Bugzilla and JIRA have been widely used in soft-
ware projects, especially in large open source software projects like Firefox and
Eclipse.

Fig. 1. General Workflow for Bug Processing

Figure 1 displays a general workflow for processing a bug in web-based bug

tracking system. At first a reporter issues a bug for a special component. Then
712 L. Feng et al.

the bug is automatically sent to triager to check the quality of bug. If the triager
judges that bug information is insucient, the bug will be sent back to submit-
ter to complete. If the bug is invalid, never fixed or duplicate, triage will set
resolution and close bug. If triage have completed the entire checklist for a bug,
bug will be marked as triaged and assigned to related developer. Next, bug is
fixed by developer. After the bug is resolved, QAs verify the resolution of the
bug, and close the bug when it is fixed.

2.2 Bug Report

A bug report is a structured file which is composed by multiple fields. These
fields can be roughly categorized into three types: text data, metadata and at-
tachments.
Text data include summary, description and contents. Summary is a one-
sentence statement that presents problem; Description outlines problem in de-
tail. It usually includes detailed steps to reproduce issue, the actual and expected
result after performing the above steps; Comment is the opinion for this bug ex-
pressed by bug owner or others.
Metadata usually have several fields and generally contain following fields:
Product and component fields indicate the artifact and module where the fault
exist. Assignee is the person who is assigned to fix the bug. Version field shows
the versions of product which the bug appears. Priority is the importance of the
bug. Reporter is the person who submits the bug. Create time is when the bug
is submitted. Status field indicates the processing state which the bug stays at.
Resolution shows the result of bug, once it is resolved.
Attachments are usually added to bug reports in order to provide more data
help to understand bug. They usually are screenshots, patches or logs.

2.3 Duplicate Bug Report Detection

As figure 1 shows, triage is a key procedure to assess the quality of bugs and
prefilter some unnecessary bugs, which reduces nonessential eort, and improves
the eciency of project team. The checklist for bug triage usually contain several
tasks. One main task is to mark duplicate bugs.
Bugs are duplicate when they have the same root cause. When labeling them
duplicate, engineers will set one of them as master bug and others as duplicate
ones. The master bug represents the whole group of duplicate ones with it.
Duplicate bugs occupy a large proportion in many projects, especially in
large open source projects for its loose management. As [3] reported, issued
from November 2006 to March 2008, about 38% of bugs are duplicate out of
60233 bugs in Firefox; about 49% of bugs are duplicate out of 19204 bugs in
Thunderbird; Generally these duplicate bugs cause redundant work and increase
workload for engineers. Meanwhile, as [1] pointed out that duplicate bug pro-
vides additional information about original bug, which is useful for developers
to understand the bug. So identifying duplicate bugs and aggregating them will
significantly improve the eciency of bug processing.
 Practical Duplicate Bug Reports Detection 713

However, it is not trivial to detect duplicate bugs manually. Several factors

hinder the feasibility and eciency of this activity in the large projects:

Firstly, in general it costs much time to analyze a bug report and judge
whether it is duplicate manually. The situation gets worse with the huge
number of bug reports and limited manpower.
Secondly, same error causes failure at dierent level of program. In some
cases, duplicate bugs show completely dierent symptoms. Due to short of
global knowledge, it is hard for triage to detect the relation of these bugs.
Thirdly, a bug is not filed against appropriate component, which might mis-
lead triagers.

Therefore, it need an automatical duplicate bug detection tool to help triagers.

We propose PDB, a practical duplicate bug detection method. It is designed to
complete two tasks and combine them: one is to make prediction whether one
bug is duplicate, the other is to return top-k most possible duplicate bugs given
a bug.

3 Practical Duplicate Bug Report Detection

In this section we describe the techniques in PDB in details; Overall, PDB adopts
three stages of classification in duplicate bug detection. The first one is to predict
how much contribution a comment makes to define a bug. The second one is to
predict the possibility that two bugs are relevant and duplicate. The possibility
is used to be score for ranking. The third one is to predict whether one bug is
duplicate.

3.1 Modeling Comment Similarity

Comments in bugs are the points of view and discussions about the bugs issued
by owner or others. Among these, a large proportion of comments are supplement
to the original content, and help people better understand the appearing result
and find root cause of bug. In some cases when the summary and description
in a bug are vague and ambiguous, some key comments are critical to define to
bug, and also contain important clues in finding duplicate bugs. So PDB puts
comments into consideration when calculating the degree of relevance between
bugs.
Comments in a bug report are issued by various roles with various kinds
of intents. Some comments make some supplements to the description of bug
phenomenon or possible rule causes, which can assist in identifying the relevance
of bugs. But other comments just talks about bug fixing procedure or the results
of bug processing, which deviate from the topic of main body for bug report and
are not crucial in identifying the relevance of bugs. So it is intuitive to put
comments weights dierentially based on the type of comments, and set the
former type of comments higher weight.
714 L. Feng et al.

In order to measure the weight of comment, we divide all the comments into
two classes: usefulness and less usefulness. The first class of comment is the one
which describes bug phenomena, discusses root cause of bugs and solution of
bug. The remaining of comments are categorized to the second class. We set
comment weight as probability that the classifier put the content into usefulness
class. For comment ci in bug Y , we define the weight of this comment as follows:
WY (ci ) = p(usef ulness|ci) (1)
SVM with RBF kernel is adopted as classifer. There are four types of features to
classify the types of comments: average-idf, bug status, length of comment text,
and unigrams of a term within comments.
Average-IDF: From our observation, those useful comments usually talks about
the technical details of programs, and contain many professional vocabulary.
These professional vocabularies generally have high idf value. On the contrary,
the less useful comments usually are full of common words. Its much reason-
able that these comments might appear in other bug reports, like please fix
as soon as possible.

Bug status: For a comment, the bug status indicates current state when the
comment is submitted. Usually the intent of comment has strong relation-
ship with the current bug status. E.g. when a bug stands on state of NEW,
QAs usually supplement some source and test environment of bug to facili-
tate triagers to understand bug.

Length of comment text: In generally, the length of comment text is also a

useful measure for comment. We define the length of comment as the number
of terms in comment. A comment with a large number of terms indicates
more information and content. A comment with large information is more
likely to be supplement to bug and be a useful one.

Unigrams of a term within comments: If a comment contains root,

cause and likely, this comment is much likely to discuss phenomena
or root cause of a bug. It is useful. On the other hand, comments containing
words such as close and verify are likely to discuss the result of bug
processing and be less useful. So we use vocabulary clues to determine type
of comments.
Because summary and description field of a bug are representative of defining
bug and tend not to contain unessential information, The weights of summary
and description are set respectively with 2 and 1.
As discussed above, usually a comment in a bug report has its own intent
and topic, the text content of comments in one report diers a lot. So it is
common that it is only one comment has high similarity given a text content,
and others comment in this bug are unmatched. We assume a new submitted
bug have no comment. Given a query bug Q and a candidate bug Y , we propose
W eighted Comment similarity P s(Q, Y ) to calculate the similarity between
 Practical Duplicate Bug Reports Detection 715

summary and description in query bug and comments in candidate bug. Given a
set of comments C in Bug Y and query text q, similarity function between them,
ScoreY (q, C), is the maximum score in similarity between text and comment. It
is defined as follows:
S T
ScoreY (q, C) = max Sim(q, ci ) WY (ci ) | ci C (2)
Where ci denotes the ith comment in a bug report, ci and query text q is modeled
as vector space model, and Sim(q, ci ) is cosine similarity between query text q
and comment ci .
Given query bug X, text set(X) denotes the text set of bug X including
summary and description. C denotes the set of comments in one bug Y , P s(Q, Y )
is defined as follows:
#
qi text set(Q) WQ (qi ) ScoreY (qi , C)
P s(Q, Y ) = # (3)
qi text set(Q) WQ (qi )

Where WQ (qi ) denotes weight of qi in bug Q. It has been discussed above.

3.2 Duplicate Bug Reports Retrieval Model

PDB utilizes a discriminative model to calculate the possibility that two bug
reports are duplicate. The possibility is used as score in ranking. Apart from
weighted comment similarity and features proposed in [10], here we introduce
several new features:
Topic model for document representation has widely used in machine learning
and information retrieval. It makes excellent performance in many applications
[14]. Topic model represents a document as a mixture of multiple topics. For all
bugs report in a project, we learn LDA[2] topics from the corpus of past submit-
ted bugs through gibbs sampling[4]. We put summary and description together
as a bug document, and measure topic distance between two bug documents via
both Jensen-Shannon and symmetric KL divergences[9] between two topic dis-
tributions. Finally, binary match/nonmatch of the most probable topic in each
distribution is also feature. That is, 1 if the match of the most probable topic of
two reports are the same; 0, otherwise.
Generally, the author of bug is much familiar with the bugs which are issued
by him, and can roughly find the cause of bugs. Accordingly, same author is not
likely to issue duplicate bugs. We use an indicator of the same author relationship
as a feature, that is, 1 if the authors of two posts are the same, 0, otherwise.
Table 1 shows all the features in our algorithm.

3.3 Duplicate Bug Classifier Model

PDB leverages classifier to predict whether query bug is duplicate. The features
in use consist of two aspects: user profile and query feedback.
User profile includes both demographic information, eg. name, age, etc, and
his individual preference(technical background, interests and working styles).
716 L. Feng et al.

Table 1. All the features used in bug retrieval model, CS is abbreviation for Cosine
Similarity

# Feature name # Feature name

1 CS : summary 10 Priority Similarity
2 CS : description 11 Version Similarity
3 CS : summary+description 12 Create date distance
4 CS : summary 2gram 13 Jensen-Shannon(LDA): summary+description
5 CS: description 2gram 14 Symmetric KL(LDA): summary+description
6 CS : summary+description 2gram 15 Maximum topic(LDA): summary+description
7 Product distance 16 Reporter distance
8 Component distance 17 CS: summary+description+comments
9 Severity Similarity 18 Weighted Comment similarity

The latter is hard to get, but valuable for predicting the resolution of submitting
bugs. Intuitively novice users as well as expert or advanced users tend to submit
bugs with dierent quality.
PDB takes some user profiles features into consideration. They include the
total number of issuing bugs, the portion of issuing bugs with resolution as
FIXED and the portion of issuing bugs with resolution as DUPLICATE. The
total number of issuing bugs tries to identify the participation and familiarity
of submitter in a project. Two portions shows submitters habit of issuing bugs.
These features are helpful to predict the resolution of a new query bug. For
example, high total number of issuing bugs indicates that the submitter might
have profound knowledge of project and program, and thus is less likely to submit
duplicate bugs; A low portion of bugs with resolution as FIXED and a high
portion of bugs with resolution as DUPLICATE mean the submitter might have
not habit of searching relevant bugs before issuing a new bug, the resolution of
the bug which this author issues has good chance of being duplicate.
The idea of query feedback is to judge the features of query from the returning
sorted results by retrieval engine. Suppose that a user issues query bug Q to a
duplicate bug retrieval system and a ranked list L of candidate bugs is returned.
Intuitively, the more similar top one candidate bug is, the more probably query
bug Q is duplicate bug. However, Due to convenience, submitters usually employ
many same words and very similar writing style when filling in some relevant
bug reports. It will lead to very high similarity score between these bugs which
are submitted by same author, but not duplicate. In order to catch these situ-
ations, PDB consider several query feedback information. The features include
the similarity score with first, fifth, tenth bug report, whether the first, fifth,
tenth returning bug is submitted by same author, the portion of same author in
top five returning bug reports, the portion of same author in top ten returning
bug reports, the average similarity score for in top five returning bug reports and
the average similarity score for in top ten returning bug reports. Similarity score
of bugs is the probability that two bugs are duplicates, described in previous
subsection.
 Practical Duplicate Bug Reports Detection 717

4 Experiments Study

4.1 Dataset
MeeGo is a Linux-based free mobile operating system project. We crawled a
dataset of bugs report from public MeeGo bugzilla1 through open RPC/XML
API. The dataset contains 22486 bugs, which are issued from March 2010 to
August 2011. There are 1613 bugs with the resolution as DUPLICATE.

4.2 Evaluation and Discussion

The first experiment concerns at the performance of PDBs finding duplicate bug
reports We compare PDB with the REP[10]. Several standard measures (MAP,
Precision, Recall) are used in evaluation.
PDB adopts SVM classifer2 with linear kernel. The number of topics in topic
model is set to be 100. The ids of bugs for training set range from 1 to 5000,
and the others are used for test. The query bug set contain both duplicate and
master bugs. The result set are expected to contain all the duplicate bug reports.
Table 2 displays the performance on recall, precision and MAP respectively
for these two approaches. PDB outperforms REP in all measures. PDB achieves
higher recall rate by about 1%-8% for the resulting lists of top 1-20 bug reports.
The gain gap grows as the size of list increases. In precision, PDB outperforms
REF by 1%-8% in top 1-20 resulting bug reports. PDB improves MAP from 36%
to 46%.
We analyze the contribution of the various features to the model by measuring
the F-score through python script in LIBSVM. From all new features, weighted
comment similarity has highest F-score, which means that comments in a bug
play a major role in duplicate bug retrieval. The features related to topic model
follow. Reporter distance is lowest-ranking.

Table 2. Performance results in Recall, Precision and MAP between two methods

Method R@1 R@5 R@10 R@15 R@20

PDB 0.2644 0.5089 0.6117 0.6623 0.6967
REP 0.2569 0.4526 0.5307 0.5780 0.6095
Method P@1 P@5 P@10 P@15 P@20 MAP
PDB 0.3650 0.1411 0.0844 0.0607 0.0479 0.4677
REP 0.2846 0.1064 0.0637 0.0467 0.0371 0.3615

The second experiment is to evaluate the performance of identifying whether

one bug is duplicate bug. We randomly picked up 534 duplicate and non-duplicate
1
Available at http://bugs.meego.com
2
Implemented by LIBSVM. Available at
http://www.csie.ntu.edu.tw/cjlin/libsvm/
718 L. Feng et al.

bugs respectively, and then apply the Naive Bayes, Decision Tree3 and SVM for
classification. RBF kernel is used in SVM classifier. 10-fold cross validation is
proceeded to measure accuracy, true positive rate and true negative rate. The
parameters are grid searched to maximize accuracy.
We make comparison on the accuracy of three classifiers with PDB and state-
of-the-art[12]. TIAN stands for [12] in the following. The best classifier refers to
classifier that achieves the highest accuracy. Figure 2 shows the results. PDB
outperforms TIAN consistently in all three classifiers. SVM and Decision Tree
achieve close accuracy with both PDB and TIAN, but make increase about 10%
on accuracy compared to the Naive Bayes.

Fig. 2. Comparisons on the accuracies of three classifiers with PDB and TIAN methods

We define true positive(TP) rate as the fraction of duplicate bugs which are
correctly classified, and true negative(TN) rate as the fraction of non-duplicate
bug which are correctly classified. The higher are both these two rates, the better
the classifier works.
Table 3 shows the comparison on the TP and TN rate for three classifiers with
PDB and TIAN. In all three classifiers, PDB outperforms TIAN by 2%-31% on
both TP rate and TN rate, except on TN rate in Decision Tree. Naive Bayes
classifier shows best performance on TP rate, but get worst result on TN rate
in three classifers. Similar with accuracy, SVM and Decision Tree achieve close
performance on TP rate and TN rate overall. SVM outperforms Decision Tree
on TP Rate and TN rate by a narrow margin. Therefore, Overall SVM do most
stable and accurate work in three classifiers.

3
Naive Bayes and Decision Tree are implemented by WEKA. Available at
http://www.weka.net.nz
 Practical Duplicate Bug Reports Detection 719

Table 3. Comparisons on the TP rate and TN Rate for three classifiers with PDB and
TIAN methods

PDB TIAN
Classifer TP Rate TN Rate TP Rate TN Rate
Naive Bayes 89.9% 33% 87.2% 12.2%
Decision tree 77.9% 61.6% 45.4% 66.7%
SVM 76.2% 68% 52% 57.2%

5 Related Works

In recent years, there have been a number of literatures on detecting duplicate

bugs. These literatures can be roughly divided into two categories: one is to
find most relevant bugs in bug report collection given a bug; The other aims to
identify whether one bug is duplicate.
In the first category, Runeson et al.[8] work is the first one to address this
problem. They make natural language process on the raw text data, use vec-
tor space model as computational framework, and set term weighting based on
term frequency. Wang X et al.[13] combine text similarity and execution trace
between bugs to rank bug reports. However, as Sun et al.[11] point out, the
proportion of bugs contains execution information is fairly low in some projects
like OpenOce, Firefox. So is MeeGo dataset. Sun et al.[10] propose to train
a support vector machine classifier to detect duplicate bugs with 54 text simi-
larities feature including multiple text field. Sun et al.[10] propose REP, which
combines BM25F with some meta-data like reported product, component, ver-
sion. They make extensive experiment to show that REP outperforms the other
works in eectiveness. Kaushik, N. et al.[6] make extensive comparison on dif-
ferent IR models. The experiment shows Log-Entropy based weighting scheme
outperformed other models in eectiveness.
In second category of works, Jablert et al.[5] use surface features, textual se-
mantics, and graph clustering to predict duplicate status. More recently, Tian,
Y et al.[12] work based on [5]. They used maximum REP similarity, and add an-
other category information, the product dierence. Also they bring into relative
similarity to determine the importance of the text similarity between bugs.
Our work aims to solve both two categories. In finding relevant bug report, we
adopt discriminative model, and be first one to utilize comments to improve the
eectiveness of duplicate bug detection. In identifying duplicate bug, we leverage
user profile and query feedback to help predict query bug report resolution.

6 Conclusion

Detection of duplicate bugs is a necessary step to optimize resource allocation

and improve working eciency in the procedure of triage. In this work, we
propose a practical duplicate bug reports detection method, which will help
720 L. Feng et al.

triagers to reduce their workload by combining existing two categories of meth-

ods. PDB adopts three stage to detect duplicate bug reports. In modeling the
similarity of two bugs, PDB is the first one to utilize comments to enhance the
eectiveness. In identifying duplicate bug, PDB makes use of user profiles and
query feedback. We have evaluate our algorithm on MeeGos bug reports which is
issued in about one year. The experiment shows that PDB outperforms existing
works [10] and state-of-the-art[12] in respective category of thread.
In future, we plan to investigate more bug reports from various software
projects. Another idea is to utilize other text data like use cases and posts
on online technical forums to enhance eectiveness of duplicate bug detection.
Some of these are closely associated to and supplement to bug reports.

Acknowledgements. This work is partially supported by National Science

Foundation of China under grant number 60803022.

References
1. Bettenburg, N., Premraj, R., Zimmermann, T., Kim, S.: Duplicate bug reports
considered harmful... really? In: ICSM, pp. 337345 (2008)
2. Blei, D.M., Ng, A.Y., Jordan, M.I.: Latent dirichlet allocation. Journal of Machine
Learning Research 3, 9931022 (2003)
3. Cavalcanti, Y.C., de Almeida, E.S., da Cunha, C.E.A., Lucredio, D., de Lemos
Meira, S.R.: An initial study on the bug report duplication problem. In: CSMR,
pp. 264267 (2010)
4. Griths, T.: Gibbs sampling in the generative model of latent dirichlet allocation.
Unpublished note (2002), http://citeseerx.ist.psu.edu/viewdoc/summary
5. Jalbert, N., Weimer, W.: Automated duplicate detection for bug tracking systems.
In: DSN, pp. 5261 (2008)
6. Kaushik, N., Tahvildari, L.: A comparative study of the performance of ir models
on duplicate bug detection. In: CSMR, pp. 159168 (2012)
7. Nguyen, A.T., Nguyen, T.T., Nguyen, T.N., Lo, D., Sun, C.: Duplicate bug report
detection with a combination of information retrieval and topic modeling. In: ASE,
pp. 7079 (2012)
8. Runeson, P., Alexandersson, M., Nyholm, O.: Detection of duplicate defect reports
using natural language processing. In: ICSE, pp. 499510 (2007)
9. Steyvers, M., Griths, T.: Probabilistic topic models. Handbook of Latent Seman-
tic Analysis 427(7), 424440 (2007)
10. Sun, C., Lo, D., Khoo, S.-C., Jiang, J.: Towards more accurate retrieval of duplicate
bug reports. In: ASE, pp. 253262 (2011)
11. Sun, C., Lo, D., Wang, X., Jiang, J., Khoo, S.-C.: A discriminative model approach
for accurate duplicate bug report retrieval. In: ICSE (1), pp. 4554 (2010)
12. Tian, Y., Sun, C., Lo, D.: Improved duplicate bug report identification. In: CSMR,
pp. 385390 (2012)
13. Wang, X., Zhang, L., Xie, T., Anvik, J., Sun, J.: An approach to detecting duplicate
bug reports using natural language and execution information. In: ICSE, pp. 461
470 (2008)
14. Wei, X., Croft, W.B.: Lda-based document models for ad-hoc retrieval. In: SIGIR,
pp. 178185 (2006)
 Selecting a Diversified Set of Reviews

Wenzhe Yu1 , Rong Zhang1 , Xiaofeng He1 , and Chaofeng Sha2

1
East China Normal University, Shanghai, China
2
Fudan University, Shanghai, China
[email protected],
{rzhang,xfhe}@sei.ecnu.edu.cn,
[email protected]

Abstract. Online product reviews provide helpful information for user

decision-making. However, since user-generated reviews proliferate in
recent years, it is critical to deal with the information overload in e-
commerce sites. In this paper, we propose an approach to select a small
set of representative reviews for each product, which shall consider both
the attribute coverage and opinion diversity under the requirement of
providing high quality reviews. First, we assign weights to each attribute,
which measure the attribute importance and help realize useful review
selection; second, we cluster reviews into dierent groups representing
dierent concerns which lead to better diversification results especially
for selecting smaller sets of reviews; finally, we perform a set of experi-
ments on real datasets to verify our ideas.

1 Introduction
Online user reviews have great eects upon decision-making process of users.
Customers would like to read reviews to get a full picture of an item (i.e., a
product or service) when they are selecting items. For example, according to
the reviews of a camera, a customer can have knowledge of the size, color and
convenience of usage though they have never bought it.
However, since user-generated reviews proliferate in recent years, e-commerce
sites are facing challenges of information overload. In eBay.com, there are typi-
cally several hundreds of reviews for popular products such as the latest model
of iPhone. The massive volume of reviews prevent users from catching the clear
view of products. On one hand, it is not easy for users to read all of the re-
views. Users may be unable to go through all the reviews for each item, for it is
too time-consuming. On the other hand, an increasing number of users choose
to shop online via mobile phones, such as reserving a table at a restaurant or
picking a movie to watch when they are outside. Those users have to make their
purchase decision in a short time, so they can only read a small fraction of re-
views for each item. In addition, due to the limited screen size of mobile phones,
it is inconvenient for users to scan more reviews.
To address the problem of information overload, e-commerce sites have adopted
several kinds of review ranking and selection methods. Review ranking ranks re-
views according to their helpfulness votes, so as to provide top-k reviews to users.

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 721733, 2013.
c Springer-Verlag Berlin Heidelberg 2013

722 W. Yu et al.

Helpfulness votes are evaluated by users to those reviews that are helpful to them.
And there are also a number of researches on automatically estimating the quality
of reviews [8][11][15][6]. However, there are two drawbacks in these review ranking
work. First, the resulting top-k reviews of an item may contain redundant infor-
mation while some important attributes may not be covered. For example, top-k
reviews of a mobile phone may comment on display, camera and battery life, but
mention nothing about whether it is easy to use or carry. Second, since previous
experiments [3] showed that users tend to consider helpful the reviews that fol-
low the mainstream, the resulting top-k reviews may lack diversity of opinions.
By these two observations, review selection based on attributes coverage [14] is
proposed. It prefers to select reviews covering as many attributes as possible. But
it is found that such kind of method does not reflect the original customer opinion
distributions and may lead to unfair picture to users. Then [9] comes to solve the
problem by selecting a set of diversified reviews that keep the proportion of posi-
tive and negative opinions. But we have found that this work does not perform well
especially for selecting a smaller set of reviews, which is caused by the overlooking
of attribute importance. For example, for a T-shirt, there are a lot of reviews com-
plaining about its delivery speed while a smaller number of reviews complaining
on the product quality. In such a case, it may be better to show reviews for clothes
quality.
Hence, to improve the overall value of the top-k reviews, we view the top-k
reviews as a review set rather than simple aggregation of reviews. We propose
an approach to select a small set of reviews that cover important attributes
with high quality, while at the same time have better diversified attributes and
opinions by clustering. So our contributions in this paper are as follows:
1. We propose to evaluate attribute importance by calculating weights to
them. We suppose to return users reviews covering attributes to their concerns.
2. In order to improve the diversity of the top-k results, we propose to cluster
reviews to dierent topic groups. We design an algorithm to select diversified
reviews from dierent groups which can help improve diversification results.
3. We perform experiments on real data crawled from e-commerce site to
evaluate our algorithm.
The rest of the paper is organized as follows. Section 2 introduces the re-
lated work. Section 3 defines the problem. Section 4 gives the algorithms to the
problem. Section 5 presents the experimental evaluation. Section 6 reaches the
conclusion.

2 Related Work
Our work is related to the existing work that addresses the problem of review
selection, review assessment and ranking and review summarization. A sequence
of recent work has focused on these problems.
Review Selection: The review selection problem is to select a set of reviews
from a review collection. Lappas and Gunopulos [10] proposed to select a set
of reviews that represent the majority opinions on attributes. The drawback of
 Selecting a Diversified Set of Reviews 723

this approach is that it reduces the diversity of opinions, regardless of the fact
that users tend to make purchase decision after viewing dierent opinions on an
item. Tsaparas et al. [14] intended to select a set of reviews that contain at least
one positive and one negative opinion on each attribute. This method fails to
reflect the distribution of opinions in the entire review collection, thus misleading
users. Lappas et al. [9] proposed the selection of a set of reviews that capture the
proportion of opinions in the entire review collection. The shortcoming of this
approach falls in the lack of consideration on the quality of reviews. Furthermore,
the three methods view all the attributes as the same, which may lead to the
overload of attributes.
Review Assessment and Ranking: The review assessment and ranking prob-
lem is to rank reviews according to their estimated quality. Kim et al. [8] trained
a SVM regression system on a variety of features to learn a helpfulness function
and applied it to automatically rank reviews. They also analyzed the importance
of dierent feature classes to capture review helpfulness and found that the most
useful features were the length, unigrams and product rating of a review. Hong
et al. [6] used user preference features to train a binary classifier and a SVM
ranking system. The classifier divides reviews into helpful and helpless reviews,
and the ranking system ranks reviews based on their helpfulness. Their evalu-
ation showed that the user preference features improved the classification and
ranking of reviews, and jointly using textual features proposed by Kim et al.
[8] can achieve further improvement. However, these methods suer from low
coverage of attributes as the resulting top-k reviews may contain redundant in-
formation. It is in that they estimated the review quality separately and did not
consider the overall quality of the top-k reviews.
Review Summarization: The review summarization problem is to extract
attributes and opinions from the review collection and summarize the overall
opinion on each attribute. Hu and Liu [7] proposed a three-step approach: first
extract attributes from reviews, then identify opinion sentences and whether
they are positive or negative, and finally summarize the number of positive and
negative opinion sentences for each attribute. Yu et al. [12] intended to pre-
dict rating for each attribute from the overall rating, and extract representative
phrases. Although these methods extract useful information from the review col-
lection, they fail to provide the context of reviews from which users can view
the opinions in a more comprehensive and objective way.

3 Problem Definition

In this section, we formally define our problem and the properties of our problem.
Given an item i I, where I denotes a set of items in a certain category, let
R = {r1 , r2 , . . . , rn } be a set of reviews commenting on i which cover a set of at-
tributes A = {a1 , a2 , . . . , am } and express a set of opinions O = {o1 , o2 , . . . , oz }.
We assume that reviewers may express two types of opinions on each attribute:
positive or negative opinion. Hence, the size of O is z = 2m. We also say that a
724 W. Yu et al.

review r covers a subset of attributes Ar A and expresses a subset of opinions

Or O.
Our task is to select a small set of reviews that have high quality, while at
the same time cover important attributes and have dierent opinions. We now
formalize our problem as shown in Problem 1.
Problem 1 (Diversified Review Selection Problem). Given a set of reviews R of
an item i that cover a set of attributes A and express a set of opinions O, find
a subset of reviews S R of size |S| k that maximize the overall value of S:

F (S) = (1 ) Q (S) + D (S) (1)

where Q (S) measures the quality of S, D (S) rewards the diversity of S, k is an
integer number, and [0, 1] is a diversity factor. The definition of Q (S) and
D (S) are given in the rest of the section.

3.1 The Quality of Review Set

Since the overall quality of a review set is impacted by the quality of every single
review in the set, we regard the quality of a review set as the average quality of
reviews in the set. Therefore, the review set quality function Q (S) is defined as
follows: #
q (r)
Q (S) = rS (2)
|S|
where q (r) [0, 1] measures the quality of review r.
Assessment of Review Quality. Motivated by the helpfulness vote mecha-
nism adopted in e-commerce sites, we view the quality of a review as its propor-
tion of helpfulness votes. Therefore, the review quality function q (r) is defined
as follows:
q + (r)
q (r) = + (3)
q (r) + q (r)
where q + (r) (or q (r)) represents the number of users that consider helpful (or
helpless) the review r.
However, some reviews such as newly-written reviews have few votes, which
cannot properly estimate the quality of reviews, so instead of computing q (r)
directly, we use SVM regression to learn the function q (r). It takes the following
two steps to finish the task.
Step 1: define features impacting review quality. We consider two types of fea-
tures: textual features [8] and user preference features [6], as shown in [6] that the
combination of the two feature classes can improve the assessment. The textual
features include the length and unigrams of a review, while the user preference
features include the coverage of important attributes and the divergence from
mainstream viewpoint of a review. The user preference features are defined as
follows.
 Selecting a Diversified Set of Reviews 725

Attribute Coverage: Covering as many attributes as possible is not enough,

users pay more attention to those key attributes, so we consider the coverage of
important attributes. Let Ar A be a set of attributes covered in review r, the
coverage of important attributes cov (r) is defined as follows:
#
w (a)
cov (r) = # aAr
(4)
a A w (a )

where w (a) (0, 1) represents the weight of attribute a. We will introduce the
attribute weight later.
Mainstream Divergence: Users tend to consider helpful the reviews that fol-
low the mainstream, so we measure the gap between the viewpoint of a review
and the mainstream viewpoint. In e-commerce sites, users are required to give
an item an overall rating ranging from one to five along with their reviews, so we
can say that the overall rating expresses the same viewpoint as the review, and
the average rating reflects the mainstream viewpoint. Therefore, the divergence
from mainstream viewpoint div (r) is defined as follows:
8 " 8
8 8
8v (r) r R v(r ) 8
8 |R| 8
div (r) = (5)
maxr R v (r )
"
v (r )
where v (r) {1, 2, . . . , 5} is the viewpoint of review r, r R
|R| is the main-
stream viewpoint, and maxr R v (r ) is the maximum rating.
Step 2: construct training data for SVM regression. We randomly choose reviews
with a certain amount of votes and estimate their quality according to q (r) so
as to form a labeled dataset of {r, q (r)} pairs. We employ the labeled dataset
to train our regression model and use the model to estimate review quality.
Assignment of Attribute Weight. Users read reviews to acquire information
on attributes. However, the review collection of an item may cover dozens of
attributes, while users are only interested in part of them. In order to provide
key information to users, we measure the importance of attributes and assign
relevant weights.
According to user behaviors of review-writing, we regard attributes that are
frequently referred in the reviews of an item and items in the same category as
important attributes. Here, attributes that appear in the reviews of most items
in a category are taken into consideration since they are common attributes
for a category. When users are selecting items of this category, they are more
likely to refer to those attributes first. Therefore, let Ra R be a set of reviews
commenting on an item i I that cover attribute a, where I denotes a set of
items in a certain category, and Ia I be a set of items that have reviews
covering a, the attribute weight w (a) is defined as follows:
726 W. Yu et al.

|Ra | |Ia |
w (a) = (6)
maxa A |Ra | maxa A |Ia |
where max |R a|
measures the importance of a for an item, and max |I a|
a A |Ra | a A |Ia |
measures the importance of a for a category.
The definition of w (a) is similar to that of TF-IDF, as an attribute can be
viewed as a term, and the review collection of an item as a document. Hence, the
importance of an attribute for an item is equal to TF, while the importance of an
attribute for a category is equal to DF instead of IDF since we value attributes
that are frequently referred in the reviews of items in the same category.

3.2 The Diversity of Review Set

Since each time we select one review to add into the review set, we regard the
diversity of a review set as the sum of the diversity of each review in the set
when it is selected. Therefore, the review set diversity function D (S) is defined
as follows: "
D (S) = d (r) (7)
rS

where d (r) rewards the diversity of review r.

Diversification of Opinions. To diversify opinions in a review set, we take

the following three-step method.
In step 1, we cluster reviews with similar concerns. We assume that there
are l clusters, and randomly select l reviews to each initialize a cluster. Since
there may exist opinion sparsity in the review collection, we represent a review
as a vector of TF-IDFs for the unigrams in the review. We compute the cosine
distance between two review vectors as the similarity between them. Each time a
review is added into a cluster based on the similarity between the review vector
and the cluster centroid vector, and the cluster centroid vector will be updated
accordingly.
In step 2, we decide the number of reviews selected from each cluster. We
compute the proportion of each cluster in the review collection as the propor-
tion of reviews from dierent clusters in the selected review set. Therefore, the
proportion of reviews from cluster Ci in the selected review set is defined as
follows:
|Ci |
p (Ci ) = (8)
|R|
Given an integer k as the size of the selected review set, the number of reviews
selected from each cluster is defined as follows:

n (Ci ) = k p (Ci ) + 1/2 (9)

In step 3, we use MMR [1] to select reviews from dierent clusters. That is,
in addition to the assumption that review r has not been added into S, it is in
 Selecting a Diversified Set of Reviews 727

cluster Ci , and the number of reviews selected from Ci is less than n (Ci ), select
r if it is similar with Ci , while at the same time not similar with S. Therefore,
the review diversity function d (r) is defined as follows:

d (r) = sim (r, Ci ) (1 ) sim (r, S) (10)

where [0, 1] is a trade-o coecient.
Here, let Or O (or OR O) be the set of opinions expressed by review r
(or review set R ), the review similarity sim (r, R ) is defined as follows:
#
oO O w (ao )
sim (r, R ) = # r R (11)
o OR w (ao )

where ao represents the attribute corresponding to opinion o.

4 Review Selection Algorithm

Algorithm 1. Diversified Review Selection Algorithm

Input: a set of reviews R = {r1 , r2 , . . . , rn };
a set of attributes A = {a1 , a2 , . . . , am }; a set of opinions
O = {o1 , o2 , . . . , oz }; an integer k
Output: a subset of reviews S R of size |S| k
Review Clustering
1: randomly select {r1 , r2 , . . . , rl } R
2: for i = 1, . . . , l do
3: Ci = {ri }
4: end for
5: for all r R\ {r1 , r2 , . . . , rl } do
6: cluster r
7: end for
8: for i = 1, . . . , l do
9: compute p (Ci ), n (Ci )
10: end for
Review Selection
1: S =
2: for i = 1, . . . , k do
3: r = argmaxrR\S,rCj,n(Cj )>0 F (S {r})
4: S = S {r}
5: n (Cj ) = n (Cj ) 1
6: end for
7: return S

In this section, we give the algorithms to our problem including the algo-
rithms for opinion extraction and review selection. Before we select reviews, we
apply opinion extraction algorithm that extracts attributes and opinions from
the original review collection to generate input for review selection algorithm.
728 W. Yu et al.

4.1 Opinion Extraction

There are a number of researches [7][5][12] on extracting attributes and mining

opinions from reviews. Our method is lexicon-based as well as rule-based to
handle context-dependent opinion words. Given a review, it takes the following
three steps.
First, extract opinion phrases. We define an opinion phrase {a, o} as a pair of
attribute a and opinion word o. We refer to the approach proposed in [13]. They
employ the grammatical relations produced by Stanford Dependency Parser [4]
to define a set of dependency patterns, and generate opinion phrases according
to the patterns.
Second, identify opinions expressed by opinion words in opinion phrases.
There are two types of opinions: positive or negative opinion. Opinions are de-
cided by opinion word lists [7], including a list of positive words and a list of
negative words. If opinions cannot be determined according to the opinion word
lists, we apply intra-sentence rule and inter-sentence rule in [5].
Intra-Sentence Rule: given an opinion word o and the sentence s where o
is, identify its opinion by the other opinion words in s. This rule is based on the
fact that if there are conjunctions such as and and but near o in s, we can
decide its opinion by the opinion word o on the other side of the conjunction.
Inter-Sentence Rule: given an opinion word o and the sentence s where o
is, if s (or the next sentence to s) begins with a word like but or however,
it means o has dierent opinion from the opinion word in the last sentence (or
the next sentence), otherwise they have the same opinion.
Finally, summarize opinions of attributes. Since an attribute may appear
several times in a review, we aggregate all the opinions to produce an overall
opinion. Given an attribute, if the number of positive opinions is larger than
that of negative opinions, the attribute gains a positive opinion, or otherwise it
gains a negative opinion.

4.2 Review Selection

We use a greedy algorithm to implement review selection as shown in Algo-
rithm 1. It first clusters reviews according to their concerns, and then selects
reviews proportionally from dierent clusters to maximize the overall value of
the selected review set.

5 Experimental Evaluation

In this section, we evaluate our algorithm for review selection. We set the
size of the selected review set as 5. We apply the diversity factor with
{0, 0.1, . . . , 0.9} to observe the eects on the quality and diversity of the selected
review set with the increase on .
 Selecting a Diversified Set of Reviews 729

5.1 Datasets

In our experiments, we use data crawled from e-commerce site eBay.com. This
dataset includes reviews from three categories: tablets and e-book readers, portable
audio and digital cameras. There are 4789 items and 110480 reviews in all. Af-
ter pruning items having less than 20 reviews, there are 121 items on tablets
and e-book readers with 9555 reviews, 286 items on portable audio with 38799
reviews, and 583 items on digital cameras with 42783 reviews.

5.2 Experiment Setup

In assessment of review quality, we select reviews having at least 5 votes as

labeled data and use LIBSVM [2] to train regression model. To measure our
algorithm for quality assessment, we select 60% of the labeled data as training
data and the rest as test data to predict review quality. We adopt Pearson
correlation coecient as the metric to the correlation between the estimated
quality and the labeled quality, so as to find whether the quality function is well
learned. According to the evaluation in [8], the resulting coecient is around 0.5,
so we set 0.5 as the borderline between well-learned quality functions and poorly-
learned quality functions. Our algorithm results in a coecient of 0.6405749,
which is a little higher than 0.5. In diversification of opinions, since the size of
the selected review set is small, we set the number of clusters as 3 to ensure the
selection of reviews from each cluster. We set the trade-o coecient as 0.7.

5.3 Result Analysis

We are interested in how quality behaves when raising diversity factor from 0.1
up to 0.9, we hypothesize that the quality will reduce with the increase on . On
the other hand, we will focus on whether attributes and opinions are diversified
through our algorithm. In addition, we will compare our algorithm with other
algorithms for review selection in both the quality and diversity of the selected
review set.

Review Set Quality Analysis. We apply our algorithm on all three categories
of items and compute the average quality of all the selected review sets. As shown
in Fig. 1, the quality goes down smoothly when raising by 10% each time, which
agrees with our hypothesis, and reveals that the diversification of review set have
detrimental eects on the quality.

Review Set Diversity Analysis. To decide whether attributes and opinions

are diversified through our algorithm, we compute the average opinion coverage
and the average cluster coverage of all the selected review sets.
The opinion coverage includes opinion coverage and weighted opinion cover-
age. Let OS O be the set of opinions expressed by the selected review set
S, the opinion coverage is defined as: covop (S) = |O S|
|O| . The weighted opinion
730 W. Yu et al.

Fig. 1. Average Quality for Increasing

"
w(ao )
S O
coverage is defined as: covwop (S) = "oO , where ao represents the at-
o O w(ao )
tribute corresponding to opinion o. The opinion coverage measures the diversity
of opinions while the weighted opinion coverage measures the coverage of impor-
tant attributes.
The cluster coverage includes cluster coverage and weighted cluster coverage,
the definitions of which are similar with that of covop and covwop , except that
O in the denominator is replaced by OCi O, the set of opinions expressed by
cluster Ci , and after computing the cluster coverage for each cluster, use the
average cluster coverage as the result.
As depicted in Fig. 2 and Fig. 3, the opinion coverage grows rapidly at first,
peaking at 31.5% when is 0.3, while the weighted opinion rises up to around
70% and remain so afterwards. Hence, with the introduction of diversity factor
, the opinion coverage improves.
In Fig. 4 and Fig. 5, the trend of cluster coverage is in line with that of opinion
coverage, while the weighted cluster coverage goes up faster than the weighted
opinion coverage in the first step. Hence, the diversity factor enhances the
cluster coverage as well.
In conclusion, the attributes and opinions of the selected review set are diver-
sified through our algorithm.

Comparison of Algorithms. We compare our algorithm with the charac-

teristic method proposed in [9] that selects a set of reviews that capture the
proportion of opinions in the review collection of an item. We implement the
characteristic method with a greedy algorithm.
As presented in Review Set Diversity Analysis, the diversity of the selected
review set reaches its peak when is 0.3, so we set as 0.3, and compare the
average quality QLTY, the average (weighted) opinion coverage (W)OP-COV
and the average (weighted) cluster coverage (W)C-COV of all the selected
review sets produced by the two algorithms. We will also consider the average
error of the proportion of opinions PROP-ERR defined in [9].
In Table 1, our algorithm outperforms the characteristic algorithm in all as-
pects except PROP-ERR. Since the characteristic algorithm aims at reflecting
the distribution of opinions in the review collection of an item, it is as expected
 Selecting a Diversified Set of Reviews 731

Fig. 2. Average Opinion Coverage Fig. 3. Average Weighted Opinion

for Increasing Coverage for Increasing

Fig. 4. Average Cluster Coverage Fig. 5. Average Weighted Cluster

for Increasing Coverage for Increasing

to perform better than our algorithm in PROP-ERR. However, since the size
of the selected review set is small and the characteristic algorithm diversifies the
selected review set by maintaining the proportion of opinions in the original re-
view collection, some rarely mentioned attributes may not be covered, resulting
in the low coverage of opinions and important attributes as well.
In our work, we define important attributes as those attributes that are fre-
quently referred in the reviews of an item and items in the same category. Some
attributes may only appear in a small fraction of reviews of an item, but they are
likely to be important for the category where the item belongs to. For example,
the reviews selected by our algorithm for Canon EOS 1D 4.2 MP Digital
SLR Camera cover not only those common attributes such as:
picture quality: . . . It delivers excellent image quality, with great colors
and detail. . .
shooting: . . . I also like the 6.5fps shooting, which is handy for my wildlife
shooting. . .
price: . . . This is one of the best cameras out there for the price in my
opinion. . .
but also attributes that are not frequently mentioned in the reviews of this
item while in fact are of great importance to the category of cameras, such as:
732 W. Yu et al.

screen and interface: . . . What I love in it? Large viewfinder, great ISOs
range, nice speed, large screen and comfortable interface. . .
menu system: . . . The menu system is also very intuitive and easy to learn
and the my menu feature is good for me about 90% of the time and keeps me
from having to go through the whole menu interface. . .
Users may concern about screen, interface and menu system when selecting
cameras as they are key to user experience. Though they are not frequently
referred in the reviews of this item, they are definitely important for the category
of cameras.

Table 1. Comparison between Characteristic Algorithm and Diversified Algorithm

Characteristic Diversified
QLTY 0.86128730 0.95268065
OP-COV 0.10075375 0.31482820
WOP-COV 0.46727788 0.70877750
C-COV 0.17052619 0.37709308
WC-COV 0.60632420 0.80178090
PROP-ERR 0.21768339 1.30870930

6 Conclusion
We have proposed an approach to select a small set of reviews from the review
collection of an item that have high quality, while at the same time have diversi-
fied attributes and opinions. We first train a regression model to estimate review
quality. Then, to cover important attributes, we assign weights to attributes ac-
cording to their importance, and to cover dierent opinions, we cluster reviews
according to their opinions and select reviews proportionally from dierent clus-
ters. Finally, we use a diversity factor to balance the quality and diversity of
the selected review set, and implement our method with a greedy algorithm to
maximize the overall value of the selected review set. The experimental results
show that our algorithm helps diversify the selected review set.

Acknowledgements. The work is partially supported by the Key Program

of National Natural Science Foundation of China (Grant No. 61232002), Na-
tional High Technology Research and Development Program 863 (Grant No.
2012AA011003), National Natural Science Foundation of China (Grant No.
61103039, No.61021004, No.60903014).

References
1. Carbonell, J., Goldstein, J.: The use of mmr, diversity-based reranking for reorder-
ing documents and producing summaries. In: Proceedings of the 21st Annual In-
ternational ACM SIGIR Conference on Research and Development in Information
Retrieval, pp. 335336. ACM (1998)
 Selecting a Diversified Set of Reviews 733

2. Chang, C.C., Lin, C.J.: Libsvm: a library for support vector machines. ACM Trans-
actions on Intelligent Systems and Technology (TIST) 2(3), 27 (2011)
3. Danescu-Niculescu-Mizil, C., Kossinets, G., Kleinberg, J., Lee, L.: How opinions are
received by online communities: a case study on amazon. com helpfulness votes.
In: Proceedings of the 18th International Conference on World Wide Web, pp.
141150. ACM (2009)
4. De Marnee, M., MacCartney, B., Manning, C.: Generating typed dependency
parses from phrase structure parses. In: Proceedings of LREC, vol. 6, pp. 449454
(2006)
5. Ding, X., Liu, B., Yu, P.S.: A holistic lexicon-based approach to opinion mining. In:
Proceedings of the International Conference on Web Search and Web Data Mining,
pp. 231240. ACM (2008)
6. Hong, Y., Lu, J., Yao, J., Zhu, Q., Zhou, G.: What reviews are satisfactory: novel
features for automatic helpfulness voting. In: Proceedings of the 35th International
ACM SIGIR Conference on Research and Development in Information Retrieval,
pp. 495504. ACM (2012)
7. Hu, M., Liu, B.: Mining and summarizing customer reviews. In: Proceedings of the
Tenth ACM SIGKDD International Conference on Knowledge Discovery and Data
Mining, pp. 168177. ACM (2004)
8. Kim, S.M., Pantel, P., Chklovski, T., Pennacchiotti, M.: Automatically assessing
review helpfulness. In: Proceedings of the 2006 Conference on Empirical Methods
in Natural Language Processing, pp. 423430 (2006)
9. Lappas, T., Crovella, M., Terzi, E.: Selecting a characteristic set of reviews. In:
Proceedings of the 18th ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining, pp. 832840. ACM (2012)
10. Lappas, T., Gunopulos, D.: Ecient confident search in large review corpora. Ma-
chine Learning and Knowledge Discovery in Databases, 195210 (2010)
11. Liu, J., Cao, Y., Lin, C., Huang, Y., Zhou, M.: Low-quality product review detec-
tion in opinion summarization. In: Proceedings of the 2007 Joint Conference on
Empirical Methods in Natural Language Processing and Computational Natural
Language Learning (EMNLP-CoNLL), pp. 334342 (2007)
12. Lu, Y., Zhai, C.X., Sundaresan, N.: Rated aspect summarization of short com-
ments. In: Proceedings of the 18th International Conference on World Wide Web,
pp. 131140. ACM (2009)
13. Moghaddam, S., Ester, M.: On the design of lda models for aspect-based opinion
mining (2012)
14. Tsaparas, P., Ntoulas, A., Terzi, E.: Selecting a comprehensive set of reviews. In:
Proceedings of the 17th ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining, pp. 168176. ACM (2011)
15. Tsur, O., Rappoport, A.: Revrank: A fully unsupervised algorithm for selecting
the most helpful book reviews. In: International AAAI Conference on Weblogs and
Social Media (2009)
 Detecting Community Structures
in Microblogs from Behavioral Interactions

Ping Zhang1, Kun Yue1,*, Jin Li2, Xiaodong Fu3, and Weiyi Liu1
1
Department of Computer Science and Engineering, School of Information Science
and Engineering, Yunnan University, Kunming, China
2
Department of Software Engineering, School of Software,
Yunnan University, Kunming, China
3
Faculty of Information Engineering and Automation, Kunming University of Science
and Technology, 650500, Kunming, China
[email protected]

Abstract. Microblogs have created a fast growing social network on the Inter-
net and community detection is one of the subjects of great interest microblogs
network analysis. The follow relationships highlighted in many existing me-
thods are not sufficient to capture the actual relationship strength between users,
which leads to imprecise communities of microblog users. In this paper, we
presented the graph-based method for modeling microblog users and corres-
pondingly proposed a new metric for obtaining the relationship quantitatively
based on interaction activities. We then gave a hierarchical algorithm for com-
munity detection by incorporating the quantitative relationship strength and
modularity density criterion. Experimental results showed and verified the ef-
fectiveness, applicability and efficiency of our method.

Keywords: Microblog, Community detection, Relationship strength, Graph

model, Modularity density.

1 Introduction

Microblogging is a relatively new phenomenon defined as a form of blogging that

lets you write brief text updates, usually less than 200 characters, about our life on the
go and send them to friends and interested observers via text messaging, instant mes-
saging, email or the web [1]. Provided by several services including Twitter, Sina,
Tencent, etc., microblog has become one of the most popular applications in the real
world. The microblog services with user-generated contents have made a staggering
amount of available online information. In 2010, Twitter generated an estimated 55
million tweets a day [2].
The analysis and the corresponding knowledge discovery of microblog data have
received more and more attention from government, business, sociology etc. with
the increased volume of microblogging data generated by more and more users.

*
Corresponding author.

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 734745, 2013.
Springer-Verlag Berlin Heidelberg 2013
 Detecting Community Structures in Microblogs from Behavioral Interactions 735

For example, many companies adopt microblogging social media services as ideal
places to obtain timely opinions from customers and potential customers directly.
Among the various perspectives of knowledge discover in microblog related applica-
tions, the analysis of the so-called community structure of online social networks has
acquired much attention during the latest years. For instance, Guan et al. [3] found
that it is likely to stimulate new purchases effectively by targeting a product on poten-
tial users of a community that has a significant number of users who have already
purchased the product.
Naturally, in a specific physical domain or logical range, the users with communi-
cations of microblog messages can be described by a network, from which their
associations can be represented globally. To identify the community structures in
microblogs could help us understand and exploit these networks more effectively,
centered on which various studies on community discovery or detection were con-
ducted [4, 5]. For this purpose, many scientists cluster users by the follow relation-
ships [1, 6] based on the hypothesis that microblog users are likely to follow (or be
followed by) other similar ones if their relationships of message communication could
be reflected in a community. Intuitively, by comparing the lists of followers of the
users in small communities, large communities can be generated by merging the small
ones that share a certain number of users according to a pre-defined threshold. How-
ever, the follow relationships actually do not reflect the strength of real relation-
ships from weak to strong. Thus, these methods [1, 6] might group the users into a
community mistakenly, since the close or remote relationships not only depend on the
number of associations, but also on the strength of these associations.
Consequently, some sociologists [7, 8] proposed the method for measuring the re-
lationship strength based on communication reciprocity and interaction frequency. By
this way, the strength was predicted by analyzing the behavioral interaction data [9,
10], but just two levels of strength, namely weak and strong, were considered. There-
fore, from the community discovery point of view, it is desired to develop a strength
metric to describe the relationships both qualitatively and quantitatively.
As stated above, the relationships among microblog users can be represented by a
weighted network composed of vertices and weighted edges, on which the relation-
ship strengths can be looked upon as the weight. The property of community structure
is as follows: network nodes are joined together in tightly knit groups with strong
relationship strength, while there are only looser connections with weak relationship
strength between them. Starting from this observation, many clustering algorithms in
social network analysis, such as GN, CNM, have been applied for community discov-
ery on weighted networks [11, 12]. But only the coarse structure of the communities
can be discovered by these algorithms. As the improvement, Li [13] introduced the
concept of modularity density criterion and proved the k-means algorithm which can
optimize the criterion to obtain finer results. Actually, in the clustering algorithms like
k-means, the number of clusters is to be specified in advance, which cannot be made
for the microblog network since the actual topology and the scale of the concerned
microblog users cannot be known clearly.
Therefore, inspired by the basic idea of clustering algorithms and starting from
the relationship strengths of microblog users, we proposed the precise method for
736 P. Zhang et al.

hierarchical community discovery by statistical analysis on the behavioral interactions

while avoiding predefined parameters. More specifically, the main contributions of
this paper can be summarized as follows:

To measure the actual relationships between users by a finer manner, we presented

the graph-based method for modeling the microblog users. Then, we proposed a
new metric based on the interaction activities to obtain the relationship strength
quantitatively, as the generalization of current metric including just weak and
strong states.
To discover the communities precisely, we proposed the concept of subgraph
proximity and gave a hierarchical algorithm by incorporating the quantitative rela-
tionship strength and the modularity density criterion. By this algorithm, we can
discover the communities precisely with high resolution ratio.
To evaluate the feasibility of our idea given in this paper, we implemented the
algorithm and made preliminary experiments. By applying our algorithm to analyz-
ing some classic benchmark data sets, we tested and verified the effectiveness, ap-
plicability and efficiency of our algorithm.

The remainder of this paper is organized as follows: In Section 2, we introduce the

weighted graph model for simulating user relationships. In Section 3, we give the
method for community identification. In Section 4, we show experimental results and
performance analysis. In Section 5, we conclude and discuss future work.

2 Weighted-Graph Based Modeling of User Relationships

Suppose that user i calls user j for wij times and user j calls user i for wji times. What
is the degree of relationship strength between them? In this section, we discuss the
metric to quantify the relationship strength between a given pair of users.
Microblog has two special tweet types for communication among users, called rep-
lies and retweets. In the network of microblog users, the edges correspond to the di-
rected interactions of replies or retweets among users.
A reply is a tweet sent in response to an earlier tweet and it includes the earlier
tweet ID. This type of tweet typically includes @account_name followed by the
new message. All of the users following either the replying user or the replied-to user
can read this tweet. When a tweet is retweeted (abbreviated as RT) by a user, it is
broadcast to the users followers. For an official RT with twitters retweet command,
the tweet is simply reposted without changing the original message. An unofficial RT
is created by retweet users using the RT @account_name_original message nota-
tion, possibly followed by additional text.
It is well known that the more reciprocal and frequent of interaction of a pair, the
closer their relation will be. Therefore, we can simply define the relationship strength
as follows:

Fij Recij * Freij (1)

 Detecting Community Structures in Microblogs from Behavioral Interactions 737

where Fij, Recij, Freij represent the relationship, reciprocity and frequency strength
between i and j respectively. It is worth noting that Fij is computed by adopting the
weights in the directed micoblog behavioral network. The most intuitionistic metric to
quantify the Fij can be computed as follows:

min( wij , w ji )
Fij Recij * Freij ( wij w ji ) (2)
max( wij , w ji )

However, Eq. (2) is not suitable for measuring the relationship strength. For example,
the values of Eq. (2) for different pairs (1, 10) and (1, 1000) are equal. Actually in the
microblog applications, the information on some nodes is always transmitted enorm-
ously by others nodes, but these nodes themselves transmit little from other nodes. This
appears frequently between the nodes corresponding to stars and their fans. Although
the sum of interactions is very high, stars and their fans are generally not very close
from each other. From this standpoint, when Recij is large enough, the value of Fij
should mainly depend on Recij. Thus, the relationship between 1 and 10 should be
higher than that between 1 and 1000.
To make a distinction of the relationships as illustrated above, the increase of Freij
should tend to flat with the increase of total interactions. So, we can compute the fre-
quency strength as Freij* wij w ji , and Eq. (2) can be improved as follows:

min( wij , w ji )
Fij Recij * Freij* wij w ji (3)
max( wij , w ji )

!
! Fig. 1. Microblog network Fig. 2. Quantitative microblog network

By using Eq. (3), we can transform the directed weighted network to an undirected
one while deriving the general relationship strength between each pair of users. For
example, we can transform the directed weighted network shown in Fig. 1 into the
undirected one shown in Fig. 2. In Section 3, we will present our method for commu-
nity detection based on the undirected network.

3 Community Detection in Microblog Networks

A widely used quantitative measure for evaluating the partition of a network is called
modularity (known as Q), which was introduced in [14]. Recently, Fortunato et al. [15]
738 P. Zhang et al.

claimed that modularity contains an intrinsic scale that depends on the total size of
links in the network, and stated that Q-based method only can detect the coarse struc-
ture of the communities. To improve the situation, Li et al. [13] introduced modularity
density (D) which takes both edge and node into account, and was proved to be able
to recognize the various sizes communities in the undirected network. In our algo-
rithm, we exactly take modularity density as the quantitative measure of partition.
First, we introduce some related concepts.

Definition 1. [13] The density of subgraphs is defined as:

L(Vi ,Vi ) L(Vi ,Vi ) (4)

d (Gi )
Vi

where V is the vertex set and E is the edge set of a network G=(V, E), and A is the
adjacent matrix of G. Given the partition of a network G1=(V1, E1), , Gm=(Vm, Em),
where Vi and Ei are the node set and edge set of Gi respectively, for i=1, ... , m
and m is not larger than n. d(Gi) is the density of subgraph Gi. Vi is the number of
vertexes in Vi, L (Vi , Vi ) Aij (representing the total number of edges that
i Vi , j Vi

fall in the subgraph Gi), and L (Vi , Vi ) Aij , where Vi V Vi .

i Vi , j Vi

Given a valid partition of a network G the summation of the density value of all
communities, computed by Eq. (4), should be as large as possible. Li [13] also gave
the definition of modularity density.

Definition 2. [13] Modularity density of G is defined as

m m L (V , V ) L(Vi , Vi )
i i
D d (Gi ) (5)
i 1 i 1 Vi

This measure provides a way to determine whether a certain microscopic description

of the graph is accurate in terms of communities. The larger the value of D, the more
accurate a partition will be. So the problem of community detection can be looked
upon as the problem of finding a partition of a network such that its modularity densi-
ty D is maximized.
Naturally, this can be done by optimizing the quantitative measure or adopting the
quantitative measure into a certain clustering algorithm. The FN algorithm [14] used
the optimized idea to find communities, which is based on the greedy algorithm to
choose a pair of subgraphs that can maximum the increase of Q with respect to the
previous configuration. However, this optimized idea does not suit for the modularity
 Detecting Community Structures in Microblogs from Behavioral Interactions 739

density in Definition 2. The problem is that the subgraph containing a single node has
arbitrary density, so we cannot compute which pairs can maximum D (denoted as
the increase of D). To obtain the communities, instead of optimizing D directly, we
suppose that Ds value of all single node subgraphs is 0, and define the subgraph prox-
imity for merging nodes. Our basic idea is as follows: two subgraphs will be merged
if the combination of them could decrease the proximity by a maximal degree and
make D>0, specified in Definition 3.

Definition 3. The subgraph proximity between Vi and Vj is defined as

2 L(Vi , V j )
proij (6)
L(Vi , Vi ) L(V j ,V j )

The greater the value of proij, the closer the relationship between Gi and Gj. If there is
no connection between Gi and Gj, then proij equals 0, and if Gi and Gj are only con-
nected by each other, then proij equals 1.
In some network, the idea of joining the pair of subgraphs with the maximal prox-
imity may lead to the asymmetry growth: large subgraphs will be more likely to be
merged than small ones. In extreme cases, the whole network may be clustered as just
one community. The reason is that large subgraphs have more nodes, even if they are
linked with others by a very small rate, and the number of edges between them is
also more than those in small ones. For example, w.r.t. the two communities in Fig. 2, the
result of subgraph merging is shown in Fig. 3, where all the nodes will be clustered
together as one community in each time when D is increased. To avoid this situation,
we choose subgraphs randomly during the merging process. According to this idea,
the large communities will have some possibilities that will not be chosen to merge.
Fig. 4 illustrates the random choose, from which the correct result can be obtained.

i j k p o q i j k p o q

Fig. 3. General combination of pairs Fig. 4. Combination of random choose

740 P. Zhang et al.

Algorithm HCD: CommuDis (G)

Input: the relationship network G
Local variables:
C: adjacent matrix of a subgraph of G. The value of
each cell Cij is the number of edges connecting vertices
in subgraph i to those in subgraph j.
V: node set of a subgraph. Vi records the nodes in sub-
graph i. Initialize each node as a subgraph: V[i] node
i.
N: the number of nodes in G
F: termination of array. The value of Fi is the index of
Vi which is not a community. Initialization, all the in-
dexes of the subgraphs in V are added into F. When F is
empty, the algorithm could be stopped.
for i 0 to N1 do
while true do
if F.length = 0 then
return V
end if
qrandom(0, F.length1)
m Fq; Vk
for i 0 to V.length1 do
if im and Cim 0 then
compute Pim and D if merging them
Vk the subgraph Vi that maximizes Pim when D>0
end if
end for
if Vk = then //Vm is already a community, delete the
//mth element from F
continue // select m again
else
delete the kth element from F
Vm Vm + Vk // combine the node set
delete the kth element from V
C[m] C[k] + C[m]
//update the adjacent matrix of a subgraph
delete the kth column and row from C
end if
end while
end for
return V

Algorithm HCD starts from a state in which each vertex is the sole member of one
of n subgraphs. At each step for the randomly chosen subgraph m, joining subgraph i
with the greatest promi, which makes modularity density D increased. The value of D
 Detecting Community Structures in Microblogs from Behavioral Interactions 741

will be maximized, if all the nodes are partitioned correctly. Randomly choosing a
subgraph in F will take O(1) time. Since the joining of a pair of communities between
which there are no edges at all can never result in an increase in D, so computing
promi and D only depends on k (the mean of the node degree of the whole network)
in the worst case. Updating the values of V, F and C totally takes O(n) time. Thus,
each step of the algorithm takes O(k+n) time in the worst case. There is a maximum
of n1 join operations necessary to construct the complete dendrogram and hence the
entire algorithm runs in O(kn+n2) time on a dense graph or O(n2) time on a sparse one..

4 Experimental Results

To evaluate the performance of the proposed method, we made experiments on reso-

lution ratio, precision and efficiency. The resolution ratio was tested to represent the
granularity of discovered communities, which indicates the range of communities that
can be identified. The precision was used to indicate the number of users included in
the communities that are discovered correctly. The efficiency was tested by compar-
ing the execution time of our algorithm with those of other 2 existing ones. All
experiments were run on a PC with 3.0 GHz processor, 2.0 GB memory, running
Windows XP Professional. The codes were written in JAVA.

4.1 Resolution Ratio

Due to the small network can be clearer to illustrate, we tested the resolution ratio of
our result on well-studied small networks. The Zacharys karate club network [16] is
one of the classic studies in social network analysis. It consists of 34 members of a
karate club as nodes and 78 edges representing friendship between members of the
club which was observed over a period of two years. Due to a disagreement between
the clubs administrator and the clubs instructor, the club can be split into 2 small
ones. Fig. 5 shows the Karate club network, where the shapes of the vertices represent
2 groups: the instructor and administrator factions respectively.
Table 1 shows that the results of the different algorithms are reasonable from the
topology of the network. However, only two coarse structures were found by the fam-
ous algorithm based on Q proposed by Clauset, Newman and Moore (CNM) [17]. The
Community 1# and 2# were split into 4 subgraphs by the density-based k-means algo-
rithm [13]. The Community 1# was split into 2 subgraphs by Algorithm HCD we
proposed in this paper. This means that the resolution ratio of the result of Algorithm
HCD is better than that of the CNM algorithm and not far from the density-based k-
means algorithm. Furthermore, we note that the underlined node 10 cannot be as-
signed properly, since it should be assigned to Community 1# or Community 2# di-
rectly from the topology of the network. By the density-based k-means algorithm, 4
factions can be detected perfectly, since the number of groups is known as human
knowledge in advance.
742 P. Zhang et al.

Fig. 5. Karate club network. Square nodes and circle nodes represent the instructors faction
and the administrators faction, respectively. The dotted line represents our partition, and solid
line represents the partition by the method in [13]. Shaded and open nodes represent the parti-
tion by CNM [17].

Table 1. Community structure in the Karate club network

Community 1# Community 2# Community 3# Community 4#

CNM algorithm 3, 10, 4, 13, 8, 14, 34, 30, 27, 24, 28,
2, 22, 18, 20, 1, 7, 32, 25, 26, 29, 33, 9,
6, 17, 5, 11, 12 31, 23, 19, 21, 16 15
k-means algorithm 18, 20, 22, 2, 12, 1, 29, 30, 31, 9, 34, 27, 28,32,26,25,24 11, 5, 6, 17, 7
based on density 3, 8, 13, 14, 4 15, 16, 19, 21, 33,
23, 10
Algorithm HCD 18, 20, 22, 2, 12, 1, 32, 29, 26, 28, 25, 11, 5, 6, 17, 7
10, 3, 8, 13, 14, 4 30, 24, 31, 9, 34, 27,
15, 16, 19, 21, 33, 23

4.2 Precision
We tested the precision of our algorithm on different generated networks with the
community structure from apparent to fuzzy. Our method was applied to the bench-
mark computer-generated networks (denoted FLR) proposed by Lancichinetti et al.
[18]. The networks have a heterogeneous distribution of node degree, edge weight
and community size, which are exactly the features of real microblog networks [1,
19]. Thus, it is generally accepted as a proxy of a real network with community struc-
ture and appropriate to evaluate the performance of the community detection algo-
rithms. FLR allows the users to specify distributions for the community sizes, the
vertex degrees and the edge weights, with exponents , and muw. Users can also
adjust the average ratio to alter the community structure from apparent to fuzzy.
 Detecting Community Structures in Microblogs from Behavioral Interactions 743

To evaluate the precision of the community identification, we used the normalized

mutual information (NMI) [20] to quality the precision of Algorithm HCD. Fig. 6
shows the variation of the normal mutual information obtained by our method on the
benchmark networks, with the mixing parameter from 0.1 to 0.6, k = 20, = 2.5, =
1.5 and muw = 0.1. The 2 curves in Fig. 6 correspond to the 2 different scales of net-
works which have 1000 nodes and 10000 nodes respectively. With the increase of ,
the community structure in the networks becomes more fuzzy and harder to identify.
Comparing the results obtained by Algorithm HCD and CNM, we note that Algo-
rithm HCD is more precise than CNM, and the precision of Algorithm HCD declines
slower than that of CNM. Moreover, as the number of nodes increasing, from 1000 to
10000, the curves of CNM go down more quickly with respect to larger networks. On
the contrary, the rangeability of the curves of Algorithm HCD is smaller when the
size of network gets larger.

Fig. 6. Comparison of the normalized mutual information between Algorithm HCD and the
CNM algorithm [17] on the benchmark computer-generated network

4.3 Efficiency

We tested the efficiency of our algorithm on the generated networks with different
scales by comparing the execution time of Algorithm HCD proposed in this paper and
those of FN and CNM as the improvement of FN. The comparisons were made with
the increase of nodes in the generated networks and shown in Fig. 7. It can be seen

Fig. 7. Execution time of Algorithm HCD, CNM and FN

744 P. Zhang et al.

that the execution time of Algorithm HCD(i.e., CommuDis) is larger than that of
CNM but smaller than that of FN when the network gets larger and larger.
To sum up, we can conclude from the above experimental results that our algo-
rithm for community detection can achieve more precise results than CNM, but is less
efficient than CNM although more efficient than FN. This makes us improve our
algorithm to achieve better scalability inspired by some ideas of the improvement
from FN to CNM, which is our future work.

5 Conclusions and Future Work

In this paper, we used the weighted graph to model microblog users and their relation-
ships by user interaction activities, and gave a hierarchical algorithm which intro-
duced the modularity density D successfully to discover the communities precisely.
From a certain extent, experimental results showed that our method is effective and
efficient for community detection in the microblog networks.
The methods proposed in this paper also leave open some other research issues.
For example, different types of behavioral interactions may have different effects on
relationships, but our method does not make a distinction between them. According to
the fact that different interactions correspond to different weights, it may be possible
to make better predictions on relationships. Actually, the scales of real microblog net-
works are frequent large, so the efficiency and scalability of our method is desired to
be improved. In addition, the interaction in microblog occurs over time gradually, the
features we used in this work have not concerned the dynamical aspects of data.
These issues are exactly our future work.
Acknowledgement. This work was supported by the National Natural Science Foun-
dation of China (Nos. 61063009, 61163003, 61232002), the Ph. D Programs Founda-
tion of Ministry of Education of China (No. 20105301120001), the Natural Science
Foundation of Yunnan province (Nos. 2011FZ013, 2011FB020), and the Yunnan
Provincial Foundation for Leaders of Disciplines in Science and Technology (No.
2012HB004).

References
1. Song, X., Finin, T., Tseng, B.: Why we twitter: understanding microblogging usage and
communities. In: Zhang, H., Spiliopoulou, M., Mobasher, B., Lee Giles, C., McCallum,
A., Nasraoui, O., Srivastava, J., Yen, J. (eds.) Proc. of the 9th WebKDD and 1st SNA-
KDD 2007 Workshop on Web Mining and Social Network Analysis, pp. 5665. ACM
Press (2007)
2. Kavanaugh, A., Fox, E., Sheetz, S., Yang, S., Li, L., Whalen, T., Shoemaker, D., Natsev,
P., Xie, L.: Social Media Use by Government: From the Routine to the Critical. Govern-
ment Information Quarterly 29(4), 480491 (2012)
3. Guan, Z., Wu, J., Zhang, Q., Singh, A., Yan, X.: Assessing and Ranking Structural Corre-
lations in Graphs. In: Sellis, T., Miller, R., Kementsietsidis, A., Velegrakis, Y. (eds.) Proc.
of the 2011 ACM SIGMOD International Conference on Management of Data, pp. 937
948. ACM Press (2011)
 Detecting Community Structures in Microblogs from Behavioral Interactions 745

4. Ferrara, E.: A Large-scale Community Structure Analysis in Facebook. EPJ Data Science
1(9) (2012), doi:10.1140/epjds9
5. Liao, Y., Moshtaghi, M., Han, B., Karunasekera, S., Kotagiri, R., Baldwin, T., Harwood,
A., Pattison, P.: Mining Micro-Blogs: Opportunities and Challenges. Computational Social
Networks, Part 1, 129159 (2013)
6. Enoki, M., Ikawa, Y., Rudy, R.: User Community Reconstruction using Sampled Micro-
blogging Data. In: Mille, A., Gandon, F., Misselis, J., Rabinovich, M., Staab, S. (eds.)
Proc. of the 21st International Conference Companion on World Wide Web, Workshop,
pp. 657660. ACM Press (2012)
7. Burt, R.: Structural Holes: The Social Structure of Competition. Harvard University (1995)
8. Marsden, P.: Core Discussion Networks of Americans. American Sociological Re-
view 52(1), 122131 (1981)
9. Gilbert, E., Karahalios, K.: Predicting tie strength with social media. In: Dan, O., Ken, H.
(eds.) Proc. of the SIGCHI Conference on Human Factors in Computing Systems, pp.
211220. ACM Press (2009)
10. Kahanda, I., Neville, J.: Using transactional information to predict link strength in online
social networks. In: Cohen, W., Nicolov, N., Glance, N. (eds.) Proc. of the Third Interna-
tional AAAI Conference on Weblogs and Social Media, pp. 7481. AAAI Press (2009)
11. Coscia, M., Giannotti, F., Pedreschi, D.: A classification for Community Discovery Me-
thods in Complex Networks. Statistical Analysis and Data Mining 4(5), 512546 (2011)
12. Huang, L., Li, R., Li, Y., Gu, X., Wen, K., Xu, Z.: 1-Graph Based Community Detection
in Online Social Networks. In: Sheng, Q.Z., Wang, G., Jensen, C.S., Xu, G. (eds.) APWeb
2012. LNCS, vol. 7235, pp. 644651. Springer, Heidelberg (2012)
13. Li, Z., Zhang, S., Wang, R., Zhang, X., Chen, L.: Quantitative Function for Community
Detection. Physical Review E 77(3) (2008), doi:10.1103/PhysRevE.77.036109
14. Newman, M.: Fast Algorithm for Detecting Community Structure in Networks. Physical
Review E 69(6) (2004), doi:10.1103/PhysRevE.69.066133
15. Fortunato, S., Barthlemy, M.: Resolution limit in community detection. The National
Academy of Sciences of the United State of America, PNAS 104(1), 3641 (2007)
16. Zachary, W.: An information flow model for conflict and fission in small groups. Anthro-
pological Research 33(4), 452473 (1977)
17. Clauset, A., Newman, M., Moore, C.: Finding community structure in very large networks.
Physical Review E 70(6) (2004), doi:10.1103/PhysRevE.70.066111
18. Lancichinetti, A., Fortunato, S., Radicchi, F.: Benchmark graphs for testing community de-
tection algorithms. Physical Review E 78(4) (2008), doi:10.1103/PhysRevE.78.0461-10
19. Kwak, H., Lee, C., Park, H., Moon, S.: What is twitter, a social network or a news media?
In: Proc. of the 19th International Conference on World Wide Web 2010, pp. 591600.
ACM Press (2010)
20. Danon, L., Daz-Guilera, A., Duch, J., Arenas, A.: Comparing community structure identi-
fication. J Statistical Mechanics 2005(9) (2005), doi:10.1088/1742-5468/2005/09/P09008
Towards a Novel and Timely Search and Discovery
System Using the Real-Time Social Web

Owen Phelan, Kevin McCarthy, and Barry Smyth

CLARITY: Centre for Sensor Web Technologies

School of Computer Science & Informatics, University College Dublin
[email protected]

Abstract. The world of web search is changing. Mainstream search

engines like Google and Bing are adding social signals to conventional
query-based services while social networks like Twitter and Facebook are
adding query-based search to sharing-based services. Our search and dis-
covery system, Yokie, harnesses the wisdom of the crowd of communities
of Twitter users to create indexes of proto-content (or recently shared
content) that is typically not yet indexed by mainstream search engines.
The system includes an architecture [13] for a range of contextual queries
and ranking strategies beyond standard relevance. In this paper, we fo-
cus on evaluating Yokies ability to retrieve timely, relevant and exclusive
results with which users interacted and found useful, compared to other
standard web services.

1 Introduction

Yokie is a novel search and discovery service that harnesses the social content
streams of curated sets of users on Twitter. The system aims to provide timely,
novel, useful and relevant results by streaming content from Twitter. It also
allows for partitioning of the Twitter social graph into topical or arbitrary groups
of users (called Search Parties). While query-based search is at the core of both
Yokie and traditional search services, our system is not designed or conceived
to compete with or replace the latter. Yokie gathers its content from entirely
dierent sources for a start, and this content is indexed in response to real-time
search needs. As a result, the Yokie use-case of returning high-quality, real-time,
breaking or trending content is very dierent from archival search approaches of
traditional search engines.
Our work is motivated by the fact that the nature of online content is chang-
ing. The conventional web of links and authored content is giving ground to a
web of user-generated content, opinions and ratings. At the same time, users
who have previously relied on traditional search as their basic discovery service
are finding more content through sharing on social networks. The advent of the
social web has seen an important shift in the role of users from passive con-
sumers of content to active contributors of content creation and dissemination.

Supported by Science Foundation Ireland Grant No. 07/CE/11147 CLARITY CSET.

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 746757, 2013.
c Springer-Verlag Berlin Heidelberg 2013

 Towards a Novel and Timely Search and Discovery System 747

New technologies and services have emerged that allow users to directly manip-
ulate and contribute diverse web content [7]. The Twitter microblogging service
has, like the web as a whole, grown significantly in recent years [6][12].
While emphasis is placed on simple 140-character messaging, tweets can con-
tain a wealth of contextual metadata relating to the time, location content and
the user profile. Both Twitter and Facebook have become vital link-sharing net-
works [13][4] our previous work has shown that approximately 22% of Twitter
tweets contain a hyperlink string [13]. This prevalence of link sharing on so-
cial networks motivates an alternative approach to sourcing fresh, relevant and
human-selected search engine content, instead of crawling the entire web-sphere.
Also, by focussing on sourcing content from humans rather than the entire web,
we can subdivide the landscape of content sharers into useful groups. This has
numerous benefits, including the opportunity for users to curate their own ad-
hoc lists, algorithmically group topical producers, and to aid in eliminating spam
contamination.
A proof-of-concept prototype of Yokie was described and presented at [13], and
a full system demonstration was presented at [14]. We have since added several
user-feedback features to gain insight into the systems usefulness, and propose
a per-item global interestingness score; both of which are described below. In
this paper, we are particularly interested in evaluating the performance of the
system at providing timely, relevant and useful content. This is done by a live
user trial, and an attempt at capturing comparative data with other established
online systems.

1.1 System Background and Related Work

Information Retrieval and Search systems are rich and well-established research
landscapes. The technique of Web Crawling [1] has been widely lauded and
adopted, however it is computationally expensive and time consuming because
it attempts to operate over the entire web. As such, some of the major research
themes and challenges include introducing social signals for better relevance,
and obtaining better content freshness in their indexes [4]. More recently, eorts
have been made in applying social signals to search and content recommendation
results [10][4]. For example, Dong et al. successfully exploit features of recent
Twitter content to improve the performance of recency-sensitive queries of a
standard portal search engine such as Yahoo [4]. Alternatively, recent work in
retrieval has proven temporally-sensitive microblogs can be ranked and retrieved
with reasonable precision and recall, while ignoring profile data altogether [9].
Yokies content is entirely mined from social streams. Increasingly, web users
online time is spent mostly on social network activity [7]. Many of these services
allow relatively open access to user content streams. It is no surprise then that
many instances of recent literature includes analyses of Twitters real-time data
and social graph, largely with a view to developing an understanding of why
and how people are using such services; see for example [4][5][7][12]. The free
and abundant access to the real-time web has also given rise to applications in
other related fields beyond search. Phelan et al. describe useful recommendation
748 O. Phelan, K. McCarthy, and B. Smyth

techniques for harnessing recently shared content from Twitter for a news dis-
covery and recommendation service [12]. Chen et al. and Bernstein et al. have
explored using the content of messages to recommend topical content [3]. Garcia
et. al. successfully apply microblog data to product recommendation [5].
A main focus of applying social network properties to search systems should be
on the users and contributors. Indeed, a core component of Yokie is the selection
of groups of users for sourcing content (explained further below). Krishnamurthy
et al. demonstrate that the user base of Twitter can be split into definitive classes
based on behaviours and geographical dispersion [6]. They highlight the process
of producing and consuming content based on retweet actions, where users source
and disseminate information through the network. Several services have emerged
that take advantage of the open access to profile data with a view to providing a
score of a users online activities (provided the user has given them permission).
One such service, Klout (http://www.klout.com), aims to provide an engagement
and influence score based on an aggregation of a users activities across a broad
set of online social networks. Other work has focussed on applying user-level
reputation systems to social search contexts [10] and recommender systems [15].
Grouping users can have numerous benefits including the opportunity for users
to curate or generate their own preferred lists of personal or topical sources from
which to receive content. This would, in turn, allow for vertical searching across
multiple domains [2]. This may be favourable compared to a catch-all attempt to
gather all shared content. Previous work by [6][18] postulate a large and active
subgraph of social networks like Twitter exist that exclusively publish spam. As
such, any steps towards algorithmic subdivision of the social graph may need to
consider the potential of irrelevant or spam-laden accounts.

1.2 The Yokie System

To better understand the system and our goals, we must briefly review the
system architecture. Yokie brings together a number of key ideas to provide
an alternative discovery service that emphasizes real-time, trending information
rather than more static and archival content of traditional web search engines.
The end goal is a service that allows users to query to find content from specified
groups of users, with a variety of contextual querying and ranking functions.

Search Parties and Content Acquisition. A key aspect of the system

is the curated list of content sources, what weve termed a Search Party. Yokie
sources its information from topical communities of Twitter users. These commu-
nities are called Search Parties, and are intended to be either manually curated
or automatically created. Links each member share are indexed in Yokie using
the text of the Tweet with which it is shared. The system uses Twitters API to
acquire content as defined by the search parties. These messages are then stored
and indexed using the service described in the following subsection. This compo-
nent also carries out real-time language classification and finds other messages
that contain the same URL so the system can calculate item popularity, and
extracts tags so as to allow collaborative tagging.
 Towards a Novel and Timely Search and Discovery System 749

Keyword + Time-window Query Fields

Query String Date From Date To Peeked Item Metadata
Shows original tweet + sharer, other
Re-rank users who have mentioned URL
Menu

Peeked
Item

= Liked
Item

Fig. 1. Yokies UI includes functions for viewing extra explanatory metadata related to
the item, and unary relevance ratings. This example resultlist is based on links labelled
Instagram in the past week by high-value technology people, ranked by popularity.

Storage and Indexing. This subsystem is responsible for extracting meta-

data regarding the tweets, for instance timestamp data, hashtags (#election2012,
etc.), user profile information, location, etc. as well as the message content it-
self. The main content, the urlID of the URL mentioned in the message, and the
timestamp is pushed to an indexer for storage and querying. In our current imple-
mentation we use Apache Solr (http://lucene.apache.org/solr/) for this. We also
store the remaining extracted metadata in a MongoDB (http://mongodb.org)
NoSQL database for easy retrieval and fast MapReduce functionality.
UI and Querying. The system is presented with a query interface, currently
comprising of a query string field, and two temporal fields, date from and date
to. The system takes as input a query string with an associated time window,
which can be either a natural language query (eg. 1 day ago, now, last
week, etc) or a fixed date (3 March 2013), and then parses these into UNIX
timestamps. The querying subsystem represents user queries, Q, based on the
notation Q=(QueryString, SearchParty, Tmax , Tmin ). The user can click
through to visit a webpage and peek at an item, and drill-down on results to
explore related content such as the original tweet that the URL was shared
with, the time and date it was shared, and the related Tweet mentions (if any).
Each item UI element contains a unary rating like, which users select when they
believe an item is relevant to the query.
750 O. Phelan, K. McCarthy, and B. Smyth

User and Item-Based Result Ranking. Users can rank using typical Rel-
evance, which is vanilla Term Frequency Inverse Document Frequency scoring
(TFxIDF) [16]. However, a set of ranking strategies beyond standard relevance
have been devised using the added contextual metadata of the microblogs. These
include ranking using temporal features such as item age (Newer-First and
Older-First) and Longevity (the size of the time window between the first
and last mentions of the given hyperlink). Popular items can be ranked as we
capture a number of sample tweets that mention the hyperlink (labelled Men-
tions in the UI). Potential Audience is a user-based sum of follower-counts
for each search-party member who have mentioned the link. The premise of this
strategy is high-volume in-links for users can be a sign of high user reputation.
It is easily possible to derive a range of user-based scores as we are specifically
listening to people for content, as opposed to relying on a document graph
that ignores content publishers.
We have recently implemented two new ranking strategies, both of which are
part of our evaluation. Klout is a service that provides users of social networks
an influence score based on user reach, engagement and their ability to drive
other interactions. Using the Klout API, we can gather scores for each user (once
Klout has a score computed for them). It is possible to rank content based on
the original publishers/sharers Klout score. Interestingness, is an attempt to
measure how interesting an item is based on activity of the item on the social
network, and in other Yokie query sessions. Since we are privy to a multitude of
content, context and user interaction data for each item, and much of this data
is dependent on the conditions of the query, we can create a novel interestingness
algorithm. Given the Query Q tuple, Interestingness is defined as:
& ' & '
P op(Ui , Q) |ClkUi |
Int(Ui , Q) = |LkUi | (1)
Lng(Ui , Q) |P kUi |
P op(Ui , Q) and Lng(Ui , Q) are the popularity and longevity of the current item
Ui , given the parameters of the query tuple, (meaning their value is dependent
on the query Tmax and Tmin values). |ClkUi |, |P kUi | and |LkUi | represent
the total number of clicks, peeks and likes for the item, irrespective of the pa-
rameters of Q. These values have a default value of 1 so as to avoid null values
for interestingness of items with no user engagement. This algorithm utilizes the
contextual features of the item, benefits from prior user engagement with the
item, and ignores the profile data of the user. This allows the items score to be
potentially independent of the original sharers explicit profile data.

2 Evaluating Yokie

In the following evaluation of Yokie we consider two main objectives. First, we

explore how users engage and interact with Yokie, paying particular attention
to their engagement with the dierent ranking strategies and their subsequent
click-through rates. Second, we consider the potential for Yokie to succeed where
 Towards a Novel and Timely Search and Discovery System 751

mainstream search engines fall short as an exclusive source of timely or break-

ing content by comparing Yokie usage to Bing Search and Bing News. In
order to explore these evaluation objectives, we have conducted a live user trial
of Yokie, to explore levels of engagement and interactivity. We use the data
generated during this trial (queries and interaction data) as the basis for a com-
parative evaluation between Yokie, Bing Search and Bing News, focussing on the
ability of the former to identify novel and timely content. Broadly, we have three
main objectives we wish to evaluate from these two trials: Objective 1 - Users
engage and interact with user-based and contextually-ranked results. Objective
2(A) - Yokie locates exclusive and timely results compared with a mainstream
Search Engine. Objective 2(B) - Yokie Users engage and interact with timely
and exclusive content, compared with a mainstream Search Engine.

2.1 Open User Trial

We curated a Search Party of 1000 Twitter users whose profiles were des-
ignated as being Technology oriented. These were gathered by mining for
names listed in 10 Technology labelled lists of users on the Listorious
(http://www.listorious.com) Twitter list website. To garner high-quality users,
each unique user found in these lists was assigned a Klout score that we used to
determine the top 1000 users.
The search index was initially seeded by retroactively mining links from these
user profiles and during the course of the evaluation these profiles continued to
be mined for current links. As mentioned previously, the user interface allows
users to perform three kinds of item-based interactions namely click through
(visit the webpage), peek for more metadata, and like using the star-shaped
button beside each item. Users were also encouraged to re-rank results using
the re-ranking UI. While queries retrieved an average of 50 results, users were
presented with a maximum of 30 results at any given time. There was a possibility
of being presented with dierent results when re-ranked, however we did not
consider any particular ranking strategy as a benchmark to compare with others.
Rather, we preferred to gain insight into the performance of the complete system.
Each time the user began a new search session, a random ranking strategy was
chosen so as to eliminate bias towards any individual options.

Summary Statistics. During the course of the two month trial, a total of
365,735 unique hyperlinks were mined from the Twitter profiles of our 1000
technology influencers. The user study attracted 125 unique users who submitted
1100 queries across 327 sessions. These queries led to the retrieval of 44,856
URLs, and in 95% of the cases, had a temporal window Tmin set to now
indicating users interest in up-to-the-minute results. Users interacted with 1061
items 203 of which were click-throughs, and 858 were exploratory peeks.
While we encouraged users to click on the Like button for items, we found it
dicult to elicit meaningful volumes of user feedback during the trial. As such,
we will focus on user engagement of items based on Click-through and Peek
actions. Further dataset and usage statistics are discussed below.
752 O. Phelan, K. McCarthy, and B. Smyth

A: Frequency of User-selected RERANKS per Strategy (Used to Normalize data in B) C: No. of USER INTERACTIONS per Result Position (overall) CLICK PEEK

400 250.0

# User Interactions
Freq. of Selection

300 332 187.5

279 271
200
218 241 125.0
199
100 160 62.5
124
0 0
Longevity Klout Newer F. Older F. Interest. Pot. Audien. Relev. Mentions 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29
Result-list Position
B: Normalized frequency of INTERACTIONS per strategy
CLICK PEEK
D: MEDIAN POSITIONS for USER INTERACTIONS per Strategy CLICK PEEK
Norm. Freq. of User Interactions

50.0
10.0

Median Result Position

9.5
43.9

37.5
7.5

25.0 5.0
26.7

24.8

5
23.4
21.8

4 4 3.5 4 4
17.8

6.5
16.6

12.5 15.1 2.5

14.7

2.5 3 3 3 3.0 3.0

11
10.8
11.6

10.4
10.4
2.0 2.0 2.0
9.5

0 0
Longevity Klout Newer F. Older F. Interesting. Pot. Aud. Relev. Mentions Longevity Klout Newer F. Older F. Interesting. Pot. Aud. Relevance Mentions

Fig. 2. Chart (A) shows frequency of user selected ranking strategies. Charts (B) shows
normalized User Interaction Data for Clicks and Peeks. Charts (C) and (D) show Click
Position Data for User Interactions.

Objective 1 Users Engaged and Interacted with User-Based and

Contextually-Ranked Results. Unlike more routine IR and Web Search eval-
uations which tend to focus on exclusivity of relevance-based rankings of results,
in this work our aim is to explore the utility of dierent ranking strategies. As
shown in Figure 2, users tended to actively test a variety of dierent strate-
gies with a slight preference for strategies such as Interestingness, Relevance
and Mentions. We use this data as a basis for normalizing per strategy user
engagement of items.

Interactions vs. Ranking Strategies. While Relevance was the most fre-
quently selected strategy by users, it performed poorly in terms of normalized
user interactions. As presented in Figure 2 (B) users performed 2.6 times more
click-throughs on items ranked by Newer First, and over twice as many items
ranked by Longevity. Post-trial, we discovered that over 90% of the queries per-
formed had a Tmax of now (i.e. the query time). As such, it is conceivable the
items selected in Newer First results lists were more topical or new to the user,
or contained breaking news. Nevertheless, this is an interesting trend towards
a preference for novel context-sensitive ranking. While Peek engagement of
Relevance ranked items remains low, the trends are starkly dierent compared
to the corresponding frequencies of selected ranking strategies. Users peeked at
more items ranked by Longevity and Klout scores.

Result Positioning. Figure 2(C) represents median positions for Clicks and
Peeks across all strategies. Both interaction methods follow a standard F-pattern
of user interactivity as observed in other search result evaluations [11][8][17]. This
chart presents a clear peak of user interaction for items on average above the 7th
element in the Yokie result-lists. It also illustrates higher preference for users to
perform exploratory peek actions on items instead of visiting the hyperlinks
themselves. Finally, Figure 2(D) presents the median positions per strategy for
 Towards a Novel and Timely Search and Discovery System 753

each interaction method. Here, we can interpret the positioning of items lower
in the list as the system performing well at capturing interactions in the first
few result items. In this case, Potential Audience, Klout, Interestingness and
Mentions gained user click-throughs with a median position of the second item
in the list, while the rest slightly higher. While it can be assumed that a better-
performing engine would present higher-frequency interactions at the earliest
position in the list, we can also argue the benefits of users exploring further
beyond the top-n results. Recency-based (Newer-first) results garnered a higher
median positioning users explored more recent items lower in the list.

Discussion. There are a number of take-home messages that arise from these
results. First, users do appear to be motivated to try a variety of dierent ranking
approaches. There is a tendency to prefer those such as Relevance and Interest-
ingness, at least with relation to their frequency of selection, perhaps because
they align well with conventional Search Engine ranking. However, based on re-
sult engagement, we can see our novel strategies such as Interestingness, Klout
and temporal strategies seem to outperform others.

2.2 Comparative Study with Bing

For every Yokie query performed during the open user trial, we also gathered
search results from Bing Web and Bing News using a vector of the same key-
word queries users submitted to Yokie (Bings APIs dont support contextual
features of the Yokie query tuple). For every Yokie query, we continued to collect
Bings responses every 30 minutes so that we could track how these responses
evolved over the evaluation period. In each case, Bing Web and News returned a
maximum of 50 results, and each result item was stored based on its query, page
title, timestamp of indexing by Bing, item positioning and description. Useful
pieces of metadata returned by the API included a corresponding timestamp
that represented the moment Bing captured or indexed the item, and the result-
list position of an item with respect to that query in Yokie. These two features,
along with the ability to compare and contrast overlapping content with Yokie
allow us to do comparative studies to see if Yokie has gathered exclusive content,
and whether it can present timely results higher up its result lists. Our assump-
tion that the Twitter-sourced content would be newer and easier to index in a
more timely fashion seems obvious, however little has been done to test it before.

Objective 2 (A) Yokie Locates Exclusive and Timely Results Com-

pared with a Mainstream Search Engine. Figure 3 illustrates dataset size
and item overlap between the content captured from Bing Web, Yokie and Bing
News. In the first panel, Figure 3(A), the comparison is based on the entire
datasets of unique items captured by Bing Web and News using Yokies user
trial queries, compared to the URL index of Yokie. In the second panel, Figure
3(B), represents the total coverage of user-interacted items compared to the Bing
datasets. In each case, we pay particular attention to whether overlapping items
appeared earlier in Yokie or Bing Web/Bing News, so we present a breakdown
754 O. Phelan, K. McCarthy, and B. Smyth

Comparisons between OVERALL DATASETS Comparisons between USER-INTERACTED ITEMS (Clicks & Hovers)
A YOKIE vs Bing Web vs Bing News Datasets B YOKIE vs Bing Web vs Bing News Datasets

*scale in relation to
Bing WEB YOKIE dataset size e
Bing NEWS Bing WEB YOKIE
24,662 Bing NEWS
Unique URLs
Ls
s
YOKIE 233,991 All Unique URLs 689 All Unique URLs
Unique URLs Unique Items

365,576 Clicked + Hovered

By Users

Unique URLs

170
Overlapping Items between
2,837
Overlapping Items between
34
Overlapping Ittems between
88
Overlapping Ittems between
YOKIE and BING WEB YOKIE and BING NEWS YOKIE User Item Sele
ections and BING WEB YOKIE User Item Selecctions and BING NEWS

Appeared on average Appeared on average Appeared on average Appeared on average Appeared on average Appeared on average
18.1 Days 17 hours 7 Days 8.18 Days 7 Hours 15 Hours
EARLIER in BING WEB EARLIER in BING NEWS EARLIER in BING WEB EARLIER in BING WEB EARLIER in BING NEWS EARLIER in BING NEWS

2 21
49 21 71
1178

121
1659 13 32 15 65
Users Interacted with Users Interacted with Users Interacted with Users Interacted with
Appeared on average Appeared on average 13 items on average 32 items that were 15 items on average 32 items that were
20 Days 19.5 hours 16 Days TWEETED an average 4.61 Days TWEETED an average
EARLIER in YOKIE EARLIER in YOKIE EARLIER in YOKIE 15.37 Days EARLIER EARLIER in YOKIE 1.88 Days EARLIER

Fig. 3. Yokie dataset comparison with Bing Web Search and Bing News Search. Figure
A shows a breakdown of item overlaps in the Bing Web / Yokie / Bing News datasets.
Figure B shows the overlaps of user-engaged Yokie items and the total Bing datasets,
and also highlight items that appeared earlier in respective datasets.

of the items which appeared earliest in either Yokie or Bing, and the average
time of how early it was in the corresponding dataset.

Overall Dataset Coverage and Timeliness. As illustrated in Figure 3(A),

we found only 170 Yokie results in the entire Bing Web result list representing
less than 0.1% of the total Bing Web items for the same queries. Of those 170
items, 121 (70%) appeared on average 20 days before they appeared in Bing
Web. The remaining 49 items appeared earlier in Bing Web by an average of
18.1 days. When comparing the Yokie set with the arguably more timely of the
comparative systems, Bing News, we see a much higher overlap set with 2837
items almost 1% of the total News dataset. In this case, Yokie garners a
considerably higher rate of coverage 28% more of a share of those overlapping
items earlier than Bing News. We believe that while gaining earlier items can
be important, an overall broader coverage of items trumps Bing News gaining
items on average 2 hours earlier across a smaller share.
Clearly Yokie is maintaining a very dierent type of information index com-
pared to Bing (Web and News). It is less comprehensive than Bing Web, but it
focusses on trending URLs and produces result-sets that are entirely dierent
from those produced by Bing. The results returned by Yokie are exclusive to
Yokie. If they have been indexed by Bing they are rarely retrieved, or swamped
by more conventional results. If the results do come to be recognised by Bing,
this usually happens after a significant latency period which can be measured in
days for Bing Search and tens of hours for Bing News.
 Towards a Novel and Timely Search and Discovery System 755

Item Exclusivity and Position Performance. Our next comparison is aimed

at exploring how well Yokie performs at ranking and retrieving exclusive items
that didnt appear in Bing. During the user trial, Yokie retrieved 44,856 of its
365,735 items from the 1,100 user-performed queries (not illustrated above).
If we compare this set of items with both Bing sets, we see negligible over-
laps 109 items appeared in any Bing Web result list, whereas 1,292 of them
appeared in Bing News result lists. Therefore, Yokies results contained 99.8%
exclusive items compared to Bing Web, and 97.2% compared to the Bing News
set (also of note is Bing News 10 times larger set compared to Bing Web).
A more useful metric is comparing the Bing itemsets with items with which
users interacted. To return your attention to Figure 3(B), we analyze a similar
breakdown of overlaps as before, except here we compare them with overlapping
items that are dated from when the user interacted with them. Also depicted
is another interesting trend of the 689 unique items, only a total of 122
overlapping items. Combined, this says that of all the items with which users
interacted, 82% were exclusively present in Yokies index.

Objective 2 (B) Yokie Users Interacted with Timely and Exclusive

Content. The previous objective was limited to analysing the overall compo-
sition of the datasets. Here, we examine the items explicitly selected by users
from Yokie result lists, as these represent actual indications of user interest in
the content. We separate each overlapping item based on whether its timestamp
was earlier in Yokie or Bing. The timestamp can either represent the moment
the user interacted with it, or the moment when the original item was published
on Twitter (since Yokie live-streams content, conceivably its indexing time also).

Time of User Interaction vs. Bing Indexing. In Figure 3(B), we see that
during the trial, 689 results were interacted with by Yokie users. Once again,
the vast majority were exclusive to Yokie; only 34 items found by Bing web,
and 89 by Bing News. In terms of timeliness, we also see Yokie performing
comparatively well as opposed to Bing Web 13 items were interacted with by
users an average of 16 days earlier than they even appeared in the Bing Web
index. Compare this to Bing Webs larger share of 21 items which appear on
average only a week earlier in the Bing Web index than in Yokies. Yokie is
also seen to fall behind in terms of user engagement and Bing News publication.
While Yokie users had clicked on items that were on average almost 5 days
newer than their appearance in Bing News, this was only represented by 15
items. Bing News had a considerably broader coverage with 71 items with which
Yokie users interacted, however these were in its indexes only 7 hours before
Yokie. Overall, comparing the time of item interactions with Bing index time
shows Yokie under-performing in terms of coverage of items with which its users
interacted, regardless of its good performance gaining content earlier than Bing.

Interacted Items Tweet Time vs. Bing Indexing. Here, we describe a

similar comparison outlined earlier but here we compare like-for-like analyses
between the moment users interacted with an item, and its index time in Bing
756 O. Phelan, K. McCarthy, and B. Smyth

Web and News. As described in Figure 3(B), Yokie users interacted with a total
of 32 items that appeared on average 15.3 days earlier than they appeared in Bing
Web. They also interacted with the majority of the news overlaps on average
1.88 days earlier than they even appeared in Bing News.

Discussion. One could sensibly assume that Bing Web and Bing News would
garner higher coverage and timely content respectively compared to our system.
However, the vast majority of Yokie results were exclusive to the system, apart
from the 170 web items and 2837 news items that are overlapping. We can use
this overlapping data to further support the hypothesis that Yokie is able to
identify these results earlier than Bing with over 70% of Bing Web/Yokie
overlaps appearing almost 20 days earlier in Yokie, and almost 60% of the Bing
News/Yokie coverage appearing nearly a full day beforehand in Yokie.
Once again, the data supports the argument for the potential of a real-time
streaming technique for indexing web content. While Bing had more coverage of
earlier items amongst user-interacted results, Yokies interacted content was not
only vastly exclusive, but also in many cases more timely than Bing. This is a
surprising result considering Bing Web represents a capture of the entire web,
and Bing News represents a real-time capture of topical breaking news.

3 Trial Discussion, Limitations and Conclusions

Our user trial and comparative evaluations have highlighted Yokies potential
for timely, exclusive and relevant content compared with popular web and news
search services. Our approaches to contextual and user-based ranking strategies
also performed well at providing relevant content with which users interacted. In
particular, the temporally sensitive rankings that showed newer content and also
content with high longevity garnered significant interest from users. Our second
trial illustrates the shortcomings of both archival Search Engines and, surpris-
ingly, some reputable online news search services at providing timely content.
This work is not without its limitations. In building Yokie we have designed a
complex and novel information discovery service with many moving parts. This
presented something of an evaluation challenge. Unlike traditional IR systems,
there were no benchmark datasets (e.g. TREC [9], etc). Our approach has been
to attempt a live user trial. Obviously in an ideal world one would welcome the
opportunity to engage a large cohort of users perhaps thousands or tens of
thousands in such an evaluation. Unfortunately this was not possible with the
resources available, and so our trial is limited to a smaller cohort of just 125 users.
Nevertheless these users did help to clarify our primary evaluation objectives and
their usage provided a consistent picture of their common engagement patterns
to show how they interacted with dierent ranking strategies and gained benefit
from the results they provided. Moreover we bolstered this live evaluation with
an additional study to compare Yokie with more conventional search and news
discovery services by harnessing the activity of our live users as the basis for
search queries. Our focus on Bing was purely pragmatic as its APIs Terms and
 Towards a Novel and Timely Search and Discovery System 757

Conditions provided us with access to the data we needed, whereas alternatives

such as Google do not.
Yokies architecture and content source enables us to perform a broad range
of future evaluations. Indeed, a multitude of ranking and retrieval techniques
remain to be explored. Emphasis will be placed on trials in which users may
curate their own Search Parties, and scope exists for these Search Parties to be
collaboratively built. While the findings in this paper validate the systems merits
and novelties, significantly more work can be done with the general retrieval
aspects, such as the reasoning behind the comparatively lower positioning of
user interesting items compared to Bing, comparative studies between the user-
selected rankings and randomly assigned ranking strategies, and so on.
References
1. Brin, S., Page, L.: The anatomy of a large-scale hypertextual web search engine.
Comput. Netw. ISDN Syst. 30, 107117 (1998)
2. Chau, M.: Comparison of three vertical search spiders. IEEE Computer 36 (2003)
3. Chen, J., Geyer, W., Dugan, C., Muller, M., Guy, I.: Short and tweet: experiments
on recommending content from information streams. In: CHI 2010. ACM (2010)
4. Dong, A., Zhang, R., Kolari, P., Bai, J., Diaz, F., Chang, Y., Zheng, Z., Zha, H.:
Time is of the essence: improving recency ranking using twitter data. In: WWW
2010, ACM, New York (2010)
5. Esparza, S.G., OMahony, M.P., Smyth, B.: On the real-time web as a source of
recommendation knowledge. In: RecSys 2010, September 26-30. ACM (2010)
6. Krishnamurthy, B., Gill, P., Arlitt, M.: A few chirps about twitter. In: WOSP 2008,
NY, USA. ACM (2008)
7. Kwak, H., Lee, C., Park, H., Moon, S.: What is twitter, a social network or a news
media? In: WWW 2010, pp. 591600 (2010)
8. Lewandowski, D.: A framework for evaluating the retrieval eectiveness of search
engines. In: Next Generation Search Engines (2012)
9. Macdonald, C., Ounis, I., Lin, J., Soboro, I.: Microblog track. In: TREC (2011)
10. McNally, K., OMahony, M.P., Smyth, B., Coyle, M., Briggs, P.: Towards a
reputation-based model of social web search. In: IUI 2010. ACM (2010)
11. Nielsen, J., Pernice, K.: Eye tracking research into web usability,
http://www.useit.com/eyetracking/ (accessed May 2012)
12. Phelan, O., McCarthy, K., Bennett, M., Smyth, B.: Terms of a feather: Content-
based news recommendation and discovery using twitter. In: Clough, P., Foley,
C., Gurrin, C., Jones, G.J.F., Kraaij, W., Lee, H., Mudoch, V. (eds.) ECIR 2011.
LNCS, vol. 6611, pp. 448459. Springer, Heidelberg (2011)
13. Phelan, O., McCarthy, K., Smyth, B.: Yokie - a curated, real-time search and
discovery system using twitter. In: 2nd RSWEB Workshop, RecSys 2011 (2011)
14. Phelan, O., McCarthy, K., Smyth, B.: Yokie: explorations in curated real-time
search & discovery using twitter. In: ACM RecSys 2012. ACM (2012)
15. Resnick, P., Kuwabara, K., Zeckhauser, R., Friedman, E.: Reputation systems.
Commun. ACM 43, 4548 (2000)
16. Sebastiani, F.: Machine learning in automated text categorization. ACM Comput.
Surv. 34, 147 (2002)
17. Smyth, B., Balfe, E., Boydell, O., Bradley, K., Briggs, P., Coyle, M., Freyne, J.: A
live-user evaluation of collaborative web search. In: IJCAI (2005)
18. Yardi, S., Romero, D., Schoenebeck, G., Boyd, D.: Detecting spam in a Twitter
network. First Monday 15(1) (January 2010)
 GWMF: Gradient Weighted Matrix Factorisation
for Recommender Systems

Nipa Chowdhury and Xiongcai Cai

School of Computer Science and Engineering,

The University of New South Wales, Sydney, NSW 2052, Australia
{nipac,xcai}@cse.unsw.edu.au

Abstract. In this paper, we developed a new algorithm, Gradient Weighted Ma-

trix Factorisation (GWMF), for matrix factorisation. GWMF uses weights to
focus the approximation in the matrix factorisation to the higher approxmation
residual. Therefore, it improves the matrix factorisation accuracy and increases
the speed of convergence. We also introduce a regularisation parameter to con-
trol overfitting. We applied our algorithm to a movie recommendation problem
and GWMF performs better than ordinal gradient descent-based matrix factori-
sation (GMF) on Movielens dataset. GWMF converges faster than GMF and it
guarantees lower root mean square error (RMSE) at earlier iterations on both
training and testing.

Keywords: Weighted gradient descent, Matrix factorisation, Recommendation.

1 Introduction

The population of the Internet is increasing day by day. Now 50 billion webpages are
available on the Internet and the size of the Internet is about to double in every 5
years[1, 2]. This gigantic size of the Internet creates some problems too. When people
search in the Internet, they are now experiencing overwhelmed with lots of informa-
tion. A little portion of the information is relevant to the person's interest. So a re-
commender system is needed to guide the user when searching such a vast space.
Recommender systems suggest items or persons to a user based on the user's taste so
that the user can quickly reach to the point of his interest without going through irre-
levant information. Furthermore, a recommendation list itself may be large enough
and usually the user considers only the top few items. So it is a challenging problem
in recommender systems to rank top items accurately.
Recommender systems are mainly divided into two categories: content-based fil-
tering and collaborative filtering [3]. In recommender systems, an item is recom-
mended to an active user based on 1) similarity of the item with other items in which
the active user has already shown some preferences, or 2) similarity of users to the
active user, whom active user has similar preference on some set of items. In content-
based filtering, the similarity is based on the item or user profile. For example, in a
book recommendation, a book is recommended to a user based on the books profile

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 758769, 2013.
Springer-Verlag Berlin Heidelberg 2013
 GWMF: Gradient Weighted Matrix Factorisation for Recommender Systems 759

similarity (book type, writer and so on) with other books that were already rated high-
ly by that user [3-5]. Rather than computing similarity from profile, collaborative
filtering uses users choice on items to derive similarity. Collaborative filtering (CF) is
widely used in social network and also in online web [3-5]. As the size of the Internet
is increasing day by day, scalability becomes an important issue and matrix factorisa-
tion is the ultimate solution to handle this scalability problem [6].
In matrix factorisation (MF), data are placed into a matrix, which indicates user in-
teraction over item. For example, R RNM is the rating matrix represents preferences
on N users over M items. The major purpose of matrix factorisation in this context is
to obtain some form of lower-rank approximation to R for understanding how the
users relate to the items [7]. User preferences over item are modelled by user's latent
factors and item preferences by user are modelled by item's latent factors. These fac-
tor matrices are learned from observed preferences in such a way that their scalar
product equals to original rating matrix [8]. For example, the target matrix R RNM
is approximated by two low rank matrix P RNK and Q RMK

R P.Qt (1)

Here K represents the dimensionality of the approximation. After learning, these la-
tent factors are used to predict unobserved ratings. Gradient descent algorithm [8] is
widely used for learning these latent factors.
In this paper, we present a new algorithm for learning latent factors. Our main con-
tributions are as follows. Firstly, we develop a new gradient descent based matrix
factorisation algorithm, namely GWMF, by using weights to update factor matrix in
each steps to learn latent factors. We also introduce a regularisation parameter to re-
duce overfitting. Result shows our algorithm achieves converge faster with higher
accuracy than the ordinal gradient descent based matrix factorisation algorithm.

2 Literature Survey

For its scalability and efficiency, matrix factorisation has become most popular me-
thod in recommendation. A lot of research was done for further improvement of ac-
curacy and reducing computational time.
In [6] a probabilistic matrix frame (PMF) is proposed to approximate the user fea-
ture vector and movie feature vector from user movie rating matrix. The work ad-
dresses two crucial issues of recommender system: i) complexity in terms of observed
rating and ii) sparsity of ratings. With a cost function by a probabilistic linear model
with Gaussian noise, this method achieves 7% performance improvement over Netflix
own system [17]. Authors in [9] identify drawback of their previous PMF models [6]
as PMF is computationally complicated. To solve the problem, they propose Bayesian
Probabilistic Matrix Factorisation (BPMF) where user and movie features have no
zero mean which need to be modeled by some hyper-parameters. Therefore, the pre-
dictive distribution of the rating is approximated by Monte Carlo algorithm. When the
number of features is large, BPMF outperforms PMF in terms of time complexity.
760 N. Chowdhury and X. Cai

A hybrid approach is used in [10] to learn incomplete rating from a rating matrix.
Here they enforce non-negativity constrain to model matrix factorisation. At first,
Expectation Maximization algorithm is performed several iterations to select initial
values of user and item matrix. These initial values are used by Weighted Non-
negative Matrix Factorisations (WNMF). Here they use weights as indicator matrix.
For example, a weight equals to one if rating is observable or zero if unobservable.
Authors in [11] extend WNMF in [10] to Graph Regularized Weighted Nonnegative
Matrix Factorisation (GWNMF) to approximate low rank matrices of user and item.
They use both user and item's internal information, external information and also
social trust information. Experimental result shows that these internal and external
information increase recommendation accuracy.
Clustering is applied with Orthogonal Non-negative Matrix Factorisation
(ONMTF) in [12]. ONMTF is applied to decompose a user-item rating matrix into a
user matrix (U) and an item matrix (SVT). By using U and SVT, user/item clusters
and their centroids are identified. To predict unknown rating of a user to an item, at
first similarity between the test user and item is calculated and similar user/item clus-
ters are identified. Within the identified clusters, K most similar neighbours are found.
The linear combination model based on user-based approach, item-based approach
and ONMTF is applied to calculate the test user's rating on the item. Although expe-
rimental result does not show any improvement, their method requires less computa-
tional cost.
An ensemble of collaborative filters is applied in [13] to predict ratings on Netflix
dataset. As MF algorithm is sensitive to local minima, the author applies different MF
algorithms with different initialisation parameters several times. And then the mean of
their prediction is used to compute final ratings.
Authors in [14] develop non-negative matrix factorisation algorithm where they
put additional constraint that all elements of the factor matrices must be non-negative.
They develop two algorithms by using multiplicative update steps, one attempts to
minimise the least squares error and the other to minimises the Kullback-Leibler di-
vergence. Both of the algorithms guarantee locally optimal solutions and are of easy
implementations.
To the best of our knowledge, there is no research done by using user or item
weights with latent factor vectors to approximate ratings. Our algorithm, GWMF
gives larger weight values to an item and user with high approximation error, i.e.
latent factors not approximated the rating between the user and item well in the cur-
rent iteration, so that in next gradient step the algorithm can focus on those items and
users and give more update on them. We also introduce a regularisation parameter to
reduce overfitting. Result shows that GWMF guaranteed lower training and test error
and it achieves minimum RMSE faster than gradient descent.
The rest of the paper is arranged as follows: in next section we present gradient
descent-based matrix factorisation algorithm (GMF) and our gradient weighted matrix
factorisation algorithm (GWMF). Experimental settings and result of GWMF vs.
GMF is presented in section 4. Finally we draw conclusions and future work.
 GWMF: Gradient Weighted Matrix Factorisation for Recommender Systems 761

3 Algorithm

3.1 Gradient Descent Algorithm for Matrix Factorisation

Given a rating matrix R and mean square error (MSE) as optimisation function, Gra-
dient descent based matrix factorisation aims to find two low rank matrix P and Q,
which minimise the following objective.
N M
GMF= ( R P.Qt )2ij
i=1 j=1 (2)

The algorithm is presented in Algorithm 1.

Algorithm 1: GMF

Input: Rating Matrix R, learning rate , regularisation parameter , no. of itera-

tions steps, dimension of low rank matrix K.
Output: Factor matrix P and Q.

1 Initialise P and Q randomly.

2 for s=1:steps do
3 for i=1:N do
4 for j=1:M do
5 if R(i,j)>0
5.1 er R(i,j)-(P(i,:)*Q(:,j));
5.2 P(i,:) P(i,:)+( *er*2*Q(:,j)'-( * *P(i,:)));
5.3 Q(:,j) Q(:,j)+( *er*2*P(i,:)'-( * *Q(:,j)));
end for
end for
end for
6 return P,Q
Algorithm 1. Gradient Descent Algorithm for Matrix Factorisation

As shown in Algorihtm 1, GMF treats each element equally, i.e. assign equal learn-
ing rate for each element, in Q when update P and each equally in P when update Q.

3.2 Gradient Weighted Algorithm for Matrix Factorisation

In GMF, a rating R(i,j) is approximated by the product of the corresponding factor
vectors P(i,:) and Q(:,j). The approximation can be formulated as a learning problem
for each row R(i,:):
R(i,:) = P(i,:) * Q
When we fix Q and solve P(i,:), in the above equation, each rating R(i,j) is a target of
an instance, each corresponding column Q(:,j) is the feature value vector of that
762 N. Chowdhury and X. Cai

instance, and P(i,:) is the linear regression model parameter. Similarly, for each col-
umn R(:,j), we have:
R(:,j) = P*Q(:,j)
where we can fix matrix P and solve the parameter vector Q(:, j) for a linear regres-
sion model. In such a configuration, the matrix factorlisation is formulated as a set of
linear regression models: one for each row of P and one for each column of Q.
One of the common assumptions, underlying most linear and nonlinear least
squares regression, is that each data instance provides equally precise information
about the deterministic part of the total process variation. More specific, the standard
deviation of the error term is constant over all values of the predictor or explanatory
variables. However, this assumption does not hold, even approximately, in every
modelling application. In such situations, weighted least squares can be used to
maximise the efficiency of parameter estimation [18].
To this end, giving each instance its proper amount of influence over the parameter
estimates may improve the performance of the approximation since a method that
treats all of the data instance equally would give less precisely measured instance
more influence than they should have and would give highly precise points too little
influence.
Therefore, in GWMF, we use weights to optimise approximation of the target ma-
trix. The idea behind this is, if a user or item is not approximated well in previous
iterations, the algorithm focuses on reducing its errors and thus which contributes
more to the update than others. In GWMF we give more weight to instances that are
not approximated well by previous iterations. We use two weight matrices 1) user
weight WP RNM that indicates user approximation over item and item weight
WQ RNM that indicates item approximation by user in current iterations. At initiali-
sation all user weight values for a rated item and all rated item weight values for a
user are equal (for example as shown in Fig. 1(a)). Then at each gradient steps this
weight values keep tracks how error of rating matrix to factor matrix i.e. (R-PQ) is
changed for each observed user and item. We also enlarge this effect by exponential
cost function for weight updates, which results in large weight values for instances
which have higher error. Weight update formula of WP is given below. WQ follows
the same update formula. Furthermore, we also normalise the weight values by max-
imum normalisation.

WP = e|( R PQ )|
Therefore, the algorithm converges to the training objective quickly. But in real
world, there may be some outliers or noise in datasets. So this weight update formula
gives more weights to those instances constantly and thus could make the algorithm
lose generalisation ability, i.e. overfitting. In this concern, we introduce a parameter
called rbeta inspired by [15] which takes (R-rbeta*PQ) for weight update. So the
weight update formula becomes:

WP = e|( R rbeta PQ )| (3)

 GWMF: Gradient Weighted Matrix Factorisation for Recommender Systems 763

From Fig. 1 we see how weight values of WP and WQ changes according to equation 3 in
each gradient steps. After the 1st iteration (Fig. 1(b)) the weight values of each user-item
vary by a large amount. So more gradient update is performed to user-items that have high
weight values than others. When the algorithm converges, user weight values for an item
and item weight values for a user become equal. At 1000 iterations (Fig. 1(c)) and 2000
iterations (Fig.1(d)) training errors become very small, i.e. similar, and weight values of
user item become 1. The GWMF algorithm is given in Algorithm 2.
Algorithm 2: GWMF

Input: Rating Matrix R, learning rate , regularisation parameter , no. of itera-

tions steps, dimension of low rank matrix K.
Output: Factor matrix P and Q.

1. Initialise P and Q randomly. Initialise WP=1/M and WQ=1/N.

2. for s=1:steps do
3 for i=1:N do
4 for j=1:M do
5 if R(i,j)>0
5.1 er R(i,j)-(P(i,:)*Q(:,j));
5.2 P(i,:) P(i,:)+( *er*2*WP(i,j)*Q(:,j)'-( * *WQ(i,j)*P(i,:)));
5.3 Q(:,j) Q(:,j)+( *er*2*WQ(i,j)*P(i,:)'-( * *WP(i,j)*Q(:,j)));
end for
end for
end for
6 return P,Q

Algorithm 2. Gradient Weighted Algorithm for Matrix Factorisation

4 Results
To evaluate performance of GWMF, we use two datasets: (i) Small Dataset that consists
of first 100-by-50(user-by-item) rating for training from Movielens u1.base dataset[16].
Testing is done by those users-items rating from u1.test dataset. (ii) The Movielens Data-
set: which consists of 100,000 ratings from 1000 users on 1700 movies. We perform
experiments separately on each five Movielens dataset and average the results. We com-
pare the proposed GWMF with the standard GMF algorithm. For fair comparison, we
need to make learning rates of these two algorithms equal. To achieve this, we set learn-
ing rates of two algorithms in two different ways. For the smaller dataset, we divide
learning rate of GWMF by initialisation weight values. Here initialisation weight values
means if a user rated 5 items, all items receive equal weight values 1/5=0.2. This is be-
cause in GWMF we multiplied weight with errors so it decreases the learning rate. To
make it the same as GMF, we divide learning rate of GWMF again by initialisation
weight. Here initialisation weight values remain the same for all iterations whereas other
weight values (i.e multiplied WP and WQ) change according to GWMF. In the settings
for large Movielens dataset, we keep learning rate of GWMF unchanged but multiplied
764 N. Chowdhury and X. Cai

5 3 ? 1 0.33 0.33 0 0.33 0.25 0.33 0 0.2

4 ? ? 1 0.5 0 0 0.5 0.25 0 0 0.2
1 1 ? 5 0.33 0.33 0 0.33 0.25 0.33 0 0.2
1 ? ? 4 0.5 0 0 0.5 0.25 0 0 0.2
? 1 5 4 0 0.33 0.33 0.33 0 0.33 1 0.2
R WP WQ

(a)

0.31 0.86 0.53 0.46 1 0.08 0 0.02 1 1 0 0.03

0.55 1.71 0.85 0.89 1 0 0 0.04 0.29 0 0 0.02
0.45 1.38 0.59 1.03 0.03 0.03 0 1 0.02 0.17 0 1
0.25 0.84 0.33 0.43 0.06 0 0 1 0.02 0 0 0.67
0.36 1.23 0.47 0.67 0 0.01 1 0.3 0 0.15 1 0.53
P*Q WP WQ

(b)

4.2 2.23 3.27 1.8 1 0.98 0 0.98 1 0.85 0 1

3.42 2 2.91 1.5 1 0 0 0.97 0.8 0 0 0.79
1.79 1.2 3.05 4.3 1 0.56 0 0.96 0.99 0.48 0 0.96
1.66 1 2.89 3.4 1 0 0 0.97 0.87 0 0 0.86
2.95 1.94 4.21 4 0 1 1 0.41 0 1 1 0.48
P*Q WP WQ

(c)

5.05 2.71 3.91 1 0.78 1 0 0.78 0.96 0.88 0 0.97

3.92 2.1 3.37 1 1 0 0 0.94 0.99 0 0 0.95
1.09 0.74 3.65 4.9 0.84 1 0 0.82 1 0.86 0 0.99
1.02 0.65 3.39 3.9 0.95 0 0 1 0.93 0 0 1
2.48 1.41 4.81 4.1 0 1 0.8 0.71 0 1 1 1
P*Q WP WQ

(d)

Fig. 1. Rating matrix R and initialization weight values of WP and WQ is shown in (a). PQ,
WP, WQ is shown after the 1st (b), after 1000 (c) and after 2000 (d) iterations.

weight value with learning rate of GMF. So this rate is also decreased and learning rate
of GWMF and GMF are comparable.
RMSE is used to measure performance. For the small dataset, as the size of our
training set is smaller and thus more sparse, rating prediction is more challenging than
that of the original Movielens dataset.
To see how weights work on matrix factorisation, we perform two experiments on
small dataset. In one experiment we use one weight matrix to examine how instances
are approximated in previous iterations and give more weight to instances that are not
approximated well. In other experiment, we use two weight matrices: the user weight
 GWMF: Gradient Weighted Matrix Factorisation for Recommender Systems 765

matrix WP which indicates how items are approximated for a user in previous itera-
tions and the item weight matrix WQ which indicates how users are approximated
for an item. The idea behind this experiment is: the user factor (item factor) is multip-
lied by rated item factors (user factor) in matrix factorisation. So these two weight
matrices give an idea about how well users and items are approximated in previous
iterations. RMSE of two weight matrices vs. one weight matrix on smaller version of
Movielens dataset is given in Table 1. In Fig. 2, RMSE on test dataset is shown.

Table 1. RMSE of two weight matrix vs. one weight matrix in training and testing

Iterations MF with one weight matrix MF with two weight matrices

Train Test Train Test
10 1.0155 1.05 1.04 1.0772
20 0.8918 0.9676 0.9106 0.9879
50 0.7082 0.8765 0.7216 0.8892
100 0.5306 0.8245 0.5549 0.8403
120 0.4762 0.8142 0.5044 0.8306
200 0.3258 0.797 0.3563 0.8117
250 0.2681 0.7934 0.2967 0.8074
500 0.125 0.7843 0.1499 0.7959
1000 0.0473 0.7715 0.0562 0.7777
2000 0.0326 0.757 0.033 0.7583
5000 0.0279 0.7515 0.0282 0.7481
10000 0.026 0.7632 0.0262 0.7592
20000 0.0251 0.781 0.0252 0.7773

1.1

MF by gradient descent
MF by weighted gradient descent
1.05

0.95
RMSE

0.9

0.85

0.8

0.75

0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

No. of iterations x 10
4

Fig. 2. RMSE of two weight matrices vs. one weight matrix on the small dataset
766 N. Chowdhury and X. Cai

From Fig. 2 we see that MF by two weight matrices performs better that one
weight matrix. In this concern we use two weight matrices for the rest of our experi-
ment. Next we present performance of user parameter rbeta in our algorithm. When
rbeta=1, it means that it takes PQ's opinion completely for next iterations and when
rbeta=0 it results unique weight update for all iterations. For this small dataset, we
achieve minimum RMSE for rbeta=0.5. Here in Fig. 3 we present performance of
various rbeta values on GWMF with two weight matrices.

1.2
rbeta=1
1.15 rbeta=.7
rbeta=0.5
rbeta=0.2
1.1
rbeta=0

1.05

1
RMSE

0.95

0.9

0.85

0.8

0.75

0.7
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
No. of iterations x 10
4

Fig. 3. RMSE of various rbeta values on GWMF with two weight matrices on small dataset

We also compare our algorithm to gradient descent and the result is presented in
Fig. 4 and 5 on the smaller dataset and Fig. 6 and 7 for the Movielens dataset. We set
rbeta=1 for both dataset and learning rate=.0002 for smaller dataset and 0.01 for Mo-
vielens dataset.
1.4

MF by gradient descent
MF by weighted gradient descent
1.2

0.8
RMSE

0.6

0.4

0.2

0
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
No. of iterations

Fig. 4. MF by gradient descent vs. MF by weighted gradient descent on the training set (Small).
 GWMF: Gradient Weighted Matrix Factorisation for Recommender Systems 767

1.4

MF by gradient descent
1.3
MF by weighted gradient descent

1.2

1.1
RMSE

0.9

0.8

0.7
0 1 2 3 4 5 6 7 8 9 10
No. of iterations x 10
4

Fig. 5. MF by gradient descent vs. MF by weighted gradient descent on the test set (Small).

1.2

MF by gradient descent
MF by weighted gradient descent
1.1

0.9
RMSE

0.8

0.7

0.6

0.5
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
No. of iterations

Fig. 6. MF by gradient descent vs. MF by weighted gradient descent on the training set (Movielens)

From Fig. 4 and 5, we see that GWMF achieves lower RMSE on earlier iterations than GMF
but higher RMSE than GMF at the rest of the training phase. But in testing (Fig. 5), GWMF
achieves lower RMSE (0.7334) which is 10 times earlier than GMF.
768 N. Chowdhury and X. Cai

1.15

MF by gradient descent
1.1
MF by weighted gradient descent

1.05

1
RMSE

0.95

0.9

0.85

0.8

0.75
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
No. of iterations

Fig. 7. MF by gradient descent vs. MF by weighted gradient descent on the test set ( Movielens)

For earlier iterations, it shows GWMF achieves lower RMSE and in 5000 itera-
tions GWMF achieves its lowest RMSE. However, GMF reaches its lowest RMSE at
50000 iterations that is still larger than the lowest RMSE of GWMF.
In Movielens dataset, GWMF also outperforms GMF (Fig. 6 and 7). On test set
GWMF achieves the lowest RMSE in 100 iterations (0.75) whereas GMF achieves
the lowest RMSE at 2000 iterations (0.76) that is also higher than GWMF. So our
GWMF performs better than GMF in terms of accuracy and computational cost in
both Movielens dataset.

5 Conclusion

Although Gradient descent is the widely used techniques in matrix factorisation, it has
slow convergence. Researchers use multiplicative update steps or other techniques to re-
solve this problem. However, to the best of our knowledge, there is no research was done
to optimise matrix factorisation by weighted gradient descent. GWMF is the first work in
this direction, where we use weights directly in gradients update steps. As gradient descent
moves forward to objective function in each step, putting extra weights is very challeng-
ing. There is a risk to destroy gradient descents convergence property. But using an appro-
priate cost function and regularisation parameter, we achieve this and our results are very
encouraging. In future, we plan to apply this algorithm on larger datasets and other prob-
lems where gradient descent algorithm is also applied.

Acknowledgement. This project is funded by the Smart Services CRC under the
Australian Government's Cooperative Research Centre program.
 GWMF: Gradient Weighted Matrix Factorisation for Recommender Systems 769

References
1. The size of the World Wide Web, http://www.worldwidewebsize.com/
2. The Size of Internet to Double Every 5 Years,
http://www.labnol.org/internet/
internet-size-to-double-every-5-years/6569/
3. Adomavicius, G., Tuzhilin, A.: Toward the Next Generation of Recommender Systems: A
Survey of the State-of-the-Art and Possible Extensions. In: IEEE Transaction of Know-
ledge and Data Engineering, pp. 734749 (2005)
4. Pazzani, M., Billsus, D.: Learning and Revising User Profiles: The identification of Inter-
esting web sites. J. Mac. Lea. 27, 313331 (1997)
5. Jannach, D., Friedrich, G.: Tutorial: Recommender Systems. In: Joint Conference on Ar-
tificial Intelligence, Barcelona (2011)
6. Salakhutdinov, R., Mnih, A.: Probabilistic Matrix Factorization. In: Neural Information
Processing Systems (2008)
7. Hubert, L., Meulman, J., Heiser, W.: Two Purposes for Matrix Factorization: A Historical
Appraisal. Society for Industrial and Applied Mathematics 42, 6882 (2000)
8. Koren, Y., Bell, R.M., Volinsky, C.: Matrix Factorization Techniques for Recommender
Systems. IEEE Computer 42(8), 3037 (2009)
9. Salakhutdinov, R., Mnih, A.: Bayesian Probabilistic Matrix Factorization using Markov
Chain Monte Carlo. In: 25th International Conference on Machine Learning (ICML), Hel-
sinki (2008)
10. Zhang, S., Wang, W., Ford, J., Makedon, F.: Learning from Incomplete Ratings Using
Non-negative Matrix Factorization. In: Proceedings of the SIAM International Conference
on Data Mining, SDM (2006)
11. Gu, A., Zhou, J., Ding, C.: Collaborative Filtering: Weighted Nonnegative Matrix Factori-
zation Incorporating User and Item Graphs. In: 10th SIAM International Conference on
Data Mining (SDM), USA (2010)
12. Chen, G., Wang, F., Zhang, C.: Collaborative Filtering Using Orthogonal Nonnegative
Matrix Trifactorization. J. of Inf. Pro. and Man. 45, 368379 (2009)
13. Wu, M.: Collaborative Filtering via Ensembles of Matrix Factorizations. In: Proceedings
of KDD Cup and Workshop, pp. 4347 (2007)
14. Lee, D.D., Seung, H.S.: Algorithms for Non-Negative Matrix Factorization. In: Advances
in Neural Information Processing, pp. 556562 (2000)
15. Jin, R., Liu, Y., Si, L., Carbonell, J., Hauptmann, A.G.: A New Boosting Algorithm Using
Input-Dependent Regularizer. In: Proceedings of the 20th International Conference on Ma-
chine Learning, Washington (2003)
16. MovieLens Data Sets, http://www.grouplens.org/node/73
17. Netflix Prize, http://www.netflixprize.com/
18. Gentle, J.: 6.8.1 Solutions that Minimize Other Norms of the Residuals. In: Matrix
Algebra. Springer, New York (2007)
 Collaborative Ranking
with Ranking-Based Neighborhood

Chaosheng Fan and Zuoquan Lin

School of Mathematical Sciences, Peking University, Beijing 100871, China

{fcs,lz}@pku.edu.cn

Abstract. Recommendation system is a very important tool to help

users to find what they are interested in on the web. In many commer-
cial recommendation systems, only the top-K items are shown to users,
and recommendation becomes a ranking task rather than a classical rat-
ing prediction task. In this paper, we propose a new collaborative ranking
algorithm based on learning to rank framework in information retrieval.
For a given user-item pair (u, i), we use Kendall Rank Coecient as
similarity metric to choose neighborhood for user u and use the rank-
ing statistical information of item i from user us neighborhood as the
feature representation of pair (u, i). We apply LambdaRank to learn the
ranking model and experimentally demonstrate the eectiveness of our
method by comparing its performance with several collaborative ranking
approaches. Moreover, we can address scenarios where users feedbacks
are non-numerical scores.

Keywords: collaborative ranking, collaborative filtering, recommenda-

tion system.

1 Introduction
Electronic retailers and content providers oer a huge selection of products for
modern consumers. Matching consumers with products they like is very impor-
tant for users satisfaction on the web. Recommendation systems, which provide
personalized recommendation for users, become a very important technology to
help users to filter useless information. Through automatically identifying rele-
vant information by learning users tastes from their behavior data in the system,
recommendation systems greatly improve users experience on the web. Inter-
net leaders like Amazon, Google, Netflix and Yahoo increasingly adopt dierent
kinds of recommendation systems, which generate huge profits for them.
Recommendation systems are usually divided into two classes: content-based
filtering and collaborative filtering. Content-based filtering approaches use the
content information of the users and items (products) to make recommenda-
tions, while collaborative filtering approaches just collect users ratings on items
to make recommendations. Recommendation has been treated as a rating pre-
diction task for a long time. Two successful approaches to rating prediction task
are neighborhood model and latent factor model. Neighborhood models predic-
tion a target users rating on an item by (weighted) averaging the items ratings

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 770781, 2013.
c Springer-Verlag Berlin Heidelberg 2013

 Collaborative Ranking with Ranking-Based Neighborhood 771

of her/his k-nearest neighbors, who have the similar rating pattern with the
target user. Latent factor models transform both items and users to the same
latent factor space and predict a users rating on an item by multiplying the
users latent feature with the items latent feature. The combination of neighbor-
hood model and latent factor model successfully completes the rating prediction
task [1].
However, in many recommendation systems only the top-K items are shown
to users. Recommendation is a ranking problem in the top-K recommenda-
tion situation. Ranking is more about predicting items relative orders rather
than predicting rating on items and it is broadly researched in information re-
trieval. The problem of ranking documents for given queries is called Learning
To Rank(LTR) [2] in information retrieval. If we treat users in recommendation
systems as queries and items as documents, then we can use LTR algorithms
to solve the recommendation problem. A key problem in using LTR models for
recommendation is the lack of features. In information retrieval, explicit features
are extracted from (query, document) pairs. Generally, three kinds of features
can be used: query features, document features, and query-document-dependent
features. In recommendation system, users profiles and items profiles are not
easy to be represented as explicit features. Extracting ecient features for rec-
ommendation systems is emerged and also very important for learning a good
ranking model.
Some work tries to extract features for user-item pairs [3,4] etc. In [3], the
authors extract features for a given (u, i) pair from user us k-nearest neighbors
who rated item i. They use rating-based user-user similarity metric to find neigh-
bors for a target user. However, in some e-commerce systems users preferences
to items are perceived by tracking users actions. A user can search and browse
an item page, bookmark an item, put an item to the shopping cart and purchase
an item. Dierent action indicates dierent preference to the item. For example,
if a user u purchased item i and bookmarked item j, we can assume that user u
prefers item i to item j. Mapping the user actions into a numerical scale is not
natural and trivial. So it is hard to accurately compute the user-user similarity
in e-commerce systems where users feedbacks are non-numerical scores.
In this paper, we argue that using ranking-based neighbors to extract features
for user-item pairs is more suitable than using rating-based neighbors, because we
are facing and completing a ranking task. Choosing neighbors through a ranking-
based user-user similarity metric is natural and much closer to the essence of
the ranking. Neighbors with the similar preference over items with the target
user will give more accurate ranking information on items than rating-based
neighbors. Moreover, We can also easily use ranking-based user-user similarity
metric to find neighbors for a target user even that the users feedbacks are
no-numerical scores. Our contributions in the paper can be summarized as:

1. We use ranking-based neighbors to extract features for user-item pairs and

apply learning to rank algorithm LambdaRank for making recommendations.
Based on the characteristic of the features we used, we significantly improve
top-K recommendation performance.
772 C. Fan and Z. Lin

2. The feature extraction method we used can be applied to recommendation

systems where users feedbacks are non-numerical scores.
The paper is organized as follows: In Section 2 we discuss notations and related
works. The algorithm framework is presented in Section 3. Our experimental
steps and experimental results are discussed in Section 4. In Section 5, we sum-
marize our work and discuss the future work.

2 Notations and Related Work

In this section, we briefly discuss the notations used in the paper and some
related topics.

2.1 Notations
In recommendation system, we are given a set of m users U = {u1 , u2 , ..., um }
and a set of n items I = {i1 , i2 , ..., in }. We also use u, v to represent any user
and use i, j to represent any item for convenience. The users ratings on the
items are represented by an m n matrix R where entry ru,i represents user
us true rating on item i. ru,i represents user us prediction rating on item i. Ru
represents all the ratings of user u.

2.2 Collaborative Filtering

There are two kinds of collaborative filtering algorithms: user-based and item-
based [5]. Given a user-item pair (u, i), user-based neighborhood model estimates
the rating on item i based on the ratings by a set of neighbors who have the sim-
ilar rating pattern with u. Possible choices for user-user similarity su,v include
the Vector Space Similarity (VSS), Adjusted Cosine Similarity (ACS) and Pear-
son Correlation Coecient (PCC). For example, the VSS similarity represents a
user as a vector in a high dimensional vector space based on her/his ratings. The
cosine of the angle between two users ratings is used to measure their similarity:
#
rui rvi
iRu Rv
su,v = U # U # (1)
2
rui 2
rvi
iRu Rv iRu Rv

Since dierent users and items have dierent bias, PCC and ACC are proposed
to relieve the bias problem. An alternative form of the neighborhood-based
approach is the item-based model. Item-item similarity can also be computed
like user-user similarity. Given a user-item pair (u, i), item-based neighborhood
model estimates the rating on item i based on the k most similarity items that
have been rated by u. PCC, ACC and VSS are rating-based similarity metric.
The similarity is computed purely based on rating vector, and it ignores the ex-
plicit preference over items. In [6], the similarity between users is computed by
 Collaborative Ranking with Ranking-Based Neighborhood 773

Kendall Rank Coecient [7] and a ranking neighborhood-based approach is used

to make recommendations. The major problem with the neighborhood-based ap-
proaches is that the rating matrix R is highly sparse, making it very dicult
to find similar neighbors reliably. Matrix Factorization approach is better than
neighborhood-based approaches at minimizing rating prediction error. Matrix
Factorization approximates the observed rating matrix R as the inner product
of two low rank matrices. The prediction rating on item i by user u is equal to
the inner product of the corresponding user and item features. Classical matrix
factorization approach is more about rating prediction, but in many recommen-
dation systems only the top-K items are shown to user and ranking becomes
more important than rating prediction.

2.3 Collaborative Ranking

We will introduce ranking algorithms in recommendation system in this part. In
[8], by considering the missing entries as zeros, the authors perform conventional
SVD on sparse matrix R and make top-K recommendation based on the value
of the test items in the reconstruct matrix. Overfitting on training dataset is
avoided by only keeping high singular values. The algorithm shows good per-
formance (precision/recall) than SVD++ [1], which is a famous model based on
matrix factorization.
Some collaborative ranking algorithms are proposed based on LTR and matrix
factorization. Ordrec [9] is an ordinal rank algorithm based on ordinal regression
and SVD++. It models users rating via SVD++ and aims at minimizing an or-
dinal regression loss. It brings the advantage to estimate the confidence level in
each individual prediction and can also handle users non-numerical feedbacks.
In [10], the authors present a generic optimization criterion (Bpropt) for person-
alized ranking that is the maximum posterior estimator derived from a Bayesian
analysis of the ranking problem. Bpropt optimizes the measure of the Area Un-
der the ROC Curve (AUC) based on matrix factorization and adaptive KNN. It
uses stochastic gradient descent with bootstrap sampling to update parameters
and converges very fast. PMF-co [4] uses pairwise learning to rank algorithm
to solve recommendation problem, both user-item features and the weights of
the ranking function are optimized during learning. ListRankMF [11] aims at
minimizing the cross entropy between the predict item permutation probability
and true item permutation probability. In [6], the authors propose a probabilistic
latent preference analysis model for ranking prediction by directly modeling user
preferences by a mixture distribution based on Bradley-Terry model. Some col-
laborative ranking work tries to directly optimize evaluation measures. TFMAP
[12] uses tensor factorization to model implicit feedback data with contextual in-
formation, and directly maximizes mean average precision under a given context.
In CLiMF [13] , the model parameters are learned by directly maximizing the
Mean Reciprocal Rank and in CofiRank [14], the authors fit a maximum margin
matrix factorization model to minimize the upper bound of NDCG. In [3], the
authors extract features for a given (u, i) pair from user us k-nearest neighbors
who have rated item i. They use rating-based user-user similarity metric (e.g.
774 C. Fan and Z. Lin

Vector Space Similarity) to find neighbors for a target user. Their work has
some drawbacks. First, they can not well define user-user similarity when users
feedbacks are not numerical scores. Second, they choose neighbors by rating-
based user-user similarity metric, which is not corresponding to the ultimate
goal of ranking. Our work is dierent from their work. We choose neighbors
with similar preference pattern rather than rating pattern. The problem with
rating oriented approaches is that the focus has been placed on approximating
the ratings rather than the rankings, but ranking is a more important goal for
modern recommender systems.

3 Ranking-Based Collaborative Ranking Framework

In this section, we will discuss our ranking-based collaborative ranking frame-
work. Since users and items are not easy to be represented in term of explicit
features, we should design new method to extract ecient feature for ranking.
We extract features from neighbors of a target user to capture the local infor-
mation from the neighborhood.

3.1 Neighborhood Models

In some movie recommendation systems, users rate movies with non-numerical
values, e.g., A, B, ..., F . There is no direct resemblance this kind of feedback to
numerical or binary values. Furthermore, we argue that even when the feedback
is related to absolute numbers, taking the scores as numerical may not reflect
the users intentions well. Dierent users tend to have dierent internal scales.
For example, taking star ratings as numeric will put the same distance between
3 stars and both 4 stars and 2 stars. However, one user can take 3 stars as similar
to 4 stars, while another user strongly relates 3 stars to low quality, being similar
to 1-2 stars. Computing user-user similarity is dicult in this scene. Moreover,
the rating-oriented neighborhood models focus on rating prediction rather than
ranking. Since we should complete a collaborative ranking task, we prefer a
ranking-based user-user similarity metric to choose neighbors for users. In the
ranking-oriented approach, two users with similar preferences over the items are
more close. The preference is reflected by their ranking of the items. We modify
the Kendall Tau Correlation Coecient [6,7] to qualify the user-user similarity.
We define the user-user similarity as:
Nuv
su,v = 1 (2)
Muv
where
|Ru Rv | (|Ru Rv | 1)
Muv =
2
"
Nuv = I (ru,i ru,j , rv,i rv,j )
i,jIu Iv
 Collaborative Ranking with Ranking-Based Neighborhood 775
4
1 , if x y 0 (x = 0 y = 0)
I (x, y) =
0 , otherwise
The similarity is defined as the ratio between the number of concordant pairs
Muv Nuv and the number of all possible pairs Muv . It is negatively correlated
with the number of dis-concordant pairs Nuv and its value is in the range of
[0, 1]. We can compute the user-user similarity even that the users feedbacks
are non-numerical scores from the definition. From the ranking-based user-user
similarity definition, we can see that if two users assign on four items with ratings
{1, 2, 3, 4} and {4, 3, 2, 1} respectively, their similarity is equal to 0, so they have
totally dierent ranking-based preference. While the user-user cosine similarity
between the two users is 0.77, they will be treated as very similar users. We can
see that ranking-based user-user similarity is very dierent from rating-based
user-user similarity. Choose neighbors from ranking-based neighborhood is more
close to the essence of ranking. Given a target user u, we should carefully choose
neighbors for her/him, and those users who have high rating-based similarity
and low ranking-based similarity with the target user should not be put into
us neighborhood. The ranking-based neighbors can give more significant and
accurate ranking information about items and contribute more information for
feature extraction.

3.2 User-Item Feature Extraction

In this section, we use the ranking-based neighbors to extract the features for
user-item pairs. A feature vector (u, i) for a user-item pair (u, i) is extracted
based on the K nearest ranking-based neighbors K(u, i) = {uk }K k=1 of user u who
have rated i. The features are descriptive statistics information of the ranking
of item i. Every neighbor uk of user u has her/his opinion on the ranking of
item i. For every item j (j = i) in uk s profiles, uk has her/his preference over
item i and item j. We aggregate K neighbors opinions and extract statistics
information from their opinions as features. We summarize the preference for i
into three K 1 matrices Wui , Tui and Lui :
1 #
Wui (k) = I[ruk ,i > ruk ,i ]
|Ruk | 1
i Ruk \i

1 #
Lui (k) = I[ruk ,i < ruk ,i ]
|Ruk | 1
i Ruk \i

1 #
Tui (k) = I[ruk ,i = ruk ,i ]
|Ruk | 1
i Ruk \i

The three matrices describe the relative preference for i across all the neighbors
of u. Then we summarize all matrices with a set of simple descriptive statistics.

1"
(Wui ) = [(Wui ), (Wui ), max(Wui ), min(Wui ), I[Wui (k) = 0]
k
k
776 C. Fan and Z. Lin

1"
(Lui ) = [(Lui ), (Lui ), max(Lui ), min(Lui ), I[Lui (k) = 0]
k
k
1"
(Tui ) = [(Tui ), (Tui ), max(Tui ), min(Tui ), I[Tui (k) = 0]
k
k

where is mean function, is the deviation function, max and min are the
maximize and minimize function. The last term counts the number of neighbors
who express any preference towards i. We convert neighbors ratings on i to pref-
erence descriptive features in this way. After concatenating the three statistics
features we have the last feature representation for (u, i):

(u, i) = [(Wui ), (Lui ), (Tui )] (3)

The features are descriptive statistics information for the ranking of item i.
Since these feature are statistics of neighbors opinions on the ranking of items
and our experiment show that they are very robust. It is obvious that ranking-
based neighbors will give more significant and accurate ranking information than
rating-based neighbors in this feature extraction setting. After we have extracted
features for every user-item pairs, we can use any learning to rank algorithm
to learn a ranking model and make recommendation. We will introduce the
procedure in the next part.

3.3 Learn and Prediction

We use a simple linear function as our prediction function. The output of the pre-
diction function indicates the relative position of items. The prediction function
is:
r(u, i) = w (u, i) + w0 + I[K(u, i) = ]b0 (4)
where w and w0 is the ranking model parameters to be learned (w0 is the oset
term). I[K(u, i) = ] = 1 if there are no neighbors who rate item i (corresponding
to (u, i) = 0). The bias term b0 provides a base score for i if i does not have
enough ratings in the training dataset and we cant find neighbors who have
rated item i. We can apply any LTR algorithm to learn the ranking model, and
we use LambdaRank in our experiment because we evaluated our rank quality
by NDCG. We omit the details of LambdaRank due to the lack of space and
refer the reader to [15] for details description. Given a user u, we predict every
item i in user us test dataset by:

r(u, i) = w (u, i) + w0 + I[K(u, i) = ]b0 (5)

where w and w0 is the learned ranking model. Only ratings in the training
dataset are used to select the neighbors, and make predictions for the items in
the validating and testing dataset. After all unrated items prediction ratings
are computed, we can order them based on the prediction rating value and make
top-K recommendation. We can see that the rating is not on a scale of 1 to 5
anymore and it represents the relative position of the items. Moreover, based
 Collaborative Ranking with Ranking-Based Neighborhood 777

on the ranking-based neighbors statistical feature and the personalized ranking

function we used, we can see that if ru,i > ru,j and ru,j > ru,k , we will have
ru,i > ru,k . The prediction function keeps the symmetric attribute on items, so
we dont need use a greedy algorithm to infer the final ranking list like [6].

4 Experiments
4.1 Evaluation Measure
Since recommendation systems often show several items to the user, a good
ranking mechanism should put relevant impressions in the very top of a ranking
list extremely well. So we choose Normalized Discount Cumulative Gain (NDCG)
[16] metric in information retrieval as the evaluation measure, because it is very
sensitive to the ratings of the highest ranked items, making it highly desirable
for measuring ranking quality in recommendation system. A ranking of the test
items can be represented as a permutation : {1, 2, ..., N } {1, 2, ..., N } where
(i) = p denotes that the ranking of the item i is p and i = 1 (p) is the item
index ranked at p-th position. Items in test dataset are sorted based on their
prediction ratings. The NDCG value for a given user u is computed by:
K
" 2ru,1 (p) 1
1
NDCG(u, )@K = (6)
MDCG(u) p=1 log(p + 1)

We assume that is a permutation over the test items that are sorted on the
true ratings and MDCG@K is DCG@K value corresponding to permutation .
We can see that a perfect ranks NDCG value is equal to one. We set k to be 10
in all our experiments like [11,14,4]1 . We report the macro-average NDCG@K
across all users in our experiment results.

4.2 Datasets
We use two well known movie ratings datasets in our experiment: Movielens100K
and Movielens1M 2 . Most of the state-of-the-art ranking-based collaborative
ranking algorithms also use them. The Movielens100K dataset consists of about
100K ratings made by 943 users on 1682 movies. The Movielens1M dataset
consists of about 1M ratings made by 6040 users on 3706 movies. Since users in
the two datasets rate at least 20 rating, the cold start users wont appear. The
ratings for Movielens100K and Movielens1M are on a scale from 1 to 5. Table 1
gives statistics information associated with the two datasets.
We divide the whole dataset into training dataset and Testing dataset. We
randomly sample a fixed number N of ratings for every user for training test
on the rest data. We sample 10 ratings from training dataset for validating and
1
NDCG values at other cutos give us dierent numerical values but qualitative
almost the same results.
2
http://www.grouplens.org/node/73
778 C. Fan and Z. Lin

Table 1. Datasets Statistics

Dataset Users Items Train 20 Train 30 Train 40 Train 50

Movielens100K 943 1682 744/1682 645/1681 568/1681 497/1677
Movielens1M 6040 3706 5289/3701 4737/3694 4297/3689 3938/3677

after we found the best hyper parameters, we put validating dataset into training
dataset and retrain the model, then we report the rank performance on testing
dataset. In our experiment, we set N as 20, 30, 40 and 50, so users with less than
30, 40, 50 and 60 ratings are removed to ensure that we can compute NDCG@10
on the testing dataset. The number of users and items after filtering for dierent
N are reported in Table 1.

4.3 Parameters and Baselines

We train our algorithm (RankWLT) using user-wise stochastic gradient descent.
We update parameter with learning rate in {1, 0.1, 0.01, 0.001}. In our experi-
ment, the test performance is the best when learning rate is 0.01. Our neigh-
borhood size is chosen in range [10, 500]. We filter neighbors who have very low
similarity with the target user with threshold t = 0.05, so these users wont
appear in the target users neighborhood. When the neighborhood size becomes
too large, no performance gain are observed.
We compare our algorithm with Probability Matrix Factorization (PMF) and
PureSVD [8]. Since TOPK-ListMLE is state-of-the-art learning to rank algo-
rithm, so we also extend TOPK-ListMLE to recommendation systems based on
the framework proposed by [11]. For PureSVD algorithm, the dimension of the
singular value matrix is 10. The user and item latent dimension size D in
TOPK-ListMLE and PMF are chosen from {5, 10, 15, 20, 50} by cross validation
and the regularization term for latent factor are chosen from {1, 0.1, 0.01, 0.001}.
We report the best rank performance of TOPK-ListMLE and the rank per-
formance of PMF aiming at minimizing RMSE. We also report the best rank
performance for PMF aiming at maximizing NDCG@10 (We call it RankPMF
model). The user latent factor and item latent factor in TOPK-ListMLE, PMF
and RankPMF are initialized as a random positive value in range [0, 0.1] (we
find that initializing the latent factors with positive values converges faster than
initializing them with random values).

4.4 Experiment Results

The experiment results are listed in Table 2 and Table 3.
We have some intuitions from the experiment results:
1. NDCG@10 increases when the sample number N increases in most of the
models, because the training dataset size increases and we have more infor-
mation to learn users preference. For example, NDCG@10 of PMF increases
from 0.681 to 0.708 on Movielens100K.
 Collaborative Ranking with Ranking-Based Neighborhood 779

Table 2. NDCG@10 on Movielens100K

Algorithm Train 20 Train 30 Train 40 Train 50

PMF 0.681 0.691 0.696 0.708
RankPMF 0.706 0.705 0.707 0.712
PureSVD 0.646 0.649 0.669 0.662
TOPK-ListMLE 0.710 0.704 0.711 0.709
RankWLT 0.685 0.701 0.716 0.720

Table 3. NDCG@10 on Movielens1M

Algorithm Train 20 Train 30 Train 40 Train 50

PMF 0.729 0.733 0.731 0.737
RankPMF 0.747 0.744 0.745 0.748
PureSVD 0.699 0.704 0.708 0.710
TOPK-ListMLE 0.740 0.745 0.745 0.746
RankWLT 0.749 0.754 0.753 0.756

2. Surprisingly, PureSVD performs non-competitively compared to other meth-

ods on NDCG@10. This is very dierent from the conclusion in [8]. We
think that this is mainly due to the nature of their evaluation metric (preci-
sion/recall). They evaluate precision/recall performance on the test dataset
which is constructed by high ratings (above 5 point) and a large number
of items sampled at random. We prefer our evaluation using NDCG which
avoids random sampling and has very little bias.
3. Since improvements in RMSE (Root Mean Square Error) often dont trans-
late into accuracy improvements, model (PMF) which purses of minimizing
prediction error behaves not well in completing ranking task. It is clear
that PMF behaves worse than other ranking models (except PureSVD) and
RankPMF behaves better than PMF.
4. The results demonstrate the robust of the statistical features. Our algorithm
generalizes well when both few and many ratings are available.
5. We have better NDCG@10 than other approaches. We have many informa-
tive ranking-based neighbors, and these neighbors have the same preference
with the target user, so they contain accuracy preference information.

5 Conclusion and Future Work

In this paper, we propose a new collaborative ranking approach. We use the
statistical information of the rankings of item from users nearest neighbors as
the feature representation for user-item pairs. We use Kendall Rank Coecient
as similar metric to choose neighbors and experimentally demonstrate the eec-
tiveness of our method. Interestingly, we can address scenarios where assigning
numerical scores to dierent types of users feedbacks would not be easy.
There are some future work based on this work. Since classical collaborative
filtering algorithm ignores the social connections between users, but in real-world
780 C. Fan and Z. Lin

user often asks her/his friends for recommendations and users taste is influenced
by friends taste on the social network. Incorporating social eects in our collab-
orative ranking framework is our next work. Dierent users in recommendation
systems may vary largely. For example, the influence of neighbors on dierent
users are very dierent. Considering the large dierence between users, it is not
the best choice to use a single ranking function to deal with all users. So we also
plan to investigate how to design personalized collaborative ranking algorithm
in the future.

Acknowledgement. We thank the anonymous reviewers for helpful comments.

This work is supported by the program of the National Natural Science Foun-
dation of China (NSFC) under grant 60973003.

References

1. Koren, Y.: Factor in the neighbors: Scalable and accurate collaborative filtering.
ACM Trans. Knowl. Discov. Data 4(1), 1:11:24 (2010)
2. Liu, T.Y.: Learning to rank for information retrieval, Berlin, German, vol. 3(3),
pp. 225331. Springer (2011)
3. Volkovs, M.N., Zemel, R.S.: Collaborative ranking with 17 parameters. In: Pro-
ceedings of the Twenty-sixth Annual Conference on Neural Information Process-
ing Systems, NIPS 2012, Lake Tahoe, Nevada, United States, December 3-6, MIT
Press (2013)
4. Balakrishnan, S., Chopra, S.: Collaborative ranking. In: Proceedings of the Fifth
ACM International Conference on Web Search and Data Mining, WSDM 2012, pp.
143152. ACM, New York (2012)
5. Deshpande, M., Karypis, G.: Item-based top-n recommendation algorithms. ACM
Trans. Inf. Syst. 22(1), 143177 (2004)
6. Liu, N.N., Yang, Q.: Eigenrank: a ranking-oriented approach to collaborative fil-
tering. In: Proceedings of the 31st Annual International ACM SIGIR Conference
on Research and Development in Information Retrieval, SIGIR 2008, pp. 8390.
ACM, New York (2008)
7. Marden, J.I.: Analyzing and modeling rank data. Chapman & Hall Press, New
York (1995)
8. Cremonesi, P., Koren, Y., Turrin, R.: Performance of recommender algorithms on
top-n recommendation tasks. In: Proceedings of the Fourth ACM Conference on
Recommender Systems, RecSys 2010, pp. 3946. ACM, New York (2010)
9. Koren, Y., Sill, J.: Ordrec: an ordinal model for predicting personalized item rat-
ing distributions. In: Proceedings of the Fifth ACM Conference on Recommender
Systems, RecSys 2011, pp. 117124. ACM, New York (2011)
10. Rendle, S., Freudenthaler, C., Gantner, Z., Schmidt-Thieme, L.: Bpr: Bayesian
personalized ranking from implicit feedback. In: Proceedings of the Twenty-Fifth
Conference on Uncertainty in Artificial Intelligence, UAI 2009, Arlington, Virginia,
United States, pp. 452461. AUAI Press (2009)
11. Shi, Y., Larson, M., Hanjalic, A.: List-wise learning to rank with matrix factoriza-
tion for collaborative filtering. In: Proceedings of the Fourth ACM Conference on
Recommender Systems, RecSys 2010, pp. 269272. ACM, New York (2010)
 Collaborative Ranking with Ranking-Based Neighborhood 781

12. Shi, Y., Karatzoglou, A., Baltrunas, L., Larson, M., Hanjalic, A., Oliver, N.: Tfmap:
optimizing map for top-n context-aware recommendation. In: Proceedings of the
35th International ACM SIGIR Conference on Research and Development in In-
formation Retrieval, SIGIR 2012, pp. 155164. ACM, New York (2012)
13. Shi, Y., Karatzoglou, A., Baltrunas, L., Larson, M., Oliver, N., Hanjalic, A.: Climf:
learning to maximize reciprocal rank with collaborative less-is-more filtering. In:
Proceedings of the Sixth ACM Conference on Recommender Systems, RecSys 2012,
pp. 139146. ACM, New York (2012)
14. Weimer, M., Karatzoglou, A., Bruch, M.: Maximum margin matrix factorization
for code recommendation. In: Proceedings of the Third ACM Conference on Rec-
ommender Systems, RecSys 2009, pp. 309312. ACM, New York (2009)
15. Burges, C.J.C., Ragno, R., Le, Q.V.: Learning to rank with nonsmooth cost func-
tions. In: Proceedings of the Twentieth Annual Conference on Neural Information
Processing Systems, NIPS 2006, Vancouver, British Columbia, Canada, December
4-7, pp. 193200. MIT Press (2007)
16. Jarvelin, K., Kek
al
ainen, J.: Cumulated gain-based evaluation of ir techniques.
ACM Trans. Inf. Syst. 20(4), 422446 (2002)
 Probabilistic Top-k Dominating Query
over Sliding Windows

Xing Feng1 , Xiang Zhao2 , Yunjun Gao1 , and Ying Zhang2

1
Zhejiang University, China
{3090103362,gaoyj}@zju.edu.cn
2
The University of New South Wales, Australia
{xzhao,yingz}@cse.unsw.edu.au

Abstract. Probabilistic queries on uncertain data have been intensively inves-

tigated lately. Top-k dominating query is important in many applications, e.g.,
decision making, as it offers choices which are better than the most of others. In
this paper, we study the problem of probabilistic top-k dominating query over
sliding windows. An efficient algorithm is developed to compute the exact so-
lution. Extensive experiments are conducted to demonstrate the efficiency and
effectiveness of the algorithm.

1 Introduction

Spatial database has drawn attention in recent decades. Beckmann et al. [1] design
the R-tree as an index for efficient computation. Roussopoulos et al. [8] study nearest
neighbor queries, and propose three effective pruning rules to speed up the computation.
They also the extend nearest neighbor queries to k nearest neighbor queries, which
return top-k preferred objects. Borzsonyi et al. [4] is among the first to study the skyline
operator, and propose an SQL syntax for the skyline query.
Top-k dominating query is first studied by Yiu et al. [10], which retrieves the top-
k objects that are better than the largest number of objects in a dataset. This is quite
different from the skyline query in [4] that retrieves objects which are not worse than
other objects.
Uncertain data analysis is of importance in many emerging applications, such as
sensor network, trend prediction and moving objects management. There has been a lot
of works focusing on uncertain data management, [5,11] to cite a few.
In this work, we study probabilistic top-k dominating query over sliding windows.
We employ the data model with sliding windows as used in [9,13]. In sliding windows,
data is treated as a stream, and only recent n objects are considered.
There are some closely related work, seen in [7,12]. However, they study objects
with multiple instances. Besides, their query is to get top-k dominating objects from
the total data set. While our paper studies objects from append-only data stream. And
our main concern is to maintain top-k dominating objects from recent N objects where
N is the window size.
Probabilistic top-k dominating query is desirable in various real-life applications. Ta-
ble 1 evidences a scenario of house rental, where answering such type of query can be

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 782793, 2013.
c Springer-Verlag Berlin Heidelberg 2013
 Probabilistic Top-k Dominating Query over Sliding Windows 783

beneficial. Lessees are interested in knowing the top-1 house from all recent 4 advertise-
ments. Each advertisement is associated with its house ID, post time, price, distances
to supermarket and trustability. Trustability is derived from Lessees feedback on the
lessors product quality; this trustability value can also be regarded as the probability
that the house is the same as what it claims.
We assume lower price and closer to supermarket are preferred and lessees do not
care about other attributes. We also assume that lessees want a house that are better
than the most of others. A top-k dominating (k = 1) query would retrieve {H1} as the
result for the first 4 objects {H1, H2, H3, H4} because H1 is better than H2, H3, H4.
{H4} is the result for the next sliding window {H2, H3, H4, H5}. However, we may
also notice that H4 is of low trustability. It is more reasonable to take the trustabil-
ity into consideration. We model such on-line selection problem as probabilistic top-k
dominating query against sliding windows by regarding on-line advertisements as an
append-only data stream. This query will be formally defined in Section 2. Hence, the
probabilistic top-k dominating query over sliding windows provides a more reasonable
solution under the given scenario.

Table 1. Example of house rental

House ID Post time Rent per week Distance to supermarket Trustability

H1 100 days ago $167 100m 0.80
H2 97 days ago $205 340m 0.90
H3 63 days ago $206 820m 0.85
H4 55 days ago $185 110m 0.50
H5 50 days ago $230 700m 1.00
H6 43 days ago $200 760m 0.90

In summary, we make the following contributions in this paper:

We identify the problem of probabilistic top-k dominating query over sliding win-
dows. To the best of our knowledge, there is no prior work regarding this.
We develop novel algorithms to continuously compute the top-k dominating ob-
jects, and analyze the complexity.
We conduct extensive experiments to demonstrate the correctness and efficiency of
our algorithms.

Organization. The rest of the paper is organized as follows. In Section 2, we present

background information and formally define the problem. Sections 3 and 4 present the
techniques for processing top-k dominating query over sliding windows. Experimental
results are provided and analyzed in Section 5. Section 6 concludes the paper.

2 Preliminary
This section introduces the background information, and defines the problem.
784 X. Feng et al.

2.1 Background
When comparing two objects, if one object p is not worse than another q in all aspects
and p is better than q in at least one aspect, we say p dominates q. Formally, we have
Definition 1.
Definition 1 (Dominate). Consider two distinct d-dimensional objects p and q, and
p[i] denotes the i-th value of p. p dominates q (denoted by p q) iff p[i] q[i] holds
for 1 i d, and there exists j, such that p[j] < q[j].
Based on Definition 1, [10] develops a score function to count the number of objects
dominated by an object as
(p) = |{p D|p p }|. (1)
Consequently, top-k dominating query in a data set is to retrieve the top-k objects with
maximal (p) in the data set.
Given a sequence data stream DS of uncertain data objects, a possible world W
is a subsequence of DS; i.e., each object from DS can be either in W or not. For
example, given a data stream with 3 objects, {a, b, c}, there are totally 8 possible worlds
{, {a}, {b}, {c}, {a, b}, {a, c}, {b, c}, {a, b, c}}. The probability of W to appear is
the product of probabilities of all objects which appear ?in W and 1 ? the probabilities of
all objects which do not appear in W ; i.e., P (W ) = W P () W / (1 P ()).
We compute the function in Equation (1) for every object in the certain possi-
ble world W . For a top-k dominating query over sliding window, probability of each
qualified possible world W is accumulated in order to get the overall probability.

2.2 Problem Definition

In many applications, a data stream is append-only; that is, there is no deletion of data
element involved. In this paper, we study the top-k dominating problem by employing
the append-only data stream model. Given an uncertain data stream DS consisting of n
uncertain objects, sliding window size W , probability threshold q, we use l to measure
the dominating ability of each object in a sliding window. That is, an object dominates
at least l objects in possible worlds, of which the total probability is over q in a sliding
window of size W . Thus, top-k dominating over sliding windows maintains the top-k
objects which are ranked according to l over sliding windows with regards to threshold q.
Example 1. Consider the running example in Table 1, where we need the top-1 house
over a sliding window of size N = 4, and probability threshold q = 0.6. We have in
total 16 possible worlds in each window. When a window contains {H3, H4, H5, H6},
it is a little different. There are 8 possible worlds rather 16 possible worlds, since
P (H5) = 1. First, we compute the dominating number of each object in every possible
world (as shown in Table 2). When computing l value, we sum up the probability of
each possible world according to the number of objects it dominated in descending or-
der, till the probability exceeds the threshold q. Thus, H6 is 76.5%(38.25% + 38.25%),
and H4 gets 50%(38.25% + 6.75% + 4.25% + 0.75%). H3 and H5 are both 0, since
they do not dominate others. Therefore, we have H3.l = 0, H4.l = 0, H5.l = 0, and
H6.l = 1; and hence, H6 is the top-1 object in the probabilistic top-1 dominating query
with probability threshold = 0.6 over the sliding window {H3, H4, H5, H6}.
 Probabilistic Top-k Dominating Query over Sliding Windows 785

Table 2. Dominating number in possible world

Probability Possible world H3 H4 H5 H6

0.75% H5 0 0 0 0
4.25% H3, H5 0 0 0 0
0.75% H4, H5 0 1 0 0
6.75% H5, H6 0 0 0 0
4.25% H3, H4, H5 0 2 0 0
38.25% H3, H5, H6 0 0 0 1
6.75% H4, H5, H6 0 2 0 0
38.25% H3, H4, H5, H6 0 3 0 1

2.3 Dominance Relationships

We implement our techniques based on the R-tree. Thus, we define the following rela-
tionships between a pair of entries E and E in the R-tree. E.min denotes the lower-
left corner of the minimum bounding box (MBB) of the objects contained by E, and
E.max denotes the upper-right corner of MBB. Note if an entry contains only one
object , E.min = E.max = .
An entry E is said to fully dominate another E , iff E.max E .min or E.max =
E .min, denoted as E E . An entry E partially dominates E , iff E.max

E .min, E.min E .max, denoted as E partial E . An entry E does not domi-

nate E , iff E.min E .max, denoted as E E . As depicted in Figure 1, E2 fully
dominates E3, E partially dominates E2, and E1 does not dominate E.

E1 E3
0.9

0.8

0.7

E2
0.6

0.5

0.4

E
0.3

0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Fig. 1. Dominance relationship

3 Dominating Score Calculation

As aforementioned, we may calculate the dominating score by the finding maximal l

whose accumulating possibility of all the qualified possible worlds is over q. However,
this can be rather time-consuming due to the exponential number of possible worlds;
i.e., there are 2N possible worlds, where N is the number of all objects.
786 X. Feng et al.

We observe that the complexity can be reduced by viewing the problem from another
perspective. Instead of enumerating possible worlds, we retrieve the largest l in the
condition, where at least l objects exist of M objects, where M is the number of all
dominated objects.
Specifically, let mi,j denote the probability that exactly i objects out of j objects
exist, and P (i) denote the probability that the i-th object exists. We have the following
relations.

m0,j1 (1 (P (j))) i=0
mi,j = mi1,j1 P (i) i=j (2)

mi1,j1 P (j) + mi,j1 (1 P (j)) otherwise
Equation (2), known as Poisson-Binomial Recurrence [6], is widely used in uncertain
database analysis [2,3]. Applying Equation (2), we propose a dynamic programming al-
gorithm that efficiently computes l value of a certain object, encapsulated in Algorithm 1.
Algorithm 1 takes and as inputs, where is the object of which we need to
compute l value, and contains all the objects which are dominated by . A matrix is
used to store the intermediate result, initialized in Line 2. Then, we use Equation (2) to
fill in the cells of the matrix (Lines 4 13). Finally, we remove mi,size from f at a time
till f is less then q (Lines 15 - 17), and i 1 is returned as the result.

Algorithm 1. calDS(, )
1 if P () < q then return 0;
2 m0,0 = 1; i = 1;
3 size = ||; f = P ();
4 while j size do
5 m0,j = m0,j1 (1 P (j));
6 j = j + 1;
7 while i size do
8 mi,i = mi1,i1 P (i);
9 j = 1;
10 while j < i do
11 mi,j = mi1,j1 P (j) + mi,j1 (1 P (j));
12 j = j + 1;
13 i = i + 1;
14 i = 0;
15 while f > q do
16 f = f mi,size P ();
17 i = i + 1;
18 return i 1;

Correctness and Complexity Analysis. It is immediate that Algorithm 1 correctly

computes the l value. The most time-consuming part is the loop in Lines 7 13, and
thus, the complexity of algorithm is O(||2 ).
 Probabilistic Top-k Dominating Query over Sliding Windows 787

Example 2. Consider the running example in 2.2. We only compute H4 and H6, since
of H3 or H5 is empty. For H4, P (H4) = 0.5 < 0.6, and hence , H4.l = 0. For H6,
P (H6) = 0.9 > 0.6, m0,1 = 0.15, m1,1 = 0.85. As 0.9 0.15 P (H6) = 0.765 >
0.6, 0.765 0.85 P (H6) = 0 < 0.6, the H6.l is 1. Therefore, top-1 object among
{H3, H4, H5, H6} is H6.

4 Updating Technique
A naive solution to retrieving top-k dominating objects over sliding windows is to visit
each object every time the window slides. When visiting an object, it also gets all the ob-
jects it dominates by traversing R-tree and then computes its l value using Algorithm 1.
Eventually, it chooses k objects with maximal l values. We observe that this tedious
process can be accelerated. Given a probability threshold q and a sliding window with
size N , Algorithm 2 shows how we process every arriving object.

Algorithm 2. Probabilistic top-k dominating over Sliding Windows

1 while a new object new arrives do
2 if |S| N then
3 insert(new );
4 else
5 remove(old );
6 insert(new );
7 collect top-k dominating objects;

When a new object new comes, if there are exactly N objects in the window, we
remove the oldest object old . Then, we use function insert(new ) to update the l
value of objects. Finally we collect the top-k dominating objects using min-heap.
In the following subsections, we present novel algorithms to efficiently execute Al-
gorithm 2. We first introduce the data structure, then provide our algorithm to handle
the situation when a new object is inserted (insert(new )), followed by an algorithm
to deal with a removed old object (remove(old )).

4.1 Aggregate R-Trees and Heap

We store all the N objects in an aggregate R-tree. We store P , l and top-k dominating with
each object as aggregate information. Specifically, P is the probability of the object, l
is the measure of dominating ability regarding probabilistic threshold q. Additionally,
top-k dominating is the last column of matrices that we use to compute l by Algorithm 1.
Note that each column only depends on the previous column. In another word, when
a new object new comes, keeping last column of the matrix is enough to update the l
value.
By storing these aggregate information, we do not have to recalculate the l value of
all objects when a new object comes. We only need to update top-k dominating and l value
every time a new object comes.
788 X. Feng et al.

In addition, we maintain the top-k dominating objects in a min-heap; i.e., we insert

the first k objects into a min-heap. For each following object having larger l than the
minimum object in the heap, we first dequeue the minimal object, and then insert the
new object into the min-heap.

4.2 Insert
When a new object new comes, we need to consider all the objects dominating it
(denote as S1) and being dominated by it (denote as S2). To this end, we process the
following tasks: 1) compute S1 and S2 by traversing the aggregate R-tree, 2) update l
of each object in S1; and 3) compute l of the new object.
Algorithm 3 describes the steps we take to handle an insertion. R is the aggregate
R-tree we used to store all objects in current sliding window. S3 is used to collect all
the entries which partially dominate new , S4 is used to collect all the entries which
are partially dominated by new , and S34 is used to collect the entries which partially
dominate new and are partially dominated by new . First, we classify all the entries to
different sets (Lines 1 - 6). Then, we refine the result and get all the entries of interest,
namely S1 and S2, using probe1, probe2 and probe3 (Lines 7 - 9). So far, we have
finished task 1). Then, task 2) is done by updating all the entries in S1 using update
(Line 10). Afterwards, we achieve task 3) by computing l of new (Line 11).

Algorithm 3. insert(new )
1 for each E R.root do
2 ifE new then add E to S1;
3 else ifnew E then add E to S2;
4 else ifE partial new &new E then add E to S3;
5 else ifE partial new &new partial E then add E to S34;
6 else ifnew partial E&E new then add E to S4;
7 ifS3 = then probe1(S3);
8 ifS4 = then probe2(S4);
9 ifS34 = then probe3(S34);
10 ifS2 = then update1(S1);
11 ifS1 = then new .l = calDS(S2, new );

Algorithm 4 shows how we refine S3. Note that all the entries in S3 cannot be
dominated by new . We add entries dominating new to S1 (Line 4 in Algorithm 4),
and leave entries which partially dominate new in S3 (Line 5 in Algorithm 4).

Algorithm 4. probe1(S3)
1 while S3 = do
2 E = Dequeue(S3);
3 for each child E of E do
4 if E new then add E to S1;
5 else if E partial new then add E to S3;
 Probabilistic Top-k Dominating Query over Sliding Windows 789

The ideas behind Algorithm 4 and Algorithm 5 are similar. We add entries which are
dominated by new to S2 (Line 4 in Algorithm 5), and leave entries which are partially
dominated by new in S4 (Line 5 in Algorithm 5).

Algorithm 5. probe2(S4)
1 while S4 = do
2 E = Dequeue(S4);
3 for each child E of E do
4 if new E then add E to S2;
5 else if new partial E then add E to S4;

S34 is special in that it may contain entries which dominate new and entries which
are dominated by new . Thus, we first partition them (Lines 1 - 8), and apply probe1
and probe2 thereafter.

Algorithm 6. probe3(S34)
1 while S34 = do
2 E = Dequeue(S34);
3 for each child E of E do
4 if new E then add E to S1;
5 else if E new then add E to S2;
6 else if new partial E &E partial new then add E to S34;
7 else if E partial new &new E then add E to S3;
8 else if new partial E &E new then add E to S4;

9 if S3 = then probe1(S3);
10 if S4 = then probe2(S4);

Algorithm 7 describes the update of S1. For each object at leaf node, we update its l
and top-k dominating by using new . As a consequence, the time complexity of the update
is O(l).

Algorithm 7. update1(S1)
1 while S1 = do
2 E = Dequeue(S1);
3 for each child E of E do
4 if E is at leaf level then
5 Read E ;
6 for each child E of E do
7 updateLambda(E , new );
8 updateL(E );

9 else
10 for each child E of E do
11 add E to S1;
790 X. Feng et al.

We analyse the worst case complexity of Algorithm 3; i.e., when every object is
dominated by all the previous objects. Let N denote the window size. In this case, S1
contains all N objects. Immediately, the most time consuming part is in Line 11, and
thus, the time complexity is given by O(N 2 ).
Since Algorithm 3 is complicated, we give an example below.
Example 3. Regarding the example in Table 1. We assume windows size N = 4, prob-
abilistic threshold q = 0.6, and there are exactly H1, H2, H3 in current window. H4
comes as new . Before H4 comes, we use Algorithm 1 to compute the corresponding l
values of H1, H2, H3. As an example, the first three columns in Table 3 shows how to
compute the l value of H1. And we keep {0.015, 0.22, 0.765} as top-k dominating of H1.
The aggregate information of other objects are listed in the first three rows of Table 4.
When H4 comes, according to Algorithm 3, we first collect all objects dominating H4
in S1 and all objects dominated by H4 in S2. Here, we have S1 = {H1}, S2 =
{H2, H3}. Then, we update l and top-k dominating of H1. Using Equation (2), we get
the 4-th column in Table 3. Because (0.3825 + 0.4925) P (H1) > q = 0.6, the l
value still equals 2. And top-k dominating is updated as {0.0075, 0.1175, 0.4925, 0.3825}.
Finally, we compute l and top-k dominating for H4. Because P (H4) < q = 0.6, we
have l = 0 and top-k dominating = . Therefore, the top-1 dominating query for window
{H1, H2, H3, H4} is H1.

Table 3. The matrix of H1 Table 4. Aggregate information

H1(0.80) H2(0.90) H3(0.85) H4(0.5) House ID l top-k dominating

1 0.1 0.015 0.0075 H1 2 {0.015, 0.22, 0.765}
H2 1 {0.15, 0.85}
0.9 0.22 0.1175
H3 0
0.765 0.4925 H1(updated) 2 {0.0075, 0.1175, 0.4925, 0.3825}
0.3825 H4 0

4.3 Expire
When old expires, the l value of each object dominated by old will not change. There-
fore, we only need to collect all the objects dominating old , re-calculate their l values.

Algorithm 8. remove(old )
1 for each E R.root do
2 if E old then add E to S1;
3 else if E partial old then add E to S3;
4 while S3 = do
5 E = Dequeue(S3);
6 for each child E of E do
7 if E old then add E to S1;
8 else if E partial old then add E to S3;

9 update2(S1);
10 delete old from R;
 Probabilistic Top-k Dominating Query over Sliding Windows 791

The update process of S1 is shown in Algorithm 9. For each object at leaf node,
we re-calculate its l and top-k dominating by traversing the R-tree to collect all the objects
which are dominated by the object in .

Algorithm 9. update2(S1)
1 while S1 = do
2 E = Dequeue(S1);
3 for each child E of E do
4 if E is at leaf level then
5 read E ;
6 for each child E of E do
7 traverse R to collect all objects dominated by E in ;
8 E .l = calDS(, E );

9 else
10 for each child E of E do
11 add E to S1;

Similarly, we provide a worst case complexity analysis of Algorithm 8. When all

objects dominate all previous objects, we need to re-calculate l value of all remaining
objects. Therefore, the time complexity is given by O(N 3 ), where N is the window
size.
Correctness. The correctness of our techniques is guaranteed by computing the exact
l value of each object in current sliding window. Besides, we get the top-k dominating
objects using a min-heap. This is also correct because the k objects are guaranteed to
have larger l values than these not in the min-heap.

5 Experiments
This section reports the experimental studies.

5.1 Setup
We ran all experiments on a MacBook Pro with Mac OSX 10.8.1, 2.26GHz Intel Core
2 Duo CPU and 4G 1333 MHz RAM. We evaluated the efficiency of our algorithm
against sliding window size, dimensionality, and probabilistic threshold, respectively.
The default values of the parameters are listed in Table 5.
Parameters are varied as follows:
sliding window size: 2k, 4k, 6k, 8k, 10k
dimensionality: 2, 3, 5
probabilistic threshold: 0.1, 0.3, 0.5, 0.7, 0.9
We ran all experiments with 100 windows, and repeated all experiments 10 times to get
an average execution time. The execution time in each following figure is the running
time for processing 100 windows.
792 X. Feng et al.

Table 5. Experimental parameters

Notation Definition(Default Values)

N sliding window size (4K)
d dimensionality of the data set (3)
D data set (Anti)
Dp probabilistic distribution of appearance (uniform)
P expected appearance probability (0.5)
q probabilistic threshold (0.3)

5.2 Evaluation
Figures 2, 3 and 4 show the results when window size is varied. In Figure 5, we vary
the threshold probability to see the running time of baseline algorithm. In Figure 6, we
vary the threshold probability to see the running time of the efficient algorithm. From
the figures, we see that the efficient algorithm outperforms the baseline algorithm in all
cases.

4
10 baseline baseline 2 baseline
efficient 2 efficient 10 efficient
Avg. Delay(s)

Avg. Delay(s)

Avg. Delay(s)

103 10
1
2 10
1 10
10
0 0
10
1 10 10

100 10-1 10-1

2k 4k 6k 8k 10k 2k 4k 6k 8k 10k 2k 4k 6k 8k 10k

Fig. 2. Varying N , d = 2 Fig. 3. Varying N , d = 3 Fig. 4. Varying N , d = 5

As shown in Figures 2, 3 and 4, when window size increases linearly, the running
time increases in a super-linear manner for of the both efficient and baseline algo-
rithms. Moreover, we conclude, by comparing Figures 2, 3 and 4, that the running time
decreases greatly with the increase of dimensionality. This is because the most time-
consuming part of the baseline algorithm is to compute l; i.e., collect all objects and
apply Algorithm 1 to compute l. With higher dimensionality, there is a smaller for
each object. Similar conclusion can be drawn for the efficient algorithm.

3 2d 3d 5d 2d 3d 5d
10 10
2
Avg. Delay(s)

Avg. Delay(s)

1
10
102
100
1 -1
10 10
0.1 0.3 0.5 0.7 0.9 0.1 0.3 0.5 0.7 0.9

Fig. 5. Baseline, Varying q Fig. 6. Efficient, Varying q

 Probabilistic Top-k Dominating Query over Sliding Windows 793

From Figures 5 and 6, when threshold grows, the running time decreases. With larger
threshold, there are more objects being pruned, i.e. objects whose probability is lower
than threshold. In addition, this effect is more obvious with low dimensionality, as more
objects to be pruned by the algorithm in the lower dimensional space.

6 Conclusion
In this paper, we have investigated the problem of top-k dominating query in the con-
text of sliding window on uncertain data. We first model the probability threshold based
top-k dominating problem, and then present a framework to handle the problem. Ef-
ficient techniques have been presented to process such query continuously. Extensive
experiments demonstrate the effectiveness and efficiency of our techniques.

References
1. Beckmann, N., Kriegel, H.-P., Schneider, R., Seeger, B.: The r*-tree: An efficient and robust
access method for points and rectangles. In: SIGMOD Conference, pp. 322331 (1990)
2. Bernecker, T., Kriegel, H.-P., Mamoulis, N., Renz, M., Zufle, A.: Scalable probabilistic sim-
ilarity ranking in uncertain databases. IEEE Trans. Knowl. Data Eng. 22(9), 12341246
(2010)
3. Bernecker, T., Kriegel, H.-P., Mamoulis, N., Renz, M., Zuefle, A.: Continuous inverse rank-
ing queries in uncertain streams. In: Bayard Cushing, J., French, J., Bowers, S. (eds.) SSDBM
2011. LNCS, vol. 6809, pp. 3754. Springer, Heidelberg (2011)
4. Borzsonyi, S., Kossmann, D., Stocker, K.: The skyline operator. In: ICDE, pp. 421430
(2001)
5. Cormode, G., Garofalakis, M.N.: Sketching probabilistic data streams. In: Chan, C.Y., Ooi,
B.C., Zhou, A. (eds.) SIGMOD Conference, pp. 281292. ACM (2007)
6. Lange, K.: Numerical analysis for statisticians. Statistics and Computing (1999)
7. Lian, X., Chen, L.: Top-k dominating queries in uncertain databases. In: Kersten, M.L.,
Novikov, B., Teubner, J., Polutin, V., Manegold, S. (eds.) EDBT. ACM International Con-
ference Proceeding Series, vol. 360, pp. 660671. ACM (2009)
8. Roussopoulos, N., Kelley, S., Vincent, F.: Nearest neighbor queries. In: SIGMOD Confer-
ence, pp. 7179 (1995)
9. Shen, Z., Cheema, M.A., Lin, X., Zhang, W., Wang, H.: Efficiently monitoring top-k pairs
over sliding windows. In: Kementsietsidis, A., Salles, M.A.V. (eds.) ICDE, pp. 798809.
IEEE Computer Society (2012)
10. Yiu, M.L., Mamoulis, N.: Efficient processing of top-k dominating queries on multi-
dimensional data. In: VLDB, pp. 483494 (2007)
11. Zhang, Q., Li, F., Yi, K.: Finding frequent items in probabilistic data. In: SIGMOD Confer-
ence, pp. 819832 (2008)
12. Zhang, W., Lin, X., Zhang, Y., Pei, J., Wang, W.: Threshold-based probabilistic top-k domi-
nating queries. VLDB J. 19(2), 283305 (2010)
13. Zhang, W., Lin, X., Zhang, Y., Wang, W., Yu, J.X.: Probabilistic skyline operator over sliding
windows. In: Ioannidis, Y.E., Lee, D.L., Ng, R.T. (eds.) ICDE, pp. 10601071. IEEE (2009)
 Distributed Range Querying Moving Objects
in Network-Centric Warfare

Bin Ge, Chong Zhang, Da-quan Tang, and Wei-dong Xiao

Science and Technology on Information Systems Engineering Lab,

National Univ. of Defense Technology, Changsha 410073, China
[email protected]

Abstract. Sensing and analyzing moving objects is an important task for com-
manders to make decisions in Network-centric Warfare(NCW). Spatio-temporal
queries are essential operations to understand moving objects situation. This
paper proposed a new method for addressing distributed range querying moving
objects problem, which was not considered in related studies. Firstly, the paper
designed an index structure called MOI which utilized hilbert curve to mapping
moving objects into 1-dimensional space and then used DHT to index. Then
based on locality-preserved characteristic of hilbert curve, the paper devised a
high-efficient spatio-temporal range query algorithm, which greatly reduces
routing messages. Experimental results reveal that MOI outperforms other re-
lated algorithms and exhibits a low index maintaining cost.

Keywords: P2P, Distributed, MOI, Hilbert curve, Spatio-temporal Indexing.

1 Introduction

Moving objects(MO) data is a kind of important information for analysis, planning

and fire striking in NCW, especially in the future digital battlefield. Awaring battle-
field situation by sharing moving objects data, it is possible to accomplish tasks such
as effective alert action, battlefield environment protection and weapon unit precision
strikes, which is of great importance on improvement of battlefield transparency and
joint operations ability and implementation of high efficient command[1,2]. Accumu-
lating battlefield information, forming large battlefield spatio-temporal data set, new
requirements for spatio-temporal index technology in collaborative, efficiency, agility
and reliability have been put forward.
Based on Peer-to-Peer computing (P2P) technology, this paper presented a distri-
buted spatio-temporal index structure MOI (Moving Objects Index). Utilizing DHT to
organize node topology structure, adapting Hilbert curve to map spatio-temporal ob-
jects to one dimensional space and utilizing locality preservation property of the
curve, the index structure designs efficient query algorithm, with which the number
of routing messages has been greatly reduced and performance of spatio-temporal
query has been improved, ensuring efficient spatio-temporal information sharing in
battlefield.

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 794803, 2013.
Springer-Verlag Berlin Heidelberg 2013
 Distributed Range Querying Moving Objects in Network-Centric Warfare 795

2 Problem Description

Moving objects constantly change their position and shape with time passing by. This
paper considers that spatio-temporal data is a kind of data related to the position de-
tected by various detection methods in the earth reference space, expressing the gra-
dual changes over time of entitys and processs state of being (position, shape, size,
etc.) and property.
In the battlefield, there exists all kinds of objects whose properties evolves over
time, such as position changing individual, tanks, and shape changing cloud, smoke,
fire, etc. In this paper, the use of thought about traditional multidimensional index
(such as R-trees) abstracts time-space objects in the battlefield into minimum bound-
ing rectangle (MBR). The corresponding coordinates of MBR (i.e., two point coordi-
nates in left bottom corner and right-top-corner) change with the position and shape of
time-space object.
Besides moving objects, other command unit exists in battlefield. With positioning
technology, the position and shape of objects can be transmitted to these units. Mov-
ing objects (such as individual, tanks and other autonomy type objects) can register
nearby to report to corresponding units, whereas the units can collect space informa-
tion to local of the objects by continuous reconnaissance and detection.
Through the description above, the problem can be defined as: in the battlefield,
command unit can be abstracted as network nodes (i.e., peer), each node gets physical
connection through the ground link, the investigation data is position and shape
changing over time rectangular data (i.e., MBR), in the distributed environment, it is
required to set up a kind of spatio-temporal index mechanism, i.e., set up a network,
each node maintains part of the global index, each node, through the mutual commu-
nication, completes the whole index operation as query, insert, update, and etc, as
shown in figure 1. This paper here assumes that each moving object has a global
object ID.

Fig. 1. Problem Description

796 B. Ge et al.

3 MOI: Distributed Spatio-temporal Index Based on DHT

Using consistent hash, Chord[3] maps resources to one dimensional space, which
destroys the locality of resources. Namely resources nearby in the original space
through mapping are no longer neighbors, which cannot meet the needs of queries in
space range. So this paper maps the time-space information to one dimension infor-
mation using Hilbert curve[4].

3.1 Index Structure

MOI is actually a kind of overlay network. Based on DHT technology, each node has
a node identifier, and is responsible for part of key areas. Moving objects belonging to
this key area through mapping, their spatio-temporal information is maintained by
nodes responsible for. The detail description is as below:
1) This paper views the moving objects to be considered as extent objects, and
present its space state with Minimal Bounding Rectangle (MBR). This paper uses
MBR to represent moving objects space state below. This paper assumes the spatio-
temporal information format of moving objects obtained by nodes as (Oid, Xmin, Xmax,
Ymin, Ymax, t), in which Oid means the objects global ID, (Xmin, Xmax, Ymin, Ymax) means
the coordinates of MBR, and t means MBR timestamp.
2) This paper sets the whole space to be considered as square in two dimensions
with the side length equals to value E, and uses D order Hilbert curve. The whole
space is divided into cells of the same size. Through the construction of Hilbert
curve, the mapping from the set P of all the points in the whole space to the set V of
Hilbert value exists, among which and
.
3) This paper uses Chord as overlay network structure, with [0, ] as its key
space. We set m=2D, namely ensure the consistency of key space and the set of Hil-
bert value.
4) The sizes of corresponding MBR for every MO may be different. One MBR
may overlap several cells while one cell may contain several MBR. According to the
Hilbert curve,all points in each cell have unique counterpart Hilbert value, which can
be regarded as each cell has an unique corresponding Hilbert value, so that any one of
the MBR can be mapping for one or several Hilbert value based on Hilbert curve.
When an MBR maps for several Hilbert value, because of the locality preservation
property of the Hilbert curve, in these Hilbert values, some may be continuous, which
can be regarded that the MBR will be mapped to a series of 1 dimensional region.
This paper calls one dimension region for segments.
5) Each segment has a starting value and an end value, and according to the
Chords position in key space, MBR will be stored to the corresponding node, and the
specific process will be described in section 3.2. According to its different size, each
MBR can be stored to one or more nodes, which ensures the spatio-temporal query
rate. Each node establish HR-trees[5] index in local according to the spatio-temporal
information (Xmin, Xmax, Ymin, Ymax, t) of stored MBR.
 Distributed Range Querying Moving Objects in Network-Centric Warfare 797

Fig. 2. MOI Structure

Figure 2(a) reveals the space state of STO1~STO7 (for simplicity, we use O1~O7 in
the figure) respectively at t=10, t=20 and t=40 (namely MBR). Adopting 3 order Hil-
bert curve, the whole space is divided into 64 cells and numbered according to Hilbert
curve. Figure 2(b) shows the distribution of MO in the MOI index. If the segments of
O6 through mapping are [43,45],[39,40] and 34 respectively (figure 2 (a) distinguish
different segment with different shadow patterns), according to the segment value and
Chord resource placement strategy, referring data maintenance algorithm for detail,
when t=20, O6 distributed to node N38, N42, and N48. Each node establishes HR-tree
index locally, figure 2(b) samples the HR-tree index of node N21 and Finger Table of
node N8.

3.2 Spatio-temporal Query Algorithm

Two types of spatio-temporal queries are discussed: time point query and time inter-
val query. Time point query is that at time TQ, all the objects in the space range (Qxmin,
Qxmax, Qymin, Qymax) return as result. Time interval query is that in the period (TS, TE),
all the objects in the space range (Qxmin, Qxmax, Qymin, Qymax) return as result. They can
be presented respectively as ((Qxmin, Qxmax, Qymin, Qymax), TQ) and ((Qxmin, Qxmax, Qymin,
Qymax), (TS, TE)).
Spatio-temporal query algorithm described as below:

1) In accordance with Hilbert curve mapping rules that any coordinate in each cell
maps for the cell number Hilbert curve running through, the space range (Qxmin, Qxmax,
Qymin, Qymax) can be converted to a series of segments, one of which is a continuous
Hilbert value region. Each segment has the structure of (Henter, Hexit). Henter is a Hilbert
value as Hilbert curve into the segment, Hexit is a Hilbert value as Hilbert curve out
the segment, Henter Hexit;
798 B. Ge et al.

2) For each segment, according to the Chord routing rules that stated in [3] as
find_successor() function, node P routes the spatio-temporal query conditions and it
to keys successor node S;
3) When receiving spatio-temporal query request, node S carries on the judgment:
if node R, the predecessor node of node S, had identifier value equal to or greater than
the value of the received segment, the query request would be forward to node R;
4) Using of local HR-tree, node S conducts time point queries or time interval que-
ries, then routes the result to node P.

All nodes which is the equivalent of node S (successor node corresponding to Hexit
value of each segment) and node R (predecessor node meet the space overlapping con-
ditions through judgment by the equivalent node of node S) will go to step 3) and 4).
The example of spatio-temporal query algorithm is shown in figure 3. According to
the four segments, query conditions will route to node N38, N48 and N56 respectively.
After comparison, node N56 will forward query conditions to node N51. Each node
searches through time point queries or time interval queries in the local HR-tree and
consequently returns the result to the original inquiring node.

Fig. 3. Spatio-temporal Query

4 Construction and Maintenance of Index

4.1 Construction
In the battlefield, when a new command unit added, namely supposing that node Q
joins the network, implement steps as follows:
1) Node Q accesses to node identifier through connecting to match-making server
and other ways. Node Q then arbitrarily connects a node M, sending node identifier.
On the basis of Chord routing rules, namely the find_successor() function in [3], node
 Distributed Range Querying Moving Objects in Network-Centric Warfare 799

M forward the message. The join request of node Q is routed to the successor node N
of the key where node Qs identifier is.
2) Node N and node Q are respectively responsible for updating the key space
themselves (in agreement with chord protocols node adding algorithm).According to
the relationships between Hilbert mapping rules and key space itself, node N shifts
MO in the scope of node Q to node Q, by the way updating local HR-tree index.
3) Node Q receives the MO, and sets up the local HR-tree index.

4.2 Maintenance

Index maintenance is that when a node leaving or failing and data on node changing,
relevant nodes need to update the index, this kind of situation on the battlefield envi-
ronment often occur.

Leaving and Failing of Node

Node leaving is that before exiting system, the node will inform its successor node or
predecessor node that I will leave system, while node failing is that because of the
network connection instability or hardware failure, the line between nodes breaks, the
successor node or predecessor node do not know its exiting. Although in the two situ-
ations, node leaves system in different ways, the steps it takes are similar. The specif-
ic process is as below:

1) Node L sends "I will leave system" message and local MO to the successor node
K, node K updating local HR - trees, followed by the exiting of node L;
2) The predecessor node of node J keeps detecting the state of node L, once found
node L leaving the system, according to its own successor list (the same with Chord),
it would contact node K and save node K as successor node;
3) Node K saves node J for predecessor node, meanwhile, the index update com-
pletes.

For node failure, we adopt Heartbeat mechanism that periodically tests successor
nodes and predecessor nodes state, to assess whether there is a node failure, and then,
take steps 2)~3), completing index update.

Data Change
In the battlefield, position and shape of all kinds of MO change continuously as time
goes on, new situation, new object appear constantly in the battle space while some
objects possibly will quit or disappear (every MO has the only object ID), these are
corresponding to the system operations including update, insert and delete. These
operations are similar to the spatio-temporal query, specifically, when a node in the
system obtains a new object O (Oid, Xmin, Xmax, Ymin, Ymax, t) to insert, take steps as
follows:

1) Based on Hilbert mapping rules, the space range of object O (Xmin, Xmax, Ymin,
Ymax) is transformed into a series of segments (Henter, Hexit);
800 B. Ge et al.

2) For each segment, according to Chord routing rules, node I routes object O and
this segment to key Hexits successor node G;
3) After receiving object insertion request, node G start to assess: if identifier value
of node Gs predecessor node H is greater than or equal to the value Henter from the
received segment, forward this request to node H;
4) Using HR-tree insertion algorithm, node G maintains the spatio-temporal data
and returns the result of successful insertion operation to node I.

All nodes equal to node G (successor node corresponding to each segments Hexit
value) and node H (predecessor node in accordance with space overlap condition
judged by the equivalent node of node G) take steps 3)~4).
Both delete and update operations of MO are similar to insert operation, the differ-
ence is, the maintenance of local data need be shifted into HR-trees deletion and
update algorithm. Due to hardly space, no more statements is listed here.

5 Performance Evaluation

In order to evaluate performance of the proposed method in this paper, we use open
source simulation platform PeerSim[6] to conduct the simulation experiment. All
experiments are carried out on computers with Intel Core2 Quad 2.5 GHz and 2 GB
main memory. We set the whole space to be considered as 11 unit rectangular, and
set the length of time as 1 unit time region; Owing to the side length of the rectangle
in experiment is far less than that of the rectangle to be considered, we take 7 order
Hilbert curve for default.

5.1 Data Set Description

In order to simulate the battle space objects, we implement the GSTD[7] spatio-
temporal data generation method, which is widely used in spatio-temporal index ex-
periment. Based on GSTD method, two data sets are used: analog data as the base
state data (hereinafter referred as data set A) and real data as base state data (hereinaf-
ter referred as data set B).
In data set A, the center of base state data follows Gaussian distribution with
=0.5 (expectation) and =0.1 (variance), the side length presents 1000 rectangular
data which follows 0,0.1 uniform distribution; time sampling point numbers for
100, and follows uniform distribution. The variation of the center ( center) follows
(0.2, 0.2) uniform distribution and the variation of side length ( extent) follows (0.05,
0.05) uniform distribution. With time goes on, the overall change of data is that they
spread from the center to all around.
In data set B, we adopt TIGER[8] data set as the ground state data. The detailed
generate process is: choose all the ways in Maryland Montgomery county of TIGER
data set, with each road composed by limited lines. According to the starting point
and end point of the line segment, rectangular is generated, which is provided
for ground state data. Time sampling point numbers for 100, and follows uniform
 Distributed Range Querying Moving Objects in Network-Centric Warfare 801

distribution. The variation of the center ( center) follows (0, 0.4) uniform distribu-
tion, the side length variation ( extent) follows (-0.05, 0.05) uniform distribution. As
time goes on, the overall change of data is that they gradually move to northeast of the
space.

5.2 Query Performance Evaluation

We evaluate the query performance by the number of messages when querying. Mes-
sage number refers to the total message number sent by all the nodes from the net-
work for one time of spatio-temporal query. For each message number statistics, we
randomly generate 300 queries and take the average number of messages as the eval-
uation result. Compare MOI with Distributed Quadtree[9], P2P Meta-index [10], and
Service Zone[11], we conduct the query performance comparison. The size of mes-
sage number is tested under different network scales and different selection rate re-
spectively. For data set A and data set B, the experiment results are similar, here we
only take the result for data set A.
Set the querys space selection rate as 0.01. The number of nodes in the network
rises from 250 to 10000. And then, count the corresponding average message number.
As shown in figure 4(a), the result of the experiment reveals that with the increase of
the network scale, the query performance of the method proposed by this paper is
superior to the other three kinds of methods. This is because the space division me-
thod adopted by this paper using Hilbert curve with local preservation property, which
maps the query into continuous one dimensional queries. And the query algorithm
proposed by this paper considers checking the predecessor node whether it intersects
with query region by node itself, which greatly reduces the number of messages and
optimizes the query performance.

450
160 MOI MOI
Distributed Quadtree 400 Distributed Quadtree
140
Service Zone Service Zone
P2P Meta-index 350 P2P Meta-index
Number of messages

Number of messages

120
300

100
250

80 200

60 150

100
40

50
20
0
0 2000 4000 6000 8000 10000 0.00 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08
Number of peers Spatial Selectivity

(a) spatial selectivity=0.01 (b) number of peers =2000

Fig. 4. Query Performance Comparison

Set the number of node as 2000. The query space selection rate changes from 0.001
to 0.07. And then, count the corresponding average message number. As shown in
figure 4(b), the result of the experiment reveals that with the increase of the space
selection rate, the Service Zone methods and Distributed Quadtree method very rapid
802 B. Ge et al.

increase in the number of messages. And the proposed MOI method produced little
change in the number of messages, which is lower than the P2P Meta-index method.
This is because the Service Zone method uses CANs[12] thought to divide the space
binary tree, finding low efficiency. Distributed Quadtree method uses consistent hash,
where query is decomposed into a large number of one-dimensional intervals, increas-
ing traffic, resulting in query performance degradation. P2P Meta-index method with
the crawlers, need to continue to contact each server, increasing the communication
overhead.

5.3 The Evaluation of Index Maintenance Cost

Based on DHT, the processing performance for node join, exiting and the failure is
similar to DHT, so no more narrative hare. In this section, the experimental evaluation
mainly bases on data insertion cost. The number of messages caused by spatial and
temporal data insertion for data set A and data set B were tested under different net-
work size. The space default selection rate is 0.01, where spatial selectivity refers to
the size of the insertion of the MBR, the results shown in Figure 5.With the increase
in the number of nodes, for two different sets of data, the cost of maintaining indexes
showed a logarithmic growth. It illustrates that the index maintenance performance of
MOI is able to adapt to the changes in the size of network.

40000000
Dataset A
Dataset B
35000000
Number of messages

30000000

25000000

20000000

15000000

10000000

5000000

0
0 2000 4000 6000 8000 10000

Number of peers

Fig. 5. Index maintenance cost, spatial selectivity = 0.01

6 Conclusion

In future digital battlefield, the effective index of the temporal and spatial objects will
greatly help to improve the battlefield space-time query performance, which is of
great significance. For the needs of fast access to spatial and temporal information,
faced with the distributed battlefield environment, based on peer-to-peer computing,
this article proposes distributed spatio-temporal indexing mechanism MOI. Make use
of Hilbert curve combined the Chord to design spatio-temporal query algorithms,
 Distributed Range Querying Moving Objects in Network-Centric Warfare 803

indexing and maintenance algorithm. We test query performance on two different sets
of data, cost of the maintenance. The experimental results show that the local Hilbert
curves characteristics greatly improve the efficiency of the spatio-temporal query,
meanwhile the index maintenance performance is acceptable, so as to provide real-
time rapid access to the distributed spatio-temporal information in the future digital
battlefield with technical support.

Acknowledgement. This work is supported by the project of National Science Foun-

dation of China (No.71103203).

References
1. Ilarri, S., Mena, E., Illarramendi, A.: Location-Dependent Query Processing: Where We
Are and Where We Are Heading. ACM Computing Surveys 42(3), 173 (2010)
2. Wolfson, O., Xu, B., Chamberlain, S., Jiang, L.: Moving Objects Databases: Issues and
Solutions. In: Proceedings of the 10th International Conference on Scientific and Statistic-
al Database Management (SSDM), pp. 111122 (1998)
3. Stoica, I., Morris, R., Karger, D., Kaashoek, M.F., et al.: Chord: A Scalable Peer-to-Peer
Lookup Service for Internet Applications. In: Proceedings of the ACM SIGCOMM, pp.
149160 (2001)
4. Faloutsos, C., Roseman, S.: Fractals for Secondary Key Retrieval. In: Proceedings of the
8th ACM SIGACTSIGMODSIGART Symposium on Principles of Database Systems,
pp. 247252 (1989)
5. Nascimento, M., Silva, J.: Towards Historical R-trees. In: Proceedings of ACM Sympo-
sium on Applied Computing (ACM-SAC), pp. 235240 (1998)
6. PeerSim Simulator Project[EB/OL] (2005), http://peersim.sourceforge.net
7. Theodoridis, Y., Silva, J.R.O., Nascimento, M.A.: On the Generation of Spatiotemporal
Datasets. In: Proceedings of the 6th International Symposium on Spatial Databases (SSD),
pp. 147164 (1999)
8. Tiger Datasets[EB/OL], http://www.census.gov/geo/www/tiger/
9. Tanin, E., Harwood, A., Samet, H.: Using a Distributed Quadtree Index in Peer-to-Peer
Networks. VLDB Journal 16(2), 165178 (2007)
10. Hernndez, C., Rodrguez, M.A., Marn, M.: A P2P Meta-Index for Spatio-Temporal
Moving Object Databases. In: Proceedings of the 13th International Conference on Data-
base Systems for Advanced Applications, pp. 653660 (2008)
11. Wang, H., Zimmermann, R., Ku, W.-S.: Distributed continuous range query processing on
moving objects. In: Bressan, S., Kng, J., Wagner, R. (eds.) DEXA 2006. LNCS,
vol. 4080, pp. 655665. Springer, Heidelberg (2006)
12. Ratnasamy, S., Francis, P., Handley, M., Karp, R., Shenker, S.: A Scalable Content-
Addressable Network. In: Proceedings of the ACM SIGCOMM, pp. 161172 (2001)
 An Ecient Approach on Answering Top-k
Queries with Grid Dominant Graph Index

Aiping Li, Jinghu Xu, Liang Gan, Bin Zhou, and Yan Jia

School of Computer Science, National University of Defense Technology

410073 Changsha, China
[email protected], peaceful [email protected],
[email protected], [email protected], [email protected]
http://www.springer.com/lncs

Abstract. The top-k queries are well used in processing spatial data and
dimensional data stream to recognize important objects. To address the
problem of top-k queries with preferable function, a grid dominant graph
(GDG) index based on reserve data points set (RDPS) is presented.
Because of the correlation between top-k and skyline computation, when
the count of one points RDPS is bigger than or equals to the value of
k, it is for sure to be pruned the points dominant set. This approach in
used to prune a lot of free points of top-k queries. GDG used grid index
to calculate all RDPS approximately and eciently. When construction
GDG, grid index figures out k-max calculating region to prune free points
that all top-k function would not visit, which decreases the amount of
index dramatically. When travelling GDG, grid index figures out k-max
search region by top-k function which avoids travelling those free points
of ad-hoc top-k function. Because GDG uses grid index to rank those
data points in the same layer approximately by k-max search region
by top-k function, this index structure make less visited points than
traditional dominant graph (DG) structure. Whats more, grid index
needs less additional computation and storage that make the GDG index
more adaptive for top-k queries. Analytical and experimental evidences
display that GDG index approach performs better on both storage of
index and eciency of queries.

Keywords: top-k query, database, index, grid index, reverse dominant

point set.

1 Introduction
Given a record set D and a query function F , a top-k preference query (top-k
query for short) returns k records from D, whose values of function F on their
attributes are the highest. The top-k query is crucial to many multi-criteria
decision making applications. For example, autonomous robot explore a given
spatial region under dangerous environment and returns the search results as a
sequence of objects ranked by a scoring function. As another example, in self-
determinded soccer robet game, mobile robotic system list an ordered set of

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 804814, 2013.
c Springer-Verlag Berlin Heidelberg 2013

 An Ecient Approach on Answering Top-k Queries 805

objects based on how aggressive the teammate object is or how weakly the rival
objects is matching the score function used in the game.
Existing work on top-k query processing is to avoid searching all the objects
in the underlying dataset, and limiting the number of random accesses to the
spatial data. A key technique to evaluate top-k query is building index on the
dataset. They can be divided into four categories: Sorted list-based, Layer-based,
View-based and Sketch-based.
1. Sortedlists-based. The methods in this category sort the records in each
dimension and calculate the rank by parallel scanning each list until the top-k
results are returned, such as TA[6] and SA[2].
2. Layer-based. The algorithms organize all records into consecutive layers,
such as DG[9],Onion[3] and AppRI[7]. The organization strategy is based on
the common property among the records, such as the same convex hull layer in
Onion[3]. Any top-k query can be answered by up to k layers of records.
3. V iew-based. Materialized views are used to answer a top-k query, such as
PREFER[5] and LPTA[4].
4. Sketch-based. Data stream sketch-based method is used to calculate top-k
query, such as RankCube[8].
In this paper we proposed a hybrid index structure gridded dominant graph
(GDG) method to improve top-k query eciency, which integrates grid index into
dominant graph (DG) index. The rest of this paper is organized as follows. The
next section discusses related work. Section 3 describes the problem description
and definition. Section 4 gives the index structure and algorithm. Section 5
present and analyze the experiments, and Section 6 concludes.

2 Related Works

There are considerable amount of query answering algorithms proposed for top-
k queries in the literatures. The threshold algorithm (TA) [6] is one of the most
common algorithms. TA first sorts the values in each attribute and then scans the
sorted lists in parallel. A threshold value is used to prune the tuples in the rest
of the lists if they cannot have better scores than the threshold. A lot of follow-
on methods have built on the TA algorithm as it addresses a common category
of top-k queries, where the scoring function is a monotone function composed by
using various attributes associated with the objects in the input dataset.
The layer-based method divided all the tuples of the dataset into several
layers; any top-k query can be answered by up to k layers of records. There
are some layered methods, such as DG[9],Onion[3] and AppRI[7]. The Onion
method takes convex hull as origination rule. For given preference liner function,
the result only exists in the convex hull layer. The process of Onion is calculating
the convex hull of the dataset. The first convex hull is calculated and then the
next is calculated on the rest dataset, and so on, until no dataset is left. The
layer rule in AppRI method is the tuple t is put into layer l, if and only if it
satisfy two conditions: a) For any liner function, t is not in the result of top-(l-
1); b) Exists at least one function satisfy t belongs the result of top-l. The layer
806 A. Li et al.

rule in DG is each layer is the former Skyline. The first skyline is calculated,
then the second skyline is computed on the rest dataset, till no dataset is left.
The most important property in DG is the necessary condition that a record
can be in top-k answers is that all its parents in DG have been in top-(k-1)
answers. Benefiting from the property, the search space (that is the number of
retrieved records from the record set to answer a top-k query) of a top-k query is
reduced greatly. The dierence between DG and the other two methods is there
is dominant relationship in DG, so it is not necessary to access and calculate all
the former record set in front of k layer.
The DG method based on k-skyband is to keep adding the candidate points.
Our method is to prune the free points based on Reverse Dominant Point Set. In
other words, if exists k-1 points dominate the point e, all the points dominated
by e can be pruned.

3 Problem Definitions and Theorem

We begin by specifying the notation that will be used in the rest of this paper.
We use D = {O1 , O2 , , ON } to denote a set of objects, the state of each object
is denoted by < id, x1 , x2 , , xd >, where id is the identification of the object,
xi (i = 1, , d) is the ith property of the object. All the properties composed d
dimension space; each state of an object is a point the space.
Definition 1. (Dominate [1]). Given two objects O1 and O2 in a d dimension
space, we say that O1 dominates O2 if and only if they satisfy the following two
conditions: 1) in any dimension i, the value of O2 is larger than or equal to O1 ,
i.e.O2 .xi O1 .xi ; 2) there must exist at least one dimension j, O2 .xj > O1 .xj .
Definition 2. (Aggregate Monotone Function [9]). An aggregation function
f is a monotone if f (x1 , , xn ) f (y1 , , yn ) , whenever xi yi , for every
i.
Both DG and Grid index implement aggregate monotone function top-k prefer-
ence query through order the records in advance.
Definition 3. (Free). If one point or Cell can never be the top-k result, which
neednt to be calculated, then we call it is free.
Definition 4. (Reverse Dominant Point Set). In data points set P, RDP S =
{e} {t|Dominate(t, e), e, t P } , is the RDP S of e, denoted by RDP S(e).
The number of Reverse Dominant Point Set is denoted by RDP SCount(e). The
free points in the top-k query results can be pruned by RDP S, as shown in the
next lemma.
Lemma 1. Let e P , m = RDP SCount(e) , if m = k, then for any monotone
top-k preference query function f , the points dominated by e are free.
 An Ecient Approach on Answering Top-k Queries 807

The proof is omitted.

Definition 5. (Determine Point). Let e, e P , e dominate e, m =
RDSP Count(e), m = RDSP Count(e ), if m < k and m k, we call e is
the Determine Point of top-k query.
Definition 6. (Dominant Graph Index[9]). Given a set D of objects in a
d dimension space, D has k nonempty maximal layers Li , i = 1, , n. The
objects O in ith maximal layer and records O in (i + 1)th layer form a bipartite
graph gi , i = 1, , (m 1).There is a directed edge from O to O in gi if and
only if record O dominates O . We call the directed edge as parent-children
relationship. All bipartite graphs gi are joined to obtain Dominate Graph (DG
for short). The maximal layer Li is called ith layer of DG.
To build a DG oine, any skyline algorithm can be used to find each layer of
DG. Then, the parent-children relationship between records in the ith layer
and that in the (i + 1)th layer. Essentially, DG is a lattice structure, which uses
dominating relationship as partial order, as defined in Ref[9].
As a popular index in data stream management, grid index is very ecient.
The extent of each cell on every dimension is 1 and 2 . Assuming a 2-dimensional
space, the cell Ci,j at column i and row j contains all the tuples whose i 1 <
O.x1 (i + 1) 1 and j 2 < O.x2 (j + 1) 2 , denoted by IN (O, Ci,j ).
Then for each object O, the cell Ci,j it belongs to can be easily computed by i =
O.x1 /1 and j = O.x2 /2 . The right-top point of Ci,j is denoted Ci,j .M ax,
while the left-bottom point of Ci,j is denoted Ci,j .M in.
Definition 7. (Cell Dominate). Let Ca , Cb Cell, if e Ca , s Cb ,
Dominate(e, s) is right, then call Ca Dominate Cb , denoted by (Ca , Cb ). We call
Ca is parent of Cb , and Cb is child of Ca . The Cells which are neither parent
nor child of Ca are called siblings of Cb .
We can determine the dominant relationship between two Cells via Ca .M in,
Ca .M ax, Cb .M in and Cb .M ax. Building the DG index is expensive as it requires
a full scan of the entire dataset. When k is known, we can only build the k layer
skyline index, and then get the results. Due to the property of RDSP, we can
prune more records while build the k layer index.
Definition 8. (k-Max Calculating Region, k-MCR.) If there is Cell Ci ,
the sum of the points in the parent Cell of Ci is m, the sum of the points in
the siblings Cell of Ci is l, and the sum of points in Ci is n, if m < k and
m + n + l k, then Ci is Determine Cell. The Union region of Ci parent and
sibling Cells is k-Calculating Region, k-CR. All the child Cells of Ci are k-
Free Region. The Intersection Region of all the k-CR are the k-Max Calculating
Region, k-MCR.
Theorem 1. For any monotone top-k preference query function f , its results
are in the k-MCR.
Proof is omitted.
Corollary 1. If n < k, then top-n query can be calculated from k-MCR.
808 A. Li et al.

Fig. 1. Schematic diagram of computing k-MCA

Fig. 2. Architecture of GDG index

Let Ca , Cb CELL, if t Ca , s Cb , f (t) > f (s), then a dominates b by

function f , denoted by f Dom(Ca , Cb ).
Property 1. Ca , Cb CELL, if Dominate(Ca , Cb ) , then f Dom(Ca , Cb ).
Property 2. Ca , Cb CELL, for given function f , iff minscore(a)
f maxscore(b) , then f Dom(a, b).
Where f minscore(Ci ) ,f maxscore(Ci ) is the minimum and maximum f value
of Cell Ci .
Example 1. Figure. 1 is 4 4 grid. The a k points are data points, we execute
top-2 query, where a, d, e, f are skyline points, b, g, h are sk-2 points, c, k, i are
sk-3 points. C3,3 is the older brother of C2,3 . The total number of C3,3 and C2,3
is 3, for 3 > 2, then C2,3 is determine cell. The shadowed region in Figure. 1(a)
is 2-CR of C2,3 . The non-shadowed region isnt the calculated region. Similarly,
C3,2 is the determine cell in Figure. 1(b). The shadowed region in Figure. 1(b) is
2-CR of C2,3 . The non-shadowed region isnt the calculated region. The shadowed
region in Figure. 1(c) is 2-CR of intersection of C2,3 and C3,2 , i.e. 2-M CA.
Definition 9. (Max Search Region by f , k-M SRf ). For top-k function f ,
if there is Cell Ci , the sum of the f -parent Cell of Ci is m, the sum of f -siblings
Cell of Ci is l, and the sum of points in Ci is n, if m < k and m+ n+ l k, then
Ci is f -Determine Cell. The Union region of Ci f -parent and f -sibling Cells is
the k-Max Search Region by f , k-M SRf .
 An Ecient Approach on Answering Top-k Queries 809

That is, when f is given, we can improve the eciency more by pruned the
redundant Cells in the k-M CR.
Theorem 2. For any monotone top-k preference query function f , its results
are in the k-M SRf .
Proof is omitted.

4 GDG Index
4.1 Grid Index Structure
The GDG index is based on grid index and DG index, as shown in Figure. 2.
There are two parts in the GDG index, which are DG and k-MCR. We main-
tain the layer and dominate relationship between the data point dp. Each data
point dp is composed of id, cid (cell id), the pointer set of its children chd, and
the next pointer to the next data in the same layer sk-i. The sk-i( i = 1, , k)
(skyline-i) , represents the ith skyline of the DG index. The same sk-i and the
data points dp in the same cell are stored in the continuous memory via link
structure. There is one GL link table in the k-MCR, each node of the link table
is the cell in the grid. The basic information Inf and link information with DG
DG-link is stored in the cell.
Theorem 3. GCG index correctly finds top-k answers.
Proof is omitted.

4.2 Initial Computation Module

We calculated the k-MSRf through the initial computation module(CM), and
can get the approximate order relationships among the grid cells. Then we can
obtain the free cells in the top-k queries about function f , which can be pruned.
It is shown in algorithm 1.
As shown in algorithm 1, Lines 1 to 2 calculate the maximum and minimum
point of cell C from the k-MCR points. Lines 5 to 11 calculate the k-MCRf of
the cells. Because the results of the data sets function f are mapped into one
dimensional space, it is simple to calculate the f -relationship between the data
points. For lines 8 to 9 read the ordered cells, we can judge cell C is f determine
cell without judge the father of cell C.

4.3 GDG Query Module

After calculated k-MCRf , we can get the result sets of top-k query must be
located in the data points in the region of the k-MCRf . The data points outside
the region are free, can not appear in the result set.
As shown in algorithm 2, the top-k query algorithm is similar with DG, with
the kMSRf points are pruned. Line 1 read the data in the first layer Sk 1 of
DG, pruned the points without in the kMSRf region, and put the results into
810 A. Li et al.

Algorithm 1. Algorithm of Initial Computing k M SRf

Input: Data Set P ;
Output: kM CRf of P ;
1: for each cell C in k-MCR do
2: Calculate f -range [[f minscore(C), f maxscore(C)]] of C;
3: end for
4: Sort cells in k-MCR by f -range;
5: kM CRf =kM CR;
6: while kM CR is not null do
7: Read first cell C in kMCR;
8: kMCR= kMCR - C;
9: t = f F atherDP Count(C) + f ElderBrotherDP Count(C);
10: if t k then
11: kM CRf = kM CRf C.f children;
12: kM CR = kM CR C.f children;
13: end if
14: end while

the CL list. The data in CL are ordered by the f value, and the count of the
data in CL is k n, where n is the number of result set(RS). Line 1 put the
data have largest R value into RS. Lines 2 to 21 calculated the rest of the RS
one by one. The proof of algorithm 2s correctness is omitted.

Algorithm 2. k-MCR Traveler Algorithm

Input: A GDG of the database D, top-k query function f and kMSRf ;
Output: the top-k answers;
1: Move the first largest record R in CL into result set RS;
2: Set n=1;
3: while kMSRf is not NULL and RS.size() < k do
4: for each child C of R do
5: if C has at least one parent that is not in RS or C has been computed using
query function f before or not in kMSRf then
6: Continue;
7: else
8: Compute C by query function f and insert it into CL;
9: end if
10: if CL.size() > k n then
11: Only keep the first k n records in CL;
12: end if
13: for each cell Ce in kMSRf do
14: if Ce.M ax < min(CL) then
15: Delete cell Ce from kMSRf ;
16: end if
17: end for
18: end for
19: Move the first largest record R in CL into the answer set RS.
20: Set n=n + 1;
21: end while
 An Ecient Approach on Answering Top-k Queries 811

4.4 Space Complexity and Time Complexity

We firstly analysis the space complexity of the algorithm. Suppose the dimension
of data set is d, the m is the number of intervals in index. We can get the total
number of cells is M = md . Each data point needs d memory units. The ratio of
k-MCR in total cells is P , with two points in this type cells, one is total numers,
the f -value of this cell. So the number of total memory units is md P (2d + 4).
Let the ratio of k-MSRf in total cells is P , where cells f -value is mapped to
one cell ID, is md P . The number of k-MSRf access points in GDG is s d.
So, the space complexity of our method is O(md ), where m is the number of
intervals in index, d is the dimension.
Then we analysis the time complexity. Suppose the distribution of the data
is uniform distribution, then the number of points in each cell is about N/md .
During the process of building index k-MCR, the complexity of k-MCR equals
scan the data set time plus find the determine point time. Let N is the sum of
the data points in the window, then scan the data set time is O(N ). Find the
determine point time is find the cell.M in point. The worst case is find the whole
cell set, the time is O(M ), where M is the total number of cells in the Grid
Index. If N is large, and M is small, then the complexity of k-MCR is O(N ),
where N is the window size. Similarly, we can get the complexity of k-MSRf
is O(M P + log(M P )), approximately equal O(M ), where M is the total
number of cells in the Grid Index.

5 Experimental Result
All experiments were carried out on a PC with Intel Core2 6400 dual CPU
running at 2.13GHz and 2GB RAM. The operating system is Fedora with the
Linux kernel version 2.6.35.6. All algorithms were implemented in C++ and
compiled by GCC 4.5.1 with -O3 flag. Our experiments are conducted on both
real and synthetic datasets.
Synthetic datasets are generated as follows. We first use the methodologies
in [10] to generate 1 million data elements with the dimensionality from 2 to 5
and the spatial location of data elements follow independent distribution. Syn-
thetic datasets are represented with Ui , which have 1000K tuples, the granularity
is 8000, with normal distribution between [0,7999]. The real data dataset is from
http://archive.ics.uci.edu/ml/machine-learning-databases/ covtype/, 54 dimen-
sions, 10 dims are numeric. We choose 3 numeric variable as dimension, while
the granularity is 1989, 5787 and 5827 respectively.
We first compare the space and time cost during building the index between
DG and GDG.
The method we compared with is RankCube[4] and DG[9], the dataset is of-
fline. Our methods are GDG40 and GDG80, where each dimension is divided
into 40 intervals and 80 intervals. As synthetic data (3 dimension) in Figure. 3(a),
812 A. Li et al.

Fig. 3. Reduce the size of index

Fig. 4. Reduce the number of access data

when answer top-100 queries, the DG index is 10K 572K tuples, the maximum
ratio is 57.2%. While the GDG40 index is 9K 74K tuples, the maximum ratio is
7.4%, and the GDG40 index is 8K 39K tuples, the maximum ratio is 3.9%. As
real data in Figure. 3(b)(580 tuples), when answer top-100 queries, the DG index
is 216K tuples, the ratio is 37%. While the GDG40 index is 19K 34K tuples, the
maximum ratio is 6%, and the GDG80 index is 6K 24K tuples, the maximum
ratio is 4%. So, our methods can reduce the size of index obviously.
Secondly, the comparisons of the number of access data, query response time
among our method and RankCube[4] , DG[9] is completed.
As shown in Figure. 4 and Figure. 5, for each function f has dierent number
of access data and query response time, we choose 10 functions, then take average
value as the result. As shown in Figure. 4, our methods have less number of access
data than RankCube[4] and DG[9], where GDG80 is the best.
As shown in Figure. 5, our methods have better performance than the exist-
ing RankCubee[4] and DG[9] method, both using synthetic data and real data
environments.
We can draw conclusions from the experimental results as follows. Firstly,
some data sets accessed in DG method are free, which are pruned in GDG
method. So GDG method can decrease the number of access points and response
time. Secondly, the RankCube method access the data set unordered and have
 An Ecient Approach on Answering Top-k Queries 813

Fig. 5. Query time response Comparison

no hierarchical relationships, but GDG method did. So the number of access

points and response time of GDG method is better. Finally, Compared GDG80
and GDG40, the former one has more intervals, the result is more precise, so the
the number of access points is better. But the time of precalculated kMSRf is
increased a little.

6 Conclusions
The top-k query is well used in spatial dara or searching multi-dimensional data
stream to recognize important objects. A grid-index based on reserve data points
set (RDP S) is presented to calculate multi-dimensional preference top-k. Be-
cause of the correlation between top-k and skyline computation, when the count
of one points RDP S is bigger than or equals to the value of k, it is for sure to be
pruned the points dominant set. RDP S can be calculated with grid-index algo-
rithm quickly. Combining grid-index and dominant graph index, we get gridded
dominant graph (GDG) index. Analytical and experimental evidences display
that GDG index approach performs better on both storage of index and e-
ciency of queries.
The GDG index we proposed in this paper can clearly reduce the index size,
the random access of the dataset and increase the performance of top-k queries,
which can satisfy the critical performance demand in robots environments.
In our future work, we will study how apply RDSP in other layered index,
such as Onion and AppRI, to reduce random access number of datasets. Morever,
we will use our methods to analysis online data streams, which can be used in
the mobile robots exploring in the computational complicated environments.

Acknowledgments. The authors acknowledge the financial support by State

Key Development Program of Basic Research of China (No. 2013CB329601),
National Key Technology R&D Program (No. 2012BAH38B-04), 863 program
(No. 2010AA -012505, 2011AA010702) and NSFC(No. 60933005). The author is
grateful to the anonymous referee for a careful checking of the details and for
helpful comments that improved this paper.
814 A. Li et al.

References
1. Borzs
onyi, S., Bailey, M., Kossmann, D., Stocker, K.: The Skyline Operator. In:
Proc. 17th International Conference on Data Engineering, pp. 421430 (2001)
2. Chang, K.C., He, B., Li, C., Patel, M., Zhang, Z.: Structured databases on the
web: observations and implications. SIGMOD Rec., 6170 (2004)
3. Chang, Y.C., Bergman, L., Castelli, V., Li, C.S., Lo, M.L., Smith, J.R.: The onion
technique: indexing for linear optimization queries. SIGMOD Rec., 391402 (2000)
4. Das, G., Gunopulos, D., Koudas, N., Tsirogiannis, D.: Answering top-k queries
using views. In: Proc. 32nd International Conference on Very Large Data Bases,
pp. 451462 (2006)
5. Hristidis, V., Koudas, N., Papakonstantinou, Y.: PREFER: a system for the e-
cient execution of multi-parametric ranked queries. In: Proc. 2001 ACM SIGMOD
International Conference on Management of Data, pp. 259270 (2009)
6. Nepal, S., Ramakrishna, M.V.: Query Processing Issues in Image(Multimedia)
Databases. In: Proc. 15th International Conference on Data Engineering, pp. 2229
(1999)
7. Dong, X., Chen, C., Han, J.W.: Towards robust indexing for ranked queries. In:
Proc. 32nd International Conference on Very Large Data Bases, pp. 235246 (2006)
8. Dong, X., Han, J.W., Cheng, H., Li, X.L.: Answering top-k queries with multi-
dimensional selections: the ranking cube approach. In: Proc. 32nd International
Conference on Very Large Data Bases, pp. 463474 (2006)
9. Zou, L., Chen, L.: Dominant Graph: An Ecient Indexing Structure to Answer
Top-K Queries. In: Proc. 24th International Conference on Data Engineering, pp.
536545 (2008)
10. Li, A.P., Han, Y., Zhou, B., Han, W., Jia, Y.: Detecting Hidden Anomalies Using
Sketch for High-speed Network Data Stream Monitoring. Appl. Math. Inf. Sci. 6(3),
759765 (2012)
 A Survey on Clustering Techniques
for Situation Awareness

Stefan Mitsch2 , Andreas Muller1 , Werner Retschitzegger1 , Andrea Salfinger1 ,

and Wieland Schwinger1
1
Johannes Kepler University Linz, Altenbergerstr. 69, 4040 Linz, Austria
2
Computer Science Dept., Carnegie Mellon University, Pittsburgh, PA-15213, USA

Abstract. Situation awareness (SAW) systems aim at supporting as-

sessment of critical situations as, e.g., needed in trac control centers,
in order to reduce the massive information overload. When assessing sit-
uations in such control centers, SAW systems have to cope with a large
number of heterogeneous but interrelated real-world objects stemming
from various sources, which evolve over time and space. These specific
requirements harden the selection of adequate data mining techniques,
such as clustering, complementing situation assessment through a data-
driven approach by facilitating configuration of the critical situations to
be monitored. Thus, this paper aims at presenting a survey on clustering
approaches suitable for SAW systems. As a prerequisite for a systematic
comparison, criteria are derived reflecting the specific requirements of
SAW systems and clustering techniques. These criteria are employed in
order to evaluate a carefully selected set of clustering approaches, sum-
marizing the approaches strengths and shortcomings.

1 Introduction

Situation Awareness (SAW). SAW systems are increasingly used in large

control centers for air or road trac management in order to reduce the infor-
mation overload of operators induced by various data sources. This is done by
automatically assessing critical situations occurring in the environment under
control (e. g., an accident causing a trac jam) [3].
Data Mining (DM) for SAW. The definition of relevant critical situations
is provided in current SAW systems (e.g., [3,26]) explicitly by domain experts
during a configuration phase, representing a time-consuming task. This eort of
providing explicit knowledge could be complemented and eased by a data-driven
approach in terms of DM techniques making direct use of the observed data,
through detecting interesting or uncommon relationships, which the domain
experts might not be explicitly aware of (i.e., intrinsic knowledge). Furthermore,
during runtime of the SAW system ongoing changes and additional uncommon

This work has been funded by the Austrian Federal Ministry of Transport, Innovation
and Technology (BMVIT) under grant FFG FIT-IT 829589, FFG BRIDGE 838526
and FFG Basisprogramm 838181.

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 815826, 2013.
c Springer-Verlag Berlin Heidelberg 2013

816 S. Mitsch et al.

relationships can be spotted that have not even been explicitly defined so far.
Especially clustering techniques are considered to be beneficial, since they require
neither a-priori user-created test data sets nor other background knowledge and
also allow anomaly detection, such as uncommon relations. If we consider a
road trac SAW system, clustering might for example be used to detect (ST)
hotspots on a highway, i. e., road segments and time windows where atypically
many accidents occur, or to reveal current major trac flows in an urban area.
Specific Requirements of SAW. The nature of SAW systems, however, poses
specific requirements on the applicability of existing clustering techniques. SAW
systems have to cope with a large number of heterogeneous but interrelated
real-world objects stemming from various sources, which evolve over time and
space, being quantitative or qualitative in nature (as, e.g., demonstrated in [3]).
The focus of this paper is therefore on evaluating existing spatio-temporal (ST)
clustering techniques with respect to their ability to complement the specification
of critical situations in SAW systems.
Contributions. In this paper, we first systematically examine the requirements
on clustering techniques to be applicable in the SAW domain. Then, we survey
several carefully selected approaches according to our criteria stemming from
the fields of SAW and ST clustering, and compare them in our lessons learned,
highlighting their advantages and shortcomings.
Structure of the Paper. In the following section, we compare our survey
with related work (cf. Section 2). Then, we introduce our evaluation criteria
(cf. Section 3), before we present lessons learned (cf. Section 4). Due to space
limitations, the in-depth criteria-driven evaluation of each approach surveyed
can be found online1 only.

2 Related Work
Despite a plethora of work exists in the field of data mining, to the best of our
knowledge no other survey has dealt with the topic of clustering ST data for
SAW applications so far. However, there exist surveys about ST clustering (e. g.,
[21]) and ST data mining in general (e. g., [19], [29]), surveys about spatial (e. g.,
[11]) and temporal (e. g., [34]) clustering and various surveys on clustering in
other domains (e. g., [15], [36]). Furthermore the field of stream data mining has
received a lot of attention over the last few years and as a result several surveys
on stream data clustering (e. g., [20]) and on stream data mining (e. g., [8], [13])
in general have been conducted.
In the following paragraphs these mentioned surveys are briefly discussed with
respect to the contributions of our survey.
ST Clustering. Kisilevich et al. [21] propose a classification of ST data and fo-
cus their survey of clustering techniques on trajectories, being the most complex
setting in their classification. The main part of the survey is a discussion of sev-
eral groups of clustering approaches including dierent example algorithms for
1
http://csi.situation-awareness.net/stc-survey
 A Survey on Clustering Techniques for Situation Awareness 817

each group. Finally, they present examples from dierent application domains
where clustering of trajectory data is an issue, like studying movement behavior,
cellular networks or environmental studies. In contrast to our survey, Kisilevich
et al. focus on presenting various groups of ST data mining approaches, rather
than systematically analyzing dierent techniques backed by a catalog of evalu-
ation criteria. Furthermore, we considered the applicability of the techniques to
our domain on basis of SAW-specific criteria, while they conducted their analysis
in a more general way. Nevertheless, their work represented a valuable starting
point for our survey, from which we especially adopted their classification of
ST data, like trajectories or ST events.
Spatial, Temporal and General Clustering. Several surveys on spatial clus-
tering (e. g., [11]) and temporal clustering (e. g., [34]) have been conducted for
dierent application domains. Nevertheless, none of the surveyed approaches
specifically deals with the unique characteristics that arise from the domain of
SAW respectively from ST data in general (cf. Section 3.1), but take only either
spatial- or temporal characteristics into account.
Finally, there exist numerous and comprehensive surveys on non-ST clustering
(e. g., [15], [36]). These approaches cannot simply be used to work with ST data,
but have to be adopted manually to deal with its particular nature.
ST Data Mining. Kalyani et al. [19] describe the peculiarities of ST data
models and the resulting increase of complexity for the data mining algorithms.
In their survey they outline dierent data mining tasks compared to their spatial
counterparts and motivate the need for dedicated ST data mining techniques.
Geetha et al. [29] present a short overview of challenges in spatial, temporal and
ST data mining. However, none of the above focuses on concrete data mining
techniques, but rather gives an overview of the field of ST data mining.
Stream Data Mining. In their survey on clustering of time series data streams,
Kavitha et al. [20] review the concepts of time series and provide an overview
of available clustering algorithms for streaming data. More general surveys of
stream data mining were conducted by Gaber et al. [8] and Ikonomovska et al.
[13]. Both review the theoretical foundations of stream data analysis and give
a rough overview of algorithms for the various stream data mining tasks and
applications. Furthermore, they mention that stream data clustering is a major
task in stream data mining and lots of algorithms have been adapted to work
on streaming data (e. g., CluStream [1], DenStream [7], ClusTree [22]).
However, in contrast to our SAW systems which store a history of the data
received, thus allowing an arbitrary number of read accesses, such stream data
mining approaches do not store the huge amount of data they process. Even
though the elements of a data stream are temporally ordered, they might ar-
rive in a time-varying and unpredictable fashion and do not necessarily contain
a timestamp, which does not allow for the identification of temporal patterns
(especially cyclic ones). Furthermore, most stream data mining approaches sur-
veyed by these authors do not deal with spatial data at all and hence are not
applicable here.
818 S. Mitsch et al.

Consequentially, we conduct this survey on ST clustering to investigate the

applicability of various clustering techniques for the field of SAW.

3 Evaluation Criteria
In this section, we derive a systematic criteria catalog, (methodologically ad-
hering to some of our previous surveys, e.g., [35]), which we use for evaluating
selected ST clustering techniques. The criteria catalog consists of two sets of cri-
teria, as depicted in Fig. 1, comprising SAW-specific criteria (cf. Section 3.1) and
clustering-specific criteria (cf. Section 3.2), which are detailed in the following.
Each criterion is assigned an abbreviation for reference during evaluation.

Fig. 1. The evaluation criteria at a glance

3.1 SAW-Specific Criteria

The rationale behind the following SAW-specific criteria, which are illustrated
by examples from the domain of road trac control, is based on considering the
kind of input data a clustering technique has to cope with (What do we have)
and further on reflecting on the intent of the analysis (What do we want).
Spatial, Temporal and Other Input Data Properties. As already men-
tioned, SAW systems employed in control centers need to monitor a large number
of interrelated objects anchored in space and time, both either with an extent
or without. In particular the temporal properties (TP) range from instants to
intervals, whereas the spatial properties (SP) comprise points, one-dimensional
intervals (e. g., lines) or two-dimensional intervals (e. g., regions). Non-ST prop-
erties can be of qualitative (e.g., freezing temperature) or quantitative (e.g.,
0 C) nature. Regarding SP and TP, however, we assume quantitative data, as
typically required by SAW applications to monitor the ST environment.
Heterogeneity of Input Data. Another SAW-specific requirement is that
objects constituting a certain critical situation often exhibit a mixture of dierent
ST properties. This can be exemplified by a situation involving three objects,
(i) a bus, sending location points at time instants via GPS to the control center
(i. e., a trajectory), (ii) a trac jam, comprising a spatial and temporal interval,
both evolving over time (i. e., growing, shrinking, moving) and (iii) a fog area
being spatially extended over a certain region and characterized by a temporal
interval (e. g., predicted for two hours). The more dierent ST properties data
 A Survey on Clustering Techniques for Situation Awareness 819

a clustering technique supports, i.e., allows for heterogeneous input data (HID),
the better it is applicable to the SAW domain.
Evolution of Input Data. As indicated in the examples above, the temporal
dimension furthermore captures the potential evolution of the observed objects
(E) with respect to their spatial (e.g., location or length) or non-spatial (e.g.,
temperature) properties. An evolution along the spatial dimension corresponds
to a mobile, i. e., moving, object, whereas an object solely evolving with respect
to the non-spatial properties represents an immobile, i.e., static, object.
Context of Input Data. The input data used in SAW systems often com-
prises objects which are bound to a certain context (CID), enforcing constraints
on the interpretation of the input data (e. g., in road trac, the majority of
objects cannot move around in space freely, but is bound to the underlying road
network), or further requirements on the input data (e. g., necessity of velocity
information). We evaluate if and what kind of context information is required.
Fuzzy Input Data. SAW systems often have to cope with fuzzy input data
(FID), for example incidents reported by humans who only have a partial over-
view of the situation. Fuzzy input data might address the spatial properties (SP)
(e.g., uncertainty about the exact location of an object), the temporal proper-
ties (TP) (e.g., an accident has occurred within the last half an hour), or the
non-ST properties (e.g., it cannot be defined exactly what has happened).
Besides the criteria mentioned above, which detail the nature of the input
data, the following further deal with the goal of the analysis.
Intention. This criterion (I) reflects the objective of the analysis, i.e., the kind of
implicit knowledge that should be extracted by the clustering technique. Possible
intentions are clustering of events or regions with similar ST characteristics,
clustering of trajectories, or the detection of moving clusters.
Online or Oine Analysis. The requirements imposed on the clustering tech-
niques dier with respect to the phase they should be employed. During the con-
figuration phase of a SAW system, we have to perform an analysis on a complete,
historical data set, i. e., oine analysis. However, since an SAW system typically
operates on a real-time environment, we also want to perform an analysis at run-
time, i. e., online analysis. Since clustering techniques devoted to online analysis
often rely on optimizations and approximations in order to deliver fast results
(e.g., compute only locally optimal clusters), these approaches are less suited
for configuration tasks where computation time is not a major issue, whereas
an exact result is preferred. Thus, this criterion (OOA) reflects whether the clus-
tering technique is only suited for the oine configuration phase, or if it is also
applicable to or favored for runtime analysis.

3.2 Criteria Imposed by Clustering Techniques

Whereas in the previous section, the criteria have been derived from the na-
ture of data and the intention of the analysis stemming from the SAW domain,
the present section focuses on assessing how these goals can be achieved with
dedicated ST clustering techniques.
820 S. Mitsch et al.

We first give a short description of the main contribution (MC) and distinguish
the clustering techniques according to their algorithm class (AC), describing the
method how the clusters are obtained. Following [12] this can be, due to par-
titioning (i.e., the data space is partitioned as a whole into several clusters),
hierarchical clustering (i. e., clusters are created by merging close data points
bottom-up or splitting clusters top-down), density-based clustering (i. e., clus-
ters are defined as exceeding a certain density of data points over a defined
region), and grid-based clustering (i. e., the data space is overlaid by a grid, grid
cells containing similar structures are merged).
The remaining criteria are structured into functional and non-functional ones.
Functional Criteria. The distinct algorithmic methods yield dierent cluster
shapes (CS), distinguishing spherical, rectangular or arbitrarily shaped clusters.
Besides, it is investigated if a technique can handle clusters which overlap (CO).
For analysis properties, we consider spatial analysis properties (SA), i. e., what
kind of spatial data is internally processed by the algorithm, temporal analysis
properties (TA), i. e., time instants or intervals, temporal patterns (TAP), i. e.,
linear or cyclic time patterns, and the focus of the analysis (F), i. e., if spatial
and temporal aspects are handled equally or if one is favored without ignoring
the other. We evaluate if the techniques support dedicated handling of noise (N)
within the data set and furthermore investigate the employed similarity measures
(SM) and the determinism of the produced results (D).
Non-functional Criteria. These criteria comprise the configurability (C) of the
approaches, reflecting to which degree the clustering technique can be tweaked, and
the computational complexity (CC). Furthermore, we examine whether and which
optimization strategies (O) are provided, which would be beneficial in reducing run-
time, however might also aect the quality of the obtained clustering result.

4 Survey of ST Clustering Techniques

In the following we present our selection of techniques, and group and analyze
them according to our criteria.

4.1 Rationale behind the Selection of Techniques

We carefully selected ST clustering approaches ranging from very recent ones
on the one hand, to several more mature ones on the other hand to provide a
broad overview of the field of ST clustering. We did not include visually aided
approaches (e. g., Andrienko et. al [2]) in our survey as they contradict our
approach to complement the configuration of SAW systems in a more data-
driven way because a purely user-guided clustering approach again has to be
steered by a domain expert.
In the following, we structure our discussion into three groups according to the
kind of evolution the techniques consider. The rationale behind these categories
is that algorithms from the dierent groups share several properties, which is
reflected in the evaluation criteria tables.
 A Survey on Clustering Techniques for Situation Awareness 821

Input Data Intention (I)

Evolution Properties No Analysis
(E) (HID) Evolution (OOA)

Cluster Trajectories
Cluster ST-Regions
Spatial Temporal

Cluster ST-Events

Moving Clusters
Spatial (SP) (TP)

Immobile
Fuzzi-

Interval
Mobile

Instant
Region

Online

Offline
ness Context

Point
Line
(FID) (CID)
Wang, 2006 3 - 3 - - 3 - - - 3 - - - - 3
Evo.
No

Birant, 2006 3 - 3 - - 3 - - - - 3 - - - 3
Tai, 2007 3 (3) 3 - - 3 - - - 3 - - - 3 3
Gaffney, 1999 - 3 3 - - 3 - - - - - 3 - - 3
Nanni, 2006 - 3 3 - - 3 - - - - - 3 - - 3
Object
Evo.

Li, 2010 - 3 3 - - 3 - - - - - 3 - 3 3
Lu, 2011 - 3 3 - - 3 - - - - - 3 - - 3
Gariel, 2011 - 3 3 - - 3 - - - - - 3 - 3 3
Kalnis, 2005 - 3 3 - - 3 - - - - - - 3 3 3
Iyengar, 2004 3 - 3 - - 3 - - - 3 - - - - 3
Neill, 2005 3 - 3 - - 3 - - - 3 - - - - 3
Li, 2004 - 3 3 - - 3 - - requires velocity - - - 3 - 3
Evolution
Cluster

Jensen, 2007 - 3 3 - - 3 - 3 requires velocity - - - 3 3 -

Chen, 2007 - 3 3 - - 3 - - spatial network & velocity - - - 3 3 -
Rosswog, 2008 - 3 3 - - 3 - - - - - - 3 - 3
Jeung, 2008 - 3 3 - - 3 - - - - 3 3 - 3
Rosswog, 2012 - 3 3 - - 3 - - - - - - 3 - 3
Zheng, 2013 - 3 3 - - 3 - - - - - 3 3 (3) 3

Fig. 2. Situation Awareness Criteria Table

The first group comprises techniques that do not entail any evolution (NE),
like clustering ST events (i. e., grouping events in close or similar regions) or
ST groups (i. e., finding regions sharing similar physical properties over time).
Techniques for trajectory clustering (i. e., grouping similar trajectories) deal with
the evolution of objects that move along these trajectories (OE). And finally, the
area of moving-object clustering (i. e., discovery of groups of objects moving to-
gether during the same time period, so called moving clusters) and the detection
of spatio-temporal trends (e.g., disease outbreaks) deal with the evolution of
clusters over time (CE).

4.2 Lessons Learned

Our evaluation of ST clustering approaches for the field of SAW has revealed
interesting peculiarities of current clustering techniques. In the following, we
explain our findings grouped according to our evaluation criteria (cf. also Figure
2, 3 and 4). Note that a tick in parentheses means that the criteria is only partly
fulfilled by the approach.
Evolution Mostly Supported (E). Algorithms from the NE group work with
objects in fixed locations and thus do not support spatial evolution at all. All
other techniques allow for moving objects.
Spatial and Temporal Extent Not Handled (SP, TP). None of the sur-
veyed techniques can handle anything but spatial points or temporal instances
822 S. Mitsch et al.

lines, rectangles or temporal intervals are not dealt with. Hence, currently none
of the approaches is able to directly deal with the heterogeneity (HID) criterion.
Fuzziness of Data Is Not an Issue (FID). All of the approaches treat objects
as facts and do not consider any uncertainties, except Jensen et al. [16] who
consider that objects might disappear without notifying the server and reduce
the confidence in the assumed object movement as time passes.
Almost NO Context Knowledge Supported (CID). Most of the algo-
rithms work without any knowledge of context and do not allow any further
information than a data set and parameter values. Exceptions are techniques
that cluster the objects backed by a network graph (e. g., [5]), or techniques that
make use of velocity information for moving objects (e. g., [16], [23]).
Techniques Mainly Focus on Oine Clustering (OOA). While the sur-
veyed techniques mainly aim at oine clustering of closed data-sets, only few
exceptions (from each group) allow online clustering of ever-changing data-sets
(e. g., [32], [10], [16]).
Majority of Algorithms Is Density-Based (AC). Our survey comprises
clustering techniques from all classes of algorithms, whereby OE techniques are
usually density based, while CE techniques cover all algorithm classes.
Predominance of Arbitrarily Shaped Clusters (CS). OE and NE techniques
mostly result in arbitrary clusters, whereas clusters produced by CE techniques
are often restricted to spherical or rectangular shape (e. g., [14], [30]).
Cluster Overlap in CE Techniques (CO). Overlapping clusters are only
considered by few of the techniques mostly from the CE group (e. g., Kalnis et
al. [18]). These approaches are able to detect dierent clusters moving through
each other and can keep them separated until they split again.
Spatial Analysis Properties Highly Dependent on Intent of Analysis
(SA). While NE and CE techniques are focused on clustering data points, OE
techniques mostly deal with clustering of lines, since a trajectory can be inter-
polated to a line. Li et al. [23] first group the objects into micro-clusters (i. e.,
small regions) and then combine these regions to complete clusters. However, as
they cluster the micro-clusters center points, they also deal with points only.
Only Temporal Instants (TA) and Linear Patterns (TAP). Linear pat-
terns for time instants are predominant in the surveyed approaches, while other
temporal properties like cyclic time patterns or intervals are unhandled in the
majority of cases, especially by CE techniques. Only Birant et al. [4] include
cyclic time patterns, and Nanni et al. [27] consider time intervals.
Focus on ST data (F). Most of the techniques are focused on handling the
specific nature of ST data. However, there are several spatial-data dominant
techniques originally stemming from the field of spatial clustering, enriched with
the ability of dealing with temporal data aspects (e. g., [4], [32]). Only Nanni et
al. [27] presented a temporal-data dominant variation of their algorithm.
Sparse Explicit Noise Handling (N). As most of the proposed techniques
extend well-known clustering algorithms to work with ST data, the handling
 A Survey on Clustering Techniques for Situation Awareness 823

General Information Non-functional Criteria

Algorithm
Class (AC)

Hierarchical
Partitioning

Grid-based
Computational

Density-
based
Main Contribution Configurability Complexity Optimization strategies
(MC) (C) (CC) (O)
Wang, 2006 DBSCAN-based / grid based ST-clustering - - 3 3 calculated with k-dist graph O(n) / O(n) fast vs. precise
Birant, 2006 min # points density improved R-Tree, filters to
DBSCAN-based ST-clustering - - 3 - O(n*log n)
Evo.
No

similarities reduce search space

Tai, 2007
incremental ST-clustering of dual data 3 - - - mostly determined automatically not specified NeiGraph; fast vs. precise
Gaffney, 1999 O(N K I)
clustering trajectories through regression 3 - - - # clusters -
(K# clust., I# iter.)
Nanni, 2006 min # points time window O(n*log n) /
clustering trajectories with focus on time - - 3 - M-Tree
similarities O(n_iter*n*log n)
Li, 2010 incremental clustering trajectories with micro- distance threshold micro-clusters,
Object

- - 3 - not specified
Evo.

clusters parameters for macro-clustering kinetic heap

Lu, 2011
clustering trajectories through semantic density threshold distance radius
- - 3 - not specified -
regions min # points
Gariel, 2011 detection of aircraft waypoints and nominal DBSCAN parameters # PCs
- - 3 - not specified -
trajectories, aircraft anomaly detection # sample points
Kalnis, 2005 DBSCAN parameters pruning redundant cluster
detecting moving clusters - - 3 -
LQWHJULW\WKUHVKROG comb./approximations
Iyengar, 2004 max. population candidate size
detecting evolving space-time clusters - - - 3 not specified -
nr. of search iterations
Neill, 2005 temporal window size
detecting evolving space-time clusters - - - 3 O(NST+N4T) -
level of data aggregation
Li, 2004 clustering moving objects into micro-clusters O((N+U)*log(N+U)*
- 3 - - split threshold heuristics
to enable generic clustering log(N)+N*log(N))
Jensen, 2007
max. update time/cluster capacity disk-based data structure,
continuous moving-object clustering - 3 - - not specified
time window size event queue, hash tab.
Evolution
Cluster

Chen, 2007 clustering moving-objects on a network 3 - 3 - max distances not specified cluster blocks
Rosswog, 2008 detecting and tracking spatio-temporal order of the filter, filter type (flat, adaptive, results improve as history
3 - - - not specified
clusters with adaptive history filtering adaptive+stability) grows
Jeung, 2008 min # of cluster objs.
clustering of convoys, i.e., groups of objects clustering on simplified
distance value e
that travel together for at least a min. - - 3 - not specified trajectories for detection of
OLIHWLPHNOHQJWKRIWLPHSDUWLWLRQ
duration of time convoy candidates
WROHUDQFHRIWUDMHFWRU\VLPSOLILFDWLRQ
Rosswog, 2012 DBSCAN-based ST-clustering backed by a parameterized kernel DBSCAN Relationship graph:
- - 3 - R-Tree
SVM training data set required O(n*log n)
Zheng, 2013 DBSCAN parameters
YDULDWLRQWKUHVKROG diff. spatial indexing
clustering of gatherings - - 3 - not specified
lifetime threshold of crowd/participator techniques, pruning
support threshold of crowd/gathering

Fig. 3. Clustering Criteria Table (part 1)

of noise highly depends on the abilities of the underlying base algorithm. An

example for an often used algorithm is DBSCAN [6], which is generally tolerant
to noise, although it is prone to errors if clusters of dierent densities exist. Only
Birant et al. [4] and Rosswog et al. [30] explicitly address the topic of noise and
inconsistencies by proposing an additional density parameter or a stability filter.
As detection of noise might be used to find ST outliers, approaches that deal
with anomalies might be applicable in dierent ways within an SAW system.
Euclidean Distance as Similarity Measure (SM). Regarding similarity
measures, most authors make use of an adjusted version of the Euclidean distance
(ED) (e. g., weighted ED, squared ED). Obvious exceptions are Chen et al. [5]
who operate on a network graph and thus use a network distance and some
special approaches that do not require a similarity measure at all (e. g., [33]).
A border case is the similarity measure used by Tai et al. [32], who define their
own cost function looking similar to the ED, considering only (what they call)
optimization attributes, thus excluding spatial components.
Mostly Deterministic Techniques (D). Only few of the techniques use
heuristic approaches to deliver fast, but more inexact results (e. g., Iyengar et al.
[14] who propose a heuristic cluster search because of the huge search space).
Various Levels of Configurability (C). The number of parameters ranges from
none or automatically approximated to up to seven parameters. For instance,
824 S. Mitsch et al.

Functional Criteria
Cluster Analysis Properties
Temporal (TA) Deter-
Spatial Pattern Focus minism
(SA) (TAP) (F) (D)

Overlap (CO)

deterministic
spatio-temp.
Shape (CS)

temporal

heuristic
interval
Instant
Region

spatial
Linear

Cyclic
Noise Similarity Measures

Point
Line
(N) (SM)
Wang, 2006 A - 3 - - 3 - 3 - - - 3 - n.a. / arbitrary 3 -
Evo.
No

Birant, 2006 A - 3 - - 3 3 3 - 3 - - density parameter arbitrary (ED used for evaluation) 3 -

Tai, 2007 A - 3 - - 3 - 3 - 3 - - - cost function 3 -
Gaffney, 1999 A - - 3 - 3 - 3 - - - 3 - Membership probabilities - 3
Nanni, 2006 A - - 3 - 3 - - 3 - 3 3 - average ED between objects 3 -
Object
Evo.

Li, 2010 A - - 3 - 3 - 3 - - - 3 micro-cluster extent composed three-part distance 3 -

Lu, 2011 A - - 3 - 3 - 3 - - - 3 - weighted squared ED 3 -
Gariel, 2011 A 3 - 3 - 3 - 3 - - - 3 projecting trajectories to PCs dist. betw. projections on PCs 3 -
Kalnis, 2005 A - 3 - - 3 - 3 - 3 - - - spatial distance measure (e.g., ED) 3 -
Iyengar, 2004 R - 3 - - 3 - 3 - - - 3 - n.a. - 3
Neill, 2005 R - - - 3 3 3 - 3 - - 3 - n.a. - 3
Li, 2004 S - 3 - - 3 - 3 - - - 3 - ED (exchangable) 3 -
Evolution

Jensen, 2007 S - 3 - - 3 - 3 - - - 3 - weighted squared ED over time period 3 -

Cluster

Chen, 2007 A - 3 - - 3 - 3 - 3 - - - network distance 3 -

Rosswog, 2008 S 3 3 - - 3 - 3 - 3 - - - filtered ED squared over time period - (3)
Jeung, 2008 specialized measure of ED between
A - - 3 - 3 - - 3 - - 3 - 3 -
simplified trajectories
Rosswog, 2012 A 3 3 - - 3 - 3 - - - 3 SVM n.a. 3 -
Zheng, 2013 A - - 3 - 3 - - 3 - - 3 - Hausdorff distance 3 -

Fig. 4. Clustering Criteria Table (part 2)

DBSCAN-based algorithms usually require a minimum number of points per

cluster and a desired cluster radius, while grid based clustering techniques re-
quire a grid-cell border length. We highlight the approach by Wang et al. [33],
where parameters can be chosen arbitrarily, but might be approximated by an
optional technique, thus oering a hybrid approach of parameter settings.
Computational Complexity Rarely Included (CC). Most authors did not
provide the computational complexity of their approaches. Noticeable is the grid
based algorithm proposed by Wang et al. [33] that has the advantage of a linear
runtime complexity, as a single iteration of the grid is sucient to discover
ST clusters.
Few Optimization Strategies (O). Only some of the authors suggest usage
of special optimization strategies, like indexes or pruning. Some authors sug-
gest dierent variants of their algorithms, oering a trade-o between execution
eciency and quality of results. For example Tai et al. [32] suggest an exact
technique for cluster discovery, and a second variant that is more ecient in
execution, but might not yield the best results.

4.3 Conclusion

In this paper we focused on ST clustering approaches in the domain of SAW.

We proposed evaluation criteria stemming from the field of SAW on the one
hand and from the field of ST data mining on the other hand, to evaluate the
approaches with respect to their applicability to the field of SAW.
Summing up, we state that none of the surveyed approaches fulfills all the
criteria stemming from the special nature of data in SAW systems. Spatial and
 A Survey on Clustering Techniques for Situation Awareness 825

temporal extent (SP) as well as cyclic time patterns (TAP) are not the focus of
ST clustering techniques. Also fuzzy input data (FID) is not handled in the pre-
dominant number of cases and only few online approaches (OOA) exist. As long as
no appropriate techniques are available, we suggest transformation of the input
data to enable application of the clustering techniques reviewed in this survey
(e. g., only the starting points of trac jams could be used for clustering, in order
to apply techniques operating on point input data only and thus discarding the
information about their extent).

References
1. Aggarwal, C.C., et al.: A framework for clustering evolving data streams. In: Proc.
of 29th Int. Conf. on Very Large Data Bases, pp. 8192. VLDB Endowment (2003)
2. Andrienko, G., Andrienko, N.: Interactive cluster analysis of diverse types of spatio-
temporal data. ACM SIGKDD Explorations Newsletter 11(2), 1928 (2009)
3. Baumgartner, N., et al.: Beaware!situation awareness, the ontology-driven way.
Int. Journal of Data and Knowledge Engineering 69(11), 11811193 (2010)
4. Birant, D., Kut, A.: St-dbscan: An algorithm for clustering spatial and temporal
data. Data & Knowledge Engineering 60(1), 208221 (2007)
5. Chen, J., Lai, C., Meng, X., Xu, J., Hu, H.: Clustering moving objects in spatial
networks. In: Kotagiri, R., Radha Krishna, P., Mohania, M., Nantajeewarawat, E.
(eds.) DASFAA 2007. LNCS, vol. 4443, pp. 611623. Springer, Heidelberg (2007)
6. Ester, M., et al.: A density-based algorithm for discovering clusters in large spatial
databases with noise. In: Proc. of 2nd Int. Conf. on Knowledge Discovery and Data
Mining (KDD), pp. 226231. AAAI Press (1996)
7. Cao, F., et al.: Density-based clustering over an evolving data stream with noise.
In: SIAM Conf. on Data Mining, pp. 328339 (2006)
8. Gaber, M.M., et al.: Mining data streams: a review. ACM SIGMOD Record 34(2),
1826 (2005)
9. Ganey, S., Smyth, P.: Trajectory clustering with mixtures of regression models.
In: Proc. of the 5th ACM SIGKDD Int. Conf. on Knowledge Discovery and Data
Mining, KDD 1999, pp. 6372. ACM (1999)
10. Gariel, M., et al.: Trajectory clustering and an application to airspace monitoring.
Trans. Intell. Transport. Sys. 12(4), 15111524 (2011)
11. Han, J., et al.: Spatial clustering methods in data mining: A survey. In: Geographic
Data Mining and Knowledge Discovery, pp. 129 (2011)
12. Han, J., et al.: Data Mining Concepts and Techniques, 3rd edn. Morgan Kaufmann
(2011)
13. Ikonomovska, E., et al.: A survey of stream data mining. In: Proc. of 8th National
Conf. with Int. Participation, ETAI (2007)
14. Iyengar, V.S.: On detecting space-time clusters. In: Proc. of 2nd Int. Conf. on
Knowledge Discovery and Data Mining (KDD), pp. 587592. AAAI Press (1996)
15. Jain, A.K., et al.: Data clustering: a review. ACM Computing Surveys 31(3), 264
323 (1999)
16. Jensen, C., et al.: Continuous clustering of moving objects. IEEE Transactions on
Knowledge and Data Engineering 19(9), 11611174 (2007)
17. Jeung, H., et al.: Discovery of convoys in trajectory databases. Proc. VLDB En-
dow. 1(1), 10681080 (2008)
826 S. Mitsch et al.

18. Kalnis, P., Mamoulis, N., Bakiras, S.: On discovering moving clusters in spatio-
temporal data. In: Medeiros, C.B., Egenhofer, M., Bertino, E. (eds.) SSTD 2005.
LNCS, vol. 3633, pp. 364381. Springer, Heidelberg (2005)
19. Kalyani, D., Chaturvedi, S.K.: A survey on spatio-temporal data mining. Int. Jour-
nal of Computer Science and Network (IJCSN) 1(4) (2012)
20. Kavitha, V., Punithavalli, M.: Clustering time series data stream - a literature
survey. Int. Journal of Computer Science and Inf. Sec. (IJCSIS) 8(1) (2010)
21. Kisilevich, S., et al.: Spatio-temporal clustering: a survey. In: Data Mining and
Knowledge Discovery Handbook, pp. 122 (2010)
22. Kranen, P., et al.: The clustree: Indexing micro-clusters for anytime stream mining.
Knowledge and Information Systems Journal 29(2), 249272 (2011)
23. Li, Y., et al.: Clustering moving objects. In: Proc. of the 10th ACM Int. Conf. on
Knowledge Discovery and Data Mining (KDD), pp. 617622. ACM (2004)
24. Li, Z., Lee, J.-G., Li, X., Han, J.: Incremental clustering for trajectories. In: Kita-
gawa, H., Ishikawa, Y., Li, Q., Watanabe, C. (eds.) DASFAA 2010. LNCS, vol. 5982,
pp. 3246. Springer, Heidelberg (2010)
25. Lu, C.-T., Lei, P.-R., Peng, W.-C., Su, I.-J.: A framework of mining semantic
regions from trajectories. In: Yu, J.X., Kim, M.H., Unland, R. (eds.) DASFAA
2011, Part I. LNCS, vol. 6587, pp. 193207. Springer, Heidelberg (2011)
26. Matheus, C., et al.: Sawa: An assistant for higher-level fusion and situation aware-
ness. In: Proc. of SPIE Conf. on Multisensor, Multisource Information Fusion.
Architectures, Algorithms, and Applications, pp. 7585 (2005)
27. Nanni, M., Pedreschi, D.: Time-focused clustering of trajectories of moving objects.
Journal of Intelligent Information Systems 27(3), 267289 (2006)
28. Neill, D.B., et al.: Detection of emerging space-time clusters. In: Proc. of the 11th
ACM SIGKDD Int. Conf. on Knowledge Discovery in Data Mining, KDD 2005,
pp. 218227. ACM (2005)
29. Geetha, R., et al.: A survey of spatial, temporal and spatio-temporal data mining.
Journal of Computer Applications 1(4), 3133 (2008)
30. Rosswog, J., Ghose, K.: Detecting and tracking spatio-temporal clusters with adap-
tive history filtering. In: Proc. of 8th IEEE Int. Conf. on Data Mining, Workshops
(ICDMW), pp. 448457 (2008)
31. Rosswog, J., Ghose, K.: Detecting and tracking coordinated groups in dense, sys-
tematically moving, crowds. In: Proc. of the 12th SIAM Int. Conf. on Data Mining,
pp. 111. SIAM/Omnipress (2012)
32. Tai, C.-H., Dai, B.-R., Chen, M.-S.: Incremental clustering in geography and op-
timization spaces. In: Zhou, Z.-H., Li, H., Yang, Q. (eds.) PAKDD 2007. LNCS
(LNAI), vol. 4426, pp. 272283. Springer, Heidelberg (2007)
33. Wang, M., Wang, A., Li, A.: Mining spatial-temporal clusters from geo-databases.
In: Li, X., Zaane, O.R., Li, Z. (eds.) ADMA 2006. LNCS (LNAI), vol. 4093, pp.
263270. Springer, Heidelberg (2006)
34. Warren Liao, T.: Clustering of time series data-a survey. Pattern Recogn. 38(11),
18571874 (2005)
35. Wimmer, M., et al.: A survey on uml-based aspect-oriented design modeling. ACM
Computing Surveys 43(4), 28:128:33 (2011)
36. Xu, R., Wunsch II, D.C.: Survey of clustering algorithms. IEEE Transactions on
Neural Networks 16(3), 645678 (2005)
37. Zheng, K., et al.: On discovery of gathering patterns from trajectories. In: IEEE
Int. Conf. on Data Engineering, ICDE (2013)
 Parallel k-Skyband Computation
on Multicore Architecture

Xing Feng1 , Yunjun Gao1 , Tao Jiang2 , Lu Chen1 , Xiaoye Miao1 , and Qing Liu1
1
Zhejiang University, China
{3090103362,gaoyj,chenl,miaoxy,liuq}@zju.edu.cn
2
Jiaxing University, China
jiangtao [email protected]

Abstract. Given a set of elements D in a d-dimensional space, a k-skyband

query returns all elements which are worse than at most k other elements. The
k-skyband query is a fundamental analysis query. It can offer minimal candidate
set for other queries such as top-k ranking query where the ranking functions are
monotonic. With the development of multicore processors, it has been a trend
that more algorithms focus on parallel execution to exploit the computation re-
sources. In this paper, rather than use traditional methods to compute k-skyband,
we employ parallel techniques to get the result efficiently based on multicore ar-
chitecture. Extensive experiments show the efficiency and effectiveness of our
proposed algorithm.

1 Introduction
In recent decades, more manufacturers are producing multicore architecture proces-
sors due to the frequency wall. As a consequence, even personal computer has 4 cores,
8 cores or more. It is meaningful to design algorithms to leverage this feature, espe-
cially for problems which are computationally expensive. k-skyband query is a typical
problem of such kind that requires extensive computation. This paper investigates the
problem, and designs a parallel solution using multicore processors.
Given a data set D of elements in a d-dimensional numerical space, we say an ele-
ment p dominates another element q, if p is not worse than q in every dimension and
is better than q in at least one dimension. Without lose of generality, in this paper, we
assume lower value is preferred. Given a data set D, a k-skyband query retrieves the set
of elements which are dominated by at most k elements. The concept of k-skyband is
closely related to another two concepts, skyline and top-k ranking queries. Given a data
set D, skyline query retrieves the elements of a data set that are not worse than others.
Clearly, skyline query is a special case of k-skyband where k = 0. Given a data set
D and a ranking function f , top-k ranking query retrieves the k elements with highest
score according to f . Note k-skyband query keeps a minimum candidate set for top-k
query. Thus, keeping k-skyband of a data set is sufficient to answer all top-k queries.
Below, we give an example of k-skyband.

Example 1. Consider the 13 2-dimensional elements in a data set as shown in Fig-

ure 1. k-skyband query with k = 0 (or skyline) returns the set {a, i, k}. 1-skyband

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 827837, 2013.
c Springer-Verlag Berlin Heidelberg 2013
828 X. Feng et al.

Table 1. A data set with 13 points

k=0
12
11
k=1 ID Coordinate
k=2
10 others a (10, 1)
9 b (11, 2)
8 c (8, 4)
7 d (6, 6)
6 e (12, 9)
5 f (5, 7)
4 g (6, 5)
3 h (3, 4)
2 i (2, 3)
1 k (0, 9)
0 l (4, 11)
0 1 2 3 4 5 6 7 8 9 10 11 12 13 m (2, 6)
n (3, 8)
Fig. 1. A data set with 13 points

consists {a, i, k, b, h, m} which are worse than at most 1 other element. 2-skyband con-
sists {a, i, k, b, h, m, c, g} which are worse than at most 2 other elements.
To the best of our knowledge, this paper is among the first to study the problem of
parallel k-skyband computation on multicore architecture. Our main contributions are
summarized as follows.
We design a parallel algorithm to efficiently compute k-skyband query and analyse
its time complexity.
We conduct extensive experimental evaluation to verify the effectiveness and effi-
ciency of our proposed algorithm.
The rest of the paper is organized as follows. In Section 2, we formally define the
k-skyband problem and provide some necessary background information about our
parallel techniques. In Section 3, we present the BBSkyband algorithm for k-skyband
computation, which is used as the baseline algorithm. Section 4 presents the parallel
version of BBSkyband algorithm (Parallel BBSkyband). In Section 5, we examine the
efficiency of our algorithm. Related work is summarized in Section 6. This is followed
by conclusions.

2 Background Information
We present the problem definition and introduce a parallel library, OpenMP, in this
section.

2.1 Problem Definition

Definition 1 (Dominate). Consider two distinct d-dimensional elements p and q, and
p[i] denotes the i-th value of p. p dominates q (denoted by p q) iff p[i] q[i] holds
for 1 i d, and there exists j, such that p[j] < q[j].
 Parallel k-Skyband Computation on Multicore Architecture 829

Definition 2 (k-skyband query). Given a data set D of elements in a d-dimensional

numerical space, the k-skyband query retrieves the subset of D which are dominated by
at most k elements in D.

Problem Statement. In this paper, we study the problem of efficiently retrieving k-

skyband elements from a data set using parallel techniques.

2.2 OpenMP
In this paper, we use OpenMP [5] to parallelize the k-skyband computation. OpenMP
(Open Multiprocessing) is an API that supports multi-platform shared memory multi-
processing programming in C, C++, and Fortran, on most processor architectures and
operating systems. It consists of a set of compiler directives, library routines, and envi-
ronment variables that influence run-time behavior.
OpenMP is an implementation of multithreading, a method of parallelizing whereby
a master thread (a series of instructions executed consecutively) forks a specified num-
ber of slave threads and a task is divided among them. The threads then run concur-
rently, with the runtime environment allocating threads to different processors.
There are many details about the usage of OpenMP. Here, we only present the two
examples we used in our implementation. More information can be found on their
website1 .
Example 2.
pragma omp parallel for default (shared) private (i)
f or(i = 0; i < size; i + +)
output[i] = f (input[i]);
The first line of Example 2 instructs the compiler to distribute loop iterations within the
team of threads that encounters this work-sharing construct. Besides, default (shared)
private (i) means all the variables in the loop region be shared across all threads except
i. In the following sections, for a concise presentation, we use parallel instead of
pragma omp parallel for default (shared) private (i).
Example 3.
omp set num threads(NUMBER OF PROCESSORS);
The API in Example 3 instructs the compiler to set the number of threads in subsequent
parallel regions to be NUMBER OF PROCESSORS. This can be useful to see how the
running time changes against varying number of threads.

3 Baseline Algorithm
Papadias et al. develop a branch-and-bound algorithm (BBS) to get the skyline with the
minimal I/O cost [11]. They also point that their algorithms can be used to compute
k-skyband with some modifications.
First, we demonstrate the main idea behind BBS.
1
http://openmp.org
830 X. Feng et al.

Theorem 1. An element p cannot dominate#another element #d q, if the sum of qs all

d
dimension is smaller than that of p. That is, i=1 p[i] < i=1 q[i] q p.

This theorem is intuitive and can be proved using Definition 1.

Then, we modify the BBS algorithm to compute k-skyband. Algorithm 1 shows the
details of an algorithm adapted from BBS algorithm. H is a minheap. S is a buffer used
to collect all the elements of k-skyband. k denotes the k-skyband width. R is the root
of the data sets R-tree index. All elements that cannot be in S are discarded in Line 6
(dominated by more than k elements). All elements that should be in S are inserted into
S in Line 12. All entries that may contribute to final result are inserted into the heap in
Line 10.

Algorithm 1. BBSkyband
1 S = ;
2 insert all entries of the root R in the heap H;
3 while H not empty do
4 remove top entry e;
5 if e is dominated by more than k elements in S then
6 discard e;
7 else if e is an intermediate entry then
8 for each child ei of e do
9 if ei is dominated by at most k elements in S then
10 insert ei into H;

11 else
12 insert e into S

Correctness and Complexity Analysis. We prove the correctness in two steps. First,
all the elements in S must be in k-skyband. This is guaranteed by Theorem 1. Second,
all the elements that are discarded cannot be in k-skyband. This is intuitive because
these elements are already dominated by k + 1 elements. We analyse the worst case
complexity of Algorithm 1; i.e., when all the elements are k-skyband elements. Let N
denote the size of the data set. In this case, the most time-consuming parts are Line 5
and Line 9 of Algorithm 1; and thus, the time complexity is given by O(N 2 ).
Although the worst case time complexity is O(N 2 ), this algorithm is still efficient
because it discards the non-k-skyband elements at a level as high as possible. Below is
an example for Algorithm 1.

Example 4. Consider the running example in Section 1. Suppose we are going to find 2-
skyband and the R-tree is depicted in Figure 2. Applying the BBSkyband (Algorithm 1),
we have the following steps in Table 2. We omit some steps in which the top entry is an
element. In the initial step, we insert all the entries of root R, e6 and e7 , into H. Then
we remove top entry, e7 , from H. Note the distance of e7 is 0 + 3 = 3 which is smaller
than that of e6 , 5 + 1 = 6. Because no element in S dominates e7 , we expand e7 . After
expanding e7 , we have {e3 , e6 , e5 , e4 } in H. Then we remove and expand e3 . After
that, we have {i, e6 , h, e5 , e4 , g} in H. Then, we remove i. Because i is at leaf level, we
 Parallel k-Skyband Computation on Multicore Architecture 831

12
11
l
10 e4
k
9 e
8 n
e2
7 e5 f H H
e7 d
6 m
H H H H H
5 g e6
4 h e3 D E F G H I J K L O N P Q
i c
3
2 e1 b
1 a
0
0 1 2 3 4 5 6 7 8 9 10 11 12 13

Fig. 2. The R-Tree of Figure 1

Table 2. BBSkyband Steps

Action Heap content after action S after action

Initial < e7 , 3 >, < e6 , 6 >
Expanding e7 < e3 , 5 >, < e6 , 6 >, < e5 , 8 >, < e4 , 9 >
Expanding e3 < i, 5 >, < e6 , 6 >, < h, 7 >, < e5 , 8 >,
< e4 , 9 >, < g, 11 >
Expanding e6 < h, 7 >, < e5 , 8 >, < e1 , 9 >, < e4 , 9 >, i
< g, 11 >, < e2 , 11 >
Expanding e5 < m, 8 >, < e1 , 9 >, < e4 , 9 >, < g, 11 >, i, h
< n, 11 >, < e2 , 11 >
Expanding e1 < e4 , 9 >, < a, 11 >, < g, 11 >, < n, 11 >, i, h, m
< e2 , 11 >, < c, 12 >, < b, 13 >
Expanding e4 < k, 9 >, < a, 11 >, < g, 11 >, < n, 11 >, i, h, m
< e2 , 11 >, < c, 12 >, < b, 13 >
Expanding e2 < c, 12 >, < b, 13 > i, h, m, k, a, g
i, h, m, k, a, g, c, b

insert i into S instead of expanding. This process goes until there is no element or entry
left in H. Finally, S contains the 2-skyband, that is, {i, h, m, k, a, g, c, b}.

4 Parallel BBSkyband
In this section, we are going to parallelize the baseline algorithm in Section 3.
By analysing the algorithm, we find the most computation consuming parts are Line
5 and Line 9 in Algorithm 1 because the element is compared with all the elements in S.
A naive solution is to parallelize these two lines by using the parallel command
in Example 2. However, for example, when executing Line 5, a thread has found more
832 X. Feng et al.

than k dominating elements and break out, how could other threads be notified? To
solve that, we either have to let a thread waits for all other comparing threads terminate,
or use communication method between threads. Both of the two ways decrease the ben-
efit we get from parallelization.

Algorithm 2. Parallel BBSkyband

1 S = ; S = ;
2 insert all entries of the root R in the heap H;
3 while H or S is not empty do
4 if there is no intermediate S and H is not empty then
5 e = remove top entry of H;
6 for each e S and e e do
7 ek + +;
8 if ek k then
9 S = S {e};

10 else
11 parallel for each e of S do
12 for each ei S, ei e do
13 ek + +;
14 if ek > k then
15 break;

16 for each e of S do
17 if ek k then
18 if e is an intermediate node of R then
19 parallel for each child cj of e do
20 for each e S and e cj do
21 cjk + +;
22 if cjk > k then
23 break;

24 for each child cj of e do

25 if cjk k then
26 insert cj into heap;

27 else
28 S = S {e };

29 S = ;

30 return S;

Algorithm 2 shows the details about our algorithm to parallelize BBSkyband. S is

used to store all the k-skyband elements we have found. S is a buffer to store candidate
k-skyband elements. k is the width of the k-skyband. ek , ek and cjk are used to record
the number of elements dominating ek , ek and cjk respectively. Our solution is to use
a buffer S to store all candidate entries which are not dominated by more than k other
 Parallel k-Skyband Computation on Multicore Architecture 833

elements in this buffer (Lines 4 9). Then, we compare these candidate entries with
elements that are already in k-skyband, i.e. elements in S (Lines 1115). This operation
can be run concurrently. After that, if the entry is an intermediate node, we compare
each child node of the entry with k-skyband elements in S concurrently (Lines 1923);
If the element is a leaf node, we insert it into S (Line 28). By doing this, we parallelize
the most time-consuming parts, Line 5 and Line 9 in Algorithm 1.
Correctness and Complexity Analysis. The proof of correctness is quite similar to
that of Algorithm 1. All the elements in S must be in k-skyband and all the elements
that are discarded cannot be in k-skyband. Next, We analyse the worst case complexity
of Algorithm 2; Similarly, we assume all the elements are k-skyband elements. Let N
denote the size of the data set. In this case, the most time-consuming parts are Line 12
and Line 20 of Algorithm 2. And thus, the time complexity is given by O(N 2 /c) where
c is the number of cores.

Example 5. Consider the running example in Section 3. Suppose we are going to find
2-skyband using parallel BBSkyband (Algorithm 2). Table 3 shows the steps of the
algorithm. We omit some steps where there is no intermediate entry in S . In the initial
step, we insert all the entries of root R, e6 and e7 , into minheap H. Then we remove top
entry e7 . Because no element in S dominates e7 , e7 is stored in S . In the next loop,
because there is an intermediate entry (e7 ) in S , we compare e7 with every element in
S and record the number of elements that dominate e7 (Here is 0). Note that, if there
are other elements in S , we can compare them with elements in S concurrently. Then,
we expand e7 (like Lines 19 23 in Algorithm 2). Note that, this process can be done
concurrently. After expanding e7 , we add entries that are not dominated by more than
k elements to H (like Lines 24 26 in Algorithm 2). Here, we have {e3 , e6 , e5 , e4 } in
H. Then we remove and expand e3 . After that, we have {i, e6 , h, e5 , e4 , g} in H. Note
that, i will be moved to S . Similar process goes on until there is no element or entry
left in H and S . Finally, S contains the 2-skyband, that is, {i, h, m, a, k, g, b, c}.

Table 3. Parallel BBSkyband Steps

Action Heap content after action S after action S after action

Initial < e7 , 3 >, < e6 , 6 >
Expanding e7 < e3 , 5 >, < e6 , 6 >, < e5 , 8 >, < e4 , 9 >
Expanding e3 < e6 , 6 >, < h, 7 >, < e5 , 8 >, < e4 , 9 >, i
< g, 11 >
Expanding e6 < e5 , 8 >, < e1 , 9 >, < e4 , 9 >, < g, 11 >, h i
< e2 , 11 >
Expanding e5 < e1 , 9 >, < e4 , 9 >, < g, 11 >, < n, 11 >, m i, h
< e2 , 11 >
Expanding e1 < e4 , 9 >, < a, 11 >, < g, 11 >, < n, 11 >, i, h, m
< e2 , 11 >, < c, 12 >, < b, 13 >
Expanding e4 < e2 , 11 >, < c, 12 >, < b, 13 > k, a, g, n i, h, m
Expanding e2 c, b i, h, m, k, a, g
i, h, m, k, a, g, c, b
834 X. Feng et al.

5 Experiment

This section reports the experimental studies.

5.1 Setup
We ran all experiments on a DELL OPTIPLEX 790 with Windows 7, two quad-Intel
Core CPU and 2G 1333 MHz RAM. We evaluated the efficiency of our algorithm
against data set distribution, dimensionality, data set size, number of cores and k-
skyband width, respectively. Our experiments are conducted on synthetic data sets.
The synthetic data sets are generated as follows. We use the methodologies in [12] to
generate 1 million data elements with dimensionality from 2 to 5 and the spatial location
of data elements follow two kinds of distributions, independent and anti-correlated.
The default values of the parameters are listed in Table 4. Data set distribution and
number of cores do not have default values because we vary all the values of them in all
experiments.

Table 4. Experimental parameters

Notation Definition (Default Values)

D data set distribution
d dimensionality of the data set (4)
N data set size (256K)
c number of cores
k k-skyband width (4)

Parameters are varied as follows:

data set distribution: Independent, Anti-correlated.
dimensionality: 2, 3, 4, 5.
data set size: 64K, 128K, 256K, 512K, 1024K.
number of cores: 1, 2, 4, 8.
k-skyband width: 2, 4, 6, 8.

5.2 Evaluation

We evaluate our algorithm on both Independent and Anti-correlated data set. Besides, as
demonstrated in Example 3, we use the command in Example 3 to set different number
of threads which can also be regarded as different cores. Particularly, when c = 1, we
use BBSkyband as baseline algorithm rather than Parallel BBSkyband.
The first set of experiments is reported in Figure 3 where dimensionality varies. The
performance of our algorithm decreases as the dimensionality increases. This is because
when dimensionality increases, the cost of comparing two elements increases (Note the
comparing cost is O(d) where d is the dimensionality). Besides, there are usually more
elements in a k-skyband due to higher dimensionality. Because with higher dimension-
ality, an element has to be better than another one in more dimensions. Note having
 Parallel k-Skyband Computation on Multicore Architecture 835

more elements in k-skyband also increases the cost for operations such as Line 12 and
Line 20 in Algorithm 2.
Figure 4 evaluates the system scalability towards the data set size. The performance
of our algorithm decreases as the data set size increases. When data set size increases,
there is a larger R-tree as index. The operation on R-tree takes more time. What is more,
larger data set usually results in more elements in a k-skyband.
Figure 5 evaluates the impact of k-skyband width. As expected, Figure 5 shows that
processing cost increases when k increases. k is the width of k-skyband. With larger k,
we have to do more dominating test and we usually have more elements in a k-skyband.

45 40
40 c=1 35 c=1
c=2 c=2
Avg. Delay (s)

Avg. Delay (s)

35 30
30 c=4 c=4
c=8 25 c=8
25 20
20
15 15
10 10
5 5
0 0
2 3 4 5 2 3 4 5
(a) Varying d, Anti-correlated (b) Varying d, Independent

Fig. 3. Avg. Delay vs d

120 120
c=1 c=1
100 c=2 100 c=2
Avg. Delay (s)

Avg. Delay (s)

80 c=4 80 c=4
c=8 c=8
60 60
40 40
20 20
0 0
64K 128K 256K 512K 1024K 64K 128K 256K 512K 1024K
(a) Varying N, Anti-correlated (b) Varying N, Independent

Fig. 4. Avg. Delay vs N

80 70
70 c=1 c=1
c=2 60 c=2
Avg. Delay (s)

Avg. Delay (s)

60 50
c=4 c=4
50 c=8 40 c=8
40
30 30
20 20
10 10
0 0
2 4 6 8 10 2 4 6 8 10
(a) Varying k, Anti-correlated (b) Varying k, Independent

Fig. 5. Avg. Delay vs k

836 X. Feng et al.

In all figures, there is an obvious improvement when the number of cores rises. How-
ever, it is hard to get a speedup that strictly equals the number of cores. This is because
there are other costs that are not parallelized in our algorithm (I/O operation, collecting
candidate elements, insertion into heap, etc.).

6 Related Work

Spatial database has draw many attention in recent decades. Beckmann et al. [2] design
the R-tree as an index for efficient computation. Roussopoulos et al. [12] study Nearest
Neighbor Queries and propose three efficient pruning rules. They also extend Nearest
Neighbor Queries to K Nearest Neighbor Queries which return top-K preferred ob-
jects. Borzsonyi et al. [3] first study the skyline operator and propose an SQL syntax
for the skyline query. After that, the skyline query is widely studied by a lot of papers,
like Chomicki et al. [4], Godfrey et al. [7] and Tan et al. [13]. Papadias et al. develop a
branch-and-bound algorithm (BBS) to get the skyline with the minimal I/O cost [11].
The concept of k-skyband is first proposed in [11]. [10] shows that it suffices to keep
the k-Skyband of the data set to provide the top-k ranked queries for the monotone
preference function f . Reverse k-skyband query is recently studied in [9].
With the development of multi-core processors, a recent trend is to study the parallel
skyline computation. Most of them are divide and conquer algorithms. Some papers
focus on distributed skyline computation. That is, their algorithm runs on server cluster
with several servers, such as [1]. Other papers focus on parallize skyline computation
on different cores instead of servers. This is different since all the threads share the
same memory and there is much less communication cost. Park et al. design a paral-
lel algorithm on multicore architecture with an easy idea. But it distributes the most
computation-consuming part to every core. Our work is based on their algorithm. Vla-
chou et al. [14] employ the technique which partitions the data set according to the
weight of each axis. Moreover, Kohler et al. [8] use a similar idea but a faster way
to partition the data set. Besides, Gao et al. [6] design a parallel algorithm for skyline
queries in multi-disk environment.

7 Conclusion

In this paper, we investigate the problem of k-skyband query. We give a modified Al-
gorithm (BBSkyband) from the traditional algorithm and parallelize it, namely, Parallel
BBSkyband algorithm. The experimental results demonstrate that a simple design of the
parallel algorithm provides a satisfactory runtime performance. We believe that more
efficient algorithm can be further developed based on our presented method.
Acknowledgements.. This work was supported in part by NSFC 61003049, the Nat-
ural Science Foundation of Zhejiang Province of China under Grant LY12F02047, the
Fundamental Research Funds for the Central Universities under Grant 2012QNA5018,
the Key Project of Zhejiang University Excellent Young Teacher Fund (Zijin Plan), and
the Bureau of Jiaxing City Science and Technology Project under Grant 2011AY1005.
 Parallel k-Skyband Computation on Multicore Architecture 837

References
1. Afrati, F.N., Koutris, P., Suciu, D., Ullman, J.D.: Parallel skyline queries. In: Deutsch, A.
(ed.) ICDT, pp. 274284. ACM (2012)
2. Beckmann, N., Kriegel, H.-P., Schneider, R., Seeger, B.: The r*-tree: An efficient and robust
access method for points and rectangles. In: SIGMOD Conference, pp. 322331 (1990)
3. Borzsonyi, S., Kossmann, D., Stocker, K.: The skyline operator. In: ICDE, pp. 421430
(2001)
4. Chomicki, J., Godfrey, P., Gryz, J., Liang, D.: Skyline with presorting. In: ICDE, pp. 717719
(2003)
5. Clark, D.: Openmp: a parallel standard for the masses. IEEE Concurrency 6(1), 1012 (1998)
6. Gao, Y., Chen, G., Chen, L., Chen, C.: Parallelizing progressive computation for skyline
queries in multi-disk environment. In: Bressan, S., Kung, J., Wagner, R. (eds.) DEXA 2006.
LNCS, vol. 4080, pp. 697706. Springer, Heidelberg (2006)
7. Godfrey, P., Shipley, R., Gryz, J.: Maximal vector computation in large data sets. In: VLDB,
pp. 229240 (2005)
8. Kohler, H., Yang, J., Zhou, X.: Efficient parallel skyline processing using hyperplane projec-
tions. In: SIGMOD Conference, pp. 8596 (2011)
9. Liu, Q., Gao, Y., Chen, G., Li, Q., Jiang, T.: On efficient reverse k-skyband query processing.
In: Lee, S.-g., Peng, Z., Zhou, X., Moon, Y.-S., Unland, R., Yoo, J. (eds.) DASFAA 2012,
Part I. LNCS, vol. 7238, pp. 544559. Springer, Heidelberg (2012)
10. Mouratidis, K., Bakiras, S., Papadias, D.: Continuous monitoring of top-k queries over slid-
ing windows. In: Chaudhuri, S., Hristidis, V., Polyzotis, N. (eds.) SIGMOD Conference, pp.
635646. ACM (2006)
11. Papadias, D., Tao, Y., Fu, G., Seeger, B.: Progressive skyline computation in database sys-
tems. ACM Trans. Database Syst. 30(1), 4182 (2005)
12. Roussopoulos, N., Kelley, S., Vincent, F.: Nearest neighbor queries. In: SIGMOD Confer-
ence, pp. 7179 (1995)
13. Tan, K.-L., Eng, P.-K., Ooi, B.C.: Efficient progressive skyline computation. In: VLDB, pp.
301310 (2001)
14. Vlachou, A., Doulkeridis, C., Kotidis, Y.: Angle-based space partitioning for efficient parallel
skyline computation. In: SIGMOD Conference, pp. 227238 (2008)
Moving Distance Simulation for Electric Vehicle
Sharing Systems for Jeju City Area

Junghoon Lee and Gyung-Leen Park

Dept. of Computer Science and Statistics,

Jeju National University, 690-756, Jeju Do, Republic of Korea
{jhlee,glpark}@jejunu.ac.kr

Abstract. This paper measures the moving distance for relocation

strategies in electric vehicle sharing systems. The moving distance de-
notes the distance a customer actually moves to pick up a vehicle and
indicates the service quality for carsharing. For even, utilization-based,
and morning-focused relocations schemes, the experiments are conducted
using taxi trip records in Jeju City area. The measurement results show
that the morning-focused scheme generally outperforms the others, re-
locating vehicles according to the demand just in the morning time. For
a large number of vehicles, the moving distance and the serviceability
is not aected by static relocation schemes, revealing the essentiality of
proactive relocation.

Keywords: Electric vehicle, carsharing service, relocation strategy,

moving distance, analysis framework.

1 Introduction

Considering power eciency and low carbon emissions, electric vehicles, or EVs,
are hopefully expected to penetrate into our daily lives [1]. However, private
ownership is not yet aordable to customers due to their high cost. Hence, EV
sharing is considered to be a good intermediary business model to accelerate
their deployment. EV sharing has many operation models according to the man-
ageability, user convenience, serviceability, and others. In one-way rental, a cus-
tomer picks up an EV at a station and return to a dierent station. Even if this
model provides great flexibility to customers, it faces uneven EV distribution.
Some stations do not have an EV, so a pick-up request issued at those stations
cannot be admitted. This problem can be solved only by an appropriate vehicle
relocation strategy. Obviously, there is trade-o between relocation frequency
and relocation cost.

This research was supported by the MKE, Republic of Korea, under IT/SW
Creative research program supervised by the NIPA (NIPA-2012-(H0502-12-1002).
This research was also financially supported in part by the Ministry of Knowledge
Economy (MKE), Korea Institute for Advancement of Technology (KIAT) through
the Inter-ER Cooperation Projects.

Y. Ishikawa et al. (Eds.): APWeb 2013, LNCS 7808, pp. 838841, 2013.
c Springer-Verlag Berlin Heidelberg 2013

 Moving Distance Simulation for Electric Vehicle Sharing Systems 839

Practically, relocation is highly likely to be carried out during non-operation

hours to prevent service discontinuity. Then, the relocation problem is reduced
down to decide the number of EVs for each station for the next operation hours.
To begin with, service ratio, namely, the probability that an EV is available
when a sharing request arrives, is an important performance metric. Besides,
the moving distance, which denotes the distance a customer moves to take an
EV, can indicates the fitness of sharing station, the eciency of relocation policy,
and user-side service convenience. In this regard, this paper measures the moving
distance for three relocation policies, exploiting our analysis framework, which
takes trip records collected from the Jeju Taxi Telematics System [2].
There have been some related researches on EV relocation schemes and
simulations methods. First, [3] has analyzed the performance of multiple-station
shared-use vehicle systems using diverse simulation tools. For the better knowl-
edge of operation characteristics, UCR (University of California at Riverside) In-
telliShare system began its operation with tens of electric vehicles. Both
reservation-based and on-demand sharing is provided to customers. Vehicle relo-
cation can be performed through towing and ridesharing. In ridesharing, multiple
drivers take separate vehicles and then return by sharing a ride in a single ve-
hicle. In order to reduce relocation complexity, UCR IntelliShare has introduced
techniques of trip joining and trip splitting. In trip splitting, the coordinator rec-
ommends a passenger team having more than one person to take separate vehicles
when the number of vehicles at the destination station goes below the lower bound.
On the contrary, trip joining merges multiple trips into one if possible.
[4] begins with the two operation-based relocation strategies, namely, the
shortest time and inventory balancing techniques, both of which take advantage
of vehicle-to-operation center communication channels. The first one relocates
vehicles to or from a neighbor station in the shortest possible time, putting
emphasis on service levels. The second scheme moves vehicles from a station
having surplus to a station with vehicle shortage to enhance cost eciency.
Then, a time-stepping simulation model is developed, exploiting a tremendous
trip records obtained from Honda Diracc system in Singapore. Tracing a series of
dynamic events on pick-up and return patterns, sta status, and job status, the
number of vehicles in each station is checked against the predefined threshold
values for respective stations. The simulation yields performance indicators of
zero-vehicle time, full-port time, and the number of relocations.

2 Paramter Setting
The EV relocation policy must consider such factors as relocation time, the num-
ber of service men, and relocation goal. Practically speaking, relocation during
the operation hours may lead to service discontinuity, so nonoperation hours are
appropriate for relocation and other management activities. This paper considers
3 strategies and evaluates their performance using the sharing system analysis
framework implemented in our previous work [2]. First, even relocation strategy
makes the number of EVs equal for each station. Second, utilization-based re-
location scheme distributes EVs according to the pick-up ratio of each station
840 J. Lee and G.-L. Park

during the whole observation interval. Next, morning-focused relocation strategy

just focuses on the pick-ups during the morning time, namely, just the first a
few hours in the operation time.

3 Experiment Result

Figure 1 measures the moving distance according to the number of EVs. It is the
distance the customer actually moves to the sharing station. The moving distance
decreases according to the increase of the number of EVs, as more requests can
be served. The moving distance for utilization-based scheme is the worst. The
dierence between morning-focused and even relocation schemes reaches 2.5 m.
Here again, the morning-focuses scheme shows slightly shorter moving distance
than the even relocation. The gap gets smaller when the access distance is 500 m,
while both even and morning-focused schemes show almost the same distance,
as shown in Figure 1(b).

220 380
"Even" "Even"
210 "Utilization" "Utilization"
"Morning" 370 "Morning"
200
Moving distance (m)

Moving distance (m)

360
190

180 350

170
340
160
330
150

140 320
5 10 15 20 25 30 35 40 45 50 5 10 15 20 25 30 35 40 45 50
Number of EVs Number of EVs

(a) 300 m (b) 500 m

Fig. 1. Eect of number of EVs

1 1
"Even" "Even"
"Utilization" "Utilization"
Normalized moving distance

Normalized moving distance

0.9 "Morning" "Morning"

0.9

0.8 0.8

0.7 0.7

0.6 0.6

0.5 0.5
100 200 300 400 500 600 700 800 900 1000 100 200 300 400 500 600 700 800 900 1000
Access distance (m) Access distance (m)

(a) 30 EVs (b) 50 EVs

Fig. 2. Access distance eect analysis

 Moving Distance Simulation for Electric Vehicle Sharing Systems 841

Next, Figure 2 measures the moving distance according to the access distance.
As the moving distance is sure to increase according to the increase of the access
distance, the moving distance is normalized to 1.0. It denotes how far requests
happen away from sharing stations. Figure 2(a) and Figure 2(b) have almost the
same shape, indicating independence to the number of EVs. When the access
distance is 100 m, the number of trip records is not sucient, so this point is not
consistent with other access distance range. Anyway, when the access distance
is around 500 m, the sharing requests from wider range can be served. Again,
the utilization-based scheme is the worst out of the three.

4 Conclusions
This paper has measured the moving distance for three practical relocation
strategies in electric vehicle sharing systems. Combined with the serviceabil-
ity, the moving distance indicates the easiness to access the sharing system and
thus the service quality. For even, utilization-based, and morning-focused reloca-
tions schemes, the experiments are conducted using taxi trip records in Jeju City
area. The measurement results show that the morning-focused scheme generally
outperforms the others, making customers conveniently access the EV sharing
service, discovering the necessity of a proactive relocation scheme based on the
future demand forecast.

References
1. Lee, J., Kim, H.-J., Park, G.-L., Kwak, H.-Y., Lee, M.Y.: Analysis Framework for
Electric Vehicle Sharing Systems Using Vehicle Movement Data Stream. In: Wang,
H., Zou, L., Huang, G., He, J., Pang, C., Zhang, H.L., Zhao, D., Yi, Z. (eds.) APWeb
2012 Workshops. LNCS, vol. 7234, pp. 8994. Springer, Heidelberg (2012)
2. Lee, J., Kim, H., Park, G., Kang, M.: Energy Consumption Scheduler for Demand
Response Systems in the Smart Grid. Journal of Information Science and Engineer-
ing, 955969 (2012)
3. Kek, A., Cheu, R., Meng, Q., Fung, C.: A Decision Support System for Vehicle
Relocation Operations in Carsharing Systems. Transportation Research Part E,
149158 (2009)
4. Barth, M., Todd, M., Xue, L.: User-based Vehicle Relocation Techniques
for Multiple-Station Shared-Use Vehicle Systems. Transportation Research
Record 1887, 137144 (2004)
 Author Index

Aalberg, Trond 163 Fan, Chaosheng 770

Abdullah, S M 88 Fang, Qiyu 697
Abulaish, Muhammad 268 Fariha, Anna 88
Ahmed, Chowdhury Farhan 88 Feng, Liang 709
Ahmed, Faraz 268 Feng, Shengzhong 96
Ali, Muhammad Intizar 43 Feng, Xing 782, 827
Allahbakhsh, Mohammad 196 Feng, Ziming 417
Al-Sharawneh, Jebrin 323 Fletcher, George H.L. 3
Anderson, Neil 392 Foo, Norman 196
Ashraf, Jamshaid 19 Friel, Owen 43
Fu, Xiaodong 697, 734
Bai, Ge 184 Fujiwara, Toru 228
Bai, Yiyuan 132 Fujiwara, Yuya 260
Beheshti, Seyed-Mehdi-Reza 196
Benatallah, Boualem 196 Gan, Liang 804
Bertino, Elisa 196 Gao, Yunjun 782, 827
Ge, Bin 794
Cai, Xiongcai 758 Giordano, Massimiliano 506
Cao, Jian 404 Gong, Shuai 518
Cao, Keyan 449 Gong, Xueqing 550, 709
Cao, Yini 638 Gowraj, Narayan 571
Chawla, Sanjay 595 Guan, Zhi 538
Cheema, Muhammad Aamir 461 Guo, Jing 176
Chen, Deren 208 Guo, Li 332
Chen, Huahui 563
Chen, Jia 417 Han, Donghong 449
Chen, Jinchuan 609 Han, Lu 132
Chen, Lei 2, 75 Han, Yi 372
Chen, Lu 827 Hashimoto, Kenji 228
Chen, Xuexuan 360 He, Jun 120
Chen, Yueguo 240 He, Qing 384
Chen, Zhong 538, 583 He, Xianmang 563
Cheng, Jiefeng 96 He, Xiaofeng 662, 721
Chin, Eunjung 108 Hong, Jun 392
Chowdhury, Nipa 758 Hu, Jianbin 583
Cui, Bin 360 Hu, Xiangyu 63
Cuzzocrea, Alfredo 31 Hu, Yachao 449
Huang, Guangyan 404
Davis, Joseph G. 629 Hussain, Omar Khadeer 19
Deufemia, Vincenzo 506
Ding, Zhaoyun 372 Ignjatovic, Aleksandar 196
Du, Fang 240 Ishihara, Yasunori 228
Du, Xiaoyong 120, 240, 609 Islam, Md. Rafiqul 88
Duchateau, Fabien 163 Iwata, Tomoharu 348
844 Author Index

Jia, Yan 372, 804 Mileo, Alessandra 43

Jiang, Tao 827 Mitsch, Stefan 815
Jiao, Han 151 Mizumoto, Hiroki 7
Jin, Cheqing 437 Moro, Gianluca 31
Jin, Leilei 437 M
uller, Andreas 815
Jin, Simeng 63
Nadamoto, Akiyo 260
Kabutoya, Yutaka 348 Naeem, M. Asif 494
Kapelko, Rafa!l 686 Naznin, Mahmuda 674
Kitagawa, Hiroyuki 348 Nishi, Manziba Akanda 88
Konishi, Yukio 260
Kuwada, Hayato 228 Pang, Linsey Xiaolin 595
Park, Gyung-Leen 838
Lai, Wei 429 Peng, Wei 310
Lee, Junghoon 838 Peng, Zhuo 132
Leung, Carson Kai-Sang 301 Phelan, Owen 746
Li, Aiping 804 Polese, Giuseppe 506
Li, Guoqiang 654
Li, Jin 734 Qian, Weining 550
Li, Jingxuan 310 Qin, Shouke 550
Li, Jiuyong 151
Li, Ming 621 Rahman, Md. Shaifur 674
Li, Shijun 482 Ravi, Prasanna Venkatesh 571
Li, Tao 310 Retschitzegger, Werner 815
Li, Xue 280 Ruan, Wenjing 132
Li, Yujia 563
Liang, Guangcheng 654 Saleheen, Shibli 429
Liang, Zhongjun 176 Salfinger, Andrea 815
Liao, Lejian 654 Samiullah, Md. 88
Lin, Rongheng 176 Sartori, Claudio 31
Lin, Zuoquan 770 Satoh, Tetsuji 340
Liu, Chengfei 151 Scholz, Bernhard 595
Liu, Hongyan 120 Schwinger, Wieland 815
Liu, Jixue 151 Sha, Chaofeng 662, 709, 721
Liu, Peipeng 332 Shang, Tianfeng 384
Liu, Qing 827 Sharaf, Mohamed A. 280
Liu, Weiyi 697, 734 Shen, Yan 473
Liu, Zhiyong 518 Sheng, Dakui 292
Lopes, Nuno 43 Sheng, Gang 621
Lu, Mian 51, 184 Shi, Jinqiao 332
Lu, Yifei 461 Shi, Zhongzhi 384
Luo, Qiong 51, 184 Simonetti, Luigi Marco 506
Luo, Zhunchen 140 Sinnappan, Suku 323
Lv, Yanfei 360 Smyth, Barry 746
Slvberg, Ingeborg 163
McCarthy, Kevin 746 Song, Dandan 654
Mei, Hong 75 Song, Leyi 709
Meng, Jintao 96 Song, William Wei 208
Meymandpour, Rouzbeh 629 Sumalatha, M.R. 571
Miao, Xiaoye 827 Sun, Fuzhen 654
 Author Index 845

Sun, Guangyu 360 Xiong, Jinhua 518

Sun, Huiping 538 Xu, Chen 662
Sun, Shengyun 120 Xu, Fei 332
Sun, Tao 292 Xu, Guandong 404
Sun, Tong 310 Xu, Jinghu 804
Sun, Yifang 461
Suzuki, Nobutaka 7 Yamamoto, Shuhei 340
Suzuki, Yu 260 Yang, Fangchun 176
Yang, Tao 538, 583
Takhirov, Naimdjon 163 Yang, Yong 583
Tanbeer, Syed Khairuzzaman 301 Yin, Ying 621
Tang, Da-quan 794 You, Yue 404
Tang, Jie 184 Yu, Chunfeng 372
Tang, Jintao 140 Yu, Liangwen 583
Toda, Hiroyuki 348 Yu, Philip S. 75
Yu, Wei 482
Udagawa, Yoshihisa 252 Yu, Wenzhe 721
Unankard, Sayan 280 Yu, Yong 417
V, Mouniga 571 Yu, Zukun 208
Yuan, Mingxuan 75
Wang, Bingqiang 51, 96 Yuan, Xiaojie 63
Wang, Chaokun 132 Yuan, Ye 449
Wang, Guoren 449 Yue, Kun 697, 734
Wang, Haixun 1
Wang, Hongbing 220 Zeng, Li 96
Wang, Hongya 473 Zhang, Bin 621
Wang, Jingang 654 Zhang, Cheng 518
Wang, Lihong 332 Zhang, Chong 794
Wang, Shan 526, 646 Zhang, Haiwei 63
Wang, Sheng 292 Zhang, Jianfeng 372
Wang, Teng 526 Zhang, Ming 292
Wang, Ting 140 Zhang, Ping 734
Wang, Wei 563 Zhang, Rong 721
Wang, Xiaojun 220 Zhang, Tiancheng 621
Wang, Xiaoyan 609 Zhang, Yanchun 404
Wang, Xiaoye 473 Zhang, Yansong 646
Wang, Zhihui 563 Zhang, Ying 63, 782
Wang, Ziqi 292 Zhang, Yunlu 482
Wei, Jinxian 662 Zhao, Jiuxin 51, 184
Wen, Yanlong 63 Zhao, Xiang 782
Wilcox, Georgina 595 Zhao, Yuhai 621
Williams, Mary-Anne 323 Zheng, Xiaolin 208
Wu, Hong 697 Zhou, Aoying 437, 550, 662
Wu, Junkai 51 Zhou, Bin 372, 804
Wu, Xian 417 Zhou, Xiaoling 461
Wu, Yuqing 3 Zhou, Xuan 638, 646
Zhou, Yao 437
Xia, Fan 662 Zhu, Jia 108
Xiao, Wei-dong 794 Zhu, Qing 526
Xiao, Yingyuan 473 Zhu, Yuean 646
Xin, Wei 538, 583 Zhuang, Fuzhen 384
Zou, Hua 176