Big Data Analytics Using Multiple

Criteria Decision-Making Models

EDITED BY
Big Data Ramakrishnan Ramanathan
Muthu Mathirajan
A. Ravi Ravindran
MCDM Tools

Business Analytics
Tools

Analytics Reports Human Interpretation

Business Intelligence
Big Data Analytics
Using Multiple
Criteria Decision-
Making Models
 The Operations Research Series
Series Editor: A. Ravi Ravindran
Professor, Department of Industrial and Manufacturing Engineering
The Pennsylvania State University, University Park, PA

Published Titles
Multiple Criteria Decision Analysis for Industrial Engineering:
Methodology and Applications
Gerald William Evans
Multiple Criteria Decision Making in Supply Chain Management
A. Ravi Ravindran
Operations Planning: Mixed Integer Optimization Models
Joseph Geunes
Introduction to Linear Optimization and Extensions with MATLAB ®
Roy H. Kwon
Supply Chain Engineering: Models and Applications
A. Ravi Ravindran & Donald Paul Warsing
Analysis of Queues: Methods and Applications
Natarajan Gautam
Integer Programming: Theory and Practice
John K. Karlof
Operations Research and Management Science Handbook
A. Ravi Ravindran
Operations Research Applications
A. Ravi Ravindran
Operations Research: A Practical Introduction
Michael W. Carter & Camille C. Price
Operations Research Calculations Handbook, Second Edition
Dennis Blumenfeld
Operations Research Methodologies
A. Ravi Ravindran
Probability Models in Operations Research
C. Richard Cassady & Joel A. Nachlas
 Big Data Analytics
Using Multiple
Criteria Decision-
Making Models

Edited by
Ramakrishnan Ramanathan,
Muthu Mathirajan, and A. Ravi Ravindran
CRC Press
Taylor & Francis Group
6000 Broken Sound Parkway NW, Suite 300
Boca Raton, FL 33487-2742

© 2017 by Taylor & Francis Group, LLC

CRC Press is an imprint of Taylor & Francis Group, an Informa business

No claim to original U.S. Government works

Printed on acid-free paper

International Standard Book Number-13: 978-1-4987-5355-5 (Hardback)

978-1-1387-4765-4 (Paperback)

This book contains information obtained from authentic and highly regarded sources. Reasonable
efforts have been made to publish reliable data and information, but the author and publisher cannot
assume responsibility for the validity of all materials or the consequences of their use. The authors
and publishers have attempted to trace the copyright holders of all materials reproduced in this
publication and apologize to copyright holders if permission to publish in this form has not been
obtained. If any copyright material has not been acknowledged, please write and let us know so we
may rectify in any future reprint.

Except as permitted under the U.S. Copyright Law, no part of this book may be reprinted, repro-
duced, transmitted, or utilized in any form by any electronic, mechanical, or other means, now
known or hereafter invented, including photocopying, microfilming, and recording, or in any infor-
mation storage or retrieval system, without written permission from the publishers.

For permission to photocopy or use material electronically from this work, please access www.copy-
right.com (http://www.copyright.com/) or contact the Copyright Clearance Center, Inc. (CCC),
222 Rosewood Drive, Danvers, MA 01923, 978-750-8400. CCC is a not-for-profit organization that
provides licenses and registration for a variety of users. For organizations that have been granted a
photocopy license by the CCC, a separate system of payment has been arranged.

Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and
are used only for identification and explanation without intent to infringe.

Library of Congress Cataloging-in-Publication Data

Names: Ramanathan, R., 1966- editor. | Mathirajan, M., editor. | Ravindran, A.,
1944- editor.
Title: Big data analytics using multiple criteria decision-making models /
edited by Ramakrishnan Ramanathan, Muthu Mathirajan, A. Ravi Ravindran.
Description: Boca Raton : Taylor & Francis, CRC Press, 2017. | Series: The
operations research series | Includes bibliographical references and index.
Identifiers: LCCN 2016056409| ISBN 9781498753555 (hardback : alk. paper) |
ISBN 9781498753753 (ebook)
Subjects: LCSH: Big data. | Multiple criteria decision making. | Business
logistics--Decision making
Classification: LCC QA76.9.B45 B5535 2017 | DDC 005.7--dc23
LC record available at https://lccn.loc.gov/2016056409

Visit the Taylor & Francis Web site at

http://www.taylorandfrancis.com
and the CRC Press Web site at
http://www.crcpress.com
Contents

Preface..................................................................................................................... vii
Editors.......................................................................................................................ix
Acknowledgments............................................................................................... xiii
Contributors............................................................................................................xv

1. Multicriteria Leadership and Decisions: Festschrift

in Honor of Ravi Ravindran..........................................................................1
A. B. Badiru

2. Multi-Criteria Decision Making: An Overview

and a Comparative Discussion................................................................... 21
Ramakrishnan Ramanathan, A. Ravi Ravindran, and Muthu Mathirajan

3. Basics of Analytics and Big Data................................................................ 67

U. Dinesh Kumar, Manaranjan Pradhan, and Ramakrishnan Ramanathan

4. Linear Programming (LP)-Based Two-Phase Classifier for

Solving a Classification Problem with Multiple Objectives................ 87
Sakthivel Madankumar, Pusapati Navya, Chandrasekharan Rajendran,
N. Srinivasa Gupta, and B. Valarmathi

5. Multicriteria Evaluation of Predictive Analytics for Electric

Utility Service Management..................................................................... 113
Raghav Goyal, Vivek Ananthakrishnan, Sharan Srinivas,
and Vittaldas V. Prabhu

6. Multiobjective Forecasting: Time Series Models Using

a Deterministic Pseudo-Evolutionary Algorithm................................. 135
Nagulapally Venkat Ramarao, P. Y. Yeshwanth Babu,
Sankaralingam Ganesh, and Chandrasekharan Rajendran

7. A Class of Models for Microgrid Optimization................................... 155

Shaya Sheikh, Mohammad Komaki, Camelia Al-Najjar,
Abdulaziz Altowijri, and Behnam Malakooti

8. A Data-Driven Approach for Multiobjective Loan Portfolio

Optimization Using Machine-Learning Algorithms and
Mathematical Programming..................................................................... 175
Sharan Srinivas and Suchithra Rajendran

v
vi Contents

9. Multiobjective Routing in a Metropolitan City with

Deterministic and Dynamic Travel and Waiting Times,
and One-Way Traffic Regulation.............................................................. 211
Swaminathan Vignesh Raja, Chandrasekharan Rajendran,
Ramaswamy Sivanandan, and Rainer Leisten

10. Designing Resilient Global Supply Chain Networks over

Multiple Time Periods within Complex International
Environments............................................................................................... 243
Rodolfo C. Portillo

11. MCDM-Based Modeling Framework for Continuous

Performance Evaluation of Employees to Offer Reward
and Recognition........................................................................................... 269
S. S. Sreejith and Muthu Mathirajan

12. Use of DEA for Studying the Link between Environmental

and Manufacturing Performance............................................................. 303
Ramakrishnan Ramanathan

13. An Integrated Multicriteria Decision-Making Model

for New Product Portfolio Management................................................. 315
Pulipaka Kiranmayi and Muthu Mathirajan
Index...................................................................................................................... 355
Preface

The idea to produce this book originated in an International Symposium

(http://www.ravisymposium.org/) organized in honor of Professor Ravi
Ravindran in Bangalore during 12th–13th March 2015. The symposium was
organized to commemorate 70th birthday of Professor Ravindran by his for-
mer doctoral students. The aim of the symposium was to create a platform
and facilitate knowledge sharing on the applications of operations research
(OR) in multiple criteria decision making, analytics, healthcare delivery sys-
tems, supply chain engineering, and project management. At the valedictory
session, various themes for a book were brainstormed and it was decided
to pursue the first two themes of the symposium, namely multiple criteria
decision-making and analytics, in developing a book that would fittingly
honor the extraordinary achievements of Professor Ravindran for a long
time. It was also decided that the book will include chapters from Professor
Ravindran’s legacy—from his PhD students who are very successful aca-
demics/industrialists—in various parts of the world.
Multicriteria decision making (MCDM) is a subfield of operational
research and is one of the so-called decision-making tools. A decision-mak-
ing problem is characterized by the need to choose one or a few from among
a number of alternatives. A good decision-making process should not only
improve the clarity of the problem to the decision maker, but it should also
shed new light into the problem by generating newer alternatives. The field
of MCDM has been succinctly defined in the literature as making decisions
in the face of multiple conflicting objectives. Chapter 2 of this volume pro-
vides a detailed description of the various MCDM models and also a com-
parative perspective on these models.
The field of MCDM assumes special importance in this era of big data
and business analytics (BA). In the modern digital world, a wealth of so-
called big data is being generated every minute, every second, even every
nanosecond. Thanks to the astounding technological revolution, everything
around us is being captured in someway or the other, stored in some form,
and it is believed that this has the potential to make better business deci-
sions. BA involves an appropriate use of analytic tools on big data to provide
new predictive/prescriptive/descriptive insights that will allow businesses
perform better. Since big data and BA are relatively recent phenomena,
studies on understanding the power of big data and BA are rare with a few
studies being reported in the literature. Chapter 3 of this volume is dedi-
cated to address the basics of big data and BA. BA involves both modeling-
based tools and statistics-based tools. The modeling-based tools involve use
of operational research models. In this volume, the focus is primarily on

vii
viii Preface

modeling-based tools for BA, with exclusive focus on the subfield of MCDM
within the domain of operational research.
We believe that the two themes of the book, MCDM and big data, address
a very valuable research gap. While there are several textbooks and research
materials in the field of MCDM, there is no book that discusses MCDM in the
context of emerging big data. Thus, the present volume addresses the knowl-
edge gap on the paucity of MCDM models in the context of big data and BA.
There was an instant response from Professor Ravindran’s students and
colleagues for the call for contributions of the book. A total of 15 chapters
were considered in the first round of review. Though all of them were of
good quality, after careful review and evaluation for the fit for the theme
of the book, it was decided to include 13 chapters in this volume. At least
five of these chapters have been authored by students and close associates
of Professor Ravindran. There are contributions from authors based in the
United States (5 chapters), from the United Kingdom (2), and from India (6).
This volume starts with a fitting Festschrift in Honor of Professor
Ravindran by Professor Adedeji B. Badiru. The rest of the volume is broadly
divided into three sections. The first section, consisting of Chapters 2 and 3,
is intended to provide the basics of MCDM and big data analytics. The next
section, comprising Chapters 4 through 10, discusses applications of tradi-
tional MCDM methods. The last section, comprising the final three chapters,
discusses the application of more sophisticated MCDM methods, namely,
data envelopment analysis (DEA) and the analytic hierarchy process.
Due to the topical nature of the theme of big data, it has been a challenge
to ensure that the contributions of this volume, from traditional MCDM
researchers, had adequate treatment of big data. We believe that the chapters
of this book illustrate how MCDM methods can be fruitfully employed in
exploiting big data, and will kindle further research avenues in this exciting
new field. We also believe that the book will serve as a reference for MCDM
methods, big data, and linked applications.

Ramakrishnan Ramanathan
Muthu Mathirajan
A. Ravi Ravindran
Editors

Ramakrishnan Ramanathan is the Director of Business and Management

Research Institute, in the Business School of the University of Bedfordshire,
Luton, the United Kingdom. In the past, he has worked and taught in a num-
ber of countries, including the United Kingdom, Finland, the Netherlands,
Oman, and India. He has taught basic and advanced courses on operations
management, production systems management, supply chain management,
optimization theory, DEA, management science, business statistics, simula-
tion, energy and environment, energy and environmental economics, energy
and transport economics, and others. His research interests include opera-
tions management, supply chains, environmental sustainability, economic
and policy analysis of issues in the energy, environment, transport, and
other infrastructure sectors. He works extensively on modeling using tech-
niques such as optimization, decision analysis, DEA, and the analytic hier-
archy process.
Ram has successfully completed a number of research projects across the
world. He is on the editorial boards of several journals and in the technical/
advisory committees of several international conferences in his field. He is
an advisory board member of an innovative new online resource, The Oxford
Research Encyclopedia of Business and Management. He is a member of ESRC
Peer Review College in the United Kingdom. He has produced four books
(including an introductory textbook on DEA), more than 119 research publi-
cations in journals, and more than 141 conference presentations. His research
articles have appeared in many prestigious internationally refereed jour-
nals, including Omega, Tourism Economics, International Journal of Production
Economics, Supply Chain Management, International Journal of Operations &
Production Management, European Journal of Operational Research, Transport
Policy, and Transportation Research.
More details about Professor Ramanathan are available at the following
links:

Profile: http://www.beds.ac.uk/research-ref/bmri/centres/bisc/people/
ram-ramanathan
http://www.beds.ac.uk/howtoapply/departments/business/staff/
prof-ramakrishnan-ramanathan
LinkedIn: http://www.linkedin.com/pub/ramakrishnan-ramanathan/
12/50a/204
ResearcherID: http://www.researcherid.com/rid/H-5206-2012
Google Scholar: http://scholar.google.co.uk/citations?user=1CBQZA8
AAAAJ

ix
x Editors

Facebook: http://www.facebook.com/ProfRamanathan
Twitter: @ProfRamanathan

Muthu Mathirajan has been working at the Department of Management

Studies, Indian Institute of Science (IISc), Bangalore since April 1986 with var-
ious academic/faculty positions. Currently he is a Chief Research Scientist at
IISc, Bangalore. He received MS (Engineering) degree by research in applied
operations research area, and PhD degree in operations management area
from the Faculty of Engineering, IISc, Bangalore. In addition, he holds an MSc
degree in Mathematics of Madurai Kamaraj University, and Postgraduate
Diploma in operations research (OR) of Anna University, Chennai.
During May 2008–May 2010, Dr. Mathirajan was with Anna University
of Technology, Tiruchirappalli, on deputation, and he was the Professor
of Planning and Development at the University level and he was also the
Professor and Head of the Department of Management Studies of the Anna
University of Technology, Tiruchirappalli.
Dr. Mathirajan was a postdoctoral fellow at Singapore MIT Alliance
(SMA) of Nanyang Technological University, Singapore. He was also a vis-
iting consultant at Sultan Qaboos University (SQU), Oman. Dr. Mathirajan
was selected and nominated as young Indian representative of Operational
Research Society of India (ORSI) to present a paper in the 1999 Fall Annual
Conference of ORSJ, Tokyo, Japan.
Dr. Mathirajan’s PhD thesis was adjudged as best thesis for “M. N. Gopalan
Award of 2002—Annual Convention of ORSI.” Dr. Mathirajan’s research
interests are in the development of mathematical modeling and heuris-
tic methods for the problems related to industrial engineering, operations,
logistics and supply chain management in manufacturing, service, and
container terminal management areas. He has published over 155 research
articles and 3 books.
He has guided a number of graduate and postgraduate projects. So far
seven dissertations were awarded PhD degree under his guidance in IISc,
Bangalore. Currently, he has been guiding two doctoral research students at
IISc; and he has examined several PhD theses from various higher learning
Institutes (such as IITs, IIM, NITs) and Universities (such as Anna University
Chennai, Gandhigram Rural University, Madurai Kamaraj University,
Bharathiar University, Delhi University, Dr. MGR University).

A. Ravi Ravindran is a Professor and past Department Head of Industrial

and  Manufacturing Engineering at the Pennsylvania State University
since 1997. Formerly, he was a faculty member in the School of Industrial
Engineering at Purdue University for 13 years (1969–82) and at the University
of Oklahoma for 15 years (1982–97). At Oklahoma, he served as the Director of
the School of Industrial Engineering for 8 years and as the Associate Provost
of the university for 7 years with responsibility for budget, personnel, and
Editors xi

space for the academic area. He holds a BS in Electrical Engineering with

honors from BITS, Pilani, India. His graduate degrees are from the University
of California, Berkeley, where he received an MS and a PhD in Industrial
Engineering and Operations Research.
Dr. Ravindran’s area of specialization is operations research with research
interests in multiple criteria decision making, financial engineering, health-
care delivery systems, and supply chain optimization. He has published seven
books (Operations Research: Principles and Practice, Engineering Optimization:
Methods and Applications, Handbook of Operations Research and Management
Science, Operations Research Methodologies, Operations Research Applications,
Supply Chain Engineering: Models and Applications, Multiple Criteria Decision
Making in Supply Chain Management) and over 150 journal articles in opera-
tions research. His recent text book on Supply Chain Engineering received the
Institute of Industrial Engineers Book of the Year Award in 2013.
Dr. Ravindran is a Fulbright Fellow and a Fellow of the Institute of Industrial
Engineers (IIE). In 2001, he was recognized by IIE with the Albert G. Holzman
Distinguished Educator award for significant contributions to the industrial
engineering profession by an educator. In 2013, he received the Outstanding
Teaching Award in the College of Engineering from the Penn State Engineering
Alumni Society. He also received the Outstanding Faculty Award from the
IE Graduating Seniors in spring 2014. He has been a consultant to AT&T,
CNH America, General Motors, IBM, Kimberly Clark, General Electric, U.S.
Department of Transportation, the Cellular Telecommunications Industry
Association, and the U.S. Air Force.
Acknowledgments

First and foremost, the editors acknowledge the extraordinary effort, dedica-
tion, and leadership of Dr. P. Balasubramanian, Founder and CEO of Theme
Work Analytics, who was instrumental in organizing the international sym-
posium in 2015 in Bangalore to honor Professor Ravi Ravindran. Participants
of the symposium, particularly Dr. M. Mathirajan, Chief Research Scientist
at IISc, Bangalore (one of the editors of this book); Professor S. Sadogopan,
the Director of IIT-Bangalore; and Mr. Harsha Kestur, the Vice President of
National Education Foundation, Bangalore, were instrumental for the gen-
esis of this book.
Next, we would like to thank the authors, who have worked diligently in
producing the book chapters that are comprehensive, concise, and easy to
read, bridging the theory and practice. The development and evolution of
this book have also benefitted substantially from the advice and counsel of
our colleagues and friends in academia and industry, who are too numerous
to acknowledge individually.
Finally, we would like to thank Cindy Carelli, senior acquisition editor and
Ashley Weinstein, project coordinator at CRC Press for their help from the
book’s inception until its publication.

xiii
Contributors

Camelia Al-Najjar Mohammad Komaki

Case Western Reserve University Department of Electrical Engineering
Cleveland, Ohio and Computer Science
Case Western Reserve University
Abdulaziz Altowijri Cleveland, Ohio
Case Western Reserve University
Cleveland, Ohio U. Dinesh Kumar
Indian Institute of Management
Vivek Ananthakrishnan Bangalore
Pennsylvania State University Bengaluru, Karnataka, India
University Park, Pennsylvania
Rainer Leisten
P. Y. Yeshwanth Babu Department of Mechanical and
Latentview Analytics Process Engineering
Chennai, Tamil Nadu, India University of Duisburg-Essen
Duisburg, Germany
A. B. Badiru
Department of Systems Engineering Sakthivel Madankumar
and Management Department of Management Studies
Air Force Institute of Technology Indian Institute of Technology
Dayton, Ohio Madras
Chennai, Tamil Nadu, India
Sankaralingam Ganesh
Latentview Analytics Behnam Malakooti
Chennai, Tamil Nadu, India Department of Electrical Engineering
and Computer Science
Raghav Goyal Case Western Reserve University
Pennsylvania State University Cleveland, Ohio
University Park, Pennsylvania
Muthu Mathirajan
N. Srinivasa Gupta Department of Management Studies
Manufacturing Division Indian Institute of Science
School of Mechanical Engineering Bengaluru, Karnataka, India
VIT University
Vellore, Tamil Nadu, India Pusapati Navya
Department of Management Studies
Pulipaka Kiranmayi Indian Institute of Technology
Indian Institute of Science Madras
Bengaluru, Karnataka, India Chennai, Tamil Nadu, India

xv
xvi Contributors

Rodolfo C. Portillo Shaya Sheikh

Amazon School of Management
Houston, Texas New York Institute of Technology
New York, New York
Vittaldas V. Prabhu
Pennsylvania State University Ramaswamy Sivanandan
University Park, Pennsylvania Department of Civil Engineering
Indian Institute of Technology
Manaranjan Pradhan Madras
Indian Institute of Management Chennai, Tamil Nadu, India
Bangalore
Bengaluru, Karnataka, India S. S. Sreejith
Department of Management
Chandrasekharan Rajendran
Studies
Department of Management Studies
Faculty of Engineering
Indian Institute of Technology
Indian Institute of Science
Madras
Bengaluru, Karnataka, India
Chennai, Tamil Nadu, India

Suchithra Rajendran Sharan Srinivas

Department of Industrial and Department of Industrial and
Manufacturing Systems Manufacturing Systems
Engineering Engineering
and and
Department of Marketing Department of Marketing
University of Missouri University of Missouri
Columbia, Missouri Columbia, Missouri

Ramakrishnan Ramanathan B. Valarmathi

Business & Management Research Soft Computing Division
Institute School of Information Technology
University of Bedfordshire and Engineering
Luton, UK VIT University
Vellore, Tamil Nadu, India
Nagulapally Venkat Ramarao
Latentview Analytics Swaminathan Vignesh Raja
Chennai, India Department of Management
Studies
A. Ravi Ravindran Indian Institute of Technology
Pennsylvania State University Madras
University Park, Pennsylvania Chennai, Tamil Nadu, India
1
Multicriteria Leadership and Decisions:
Festschrift in Honor of Ravi Ravindran

A. B. Badiru

CONTENTS
1.1 Background...................................................................................................... 1
1.2 The Tinker Projects.........................................................................................2
1.3 Reprint of a Tinker Project Article...............................................................3
1.4 Conclusion..................................................................................................... 18
References................................................................................................................ 18

1.1 Background
I completed my PhD in industrial engineering at the University of Central
Florida (UCF) in Orlando, Florida in December 1984, which was an off-­season
PhD completion cycle for the purpose of securing academic positions. The
graduation ceremony at UCF was on Friday, December 14, 1984. Having been
offered a job, bigheartedly, by Ravi, my wife and I packed up and set out
for Norman, Oklahoma with our two young children the same afternoon
after the graduation ceremony. After 3 days of driving a combination of a
U-Haul Truck and our old family car, we arrived in Norman on Monday,
December 17 in the morning. I called Ravi that I had arrived in town and
he immediately informed me that the last department meeting of the year
was taking place that very afternoon and he would like me to attend. Casual
as that invitation might seem, it said a lot about the caliber of the go-getter
that Ravi was and still is. Ravi never missed an early opportunity to put
us to work constructively as a way to start building our faculty profiles on
our match toward earning tenure at the University of Oklahoma. Although
I declined to attend that meeting, partly because I was tired from the long
road trip and partly because I feared getting to work right away, that engage-
ment opened my eyes to the need to be ready to take advantage of Ravi’s
mentoring ways. He would cajole, persuade, entice, and or coax the young
assistant professors under his charge into pursuing the latest line of research

1
2 Big Data Analytics Using Multiple Criteria Decision-Making Models

and journal publications. He got us engaged in funded projects to expose us

to the world of pursuing funded projects, writing proposals, and executing
projects. He would come to our offices and inquired what we were working
on and whether we were aware of some latest opportunity out there for a
funded project or a journal publication. At that time, we did not see the value
of his inquisitive ways. It was later that we realized how much his gentle
prying would put on a solid platform of becoming successful tenured profes-
sors. In fact, we, the assistant professors, often joked among ourselves that if
we see Ravi coming down the hallway, we would go the other way because
each time you meet him, he would have a new idea of something new and
worthwhile for us to be doing.
Ravi’s altruistic disposition was evident in the fact that he offered me a job
at all. At that time, I was still on a student visa and not many departments
were eager to make academic appointments without an existing “green
card” or an official work permit. Ravi was among the handful of department
heads willing to go out on a limb to offer jobs to inexperienced foreign stu-
dents. Ravi took a chance with my offer and I am greatly appreciative of that
opportunity. Even today, whenever I strive to achieve an even loftier goal, I,
subconsciously, credit the endeavor to a justification of the incipient oppor-
tunity that Ravi gave me so long ago in 1984.

1.2  The Tinker Projects

The multicriteria leadership of Professor Ravi Ravindran led to many excit-
ing times in the Industrial Engineering Department at the University of
Oklahoma. The most notable of these were the series of projects we did
under contract for Tinker Air Force Base in Oklahoma City. Ravi brought
me into his Tinker research team in 1985, my first year at the University. The
project won us international acclaim and many awards and helped launch
our academic careers on a positive trajectory. The primary publications that
emanated from the projects are the six references cited in this contribu-
tion (Ravindran et al., 1988, 1989; Foote et al., 1988, 1992; Leemis et al., 1990;
Badiru et  al., 1993). Several other publications followed these six primary
articles. It is noteworthy that Ravindran et al. (1989) was recognized as one
of the 20 best papers of the decade published in TIMS Interfaces (1980–1990).
The “Tinker Projects,” as they were affectionately called, culminated in the
team’s international recognition with the 1988 Finalist Achievement Award
for the Franz Edelman Management Science Award from The Institute of
Management Sciences. Thus, the team’s accomplishment is permanently
enshrined in the annals of the award winners and is still recognized annu-
ally until today. Figure 1.1 shows an image of the Edelman Laureate Ribbon
presented to me at the INFORMS conference in Philadelphia, Pennsylvania
Multicriteria Leadership and Decisions 3

FIGURE 1.1
Edelman Laureate Ribbon recognizing Ravi Ravindran’s team accomplishment with the
Tinker Air Force Base Project in 1988.

in November 2015. So, even after 30 years, the Tinker Projects continue to
bear intellectual fruits.

1.3  Reprint of a Tinker Project Article*

This section contains a reprint of one of the seminal journal articles pub-
lished on the Tinker Project.
Abstract
We developed a large simulation model to aid reconstruction efforts after
a disastrous fire at Tinker Air Force Base (TAFB). The model, developed in
SLAM, facilitated the analysis and efficient design of the modular repair
center (cellular type) layout that replaced the precious machine-based layout
in the engine overhaul facility. It has been used extensively to determining
the appropriate number of machines to place within the repair center, the
stacker capacity for in-process inventory, the location of elevators for send-
ing parts to the conveyor, and the optimal design and routing scheme for
the overhead conveyor system. The new layout, as predicted by the simula-
tion model, has proven to be quite effective. The new design has decreased
material handling by 50 to 80 percent, decreased flow times, allowed better
management control of part transfers, saved $4.3 million from the elimina-
tion of excess machine capacity, and saved $1.8 million from higher direct
labor efficiency.
Tinker Air Force Base (TAFB), located in Oklahoma City, Oklahoma, is
one of the five overhaul bases in the Air Force Logistics Command. It over-
hauls and repairs six types of jet engines and various aircraft and engine

* Reprinted verbatim with Permission from: Ravindran, A.; Foote, B. L.; Badiru, Adedeji B.;
Leemis, L. M.; and Williams, Larry (1989), “An Application of Simulation and Network
Analysis to Capacity Planning and Material Handling Systems at Tinker Air Force Base,”
TIMS Interfaces, Vol. 19, No. 1, Jan.–Feb., 1989, pp. 102–115.
4 Big Data Analytics Using Multiple Criteria Decision-Making Models

accessories, and it manages selected Air Force assets worldwide. This engine
overhaul facility is responsible for logistical support for a series of Air Force
engines. Engines are returned from service activities for periodic overhaul
or to complete a modification or upgrade. The engine is disassembled, and
each part is inspected for wear and possible repair. Individual parts are
repaired or modified to a like-new condition or are condemned and replaced
with a new part. The majority of the parts are overhauled and returned to
service for a fraction of the cost of a new part. A major overhaul may cost less
than five percent of the cost of a new engine in terms of labor, material, and
replaced parts, Between November 11 and 14, 1984, a fire devastated Building
3001, which contained the Propulsion (Engine) Division in the Directorate of
Maintenance. The division consists of over 2800 employees and produces
over 10 million earned hours to support Department of Defense overhaul
requirements each year. In February 1985, the Air Force published a state-
ment of work requesting assistance from industry to model and develop a
simulation of the engine overhaul process to assist in the redesign and lay-
out of approximately 900,000 square feet of production floor space. Three
commercial firms attended an onsite prebid conference to learn the scope
of the project, the nature of the data the Air Force could provide, and the
time frame in which a finished product had to be delivered. The model was
expected to predict the number and type of machines, the personnel, the
queuing space required, the material-handling distribution, and the volume
between and within organizations. The facility engineers needed various
management reports to help them to lay out the plant. The project was to be
completed within 120 days. Of the three firms, one elected not to respond,
the second bid $225,000 with the first report in nine months, and the third
quoted $165,000 to study the problem with an expected projected cost of
over $300,000. Each was a highly reputable organization with considerable
expertise and success in the field. Since TAFB had budgeted only $80,000
and time was running out, TAFB contacted the University of Oklahoma. It
had not been considered earlier because of conflicts with class schedules.
As the month of May approached, the university became a potential vendor.
A contract was let on May 1, 1985, and the first product was delivered by June
15, 1985.

Project Scope
The scope of the project was to take advantage of the disaster and forge a
state-of-the-art facility for overhauling engines with the most efficient and
cost-effective organizational structure and physical layout. The relocation
team was charged with developing and implementing a total change in the
philosophy of engine overhaul that would maximize flexibility while mini-
mizing facility and plant equipment costs. Of equal importance was the task
of developing a means to predict and forecast resource requirements as work
load mixes changed.
Multicriteria Leadership and Decisions 5

The eight-member relocation team comprised midlevel managers from

engineering and production with detailed knowledge of the inner working
of the facility plus four faculty members and four graduate students from
the University of Oklahoma School of Industrial Engineering. The Corps of
Engineers was to construct the building based on specifications provided by
the internal engineering department. Mechanical and industrial engineers
designed shop layouts. The university provided the skill and knowledge to
develop a capacity-planning and material-handling simulation model using
data provided by TAFB. The university team simulated repair activities to a
level of detail never attempted or realized before. Its responsibility was to ana-
lyze the data available from TAFB and determine what, if any, additional spe-
cific data elements were required, to assist in developing techniques to obtain
that data from existing systems, to check the data for outliers, and to develop
and implement interface programs to obtain data for the simulation model.
The baseline data base consisted of 117 fields with over 2500 records used
to describe the requirements of the organization by individual type of part
being repaired. The university used this raw data to create forecasts and net
equipment requirements for each individual modular repair center (MRC)
based on variable mixes of workloads and resources. The data base provided
the following information:
A work control document (WCD), a unique identifier for each engine part;
The annual requirement of each end item (engine or subassembly) of which
the part to be repaired is a component (the WCD attached to each part car-
ried this information):

• The sequential routes of the part and resource requirements coded

by the industrial process code with labor and machine/process time
required at each resource;
• The size and weight of the part so that storage and queue space can
be estimated, and
• The number of units per assemble (UPA) required of each part to
make up the end item.

Prior to the fire, the division was organized along functional operational
lines with each department responsible for a specific process, such as machin-
ing, welding, cleaning, or inspection. This organization structure was devel-
oped in 1074 when engine overhaul functions were consolidated into one
organization. At that time, such functional shop layouts maximized equip-
ment utilization and skill concentrations since a typical long-flow part would
require 30 to 50 production operations and change organizations only seven
to 10 times. Today, the same part requires over 120 production operations
and changes organizations as many as 30 to 50 times. This increase has been
caused by incremental introduction of technology and by improved repair
procedures that offset wear of critical engine parts and reduce replacement
6 Big Data Analytics Using Multiple Criteria Decision-Making Models

costs. The additional repairs increased routing that overburdened the mecha-
nized conveyor system. Since 1974, the only major change was an experiment
three years prior to the first to consolidate one part-type family, combustion
cans, into a partially self-contained work center.
The reconstruction period after the fire gave TAFB a unique opportunity
to design a modern production system to replace the one destroyed. TAFB
manufacturing system analysts changed the repair process from a process
specialization type of operation to a family (group) type of operation. Staff
from the University of Oklahoma helped to solve the problems associated
with ling flow types, lack of clear responsibility for quality problems, and
excessive material handling. The plan for reconstruction was based on the
concept of a modular repair canter (MRC), a concept similar to the group
technology cell (GTC) concept except that it is more interrelated with other
centers than a GTC.
We created and defined the modular repair center concept as a single orga-
nization to inspect and repair a collection of parts with similar geometries
and industrial processes so as to provide the most economical assignment
of equipment and personnel to facilitate single point organizational respon-
sibility and control. An example of such a center is the blade MRC, which
repairs all turbine blades from all engine types. With the exception of initial
chemical cleaning, disassembly, plating, paint, and high temperature heat
treatment, all industrial equipment and processes were available for assign-
ment to an MRC.
Since TAFB lost an entire overhead conveyor system in the fire, imple-
menting the MRC concept required a new conveyor design in terms of rout-
ing, size, and location of up and down elevators. The new system needed
a conveyor to move parts to their respective MRCs from the disassembly
area and to special areas such as hear treatment, painting, or plating and
back to engine reassembly. When an engine arrives for repair, its turbine
blades are removed and routed via the overhead conveyor to the blade
MRC, out to heat treatment, painting, and plating, back to the MRC, and
finally returned to be assembled back into an engine. A stacker (mecha-
nized inventory storage system) in each MRC handles excess in-process
queues that are too large for the finite buffer storage at each machine. One
of the functions of the simulation model was to compute the capacity of the
buffers and stacker.

Data Analysis
Standard sources at TAFB provided the information for analysis. The first
source, the work control documents (WCD), gives the operation sequences for
all the parts. It tells which MRC a part goes to and the sequence of machines
the part will visit within the MRC. There are 2,600 different WCDs, with as
few as 11 assigned to combustion cans and as many as 700 assigned to the
general shop.
Multicriteria Leadership and Decisions 7

TABLE 1.1
Material Handling Codes
Alphanumeric Code Weight of Item (lbs.)
A 0–1
B 1–5
C 5–10
D 10–25
E 25–50
F >50
W 0–6 6–12 12–24 24–48 >48
L
0–6 1 2 4 7 11
6–12 2 3 5 8 12
12–24 4 5 6 9 13
14–48 7 8 9 10 14
>48 11 12 13 14 15

The second data source, the engine repair plan, showed how many engines
of each type were expected to be repaired each year. We used the fiscal ’85
requirements and a projected annual work load of 2000 engine equivalents
to determine how many units of each family type would enter the system.
Table  1.1 presents the material handling codes. The top part of the Table
gives the code for the six different weight categories while the lower part (the
matrix) expresses the code for 15 different categories of length and width of
the base of the part, which rests on the pallet. Each number represents a com-
bination of length (L) and width (W), measured in inches. D9, for instance,
means a part that weighs 10–25 pounds and has a base whose length is
between 12 and 24 inches and whose width is between 24 and 48 inches.
The third source of data, the TAFB standard material handling (MH) cod-
ing of each part, was based on the size and weight of each part. Parts move
on pallets at TAFB. We used the MH coding to estimate the number of parts
per pallet (see Table 1.1). TAFB engineers had decided on the shop configura-
tion and location of the MRCs but had not determining their physical dimen-
sions prior to our analysis. The configuration was based on groupings of jet
engine parts with similar geometries, metal types, and repair processes (for
example, major cases, rotating components). The MRCs are N-nozzle, S-seal,
B-bearing housing, GX-gear box, TR-turbine compressor rotor, K-combustion
can, BR-blade, AB-after burner, C-case, CR-compressor rotor, ZH-general
handwork, ZM-general machining, ZW-general welding. In addition, gen-
eral purpose shops handle painting, plating, heat treatment, blasting and
cleaning. Since several hundred units of each WCD are processed, the facil-
ity handles over one-half million units annually. Each WCD is assigned to
one of the MRCs and goes through several processes, comprising 25 to 100
8 Big Data Analytics Using Multiple Criteria Decision-Making Models

operations each. Each MRC handles from 11 to 700 WCDs and has between
19 and 83 processes assigned to it.

Material Handling Characteristics

To conserve space and energy, parts are stored on pallets; in some cases, two
different WCDs are stored on the same pallet. Parts that are large in two dimen-
sions but small in a third are stacked. The number of WCDs on a pallet is a
random variable depending on how many parts arrive when. To convert a flow
of WCDs to a flow of pallets, we used a simple formula to estimate the total
number of pallets that would flow between operations given the number of
each part type that would move between those two operations. Table 1.2 shows
the parts per pallet material handling (MH) codes. The matrix in the Table
expresses the number of parts that can be placed on a pallet, given the MH
code from Table 1.1. Two parts per pallet can be carried if the part is coded D9.
The small values represent fixed loads based on weight and size, while larger
values are averages of actual usage, since a pallet can carry many small parts.
To obtain the pallet factor estimate OPF (see appendix), we used the TAFB
material handling codes (Tables 1.1 and 1.2). If OPF = 0.2 and 50 units of all
types flow from area i to area j per half hour, then (0.2)(50) = 10 pallets will
move on the average per half hour. Other technical details can be found in
Foote et al. (1988).

TABLE 1.2
Parts per Pallet Material Handling Codes
Weight Code
Size Code A B C D–E F
1 50 30 20 10 5
2 8 8 8 4 4
3 8 8 8 4 4
4 8 6 5 3 2
5 8 6 5 3 2
6 8 6 5 3 2
7 8 6 5 3 2
8 4 4 4 2 2
9 4 4 4 2 2
10 4 4 4 2 2
11 4 4 4 2 2
12 2 2 2 1 1
13 2 2 2 1 1
14 2 2 2 1 1
15 2 2 2 1 1
Multicriteria Leadership and Decisions 9

General Model for Conveyor Design

To establish a basis for building a minimum-size conveyor system to handle
the work load, we constructed a network model of the material-handling
system. A minimum conveyor system is one that has the least length with
the most flexibility and that meets all production volume requirements with-
out logjams. In the network representation, 56 nodes represent (1) different
MRCs and their possible associated loading/unloading points, (2) the assem-
bly areas, (3) the possible transfer points in the conveyor, and (4) general
purpose shops. Arcs or links in the network represent the possible different
sections of the conveyor linking nodes. The arrows on the arcs show the
direction items flow (one way or two way). Using the conveyor system draw-
ings, we calculated the distances between all pairs of nodes to find the linear
feet of conveyor. The numbers associated with arcs represent these distances.
We used Floyd’s algorithm (Floyd, 1962), which analyzed in Dreyfus (1969),
to calculate the shortest distance between all pairs of nodes. The algorithm
also determines the shortest path, namely, the optimal sequence of arcs (con-
veyor sections) to travel in order to minimize the total travel time from any
department to any other department. Ravindran et al. (1988) cover the details
of the conveyor design, including how the random variation in pallet flow on
the conveyor was handled and how conveyor bottlenecks were eliminated.
Figure 1.2 shows the old and new layouts and the associated conveyor sys-
tems. The figure presents the conveyor system pre- and postfire. The top
layout shows Building 3001 as it was functionally laid out before the fire.
The bottom figure shows the new layout based on a cellular manufacturing
organization with conveyor routes optimized by Floyd’s Algorithm.

Computer Generated Data

From the processing sequence on the work control document and the num-
bers of engines that need to be maintained, we calculated the flow from each
MRC to other MRCs. We wrote a computer program to scan the process-
ing sequence and determine when a move out of the MRC would be made.
For example, when the process code for heat treatment appeared, the item
would move from its MRC to hear treatment and then back to the MRC. The
number of items of each type moving was the number of engines times the
number of parts of this type per engine. We then summed the movements
between each pair of locations over all part types, and converted the move-
ment in terms of parts to pallets moved per half hour.

The Simulation Model

We wrote the simulation model, called the Tinker Integrated Planning
Simulation (TIPS), using the discrete event orientation in SLAM (Pritsker,
1986); it contains approximately 1750 lines of FORTRAN code. TIPS is
designed to simulate a single MRC at a time. The entities in the model are
10 Big Data Analytics Using Multiple Criteria Decision-Making Models

Building 3001
prefire

Heat
treat

Disassembly
Inspection

Cleaning
Heavy

Weld
machining
General
machining
Plating Paint
Stacker

Rotor
Admin

Assembly mach N.C.

Blast
Gearbox

NDI Blades

Building 3001
postfire

Seals
Heat treat
NDI

Disassembly
Cleaning
Comb
Gen can
Plating Lab Paint
MRC Case
NOZ
Gearbox
Stacker

Material
housing

Turbine
Bearing

Blast

Comp
Admin

rotor

control N.C.
rotor

Blades After
Assembly burner

Conveyor with up
elevator

FIGURE 1.2
Conveyor system pre- and postfire.

the WCDs flowing through one particular MRC. Features of the TIPS model
include three shifts, transfer to other MRC operations (that is, painting, plat-
ing, and heat treatment), and stackers to model WCD storage when machine
queue lengths are exceeded. The simulation model is capable of storing 70,000
entities (concurrent WCDs) in an MRC. Despite this, three of the MRCs were
so large that they had to be broken into smaller family groups.
Figure 1.3 illustrates the system concept of the MRC and how material flows
inside and to external shops. This allows the stacker to be sized by the simu-
lation; the maximum load will determine the size of the stacker installed.
The part shown in Figure 1.3 has a 1-4-3-painting-5 machine sequence.
Multicriteria Leadership and Decisions 11

Process Legend
number
1 2 Machine available on
first shift only
3
Machine available
on first and second shifts
Q Q Q
Route out of MRC
(4-hour material handling delay)

In Painting
Q Stacker Q

Infinite Out
4 capacity 5

FIGURE 1.3
Sample five-machine MRC configuration.

The in-process queue area at the machine is limited. When this is full, the
overflow goes to the infinite capacity stacker. The simulation computed the
maximum stacker storage requirement needed. A route out after machine
number three to painting and return after an 8-hour material handling delay
is shown. The cross hatching on the machines in the diagram indicated the
shifts when they are available. For example, there are six machines of type
number one available during the day shift, and only four available during
the second shift. If a WCD is on a machine when the shift change occurs,
it is assumed that the machine completes processing the WCD prior to the
changeover.
Tinker Air Force Base supplied the data used to determine the rate of flow
of WCDs through each MRC. The data for each MRC came in two sets, the
1985 fiscal year data and the data for 2000 engine equivalents (when the facil-
ity would run at full capacity). Both data sets contained a list of the WCDs for
the MRC, the operations sequence for each WCD, the corresponding machine
process time for each WCD, the corresponding standard labor time for each
WCD, the UPA (units per assembly) number for each WCD the data included,
and a vector containing the relative frequencies of each WCD. In addition,
the projected size of each MRC (for example, number of machines of each
type) and information needed to calculate a From-To matrix (for inter- and
intra-MRC transfers). We transformed all the data to a format that allowed
SLAM to execute the discrete event model.
Two features of the TIPS simulation model make it unique. First, the model
was so large that it used the SLAM language at its maximum configura-
tion to run a single MRC. We had to consult with Pritsker and Associates to
determine how to extend SLAM’s storage limits in the source code. Second,
the model integrated both physical (machines) and skill (labor) resources
12 Big Data Analytics Using Multiple Criteria Decision-Making Models

in a single model that supported a bottle-neck analysis, space analysis, and

overhead-conveyor-routing analysis. We designed the model for managers
and held two training sessions at Tinker AFB to allow managers to use TIPS
for decision making.
A final feature of the system is its generality. Originally, 13 MRCs were to
be modeled. This expanded to 17. We had three months to develop the model
and had to meet the due date. We developed the program using a special
format that allowed the model to be restructured for any MRC. Thus, the
type of machines, their number, and their operations in an MRC were stan-
dard input. The process plan for each WCD was an input data set. With this
structure, any MRC could be simulated. Some features of the model required
special considerations as we constructed the model of the proposed shop
configuration:

• Downtime: Machine breakdown affects the flow time and through-

put for an MRC. We assumed that after each machine processes
a part, a breakdown occurs with probability that depends on the
machine. This assumption is based on the fact that uncompleted
work can be finished on other machines and breakdowns are rare.
This simplified the code, which helped meet the deadline. The time
to repair a machine is exponentially distributed. We based the dis-
tributions and parameters used in the simulation on estimated by
TAFB personnel.
• Interarrival and service time distribution: Since no data were available
on the interarrival distribution of engine inputs, we used a determin-
istic interarrival time based on the annual volume of that particular
WCD. This was reasonable in that repairs of engines are scheduled
uniformly over the year. The service time was a truncated normal
random variable with the range set at μ ± 0.05μ.
• Labor utilization: The modeling of a WCD being processed on a
machine had to incorporate the fact that both a machine and an
operator are required to service the part. In addition, sick leave,
training leave, and vacations for machine operators are modeled.

Calculating the Number of Machines Needed

A prime use of the simulation was to determine the number of machines of
a particular type needed in each MRC. The stated objective was to have 95
percent availability for each machine type; that is, 95 percent of the time a
machine will be available at a machine center when a component arrives. To
determine the smallest number of machines needed to provide 95 percent
availability, we first ran the simulation assuming ample machine capacity
so that there was no queuing at the machine center. We then used the uti-
lization statistics for the case of ample capacity to determine the minimum
number of machines necessary for 95 percent availability.
Multicriteria Leadership and Decisions 13

We had to design the simulation for three shift operations because in

emergencies all three shifts are used, and even on one-shift operations, some
equipment, such as painting, plating, and equipment designed at TAFB, runs
three shifts because it is not economical to duplicate the equipment. Thus,
even though most of the equipment is manned for only one shift, the simula-
tion must be run for three shifts per day. When considering a machine center
that would operate for only one (or two) shift(s) per day, we had to rescale
the utilization statistics from a three-shift day to reflect the shorter work day
before calculating the number of machines needed for 95 percent availability.

Problems with Large MRCs

Some MRCs were too large to be handled by the simulation model both in
terms of unacceptably long run times and memory requirements. As a result,
the large MRCs had to be broken up into smaller family groups. For example,
we broke the combustion can shop (MRC K) down into six families. (We
could do this because the six families in the K shop shared only such enter-
ing processes as inspection, and it was easy to split inputs into six groups.)
TIPS can handle approximately 70,000 entities (or WCDs) at one time. The
run times varied depending on the size of the MRC. For example, the simu-
lation for the estimated repair work load of 2000 engines for the gear box
MRC took approximately one hour to run on an IBM 3081. The run times
on a VAX 11/780 were generally eight times longer than the IBM run times.
In one specific case, MRC CC2 (one of the smallest families in MRC K), the
simulation took 1.5 minutes to run on IBM and 7.8 minutes on VAX. The CC2
shop contains a maximum of five WCDs and can handle up to 73 different
processes. It has an annual work load of about 821 parts. By comparison, the
gear box MRC handles about 329 WCD typed and up to 72 different pro-
cesses. Its annual repair volume is over 100,000 parts. We used a warm-up
period of 13 weeks (one quarter) for each simulation run, and collected statis-
tics on MRCs starting with the 14th week. Simulation outputs were printed
in 13-week time intervals to match regular production runs at TAFB. In the
early testing, we compared the outputs from different warmup periods (13
and 26 weeks) using a t-test to determine if we needed an extra warm-up
quarter. The differences in mean values were insignificant.

Verification
To determine whether the simulation model was working as intended, we
took the following verification steps:

1. We developed the model incrementally. This made it easier to debug

the programs.
2. We analyzed the outputs of each modular component of the overall
model for reasonableness (does the output seem to represent real-world
14 Big Data Analytics Using Multiple Criteria Decision-Making Models

expectations?), consistency (does the output remain about the same for
similar inputs?), reasonable run time (does the program run longer
than expected for the given MRC?), and output (a 10 percent increase
in load should show more than a 10 percent increase in waiting time).

Validation
To validate the model, we made a diagnostic check of how closely the simula-
tion model matched the actual system, taking the following steps:

1. We cross-checked the model assumptions. For example, is the

assumption of normally distributed processing time correct?
2. We compared statistically analyzed results to actual historical data
using a representative MRC simulation, checking both average out-
put and range of output.

Output
The output from TIPS consisted of two documents: the standard SLAM sum-
mary report and a custom printout generated by a FORTRAN subroutine.
The custom output presented the SLAM output in a format and at a level of
detail suitable for prompt managerial decision making. The statistics in the
output included:

1. Machine availability by shift for each process,

2. Maximum queue length in front of each process,
3. Average processing time,
4. Average waiting time for each process,
5. Number of units for each WCD type entering and leaving the system,
6. Part flow (in units) entering and leaving the stacker for each process,
7. Time spent in the stacker waiting for a specific process,
8. Utilization level for each process per shift, and
9. Total time in the system for each WCD, including waiting time, han-
dling time, and processing time. Labor time is assumed to overlap
with processing time.

We wrote supplementary FORTRAN programs to generate certain input

data for the TIPS program. For example, we used a bottleneck program to
set the initial number of machines available for each process. The TIPS pro-
gram is, in effect, the nucleus of an integrated system of management deci-
sion aids, as shown in Figure 1.4. The simulation model is used at Tinker
Air Force Base to perform the functions listed, such as capacity planning,
production planning, and analysis of part flow, and to provide data for pro-
cess design, management control, process capability analysis, and process
Multicriteria Leadership and Decisions 15

Management
process
Engineering Design
process process

Process Capacity
capability Tips planning
analysis

Process Production
monitoring planning

Parts flow
analysis

FIGURE 1.4
Central role of TIPS in shop management.

monitoring to determine if production goals are met, and to meet the needs
of engineering in designing new overhaul procedures.

Design Constraints
Because the fire was so destructive and because slowing maintenance opera-
tions for a long time would seriously affect the national defense, TAFB set
a time limit of three months for designing the analytic computer models.
We developed quick approximations of such items as pallet flow so that we
could test model validity quickly. The closeness of predicted need to actual
need showed that these approximations were acceptable based on the crite-
rion specified by TAFB (plus or minus five percent). The predicted require-
ments have ranged from 100 to 115 percent of actual need.

Input Data Verification

The data for the simulation were delivered on a magnetic tape and consisted
of at least a quarter of a million individual elements. These data were visu-
ally scanned on a random basis and anomalies were noted. We reviewed
these anomalies in joint meetings. We then developed rules for scanning
the entire data set for data errors. These rules formed part of a rudimentary
expert system to improve the quality of the data. The following are examples
of ruled to check processing sequences and time standard data:

1. The heat treatment can never follow painting,

2. Labor or machine standard time could never be zero,
16 Big Data Analytics Using Multiple Criteria Decision-Making Models

3. Machine time is always greater than or equal to labor time, and

4. Drilling times can vary between one and four minutes (these ranges
were MRC specific).

The program checked for missing processes to see if other processes were
required to follow or precede it based on technological constraints and then
determining if the process plan met this constraint.
Errors found by use of the above rules and others were reported to
TAFB personnel for analysis and correction, if necessary. TAFB printed
out violations of the rules and corrected the errors. Some errors were
transpositions and easily corrected, others necessitated quick time studies
or verification of process sequence. Some data rejected by the tests were
actually correct. This phase took two months, but it overlapped the design
of the simulation and material-handling models. The simulation had to
be developed in some detail to calculate the space needed for sequen-
tial queues as parts moved from machine to machine and to capture the
material-handling sequence and routes as we played what-if games with
resources and work-load assignments. We also needed this detail to docu-
ment and solve bottleneck problems within the flow of a single part or for
a combination of parts.

Project Summary
The Air Force started the project in January 1985, approved the organizational
concept in February 1985, developed the industrial process code concept and
started data collection for the data base in late January, and created rough-
cut capacity plans and organizations in March–April 1985, and it required
data to meet material and scheduling lead times by June 15, 1985. We had to
complete all simulations by September 1985 to finalize the resource alloca-
tions and to allow for design lead times. The simulations were used to allo-
cate personnel, machines, and floor space to the various organizations.
Our analysis to aid the transition to the MRC layout included format-
ting the TIPS simulation program, designing the overhead conveyor sys-
tem, laying out the plant, and making a routing analysis of the inter-MRC
transfers. The TIPS program proved valuable in aiding the transition to
the new layout by estimating performance measures (for example, flow-
time and queue statistics) that helped determine the number of machines
of each type to place in an MRC. In one particular instance, the nozzle
MRC, the Production and Engineering Department called for 24 work sta-
tions of a particular type. The TIPS analysis indicated that between 11 and
13 work stations were needed. Based on the TIPS results, only 12 work
stations were installed and this has proved to be sufficient. At Tinker Air
Force Base, simulation proved to be a valuable tool in assessing the effec-
tiveness of the new plans and in determining the right parameters for
each new MRC.
Multicriteria Leadership and Decisions 17

The fire at TAFB was a disaster that turned out to be a great opportunity.
It would have been very hard to justify dislocating the entire facility for over
a year and expending the sum of money required to redesign the facility. In
the long run, the benefits of the new system may pay back the cost of the fire
with interest.
The University of Oklahoma design team responded to the emergency
with accomplishments we are very proud of. We developed and verified a
general simulation model in three months, when national consulting firms
estimated four times that long. The new system met design expectations,
which is rare. Part flow times have been reduced by 35 to 50 percent depend-
ing on size. Labor savings have been $1.8 million in 1987, $2.1 million in 1988,
and continue to rise. $4.3 million was saved in equipment purchases.
Space requirements were reduced by 30,000 square feet. The percent defec-
tive had dropped three percent in 1987 and five percent further in 1988. The
conveyor system has had no jam-ups due to overloading. Finally, other Air
Logistics Centers have adopted the TIPS concept to plan redesign of their
facilities and the report (Ravindran et  al., 1986) has been distributed for
review by over 60 organizations at their request. We have proved that mod-
ern management science techniques can be applied quickly and with great
impact.

Original 1989 Paper Acknowledgments

We express our appreciation to Margarita Beneke, K. Janikiraman, Wan-Seon
Shin, Doug Stewart, and Murali Subramaniam for their assistance in devel-
oping and running the TIPS model, to Sandee Boyer for patiently typing
several drafts of this paper, and to Hussein Saber for his help with Figure 1.2.
In addition, Stephen Graves, Mary Haight, and two referees have improved
the readability of this paper.

Appendix
The overall pallet factor is computed by

{∑ } M

i =1
W1P1

{∑ }
OPF = M
W1
i =1

where
P1 = estimate of pallets/parts given its MH code. For example, if P1 has code
1A, P1 = 1/50 = /02 (see Table 1.2), i = 1,2, …, M,
W1 = (N1)(UPA1), i = 1,2, …, M,
N1 = number of WCD’s of type i per year, i = 1,2, …, M,
18 Big Data Analytics Using Multiple Criteria Decision-Making Models

UPA1 = number of units per WCD, i = 1,2, …, M,

OPF = overall pallet factor, and
M = number of WCD’s.

1.4 Conclusion
As can be seen in the case example of the Tinker Project presented in this
festschrift, Ravi Ravindran has made several multicriteria leadership con-
tributions to several people and several organizations. Under his unflinch-
ing leadership, several PhD and MSc students graduated on the basis of the
Tinker Project. Those individuals continue to make professional and intel-
lectual contributions internationally. The seminal publications that resulted
from Ravi’s leadership of the project continue to guide management science
and operations research practitioners around the world. My personal expo-
sure to Ravi’s academic leadership actually started before I went to work
for him at the University of Oklahoma. In 1982, I was offered admission
to Purdue University for my PhD studies. Ravi Ravindran was at Purdue
University at that time and he was assigned as my initial academic adviser,
as was the practice for all new incoming students at that time. Although I
opted to attend the UCF instead of Purdue University, for financial assistant-
ship package reasons, the written communications with Ravi played a key
role in mind as I continue to dedicate myself to the challenges of doctoral
education. The Purdue admission letter, dated May 17, 1982 and signed by
Professor James W. Barany, associate head, introduced me to Professor Ravi
Ravindran, assistant head. Ravi engaged with me positively through a series
of written communications to encourage me to attend Purdue. It was fortu-
itous for me that Ravi later left Purdue University to become the department
head at the University of Oklahoma, where I ended up working for and with
him for several years until he left to go to the Pennsylvania State University
around 1997. Ravi’s mentoring and nurturing ways continue to influence my
professional activities even today. I am, thus, delighted to be able to contrib-
ute this festschrift in his honor.

References
Badiru, A. B.; B. L. Foote; L. Leemis; A. Ravindran; and L. Williams 1993, Recovering
from a crisis at Tinker Air Force Base, PM Network, Vol. 7, No. 2, Feb. pp. 10–23.
Dreyfus, S. E. 1969, An appraisal of some shortest-path algorithms, Operations
Research, Vol. 17, No. 3, pp. 395–412.
Multicriteria Leadership and Decisions 19

Foote, B.; A. Ravindran; A. B. Badiru; L. Leemis; and L. Williams 1988, Simulation

and network analysis pay off in conveyor system design, Industrial Engineering,
Vol. 20, No. 6, June, pp. 48–53.
Foote, B. L.; A. Ravindran; A. B. Badiru; L. Leemis; and L. Williams 1992, An appli-
cation of simulation and network analysis to capacity planning and material
handling systems at Tinker Air Force Base. In Excellence in Management Science
Practice—A Readings Book, A. Assad, E. Wasil, and G. Lilien, eds., Prentice-Hall,
Upper Saddle River, New Jersey, pp. 349–362.
Floyd, R. W. 1962, Algorithm 97: Shortest path, Communications of the ACM, Vol. 5, No.
6, p. 345.
Leemis, L.; A. B. Badiru; B. L. Foote; and A. Ravindran 1990, Job shop configura-
tion optimization at Tinker Air Force Base, Simulation, Vol. 54, No. 6, June, pp.
287–290.
Pritsker, A and Alan B. 1986, Introduction to Simulation and SLAM II, third edition,
Halsted Press, John Wiley & Sons, New York.
Ravindran, A.; B. L. Foote; A. B. Badiru; and L. Leemis 1988, Mechanized material
handling systems design & routing, Computers & Industrial Engineering, Vol. 14,
No. 3, pp. 251–270.
Ravindran, A.; B. L. Foote; A. B. Badiru; L. Leemis; and L. Williams 1986, Job shop
configuration optimization at Tinker Air Force Base, Final Technical Report, The
University of Oklahoma, Norman, Oklahoma.
Ravindran, A.; Foote, B. L.; Badiru, A. B.; Leemis, L. M.; and L. Williams 1989, An
application of simulation and network analysis to capacity planning and mate-
rial handling systems at Tinker Air Force Base, TIMS Interfaces, Vol. 19, No. 1,
Jan.–Feb., pp. 102–115. Reprinted with permission.
2
Multi-Criteria Decision Making: An
Overview and a Comparative Discussion

Ramakrishnan Ramanathan, A. Ravi Ravindran, and Muthu Mathirajan

CONTENTS
2.1 Introduction...................................................................................................22
2.2 MCDM Terminologies.................................................................................. 23
2.3 Classification of Different MCDM Approaches....................................... 24
2.4 Multi-Objective Decision Making: Some Basic Concepts....................... 28
2.4.1 Efficient, Non-Dominated, or Pareto Optimal Solution.............. 29
2.4.2 Determining an Efficient Solution (Geoffrion, 1968)................... 31
2.4.3 Test for Efficiency.............................................................................. 32
2.5 Multi-Attribute Decision-Making Methods: Some Common
Characteristics............................................................................................... 32
2.6 Overview of Some MCDM Methods......................................................... 33
2.6.1 MADM Methods...............................................................................34
2.6.1.1 Multi-Attribute Utility Theory.........................................34
2.6.1.2 Multi-Attribute Value Theory.......................................... 35
2.6.1.3 Simple Multi-Attribute Rating Technique...................... 36
2.6.1.4 Analytic Hierarchy Process.............................................. 37
2.6.1.5 ELECTRE Methods............................................................ 39
2.6.1.6 PROMETHEE Methods..................................................... 41
2.6.1.7 Fuzzy Set Theory...............................................................43
2.6.2 MODM Methods...............................................................................44
2.6.2.1 Multi-Objective Linear Programming............................ 45
2.6.2.2 Goal Programming (Ravindran et al., 2006).................. 45
2.6.2.3 Method of Global Criterion and Compromise
Programming..................................................................... 49
2.6.2.4 Compromise Programming............................................. 50
2.6.2.5 Interactive Methods........................................................... 50
2.6.2.6 Data Envelopment Analysis............................................. 52
2.6.3 Other MCDM Methods.................................................................... 53
2.7 A General Comparative Discussion of MCDM Methodologies.............54
2.8 MCDM in the Era of Big Data..................................................................... 57
2.9 Summary........................................................................................................ 58
References................................................................................................................ 59

21
22 Big Data Analytics Using Multiple Criteria Decision-Making Models

2.1 Introduction
Multi-criteria decision making (MCDM) is a subfield of operations research.
It is a special case of the so-called decision-making problems. A decision-
making problem is characterized by the need to choose one or a few from
among a number of alternatives. The person who is to choose the alter-
natives is normally called the decision maker (DM). His preferences will
have to be considered in choosing the right alternative(s). In MCDM, the
DM chooses his most preferred alternative(s) on the basis of two or more
criteria or attributes (Dyer et al., 1992). The terms criteria, attributes, and
objectives are closely related and will be discussed in more detail later in
this chapter.
The field of MCDM has been succinctly defined as making decisions in
the face of multiple conflicting objectives (Zionts, 1992, 2000). According to
Korhonen (1992), the ultimate purpose of MCDM is “to help a decision maker
to find the ‘most preferred’ solution for his/her decision problem.” Several
perspectives are available in the literature to characterize a good decision-
making process. Stewart (1992) suggests that “the aim of any MCDM tech-
nique is to provide help and guidance to decision maker in discovering his
or her most desired solution to the problem” (in the sense of the course of
action which best achieves the DM’s long-term goals). According to French
(1984), a good decision aid should help the DM explore not just the problem
but also himself. Keeney (1992), in his famous book on value-focused think-
ing, says that we should spend more of our decision-making time concen-
trating on what is important, and that we should evaluate more carefully
the desirability of the alternatives. He also mentions that we should artic-
ulate and understand our values and using these values we should select
meaningful decisions to ponder and to create better alternatives. Howard
(1992) describes decision analysis as a “quality conversation about a decision
designed to lead to clarity of action.” Finally, Henig and Buchanan (1996)
say that a good decision process will force the DM to understand his or her
preferences and allow the set of alternatives to be expanded. Thus, a good
decision-making process should not only improve the clarity of the problem
to the DM, but it should also shed new light into the problem by generating
newer alternatives.
We are normally concerned with a single DM. If more number of DMs are
involved, it is important to aggregate all their preferences, leading to a group
decision-making (GDM) situation.
Starr and Zeleny (1977) provide a brief historical sketch of the early devel-
opments in MCDM. Some special issues of journals have been devoted to the
field of MCDM, including Management Science (Vol. 30, No. 1, 1984), Interfaces
(Vol. 22, No. 6, 1992) (devoted to decision and risk analysis), and Computers and
Operations Research (Vol. 19, No. 7, 1994). The Journal of Multi-Criteria Decision
Analysis, starting from the year 1992, publishes articles entirely devoted
Multi-Criteria Decision Making 23

to MCDM. Special issues of the Journal of the Operational Research Society

(April, 1982) and Interfaces (November–December, 1991) provide a range of
applications. Issue No. 2, Volume 133 (January 2001) of European Journal of
Operational Research is a special issue on goal programming and contains a
collection of some papers presented at the Third International Conference
on Multi-Objective Programming and Goal Programming Theories and
Applications. Given the growing popularity of MCDM approaches in spe-
cific disciplines, a number of recent articles have reviewed applications
of MCDM in supplier evaluation and selection (Govindan et  al., 2015; Ho
et al., 2010), energy planning (Pohekar and Ramachandran, 2004; Scott et al.,
2012; Wang et al., 2009), infrastructure management (Kabir et al., 2014), con-
struction (Jato-Espino et al., 2014), and municipal solid waste management
(Soltani et al., 2015).
In this chapter, attempt is made to review the basic concepts of MCDM,
present salient features of some important MCDM methods, and provide
a comparative discussion. It is not our intention here to provide a detailed
review of all MCDM methods presented in the literature.

2.2  MCDM Terminologies

• Several terminologies are normally used when dealing with a deci-
sion problem that has multiple criteria. The terms goals, objectives,
criteria, and attributes are commonly found in the MCDM literature
and can be used with interchangeable (and confusing) ease (Henig
and Buchanan, 1996). The general meaning of these words is similar
in most cases. Useful definitions of different terminologies used in
MCDM literature are available in Roy (1999).
• Alternatives form the most fundamental entities in an MCDM
model. They represent one of several things or courses of action to
be chosen by the DM. They are also called solutions, especially when
dealing with continuous variables, in the mathematical program-
ming context.

Alternatives are normally compared with each other in terms of the so-
called criteria. Identification of the criteria for a particular problem is subjec-
tive, that is, varies for each problem. Criteria are normally developed in a
hierarchical fashion, starting from the broadest sense (usually called the goal
of the problem) and refined into more and more precise sub- and sub-sub
goals. There is no unique definition for the term “criterion,” but a useful gen-
eral definition is from Bouyssou (1990), who has defined criterion as a tool
allowing comparison of alternatives according to a particular significance
axis or point of view. Edwards (1977) calls criteria as the relevant dimensions of
24 Big Data Analytics Using Multiple Criteria Decision-Making Models

value for evaluation of alternatives. Henig and Buchanan (1996) consider criteria
to be the raison d’être of the DM. The term “Criterion” is defined by Roy (1999)
as a tool constructed for evaluating and comparing potential actions accord-
ing to a well-defined point of view.
In general, some rules should be followed in identifying criteria for any
decision problem (Keeney and Raiffa, 1976; Saaty, 1980; von Winterfeldt and
Edwards, 1986). They have to be mutually exclusive or independent, collec-
tively exhaustive, and should have operational clarity of definition.
Criteria of a decision problem are usually very general, abstract, and often
ambiguous and it can be impossible to directly associate criteria with alter-
natives. Each criterion can be normally represented by a surrogate measure
of performance, represented by some measurable unit, called the attributes,
of the consequences arising from implementation of any particular decision
alternative. Thus while warmth is a criterion, temperature measured in a suit-
able (say Celsius or Fahrenheit) scale is an attribute. Attributes are objective
and measurable features of the alternatives. Thus, the choice of attributes
reflects both the objectively measurable components of the alternatives and
the DM’s subjective criteria. Attributes of alternatives can be measured inde-
pendently from DM’s desires and expressed as mathematical functions of
the decision variables.
Objectives, used in mathematical programming problems, represent
directions of improvement of the attributes. A maximizing objective refers
to the case where “more is better,” while a minimizing objective refers to the
case where “less is better.” For example, profit is an attribute, while maxi-
mizing profit is an objective. The term “criterion” is a general term compris-
ing the concepts of attributes and objectives. It can represent either attribute
or objective depending on the nature of the problem. Perhaps, that is why
MCDM is considered to encompass two distinct fields, namely, multi-
attribute decision making (MADM) and multi-objective decision making
(MODM) (e.g., Triantaphyllou, 2013). These fields are discussed in the next
section.

2.3  Classification of Different MCDM Approaches

A wide variety of MCDM methods have been reported in the literature
in the last few decades (Stewart, 1992). Some methods use rigorous math-
ematical programming approaches. Examples include the goal program-
ming (GP) (Schniederjans, 1995), compromise programming (Zeleny, 1982),
­multi-objective linear programming (MOLP) (Zeleny, 1982), and data envel-
opment analysis (DEA) (Charnes et  al., 1978). Some other methods are not
programming-based. Examples include the multi-attribute utility the-
ory (MAUT) (Keeney and Raiffa, 1976), the simple multi-attribute rating
Multi-Criteria Decision Making 25

technique (SMART) (Edwards, 1977), the analytic hierarchy process (AHP)

(Saaty, 1980), and the ELECTRE (Elimination Et Coix Traduisant la Realite or
elimination and choice translating algorithm) methods (Roy, 1996).
Classifications of MCDM methods in terms of different criteria are dis-
cussed by Hanne (1999). According to him, MCDM methods can be catego-
rized in terms of suitability of problem type (MADM for problems with a
finite set of alternatives, and MODM for problems with a continuous and
infinite set of alternatives) and in terms of solution concepts (aspiration
levels, pairwise comparisons, interactivity or non-interactivity, outrank-
ing, etc.).
Stewart (1992) has classified MCDM methods as:

1. Value- or utility-based approaches (comprising value theory, util-

ity theory, the AHP, and some interactive methods using value
functions)
2. Goals or reference point approaches (including GP, reference point
methods, the step method [STEM], etc.)
3. Methods using outranking concepts (including the ELECTRE
approaches, PROMETHEE [preference ranking organization method
for enrichment evaluations], etc.)
4. Fuzzy set theory
5. Descriptive methods (including factor analysis, correspondence anal-
ysis, principal components analysis, multi-dimensional scaling, etc.)

Figure 2.1 illustrates a possible classification scheme for MCDM methods.

Note that this classification need not be exhaustive as some methods can be
classified in more than one category. For example, though MAUT can be clas-
sified as part of MADM in the figure, when the utility function is used as an
objective function of a programming problem, it can also be classified as part
of MODM. Similarly, the aspiration-level interactive method (AIM) can be
classified under both the categories as it draws on many methods developed
under both the categories. Several references (Evans, 1984; Stewart, 1992)
were used in arriving at the structure shown in Figure 2.1.
The broad area of MCDM can be divided into two general categories,
MADM and MODM. MADM involves cases in which the set of alternatives
is defined explicitly by a finite list (Stewart, 1992) from which one or a few
alternatives should be chosen that reflect DM’s preference structure. MODM
involves cases in which the set of alternatives is defined implicitly by a math-
ematical programming structure with objective functions. Such alternatives
are usually defined in terms of continuous variables, which results in an
infinite number of alternatives. MODM is also referred to in the literature
as “multi-objective mathematical programming” (MMP) or “multi-objective
optimization” (MOO) or vector optimization or simply “multi-objective pro-
gramming” (MOP).
 26

Multi-criteria decision making

(MCDM)

Multi-attribute decision making Multi-objective decision making

(MADM) (MODM)

All decision Some discrete

Value- or utility- Others (fuzzy,
Outranking variables continuous decision variables
based approaches descriptive, etc.)

All objectives and All objectives and

At least one objective At least one objective
constraint functions constraint functions
or constraint function or constraint function
linear linear

No preference A priori Progressive A posteriori

articulation of articulation of articulation of Combination
information
preferences preferences preferences

FIGURE 2.1
Classification of MCDM methods.
Big Data Analytics Using Multiple Criteria Decision-Making Models
Multi-Criteria Decision Making 27

MADM methods can be further classified on the basis of the solution

approaches. Value- or utility-based approaches attempt to capture the util-
ity function of the DM in order to choose the alternative that maximizes
the DM’s overall utility. MAUT, multi-attribute value theory (MAVT), and
SMART belong to this group. Sometimes, the AHP is also said to belong
to this group though its proponents have argued against this classifica-
tion. Outranking methods use the concept of outranking. They are used
in ELECTRE methods and the PROMETHEE methods. Other MADM
approaches include those based on fuzzy set theory, some descriptive and
statistical approaches (such as factor analysis, principal components analy-
sis, multi-dimensional scaling, etc.).
MODM methods normally use an optimization scheme to identify the
non-dominated solutions (see the next section for a discussion of non-­
­
dominated solutions). There can be more than one, typically many, non-­
dominated solutions to an MODM problem. The DM often needs to select
one or a few from these non-dominated solutions, that best satisfies his
preferences.
Sometimes, it is easier for the DM to screen the large number of
­non-­dominated solutions using some screening methods so as to obtain
smaller number of solutions that can be managed easily by the DM. Graves
et  al. (1992) provide some statistics-based screening methods, including
­clustering and filtering.
Identification of these best compromise solution(s) from the set of all
non-dominated solution requires that at least some additional information
about the DM’s preference structure is obtained. Hence, in MODM meth-
ods, in addition to an optimization scheme, some procedure to obtain this
preference information is required. Hwang and Masud (1979) have classi-
fied solution techniques for MODM methods according to the timing of the
requirement for preference information versus the optimization. They have
classified the methods according to three pure approaches of articulation of
the DM’s preference structure:

• Prior to the optimization (a priori articulation of preferences)

• During, or in sequence with, the optimization (progressive articula-
tion of preferences, and sometimes also termed as interactive articu-
lation of preferences)
• After the optimization (a posteriori articulation of preferences)

In addition, some more categories (no preference information) are added

in Figure 2.1. MOLP and DEA are examples of MODM methods that do
not require any preference information. Of course, both these methods can
only identify the non-dominated alternatives, and cannot choose the best
one from these non-dominated alternatives that is in conformity with DM’s
preferences.
28 Big Data Analytics Using Multiple Criteria Decision-Making Models

Evans (1984) has provided references for mathematical programming arti-

cles under most of the MODM categories listed in Figure 2.1. Many methods
of MODM have been described in detail in Steuer (1986).

2.4  Multi-Objective Decision Making: Some Basic Concepts

In this section, a comparison of an MODM problem with a single-objective
decision-making problem is used to introduce certain key concepts fre-
quently used in MCDM literature. Note that many of the concepts intro-
duced in this section are also applicable to MADM problems.
A single-objective decision-making (SODM) problem is generally written
in the form:

max f1( x)
(2.1)
subject to: g j ( x) ≤ 0, for j = 1,… , m,

where x is an n-vector of decision variables and f1(x) is the single-decision

criterion or objective function to be optimized.

Let S = {x/gj(x) ≤ 0, for all j}.

Y = {y/f1(x) = y for some x ∈ S}.

S is called the decision space and Y is called the criteria or objective space.
Methods for solving single-objective mathematical programming prob-
lems have been studied extensively for the past 40 years. However, almost
every important real-world problem involves more than one objective.
A general MODM problem has the following form:

Max F( x) = { f1( x), f 2 ( x),… , f k ( x)}

(2.2)
Subject to, g j ( x) ≤ 0 for j = 1,… , m,

where x is an n-vector of decision variables and fi(x), i = 1, … , k are the k criteria/
objective functions.

Let S = {x/gj(x) ≤ 0, for all j}.

Y = {y/F(x) = y for some x ∈ S}.

S is called the decision space and Y is called the criteria or objective space of
the MODM problem. Without loss of generality, we can assume all the objec-
tive functions to be maximizing. Thus, the MODM problem is similar to an
Multi-Criteria Decision Making 29

TABLE 2.1
Hypothetical Options
Emissions (kg of
Cost ($) Time (h) Aggregated Pollutants)
Option A 50 16 10
Option B 40 20 12
Option C 60 16 12

SODM problem except that it has a stack of objective functions instead of

only one.
Let us consider an example involving choice of transport modes. Let the
objectives be as follows:

Minimize fi(x) = Cost
Minimize f2(x) = Time
Minimize f3(x) = Emissions
Subject to some constraints

Let us assume that we have three options (Table 2.1) that satisfy the con-
straints, that is, the feasible options, identified somehow.
Note that option C fares, on all objectives, worse than A and hence it
should not be considered anymore. Options A and B are incomparable, as
none of them is at least as good as the other in terms of all the objectives.
While option A results in lesser time and emissions compared with option
B, it is more expensive. Hence, options A and B are said to be non-dominated
options while option C is called a dominated option.
In any MODM exercise, we are first interested in identifying the
­non-­dominated options. Note that there may be more than one non-domi-
nated option for any MODM problem.

2.4.1  Efficient, Non-Dominated, or Pareto Optimal Solution

A solution xo∈S to MODM problem is said to be efficient if fk(x) > fk(xo) for
some x ∈ S implies that fj(x) < fj(xo) for at least one other index j. More simply
stated, an efficient solution has the property that an improvement in any one
objective is possible only at the expense of at least one other objective.
A dominated solution is a feasible solution that is not efficient.
Efficient Set: Set of all efficient solutions is called the efficient set or efficient
frontier.
Note: Even though the solution of MODM problem reduces to finding the
efficient set, it is not practical because there could be an infinite number of
efficient solutions.
30 Big Data Analytics Using Multiple Criteria Decision-Making Models

EXAMPLE 2.1
Consider the following bicriteria linear program (BCLP):
Max Z1 = 5x1 + x2
Max Z2 = x1 + 4 x2
Subject to: x1 ≤ 5
x2 ≤ 3

x1 + x2 ≤ 6

x1 , x2 ≥ 0

Solution
The decision space and the objective space are given in Figures 2.2 and 2.3,
respectively. Corner points C and D are efficient solutions, while cor-
ner points A, B, and E are dominated. The set of all efficient solutions is
given by the line segment CD in both figures.
Ideal solution is the vector of individual optima obtained by optimiz-
ing each objective function separately ignoring all other objectives. In
Example 2.1, the maximum value of Z1, ignoring Z2, is 26 and occurs at
point D. Similarly, maximum Z2 of 15 is obtained at point C. Thus, the
ideal solution is (26, 15) but is not feasible or achievable.
Note: One of the popular approaches to solving MODM problems is
to find an efficient solution that comes “as close as possible” to the ideal
solution. We will discuss this approach later in Section 2.6.2.4.

X2
B (0,3) C (3,3)
Optimal for Z2

Feasible
decision Optimal for Z1
space
D (5,1)

E (5,0)
A (0,0) X1

FIGURE 2.2
Decision space (Example 2.1).
Multi-Criteria Decision Making 31

Z2
C (18,15)
(26,15)

(3,12)
B

D (26,9)
Achievable
objective
values

E (25,5)

A (0,0) Z1

FIGURE 2.3
Objective space (Example 2.1).

2.4.2  Determining an Efficient Solution (Geoffrion, 1968)

For the MODM problem (2.2), consider the following single-objective optimi-
zation problem, called the Pλ problem. The Pλ problem is also known as the
weighted objective problem.

Max Z = ∑ λ f ( x)
i =1
i i

Subject to : x ∈ S (2.3)
k

∑λ = 1 i and λ i ≥ 0.
i =1

Theorem 2.1 (Sufficiency)

Let λi > 0 for all i be specified. If xo is an optimal solution for the Pλ problem
(Equation 2.3), then xo is an efficient solution to the MODM problem.
In Example 2.1, if we set λ1 = λ2 = 0.5 and solve the Pλ problem, the optimal
solution will be at D, which is an efficient solution.
Warning: Theorem 2.1 is only a sufficient condition and is not necessary. For
example, there could be efficient solutions to MODM problem which could
32 Big Data Analytics Using Multiple Criteria Decision-Making Models

not be obtained as optimal solutions to the Pλ problem. Such situations occur

when the objective space is not a convex set. However, for MODM problems,
where the objective functions and constraints are linear, Theorem 2.1 is both
necessary and sufficient.

2.4.3  Test for Efficiency

Given a feasible solution x ∈S for MODM problems, we can test whether or
not it is efficient by solving the following single-objective problem:
k

Max W = ∑d
i =1
i

()
Subject to : f i ( x) ≥ f i x + di for i = 1,2,… ,k
x ∈S
di ≥ 0.

Theorem 2.2

1. If Max W > 0, then x is a dominated solution.

2. If Max W = 0, then x is an efficient solution.

Note: If Max W > 0, then at least one of the di’s is positive. This implies that at
least one objective can be improved without sacrificing on the other objectives.

2.5 Multi-Attribute Decision-Making Methods:

Some Common Characteristics
In general, MADM methods have some common characteristics. These are
discussed below:

1. The persons or organizations who are the stakeholders of the deci-

sion problem being considered have to be identified. These people
will act as the DMs, whose opinions will be elicited by the MCDM
methods. This stage is equally applicable to MODM methods, espe-
cially while eliciting the preference information for choosing one or
a few from the set of all non-dominated alternatives (solutions).
2. A model of the decision problem has to be constructed. Typically, the
model will consider the main goal of the problem, relevant criteria,
attributes, and alternatives. For example, if one wants to evaluate
Multi-Criteria Decision Making 33

some selected technologies, it is important to identify the key crite-

ria that distinguish the characteristics of technologies, and all the
relevant technologies to be considered as alternatives in the model.
The model is called as a hierarchical model in some MCDM meth-
ods as it represents a hierarchy of decisions—the goal of the exercise
(say, selection of the most important energy technologies), the crite-
ria (economic significance, social significance, etc.), and alternatives
(the technologies to be evaluated).
3. Ideally speaking, criteria have to be mutually exclusive and collec-
tively exhaustive. Collective exhaustivity is critical in some meth-
ods, such as the AHP, as final rankings could change for the addition
or deletion of alternatives at a later stage.
4. Relative importance of criteria, which are implicit trade-off informa-
tion in the identification of the best alternative(s), has to be assessed.
While assessment of relative importance of criteria forms an impor-
tant step in many MCDM methods, some methods do not use this
information (e.g., DEA*). The procedure for this assessment differs for
each method, and forms generally the single most important distin-
guishing feature of the methods. The procedures will be described
in more detail in the next section for some important MCDM meth-
ods. The assessment can be performed either by an individual, or by
representatives of conflicting values acting separately or by those
representatives acting as a group. Again, some methods have special
procedures to aggregate the opinions of individuals to form a group
opinion, though in general the group members should be encour-
aged to arrive at a single unanimous assessment.
5. Assessment of alternatives with respect to different criteria.
6. Aggregation of performance of alternatives with respect to all criteria
to provide the overall performance measures of alternatives. Simple
additive or multiplicative aggregation is used in many methods, but
can also be more sophisticated (e.g., DEA uses linear programming).
The overall performance measures could also be considered as the
rankings of the alternatives.

2.6  Overview of Some MCDM Methods

Several surveys are available that describe many MCDM methods suggested
in the literature (Evans, 1984; French et al., 1983; Steuer et al., 1996; Stewart,
1992; White, 1990). A comprehensive list of some known multi-criteria

* Note that DEA can only identify the non-dominated alternatives, and not the best one captur-
ing DM’s preferences.
34 Big Data Analytics Using Multiple Criteria Decision-Making Models

aggregation procedures (MCAP) and a comparative discussion of some

MCDM methods are available in Guitouni and Martel (1998). In the next few
pages, a brief description of some selected MCDM methods will be provided.
For ease of exposition, we have divided the methods in terms of MADM
methods and MODM methods, though some methods, such as AIM, can be
classified in either category. For each method, an attempt is made here to out-
line the basic idea, specify important introductory and review articles, and
outline strengths and criticisms, and some important applications. Reader
interested in knowing more about a method should refer to the articles
referred in the appropriate subsection.

2.6.1  MADM Methods

2.6.1.1  Multi-Attribute Utility Theory
Utility measures the subjective value “worth” of an outcome, even one that
is not a money value. Often the utility function is not linear. Traditionally,
utility functions are defined for stochastic problems that involve uncertainty.
In the case of deterministic problems, the term value functions is more com-
monly used. The utility or value functions may be thought of as evaluative
mechanisms that can be used to measure the value of a particular solution
(Zionts, 1992).
As we noted above, utility functions are defined in terms of uncertainty
and thus tie in the DM’s preferences under uncertainty, revealing his risk
preference for an attribute. An uncertain situation that is faced by a DM can
be considered similar to a lottery—he can earn $X with a probability p, and
earn Y with probability (1 − p). In these situations, a rational DM is expected
to maximize his expected utility, given by $(pX + (1 − p)Y).
Utility functions are assessed by giving the DM a sequence of choice
between alternatives or between alternative lotteries. The responses are used
to generate functions.
MAUT consists of assessing and fitting utility functions and probabilities
for each attribute, and then using the functions and probabilities to come up
with rankings of alternatives. The utility function for each attribute is aggre-
gated to get the overall utility function. At least two methods of aggregation
are used in MAUT: additive and multiplicative. Certain conditions need to be
satisfied in order that these aggregations are valid: preferential independence
and utility independence.
The additive aggregation is given by the following:

U ( A) = ∑ w u (a )
i
i i i where 0 ≤ ui ( ai ) ≤ 1

∑ w = 1,
i wi ≥ 0 ,
i
Multi-Criteria Decision Making 35

where ui (ai) is the utility function describing preferences with respect to the
attribute i, ai represents the performance of the alternative A in terms of the
attribute i, wi are scaling factors which define acceptable trade-offs between
different attributes, and U(A) represents the overall utility function of the
alternative A when all the attributes are considered together. This form of
additive aggregation is valid if and only if (sometimes referred to as “iff” in
this book) the DM’s preferences satisfy the mutual preferential independence.
Suppose that there are a set of attributes, X. Let Y be a subset of X, and let Z
be its complement, that is, Z = X − Y. The subset Y is said to be preferentially
independent of Z, if preferences relating to the attributes contained in Y do
not depend on the level of attributes in Z.
The condition utility independence is a stronger assumption. More details
can be obtained from Keeney and Raiffa (1976). A brief discussion of MAUT
is available in Kirkwood (1992).
The utility functions may also be used as objective functions for solving
mathematical programming problems.

• Utility theory has been criticized because of its “strict” assumptions

which are usually not empirically valid (Henig and Buchanan, 1996).
Because of the strict assumptions, practical applications of MAUT
are relatively difficult, though there are several practical successful
applications of MAUT. This has led to some simplifications of the
MAUT concepts. For example, MAVT is a simplification of MAUT
where uncertainty and risk are not assumed. SMART is another sim-
plification that makes weaker assumptions while eliciting utilities.
They are described in the next few subsections.

2.6.1.2  Multi-Attribute Value Theory

MAVT is a simplification of MAUT: MAVT does not seek to model the DM’s
attitude to risk, while MAUT considers risk and uncertainty. As a result,
MAVT is easier to implement compared with MAUT (Belton, 1999).
Value theory assumes that it is possible to represent a DM’s preferences in
a defined context by a value function, V(⋅), such that if alternative A is pre-
ferred to alternative B, then V(A) > V(B). For this representation to be possible,
the DM’s preferences should satisfy two properties: the transitivity property
(if A is preferred to B, B is preferred to C, then A should be preferred to C),
and comparability (given to alternatives A and B, the DM must be able to
indicate whether A is preferred to B, or B is preferred to A, or is indifferent
between the two).
Note that the value function is an ordinal function, that is, it can be used to
only rank alternative. In contrast, utility function is cardinal, that is, it can be
used to measure the strength of preference among alternatives.
MAVT explicitly recognizes that the DM will use many attributes (cri-
teria) while evaluating a set of alternatives. For each attribute i, a partial
36 Big Data Analytics Using Multiple Criteria Decision-Making Models

value function vi (ai) describing preferences with respect to the attribute i is

assessed from the DM, where ai represents the performance of the alternative
A in terms of the attribute i. Then, the overall value function V(A) of the alter-
native when all the attributes are considered together is normally obtained
using the additive form: V(A) = Σivi (ai). This is more generally expressed as
follows:

V ( A) = ∑ w v (a )
i
i i i where 0 ≤ vi ( ai ) ≤ 1

∑ w = 1,
i wi ≥ 0.
i

As mentioned above, wi are scaling factors which define acceptable trade-

offs between different attributes. Again, this additive value function is
appropriate if and only if the DM’s preferences satisfy the so-called mutual
preferential independence discussed earlier.
More detailed information on the practical implementation of MAVT is
available in Belton (1999).

2.6.1.3  Simple Multi-Attribute Rating Technique

SMART (Edwards, 1977; von Winterfeldt and Edwards, 1986) follows the
steps described in the previous section for modeling a decision problem. It
uses the simple weighting technique for the assessment of importance of
criteria, and for the assessment of alternatives with respect to criteria.
To rate (i.e., assess the importance of) criteria, one will start by assigning
the least important criterion an importance of 10. Then, he has to consider
the next least important criterion, and ask as to how much more impor-
tant (if at all) is it than the least important criterion, and assign a number
that reflects that ratio. This procedure is continued till all the criteria are
assessed, checking each set of implied ratios as each new judgment is made.
The experts will be given the opportunity to revise previous judgments to
make them consistent. Once the numbers are assigned, the relative impor-
tance of criteria is obtained by summing the importance weights, and divid-
ing each by the sum. Thus, the relative importance of the criterion j (wj) is
the ratio of importance weight of this criterion to the sum. Note that Σjwj = 1
by definition.
Alternatives are rated with respect to each criterion in a similar fashion.
While MAUT requires the development of complex utility functions for each
criterion, SMART prefers to produce the rating using a more straightforward
approach: the expert is asked to estimate the position of the alternative on a
criterion on a 0–100 scale, where 0 is defined as the minimum plausible value,
while 100 is defined as the maximum plausible value.
Multi-Criteria Decision Making 37

Once the above two measures are available, the overall performance of an
alternative i can be aggregated using the simple weighted average,

Ui = ∑w u ,
j
i ij

where Ui is the overall performance rating of alternative i, wj is the relative
importance of criterion j, and uij is the rating of the alternative i with respect
to the criterion j. The alternative that has the maximum Ui is the most pre-
ferred alternative to achieve the goal of the decision problem. The values of
Ui can be used to provide the overall rankings of the alternatives.
MAUT, or its simplified versions MAVT or SMART, has been used for sev-
eral practical applications. For example, MAUT or its variants have been plan-
ning a government research program (Edwards, 1977). Jones et al. (1990) have
applied MAVT for the study of UK energy policy. Keeney and McDaniels
(1999) have used this technique for identifying and structuring values for
integrated resource planning, while Keeney (1999) has used the technique
to create and organize a complete set of objectives for a large software orga-
nization. Duarte (2001) has used MAUT to identify appropriate technologi-
cal alternatives to implement to treat industrial solid residuals. Some more
MAUT applications are discussed by Bose et al. (1997).

2.6.1.4  Analytic Hierarchy Process

The AHP (Saaty, 1980) is one of the most popular and widely employed
multi-criteria methods (Golden et al., 1989; Shim, 1989; Vargas, 1990). In this
technique, the process of rating alternatives and aggregating to find the most
relevant alternatives are integrated. The technique is employed for ranking
a set of alternatives or for the selection of the best in a set of alternatives. The
ranking/selection is done with respect to an overall goal, which is broken
down into a set of criteria.
The application of the methodology consists of establishing the impor-
tance weights to be associated to the criteria in defining the overall goal.
This is done by comparing the criteria pairwise. Let us consider two criteria
Cj and Ck. The expert is asked to express his graded comparative judgment
about the pair in terms of the relative importance of Cj over Ck with respect to
the goal. The comparative judgment is captured on a semantic scale (equally
important/moderately more important/strongly important, and so on) and
is converted into a numerical integer value ajk. The relative importance of Ck
over Cj is defined as its reciprocal, that is, akj = 1/ajk. A reciprocal pairwise
comparison matrix A is then formed using ajk, for all j and k. Note that ajj = 1.
It has been generally agreed (Saaty, 1980) that the weights of criteria can be
estimated by finding the principal eigenvector w of the matrix A. That is,

Aw = λ max w.
38 Big Data Analytics Using Multiple Criteria Decision-Making Models

When the vector w is normalized, it becomes the vector of priorities of the

criteria with respect to the goal. λmax is the largest eigenvalue of the matrix
A and the corresponding eigenvector w contains only positive entries. The
methodology also incorporates established procedures for checking the con-
sistency of the judgments provided by the DM.
Using similar procedures, the weights of alternatives with respect to each
criterion are computed. Then, the overall weights of alternatives are com-
puted using the weighted summation,

 Overall weight   Weight of alternative i with respect to C j ×

 of alternative i = ∑ 
j
Weight of C j with respect to the goal 
.

The popularity of AHP stems from its simplicity, flexibility, intuitive

appeal, and its ability to mix quantitative and qualitative criteria in the same
decision framework. Despite its popularity, several shortcomings of AHP
have been reported in the literature, which have limited its applicability.
However, several modifications have been suggested to the original AHP,
such as the multiplicative AHP (MAHP) (Lootsma, 1999; Ramanathan, 1997),
to overcome these limitations.
Some of the prominent limitations of AHP include the following:

• Rank reversal (Belton and Gear, 1983; Dyer, 1990): The ranking of
alternatives determined by the original AHP may be altered by the
addition of another alternative for consideration. For example, when
AHP is used for a technology selection problem, it is possible that the
rankings of the technologies get reversed when a new technology is
added to the list of technologies. One way to overcome this problem
is to include all possible technologies and criteria at the beginning
of the AHP exercise, and not to add or remove technologies while
or after completing the exercise. However, MAHP, the multiplica-
tive variant of AHP, does not suffer from this type of rank reversal
(Lootsma, 1999).
• Number of comparisons: AHP uses redundant judgments for checking
consistency, and this can exponentially increase the number of judg-
ments to be elicited from DMs. For example, to compare eight alterna-
tives on the basis of one criterion, a total of 28 judgments are needed. If
there are n criteria, then the total number of judgments for comparing
alternatives on the basis of all these criteria will be 28n. This is often a
tiresome and exerting exercise for the DM. Some methods have been
developed to reduce the number of judgments needed (Millet and
Harker, 1990). Also, some modifications, such as MAHP, can compute
weights even when all the judgments are not available.
Multi-Criteria Decision Making 39

AHP has been applied to a variety of decision problems in the litera-

ture. Several detailed annotated bibliographies of AHP applications are
available (Golden et al., 1989; Shim, 1989; Vargas, 1990). Some of the more
recent applications of AHP include solar energy utilization (Elkarni and
Mustafa, 1993), integrated resource planning (Koundinya et al., 1995), cli-
mate change negotiations (Ramanathan, 1998), greenhouse gas mitigation
(Ramanathan, 1999), and environmental impact assessment (Ramanathan,
2001a).

2.6.1.5  ELECTRE Methods

ELECTRE methods (Bouyssou and Vincke, 1997; Roy, 1996; Vincke, 1999)
belong to the so-called outranking approaches. PROMETHEE method that
will be discussed in the next section also belongs to this group of outrank-
ing approaches. Outranking methods are especially popular in France and
Belgium (Hanne, 1999). Sometimes, the outranking approaches are referred
to as the French or European approaches (Roy and Vanderpooten, 1996;
Vincke, 1999) (in contrast with MAUT or AHP, which are called American
approaches).
There are several versions of ELECTRE methods, including ELECTRE I
(historically the first of the versions), ELECTRE IS, ELECTRE II, ELECTRE
III, ELECTRE IV, ELECCALC, and ELECTRE TRI. We shall briefly describe
only ELECTRE I in this book, using the material presented in Vincke (1999).
The reader is referred to other references (Bose et al., 1997; Roy, 1996) for more
detailed descriptions of these methods.
It is important to understand some preliminary definitions for appreciat-
ing the ELECTRE methods. The set of alternatives is denoted by A. A binary
relation R on A is a subset of A × A. It is said to be

• Symmetric iff aRb ⇒ bRa, ∀a,b ∈A

• Asymmetric iff aRb ⇒ bRa, ∀a,b ∈A
• Complete iff aRb ⇒ bRa, ∀a,b ∈A
• Transitive iff aRb, bRc ⇒ aRc, ∀a,b,c ∈A
• A complete preorder iff it is complete and transitive
• A partial preorder iff it is transitive and not complete.

A criterion g is defined as a real valued function on A in the sense that,

 g( a) > g(b) iff a is preferred to b ,


 g( a) = g(b) iff a is indifferent to b.
40 Big Data Analytics Using Multiple Criteria Decision-Making Models

The preferences can be represented by a criterion iff the relation R defined

by aRb iff a is preferred or indifferent to b is a complete preorder.
In ELECTRE I, like any other outranking approach, the so-called outrank-
ing concept is schematized as follows: an alternative a outranks b if, given the
information about the preferences of the DM, there are sufficient arguments
to affirm that a is at least as good as b and there is no really important rea-
son to refuse this assertion. This outranking concept is operationalized for
choosing alternatives, sorting them into categories or ranking them from the
best to worst.
The basic information is a set of n criteria {g1, g2, …, gn} on A and, for each
of them:

• A weight wj expressing the relative importance of criterion gj,

• A veto threshold vj(gj) > 0.

For each ordered pair (a,b), a concordance index c(a,b) is calculated as,

∑
1
c( a, b) = wj ,
W j: g j ( a )> g j ( b )

where

W= ∑w .
j =1
j

This index varies from 0 to 1 and can be considered as a measure of the

arguments in favor of the assertion “a outranks b.” A concordance level s is
chosen to help declare that a outranks b, denoted by aSb iff:

c( a, b) ≥ s,
∀ j such that g j ( a) < g j (b), the interval (g j ( a), g j (b)) is smaller than v j ( g j ( a)).

One can multiply all the weights by the same number, and if the new
weights are integers, the building of the outranking relations in ELECTRE I
can be interpreted as a voting procedure with a special majority rule (char-
acterized by the concordance level).
ELECTRE methods have received many practical applications. Roy
et  al. (1986) have used ELECTRE for determining which Paris metro sta-
tions should be renovated. An application of the ELECTRE TRI method
Multi-Criteria Decision Making 41

for business failure prediction is presented by Zoupounidis and Dimitras

(1998). Hokkanen and Salminen (1997) have applied ELECTRE III and IV to
environmental problems. Salminen et al. (1998) have applied ELECTRE III,
PROMETHEE I, II, and SMART in the context of four different real applica-
tions to environmental problems in Finland.
According to Hanne (1999), outranking methods have been criticized on
some theoretical grounds. Despite their appeal, the lack of an axiomatic
basis makes their underlying logic unsound, which often leads to paradoxi-
cal results (Ballestero and Romero, 1998). Alley (1983) (see also Gershon and
Duckstein, 1983) has pointed out the possibility of obtaining dominated
solutions with the ELECTRE approach and considers the ranking process
to be a “mystery to the DM.” Stewart (1992) has supposed some outrank-
ing procedures to be “difficult to verify empirically as models of human
preferences.”

2.6.1.6  PROMETHEE Methods

The PROMETHEE methods also belong to the class of outranking approaches,
and were proposed by Brans et al. (1984). They are also briefly described in
Bouyssou and Vincke (1997) and Vincke (1999). The following brief discus-
sion is based on Vincke (1999).
The basic information used by the PROMETHEE methods is a set of n so-
called generalized criteria (gj,Fj) on A, and a weight wj expressing the relative
importance of criterion gj for each of them.
A valued strict preference relation is defined by calculation, for each
ordered pair (a,b), the quantity

∑ w F (a, b),
1
π( a, b) = j j
W j =1

where

W= ∑w
j =1
j

and Fj(a,b) is the degree of preference of a over b for criterion j.

Fj(a,b) is a number between 0 and 1. The possible variations of this number
in the interval gj(a) − gj(b) are given in Figure 2.4. For each criterion gj, based
on discussions with the DM, a particular function Fj is chosen and the cor-
responding parameters (q j, pj, or σj) are fixed.
42 Big Data Analytics Using Multiple Criteria Decision-Making Models

1st form
• Immediate strict preference.
1 Fj(a,b) • No parameter to be
determined.

gj(a) –
0

2nd form
• There exists an indifference
1 Fj(a,b) threshold which must be fixed.

gj(a) –
0 q

3rd form
• Preference increases up to a
1 Fj(a,b) preference threshold which
should be determined.

gj(a) –
0 p

4th form
• There exists an indifference
1 Fj(a,b) and a preference threshold
which must be fixed. Between
½
the two, preference is average.
gj(a) –
0 q p

5th form
• There exists an indifference
1 Fj(a,b) and a preference threshold
which must be fixed. Between
the two, preference increases.
gj(a) –
0 q p

6th form
• Preference increases following
1 Fj(a,b) a normal distribution, the
standard deviation of which
must be fixed.
gj(a) –
0 σj

FIGURE 2.4
Preference functions used in PROMETHEE methods. (From Vincke, P. 1999.)
Multi-Criteria Decision Making 43

Just like the ELECTRE methods, two complete preorders are built:
one ­consists of ranking the actions following the decreasing order of the
­numbers (ϕ+(a)),

φ + ( a) = ∑ π(a, b) (outgoing flow);

b ∈A
and the order following the increasing order of numbers (ϕ−(a)),

φ − ( a) = ∑ π(b, a) (incoming flow).

b ∈A
Their intersection yields the partial preorder of the PROMETHEE I method.
The PROMETHEE II method consists in ranking the actions following the
decreasing order of the numbers ϕ(a) such that ϕ(a) = ϕ+(a) − ϕ−(a) and yields a
unique preorder.

2.6.1.7  Fuzzy Set Theory

Fundamentally speaking, MCDM deals with imprecision in human prefer-
ences: Human beings are not able to specify their preferences unambigu-
ously. It is not possible to exactly specify whether one alternative is preferred
to another in terms of some criteria (and by how much). Fuzzy set theory
has been proposed to deal with imprecision (Zadeh, 1965). The usefulness
of fuzzy sets in management science has long been recognized (Bellman
and Zadeh, 1970). One of the nice reviews of fuzzy set theory in operations
research context is provided by Zimmermann (1983).
Fuzzy set theory assigns the so-called membership function (a measure of
the degree of membership) to each alternative in the fuzzy set of good (or
satisfactory, or the like) in terms of each criterion. The overall membership of
the alternative (in terms of membership of all the criteria taken together) is
normally defined as the intersection of all these single-criterion fuzzy sets.
Fuzzy sets have received several applications in MCDM literature.
Smithson (1987) has described several applications of fuzzy set theory to
behavioral and social sciences. Ammar and Wright (2000) have described
three applications of fuzzy set theory to multi-criteria evaluation. Fuzzy sets
have also been used in conjunction with other methods to incorporate the
concept of fuzziness. For example, Bisdorff (2000) has described the applica-
tion of fuzzy ELECTRE methods. Özelkan and Duckstein (2000) have used
fuzzy sets with regression for carrying multi-criteria evaluation of rainfall–
runoff relationships pertaining to hydrologic system models. Parra et  al.
(2001) have used fuzzy GP for portfolio selection.
Use of fuzzy sets has certain practical problems, including the assessment
of membership functions, and the operational definition of fuzzy intersec-
tion. See Stewart (1992) for a brief discussion of these problems.
44 Big Data Analytics Using Multiple Criteria Decision-Making Models

2.6.2  MODM Methods*

For most of the MODM problems, we can start with the following math-
ematical programming problem (presented earlier in Section 2.4).

Max F( x) = { f1( x), f 2 ( x),… , f k ( x)}

(2.2)
Subject to, g j ( x) ≤ 0 for j = 1,… , m,

where x is an n-vector of decision variables and fi(x), i = 1, … , k are the k criteria/
objective functions.

Let S = {x/gj(x) ≤ 0, for all j},

Y = {y/F(x) = y for some x ∈ S}.

S is called the decision space and Y is called the criteria or objective space of
the MODM problem.
In MODM problems, there are often an infinite number of efficient solu-
tions and they are not comparable without the input from the DM. Hence,
it is generally assumed that the DM has a real-valued preference function
defined on the values of the objectives, but it is not known explicitly. With
this assumption, the primary objective of the MODM solution methods is to
find the best compromise solution, which is an efficient solution that maximizes
the DM’s preference function.
In the last three decades, most MCDM researches have been concerned
with developing solution methods based on different assumptions and
approaches to measure or derive the DM’s preference function. Thus, the
MODM methods can be categorized by the basic assumptions made with
respect to the DM’s preference function as follows:

1. When complete information about the preference function is avail-

able from the DM.
2. When no information is available.
3. Where partial information is obtainable progressively from the DM.

In the following sections, we first talk about a simpler extension of SODM

called MOLP. We will then discuss three more MODM methods—GP, com-
promise programming, and interactive methods, as examples of categories 1, 2,
and 3 type approaches, respectively. More methods such as DEA are dis-
cussed later in this section.

* This section is based on Ravindran (2016).

Multi-Criteria Decision Making 45

2.6.2.1  Multi-Objective Linear Programming

Note that if there were only one objective function, the above problem can be
easily solved by the traditional simplex method. It is possible to conceive a
variation of the simplex method to deal with the case of multiple objectives.
It is usually called the multi-criteria simplex method. This method has been
described in detail in Zeleny (1982). The procedure is very similar to that
of the single-criterion simplex method, except that the so-called minimum
ratio rule for identifying the outgoing basic variable in any iteration is modi-
fied to take into the presence of many objective functions, and additional
checks are incorporated for identifying whether the basic solution of any
iteration is dominated or not. Note that the multi-criteria simplex method
can identify only the non-dominated solutions of an MODM problem. As it
incorporates no preference information from the DM, it cannot reduce the
set of non-­dominated solutions to one or a few that is acceptable to the DM.
MOLP has been applied to MODM problems. Zeleny (1982) discusses some
applications. Leung et al. (2001) provide a recent application of MOLP to fish
resource utilization.

2.6.2.2  Goal Programming (Ravindran et al., 2006)

GP, likely the oldest school of MCDM approaches, has been developed in
the 50s as an extension of linear programming (Charnes and Cooper, 1961;
Charnes et al., 1955). One way to treat multiple criteria is to select one crite-
rion as primary and the other criteria as secondary. The primary criterion
is then used as the optimization objective function, while the secondary
criteria are assigned acceptable minimum and maximum values and are
treated as problem constraints. However, if careful considerations were not
given while selecting the acceptable levels, a feasible design that satisfies all
the constraints may not exist. This problem is overcome by GP, which has
become a practical method for handling multiple criteria. GP falls under the
class of methods that use completely prespecified preferences of the DM in
solving the MODM problem.
In GP, all the objectives are assigned target levels for achievement and rela-
tive priority on achieving these levels. GP treats these targets as goals to aspire
for and not as absolute constraints. It then attempts to find an optimal solu-
tion that comes as “close as possible” to the targets in the order of specified
priorities.
Before we discuss the formulation of GP models, we should discuss the
difference between the terms real constraints and goal constraints (or simply
goals) as used in GP models. The real constraints are absolute restrictions
on the decision variables, while the goals are conditions one would like to
achieve but are not mandatory. For instance a real constraint given by

x1 + x2 = 3
46 Big Data Analytics Using Multiple Criteria Decision-Making Models

requires all possible values of x1 + x2 to always equal 3. As opposed to this,

a goal requiring x1 + x2 = 3 is not mandatory, and we can choose values of
x1 + x2 ≥ 3 as well as x1 + x2 ≤ 3. In a goal constraint, positive and negative
deviational variables are introduced as follows:

x1 + x2 + d1− − d1+ = 3 d1+ , d1− ≥ 0.

Note that, if d1− > 0, then x1 + x2 < 3, and if d1+ > 0, then x1 + x2 > 3.
By assigning suitable weights w1− and w1+ on d1− and d1+ in the objective
function, the model will try to achieve the sum x1 + x2 as close as possible
to 3. If the goal were to satisfy x1 + x2 ≥ 3, then only d1− is assigned a positive
weight in the objective, while the weight on d1+ is set to zero.

2.6.2.2.1  Goal Programming Formulation

Consider the general MODM problem given by Equation 2.2. The assump-
tion that there exists an optimal solution to the MODM problem involving
multiple criteria implies the existence of some preference ordering of the cri-
teria by the DM. The GP formulation of the MODM problem requires the
DM to specify an acceptable level of achievement (bi) for each criterion fi and
specify a weight wi (ordinal or cardinal) to be associated with the deviation
between fi and bi. Thus, the GP model of an MODM problem becomes:

Minimize Z = ∑ (w d + +
i i + w i− di− ) (2.4)
i =1

Subject to : f i (x) + di− − di+ = bi for i = 1,… , k (2.5)

g j (x) ≤ 0 for j = 1,… ,m (2.6)

x j , di− , di+ ≥ 0 for all i and j. (2.7)

Equation 2.4 represents the objective function of the GP model, which

minimizes the weighted sum of the deviational variables. The system of
equations (Equation 2.5) represents the goal constraints relating the multiple
criteria to the goals/targets for those criteria. The variables, di− and di+ , in
Equation 2.5 are called deviational variables, representing the under achieve-
ment and over achievement of the ith goal. The set of weights (w i+ and w i− ) may
take two forms:

1. Prespecified weights (cardinal)

2. Preemptive priorities (ordinal)
Multi-Criteria Decision Making 47

Under prespecified (cardinal) weights, specific values in a relative scale are

assigned to w i+ and w i− representing the DM’s “trade-off” among the goals.
Once w i+ and w i− are specified, the goal program represented by Equations
2.4 through 2.7 reduces to a single-objective optimization problem. The car-
dinal weights could be obtained from the DM using any of the methods dis-
cussed earlier, such as rating method and AHP. However, in order for this
method to work, the criteria values have to be scaled properly.
In reality, goals are usually incompatible (i.e., incommensurable) and some
goals can be achieved only at the expense of some other goals. Hence, pre-
emptive GP, which is more common in practice, uses ordinal ranking or preemp-
tive priorities to the goals by assigning incommensurable goals to different
priority levels and weights to goals at the same priority level. In this case, the
objective function of the GP model (Equation 2.4) takes the form

Minimize Z = ∑ P ∑ (w
p
p
i
+
ip di+ + w ip− di− ), (2.8)

where Pp represents priority p with the assumption that Pp is much larger
than Pp+1 and w ip+ and w ip− are the weights assigned to the ith deviational vari-
ables at priority p. In this manner, lower priority goals are considered only
after attaining the higher priority goals. Thus, preemptive GP is essentially
a sequence of single-objective optimization problems, in which successive
optimizations are carried out on the alternate optimal solutions of the previ-
ously optimized goals at higher priority.
In both preemptive and non-preemptive GP models, the DM has to spec-
ify the targets or goals for each objective. In addition, in the preemptive GP
models, the DM specifies a preemptive priority ranking on the goal achieve-
ments. In the non-preemptive case, the DM has to specify relative weights for
goal achievements.
To illustrate, consider the following BCLP:

EXAMPLE 2.2 (BCLP)

Max f1 = x1 + x2
Max f 2 = x1

Subject to: 4 x1 + 3 x2 ≤ 12
x1 , x2 , ≥ 0

Maximum f1 occurs at x = (0, 4) with (f1, f2) = (4, 0). Maximum f2 occurs at

x = (3, 0) with (f1, f2) = (3, 3). Thus, the ideal values of f1 and f2 are 4 and 3,
respectively, and the bounds on (f1, f2) on the efficient set will be:

3 ≤ f1 ≤ 4

0 ≤ f 2 ≤ 3.
48 Big Data Analytics Using Multiple Criteria Decision-Making Models

Let the DM set the goals for f1 and f2 as 3.5 and 2, respectively. Then the
GP model becomes:

x1 + x2 + d1− − d1+ = 3.5 (2.9)

x1 + d2− − d2+ = 2 (2.10)

4 x1 + 3 x2 ≤ 12 (2.11)

x1, x2, d1−, d1+, d2−, d2+ ≥ 0. (2.12)

Under the preemptive GP model, if the DM indicates that f1 is much

more important than f2, then the objective function will be

Min Z = P1d1− + P2d2−

subject to the constraints (2.9) through (2.12), where P1 is assumed to be

much larger than P2.
Under the non-preemptive GP model, the DM specifies relative weights
on the goal achievements, say w1 and w2. Then the objective function
becomes:

Min Z = w1d1− + w2d2−

subject to the same constraints (2.9) through (2.12).

2.6.2.2.2  Partitioning Algorithm for Preemptive Goal Programs

Linear Goal Programs: Linear GP problems can be solved efficiently by the
partitioning algorithm developed by Arthur and Ravindran (1978, 1980a). It
is based on the fact that the definition of preemptive priorities implies that
higher-order goals must be optimized before lower-order goals are even con-
sidered. Their procedure consists of solving a series of linear programming
subproblems by using the solution of the higher priority problem as the start-
ing solution for the lower priority problem. Care is taken that higher priority
achievements are not destroyed while improving lower priority goals.
Integer Goal Programs: Arthur and Ravindran (1980b) show how the par-
titioning algorithm for linear GP problems can be extended with a modi-
fied branch and bound strategy to solve both pure and mixed integer GP
problems. They demonstrate the applicability of the branch and bound algo-
rithm by solving a multiple-objective nurse scheduling problem (Arthur and
Ravindran, 1981).
Non-Linear Goal Programs: Saber and Ravindran (1996) present an efficient
and reliable method called the partitioning gradient-based (PGB) algorithm for
solving non-linear GP problems. The PGB algorithm uses the partitioning tech-
nique developed for linear GP problems and the generalized reduced gradient
Multi-Criteria Decision Making 49

(GRG) method to solve single-objective non-linear programming problems.

The authors also present numerical results by comparing the PGB algorithm
against a modified pattern search method for solving several non-linear GP
problems. The PGB algorithm found the optimal solution for all test problems
proving its robustness and reliability, while the pattern search method failed
in more than half the test problems by converging to a non-optimal point.
Kuriger and Ravindran (2005) have developed three intelligent search
methods to solve non-linear GP problems by adapting and extending the
simplex search, complex search, and pattern search methods to account for
multiple criteria. These modifications were largely accomplished by using
partitioning concepts of GP. The paper also includes computational results
with several test problems.
Some of the major GP formulations, including MINSUM GP, least absolute
value regression, MINMAX GP, preemptive GP, fractional GP, and non-linear
GP, have been reviewed by Lee and Olson (1999). Aouni and Kettani (2001)
have sketched the developments in the nearly 40 years of the history of GP.
Review articles on GP have been published in leading journals on a regu-
lar basis (Romero, 1986; Tamiz et  al., 1995). Recently, Lee and Olson (1999)
have provided a detailed overview of the applications of GP, in a variety of
fields including engineering, operations management, business, agriculture,
and public policy.
GP has been criticized by several authors (Min and Storbeck, 1991; Stewart,
1992; Zeleny, 1980). For example, it can, in some circumstances, choose domi-
nated solution (Ballestero and Romero, 1998). The criticisms could be over-
come by careful applications of the method. Some of the related issues are
reviewed in Lee and Olson (1999).

2.6.2.3  Method of Global Criterion and Compromise Programming

Method of global criterion (Hwang and Masud, 1979) and compromise program-
ming (Zeleny, 1982) fall under the class of MODM methods that do not require
any preference information from the DM.
Consider the MODM problem given by Equation 2.2. Let

S = {x/gj (x) ≤ 0, for all j}.

Let the ideal values of the objectives f1, f2, … , fk be f1*, f2*, … , fk*. The method
of global criterion finds an efficient solution that is “closest” to the ideal solu-
tion in terms of the Lp distance metric. It also uses the ideal values to normal-
ize the objective functions. Thus the MODM problem reduces to:
k p
 f i* − f i 
Minimize Z = ∑i =1
 f * 
i

subject to : x ∈ S.

50 Big Data Analytics Using Multiple Criteria Decision-Making Models

The values of fi* are obtained by maximizing each objective fi subject to the
constraints x ∈S, but ignoring the other objectives. The value of p can be 1, 2,
3, …, etc. Note that p = 1 implies equal importance to all deviations from the
ideal. As p increases, larger deviations have more weight.

2.6.2.4  Compromise Programming

Compromise programming is similar in concept to the method of global crite-
rion. It finds an efficient solution by minimizing the weighted Lp distance
metric from the ideal point as given below.
1/p
 k p p
Min Lp =  ∑ (
 i =1
)
λ i f i* − f i 
 (2.13)

subject to x ∈ S and p = 1, 2,… , ∞ ,

where λi’s are weights that have to be specified or assessed subjectively. Note
that λi could be set to 1/(fi*).

Theorem 2.3

Any point x* that minimizes Lp (Equation 2.13) for λi > 0 for all i, Σλi = 1 and
1 ≤ p < ∞ is called a compromise solution. Zeleny (1982) has proved that these
compromise solutions are non-dominated. As p → ∞, Equation 2.13 becomes

i
(
Min L∞ = Min Max λ i f i* − f i  )

and is known as the Tchebycheff metric.

Compromise programming has received some applications in the MCDM
literature. For example, Romero (1996) has used this technique in environ-
mental economics applications. Lee et al. (2001) have used compromise pro-
gramming for helping goal setting process in rural telecommunications
establishment. More than one MCDM method has been used in this article as
the weights of different objective functions have been estimated using AHP.

2.6.2.5  Interactive Methods

Interactive methods for MODM problems rely on the progressive articula-
tion of preferences by the DM. These approaches can be characterized by the
following procedure:

Step 1: Find a solution, preferably feasible and efficient.

Step 2: Interact with the DM to obtain his/her reaction or response to
the obtained solution.
Multi-Criteria Decision Making 51

Step 3: Repeat steps 1 and 2 until satisfaction is achieved or until some

other termination criterion is met.

When interactive algorithms are applied to real-world problems, the most

critical factor is the functional restrictions placed on the objective functions,
constraints, and the unknown preference function. Another important factor
is preference assessment styles (hereafter, called interaction styles). According to
Shin and Ravindran (1991), the typical interaction styles are:

1.
Binary pairwise comparison: The DM must compare a pair of two
dimensional vectors at each interaction.
2.
Pairwise comparison: The DM must compare a pair of p-dimensional
vectors and specify a preference.
3.
Vector comparison: The DM must compare a set of p-dimensional vec-
tors and specify the best, the worst or the order of preference (note
that this can be done by a series of pairwise comparisons).
4.
Precise local trade-off ratio: The DM must specify precise values of
local trade-off ratios at a given point. It is the marginal rate of substitu-
tion between objectives fi and fj: in other words, trade-off ratio is how
much the DM is willing to give up in objective j for a unit increase in
objective i at a given efficient solution.
5.
Interval trade-off ratio: The DM must specify an interval for each local
trade-off ratio.
6.
Comparative trade-off ratio: The DM must specify his preference for a
given trade-off ratio.
7.
Index specification and value trade-off: The DM must list the indices of
objectives to be improved or sacrificed, and specify the amount.
8.
Aspiration levels (or reference point): The DM must specify or adjust
the values of the objectives which indicate his/her optimistic wish
concerning the outcomes of the objectives.

Shin and Ravindran (1991) also provide a detailed survey of MODM inter-
active methods. Their survey includes the following:

• A classification scheme for all interactive methods.

• A review of methods in each category based on functional assump-
tions, interaction style, progression of research papers from the first
publication to all its extensions, solution approach, and published
applications.
• A rating of each category of methods in terms of the DM’s cognitive
burden, ease of use, effectiveness, and handling inconsistency.
52 Big Data Analytics Using Multiple Criteria Decision-Making Models

Lofti et al. (1992) have compared the performance of interactive methods with
other methods such as AHP (implemented on the computer using the software
Expert Choice), and according to them, interactive methods outperformed the
latter on numerous measures, but for no measure was the reverse true.

2.6.2.6  Data Envelopment Analysis

We have seen that SMART or AHP simply aggregates the ratings of alterna-
tives using criteria weights. However, if experts find it difficult to provide
importance weights of the criteria, it is not possible to use these methods. An
alternative in such cases is to identify non-dominated alternative, that is, the
alternatives that have been rated better than, and not rated below, others. We
have seen that the MOLP can help identify the non-dominated alternatives.
The DEA is a relatively recent technique that can be effectively used to iden-
tify the non-dominated solutions in an MODM problem.
The methodology of DEA is relatively more complex compared to SMART
or AHP as it uses linear programming for identifying the non-dominated
alternatives. The method has been first proposed by Charnes et  al. (1978)
for assessing the relative performance of a set of firms that use a variety of
identical inputs to produce a variety of identical outputs. Firms that produce
maximum possible outputs using a given set of inputs or that consume mini-
mum possible inputs for a given set of outputs are considered more efficient
than others. To identify these efficient firms, a mathematical programming
problem, that maximizes the ratio of weighted sum of outputs to weighted
sum of inputs subject to the condition that similar ratios for all the firms are
less than one, is used. Obviously, efficient firms will have the maximized
value of their ratio to be unity or 100%.
DEA was originally developed as a tool for performance measurement
(Ramanathan, 2003), and is a relatively late entrant to the field of MCDM.
However, over the last few years, the linkages between the fields of DEA and
MCDM have been explored, and DEA is now widely accepted as a tool for
MCDM. The most fundamental use of DEA in MCDM is the identification of
non-dominated alternatives. Note that identification of non-dominated alter-
natives does not require any preference information, and is purely based on
available data. A non-dominated alternative is identified as the most efficient
alternative in DEA with 100% efficiency.
Doyle and Green (1993) have highlighted that the absence of DEA in the set
of MCDM approaches presented in the critical survey of MCDM techniques
presented by Professor Stewart (1992) was an important omission, which was
later accepted by Stewart (1994). Belton and Stewart (1999) further explored
the relationship between DEA and MCDM from a decision theoretic perspec-
tive. Joro et al. (1998) provide in detail the comparisons between DEA and
MOLP. Recent reviews of multi-criteria methods (Malczewski and Jackson,
2000; Urli and Nadeau, 1999) have included DEA as an MCDM tool in their
discussion. The book on MCDM by Shi and Zeleny (2000) has included a
Multi-Criteria Decision Making 53

separate section on DEA. One of the eight parts and three of the 21 papers
included in the book are on DEA. Incorporating preference information in
a DEA analysis provides a natural extension of DEA toward MCDM. This
issue has been discussed for a long time in DEA literature (Halme et al., 1999;
Thanassoulis and Dyson, 1992).
Since DEA was proposed more than two decades ago, the methodology
has received numerous traditional as well as novel applications. Seiford
(1996) has presented one of the recent bibliographies on DEA. A very com-
prehensive DEA bibliography is maintained at the University of Warwick
by Emrouznejad (1995–2001). This bibliography is available on the Internet
at the site, http://www.deazone.com/. Some prominent applications of DEA
include the education sector (Ganley and Cubbin, 1992; Ramanathan, 2001b),
banks (Yeh, 1996), comparative risk assessment (Ramanathan, 2001c), health
sector (Bates et  al., 1996), transport (Ramanathan, 2000), energy (Lv et  al.,
2015), environment (Bi et  al., 2015; Fare et  al., 1996; Zhao et  al., 2016), and
international comparisons on carbon emissions (Ramanathan, 2002).

2.6.3  Other MCDM Methods

Several MCDM methods, including those discussed in this chapter, have
been reviewed by Jacquet-Lagrèze and Siskos (2001). They have listed sev-
eral methods, mainly based on outranking concepts, provided a list of soft-
ware support available to these methods, and reviewed their application
areas. Miettinen (1999) provides an overview of several methods (including
some described in this chapter) for non-linear MOO. As mentioned earlier,
Guitouni and Martel (1998) have provided a comprehensive list of various
MCDM methods, references to their application literature, and a comparative
study of various methods, in their effort to provide tentative guidelines for
choosing an appropriate method for a given application. Similarly, Zanakis
et al. (1998) have compared performance of eight different MADM methods
(ELECTRE, TOPSIS [the Technique for Order of Preference by Similarity
to Ideal Solution], multiplicative exponential weighting, simple additive
weighting, and four versions of AHP). A variety of applications of MCDM
methods are available in Karwan et al. (1997).
The reference point approach developed at the International Institute
of Applied Systems Analysis (IIASA) is another popular MODM method.
It is closely related to the aspiration-level concept of GP. This approach is
described in Wierzbicki (1998, 1999). The PRIME method or preference ratio
through intervals in multi-attribute evaluation (Salo and Hamalainen, 2001)
is a relatively recent one that deals with incomplete information about the
preferences of the DMs. The MACBETH (measuring attractiveness by a cate-
gorical based evaluation technique) approach by Bana e Costa and Vansnick
(1995) has been developed in Europe in the 1990s. It is an interactive approach
for cardinal measurement of judgments about the degrees of attractiveness
in decision processes.
54 Big Data Analytics Using Multiple Criteria Decision-Making Models

Other methods based on outranking approach include QUALIFLEX

method (Paelnick, 1978), ORESTE method (Roubens, 1982), and the TACTIC
method (Vansnick, 1986). Jacquet-Lagrèze and Siskos (1982) have proposed
the regression based UTA (utilitiés additives) method. An application of
UTA method for business failure prediction is presented by Zoupounidis
and Dimitras (1998). Another MCDM methodology is the ZAPROS method
(Larichev, 1999, 2000; Larichev and Moshkovich, 1995). NIMBUS (Miettinen
and Mäkelä, 2000) is an MOO method capable of solving non-differentiable
and non-convex problems.

2.7 A General Comparative Discussion

of MCDM Methodologies
In this section, various MCDM methods are compared in general terms with
reference to some vital parameters when used in practical applications:

• Uncertainty: In terms of their purpose, it is possible to make com-

parisons between MCDM methods. Some methods, such as MAUT
and PRIME, have been designed to deal with uncertainty. Initially,
SMART and the AHP were developed for the analysis of decision
problems where one of the available alternatives is to be selected
in a situation where several incommensurate criteria need to be
accounted for. Thus, they assume that the objectives and the alter-
natives are both known, whereby the uncertainties that pertain, for
example, to the performance of the alternatives with regard to the
criteria are not explicitly modelled. However, these methods, and the
AHP, in particular, have found applications beyond this relatively
narrow problem context, as they have supported the construction of
R&D project portfolios, among others. Other methods are generally
not designed to handle uncertainty.
• Incomplete information: Practical applicability of the methods is
often constrained by the unavailability of complete information.
For example, in MADM methods, DM may not want to comment
on some criteria, or may not want to rate some alternatives on the
basis of some criteria. This may arise when DM feels that he does
not have the expertise to comment on the specific comparison. In
MODM problems, the DM may not be able to compare some criteria.
Note that, in many MODM methods, the preference information on
criteria is often required as external input. For example, GP or com-
promise programming requires that the weights of criteria are avail-
able before the actual implementation of the methods, and normally
do not specify how this information should be obtained.
Multi-Criteria Decision Making 55

• When only incomplete information is available on the preference

information of the DM, it is important that the methodologies pro-
vide rankings based on the available information. Some MCDM
methods can provide rankings based on incomplete information, for
example, PRIME and MAHP.
• Number of questions: All the MADM methods require the elicita-
tion of DMs’ opinions for rating criteria and alternatives. MODM
methods also require this elicitation for obtaining the information
on importance of criteria. This may be done using written question-
naires, direct interviews, and/or by the use of suitable software.
More often, the number of questions to be answered by the expert
tends to rise with the number of elements (criteria or alternatives)
that are to be compared. The problem is the worst in methods that
are based on pairwise comparison of decision elements (criteria or
alternatives), such as the AHP or the outranking methods. For exam-
ple, for comparing n elements using AHP, the number of questions
to be answered by an expert will be n(n − 1)/2. This is because AHP
elicits many redundant judgments from experts so that the consis-
tency of their judgments can be verified. Of course, some modifica-
tions of AHP do not need so many comparisons, but they also do not
verify consistency of judgments.
• Checking the consistency of judgments: Because of the redundant judg-
ments, AHP is able to provide a check on the consistency of judg-
ments given by the experts. This is done using the so-called the
consistency index. More details are available in the last chapter.
If the expert’s judgments are not found to be consistent, then it is
not advisable to use the judgments and the expert may be asked to
provide his judgments once again. Note that many modifications of
AHP do not provide this consistency check. Similarly, most of the
other MCDM methods also do not provide a mathematically verifi-
able consistency check.
• Group decisions: Most practical uses of MCDM methods require elici-
tation of judgments from many DMs. It is then important to appro-
priately aggregate the judgments of all the DMs. This group decision
problem is integral to all the MCDM methods. As indicated earlier,
the best way to obtain the overall group judgments is to encourage
unanimous decisions by the group through discussions. However,
sometimes it may not be possible to arrive at a single group opin-
ion by consensus, and it may be necessary to mathematically aggre-
gate the opinions expressed by individual experts. In most of the
MCDM methods, mathematical aggregation methods, usually in the
form of weighted additive or multiplicative aggregation, where the
weights are the relative importance measures associated with each
DM, are used for the group preference aggregation. However, it
56 Big Data Analytics Using Multiple Criteria Decision-Making Models

must be realized that the importance weights of group members

are extremely difficult to estimate in practice. Thus, the best way to
arrive at a group opinion is through consensus.
• User friendliness: The primary purpose of the MCDM methods for
a practical application is to elicit the opinions of DMs, and the DMs
are not expected to have a thorough knowledge of these methods.
Hence, it is important that the methods are intuitive, easy to under-
stand, flexible to be adopted to any modeling problem, and easy to
execute. MADM methods are generally superior to the other meth-
ods in this sense as they do not require any sophisticated mathe-
matical programming knowledge. They can be easily understood.
Execution is often through structured questionnaires. In general,
MODM methods fare poorly on this score as they are based on the
sophisticated mathematical programming concepts, which are not
very transparent to the users (DMs here).
• Software and Internet support: Another important feature highly influ-
encing the practical utility of the MCDM methods is the availability
of suitable software support and, if possible, the Internet support.
Software can expedite the implementation of a method. It will still
be better if there is a website on the Internet where the software
can be accessed, and, if it is available as a freeware, downloaded.
Note that most of the MCDM methods available in the literature can
be easily programmed using various high-level programming lan-
guages to generate their software.
Software implementations of many of the methods are avail-
able. MAUT is implemented in computer using the software Logical
Decision; the software PREFCALC is also based on MAUT. Hiview,
HIPRE, V.I.S.A. are the softwares used for implementing MAVT
and MAUT. These and other softwares for multi-attribute util-
ity analysis are listed by Buede (1992), in the software review sec-
tion of the Journal of Multi-Criteria Decision Analysis. Expert Choice,
Criterium, HIPRE are the softwares that support AHP. Trial versions
of the latest software of Expert Choice (the software implementa-
tion of AHP) are available for free from its developers. HIPRE (or
its windows ­version WINPRE) is available from its developers at
Aalto University. Similarly, software for DEA is available from
several developers, such as the University of Warwick, UK and
Banxia Software Ltd., UK. PROMCALC software (Mareschal, 1988)
implements the PROMETHEE outranking method (as per Hanne,
1999). The GAIA software (Mareschal and Brans, 1988) is an exten-
sion of  PROMCALC. The software for PRIME is available from
its developers at Aalto University. More detailed and latest infor-
mation on  many decision analysis software has been provided by
Maxwell (2000).
Multi-Criteria Decision Making 57

Many of these methods also have Internet support. PRIME is acces-

sible from the Internet site of Aalto University (http://sal.aalto.fi/en/
resources/downloadables/prime). Several variations of AHP are avail-
able extensively on the Internet. For example, try the following sites:
HIPRE or WINPRE (http://sal.aalto.fi/en/resources/downloadables/
hipre3) and the website of Expert Choice (http://www.expertchoice.
com/download/Default.htm). Internet support for MAUT is available
at many places such as the site of Logical Decisions, Inc. (http://www.
logicaldecisions.com/). Software for implementing the ELECTRE
methods can be obtained by contacting the website of LAMSADE at:
http://www.lamsade.dauphine.fr/spip.php?rubrique64.
In general, MODM methods can be executed as extensions of soft-
ware for Linear Programming (LP). However, using LP software is
often cumbersome, and hence specialized software will be very help-
ful. Such specialized software support is available for some MODM
methods. For example, specialized software for GP, called GPSYS, has
been developed at the University of Portsmouth, UK. Several free soft-
wares on DEA are also available on the Internet. Demonstration ver-
sions of DEA software are available from Banxia Software Ltd. (http://
www.banxia.com). Free DEA softwares are available from http://
www.holger-scheel.de/ems/, and the University of Queensland,
Australia (http://www.uq.edu.au/economics/cepa/deap.php).
• Nature of outputs: The rankings of alternatives obtained by using
the MCDM methods vary, and it is important to interpret them cor-
rectly. The weights of alternatives obtained using AHP (conventional
form) will sum to unity, while it may not be the case with the weights
derived, say, from MAUT, MAVT, SMART, or the outranking meth-
ods. Similarly, DEA provides efficiency scores such that those tech-
nologies with a score of unity (or 100%) are the best technologies.
Obviously, DEA efficiency scores do not sum to unity. Thus, while
one can never expect any alternative to get a score of one in an AHP
exercise, there will be at least one alternative with a score of one in
DEA. Most of the MODM methods, such as GP or compromise pro-
gramming, select appropriate values for the continuous decision vari-
ables, and hence cannot be considered as the weights of alternatives.

2.8  MCDM in the Era of Big Data

The next chapter in this book deals with big data ideas. The field of MCDM
assumes special importance in this era of big data and business analytics.
In the modern digital world, a wealth of so-called big data is being gener-
ated every minute, every second, even every nanosecond. Thanks to the
58 Big Data Analytics Using Multiple Criteria Decision-Making Models

astounding technological revolution, everything around us is being captured

in some way or the other, stored in some form, and it is believed that this has
the potential to make business sense. Business analytics (BA) involves an
appropriate use of big data to provide new predictive/prescriptive/descrip-
tive insights that will allow businesses perform better (Chen et  al., 2012).
BA tools involve both modeling-based tools and statistics-based tools. We
believe that MCDM models have special importance in making the sense
of huge volumes of data because it is closer to the decision-making activi-
ties of DM with more realizing assumptions. MCDM tools, either alone or
along with other BA tools, can be used for making the business sense of big
data. There is a huge knowledge gap to illustrate how MCDM methods could
be advantageously employed in this context. We believe that subsequent
chapters in this volume attempt to fill the knowledge gap on the paucity of
MCDM models in the context of big data and business analytics.

2.9 Summary
In this chapter, the basic concepts and terminologies of MCDM have been
reviewed. A classification of MCDM methods is outlined. The broad area of
MCDM can be divided into two general categories, MADM and MODM. The
former involves cases in which the set of alternatives is defined explicitly by
a finite list from which one or a few alternatives should be chosen that reflect
DM’s preference structure, while the latter involves cases in which the set of
alternatives is defined implicitly by a mathematical programming structure
with objective functions. Some of the basic concepts of MODM and some
common characteristics of MADM methods have then been presented. Then,
an overview of some important MCDM methods has been presented.
The methods covered in the overview include MAUT, AHP, the ELECTRE
methods, PROMETHEE methods, fuzzy set theory (all MADM methods),
MOLP, GP, AIM, compromise programming, and DEA (MODM methods).
The overview has provided the basic idea behind each method along with
important introductory and review articles, strengths and criticisms, and
some important applications. Other methods, such as TOPSIS, reference
point approach, PRIME, and MACBETH, have been briefly mentioned in
the overview. Finally, the different MCDM methods have been compared
in terms of several vital parameters, including their ability to handle uncer-
tainty, incomplete information, number of responses required from the DM,
ability to be used in group decisions, user friendliness, software and Internet
support, and the nature of their outputs. The authors hope that this over-
view provides enough information about the field of MCDM to kindle fur-
ther interest in this exciting field, and encourage more practical applications.
Multi-Criteria Decision Making 59

References
Alley, W. M. 1983. Comment on Multiobjective river basin planning with qualitative
criteria by M. Gershon, L. Duckstein, and R. McAniff. Water Resources Research.
19(1): 293–294.
Ammar, S. and R. Wright. 2000. Applying fuzzy-set theory to performance evalua-
tion. Socio-Economic Planning Sciences. 34: 285–302.
Aouni, B. and O. Kettani. 2001. Goal programming: A glorious history and a promis-
ing future. European Journal of Operational Research. 133: 225–241.
Arthur, J. L. and A. Ravindran. 1978. An efficient goal programming algorithm using
constraint partitioning and variable elimination. Management Science. 24(8):
867–868.
Arthur, J. L. and A. Ravindran. 1980a. PAGP: An efficient algorithm for linear goal pro-
gramming problems. ACM Transactions on Mathematical Software. 6(3): 378–386.
Arthur, J. L. and A. Ravindran. 1980b. A branch and bound algorithm with con-
straint partitioning for integer goal programs. European Journal of Operational
Research. 4: 421–425.
Arthur, J. L. and A. Ravindran. 1981. A multiple objective nurse scheduling model.
Institute of Industrial Engineers Transactions. 13: 55–60.
Ballestero, E. and C. Romero. 1998. Multiple Criteria Decision Making and its Applications
to Economic Problems. Kluwer Academic Publishers, Boston.
Bana e Costa, C. A. and J. C. Vansnick. 1995. General overview of the MACBETH
approach, In: P. M. Pardalos, Y. Siskos and C. Zopounidis (Eds.), 93–100.
Advances in Multicriteria Analysis. Kluwer Academic Publishers, Dordrecht, The
Netherlands.
Bates, J. M., D. B. Baines, and D. K. Whynes. 1996. Measuring the efficiency of pre-
scribing by general practitioners. Journal of the Operational Research Society. 47:
1443–1451.
Bellman, R. E. and L. A. Zadeh. 1970. Decision making in a fuzzy environment.
Management Science. 17: B141–B164.
Belton, V. 1999. Multi-criteria problem structuring and analysis in a value theory
framework, In: T. Gal, T. J. Stewart, and T. Hanne (Eds.), Multicriteria Decision
Making: Advances in MCDM Models, Algorithms, Theory, and Applications, Kluwer
Academic Publishers, Boston, pp. 12.1–12.32.
Belton V. and T. Gear. 1983. On a shortcoming of Saaty’s method of analytic hierar-
chies. Omega. 11/3: 228–230.
Belton, V. and T. J. Stewart. 1999. DEA and MCDA: Competing or complementary
approaches? In: N. Meskens and M. Roubens (Eds.), 87–104. Advances in Decision
Analysis. Kluwer Academic Publishers, Dordrecht, Netherlands.
Bi, G., Y. Luo, J. Ding, and L. Liang. 2015. Environmental performance analysis of
Chinese industry from a slacks-based perspective. Annals of Operations Research.
228(1): 65–80.
Bisdorff, R. 2000. Logical foundation of fuzzy preferential systems with application to
the electre decision aid methods. Computers & Operations Research. 27: 673–687.
Bose, U., A. M. Davey, and L. O. David. 1997. Multi-attribute utility methods in group
decision making: Past applications and potential for inclusion in GDSS. Omega.
25(6): 691–706.
60 Big Data Analytics Using Multiple Criteria Decision-Making Models

Bouyssou, D. 1990. Building criteria: A prerequisite for MCDA, In: C. A. Bana e Costa
(Ed.), 58–80. Readings in Multiple Criteria Decision Aid. Springer, Berlin.
Bouyssou, D. and Ph. Vincke. 1997. Ranking alternatives on the basis of preference
relations: A progress report with special emphasis on outranking relations.
Journal of Multi-Criteria Decision Analysis. 6: 77–85.
Brans, J. P., B. Mareschal, and Ph. Vincke. 1984. Promethee: A new family of outrank-
ing methods in multicriteria analysis, In: J. P. Brans (Ed.), 408–421. Operational
Research ’84. Elsevier Science Publishers B.V, North-Holland.
Buede, D. M. 1992. Software review: Overview of the MCDA software market. Journal
of Multi-Criteria Decision Analysis. 1: 59–61.
Charnes, A. and W. W. Cooper. 1961. Management Models and Industrial Applications of
Linear Programming. Wiley, New York.
Charnes, A. W. and W. Cooper, and R. O. Ferguson. 1955. Optimal estimation of
­executive compensation by linear programming. Management Science. 1(2):
138–151.
Charnes, A., W. W. Cooper, and E. Rhodes. 1978. Measuring the efficiency of decision
making units. European Journal of Operational Research. 2: 429–444.
Chen, H., R. H. Chiang, and V. C. Storey. 2012. Business intelligence and analytics:
From big data to big impact. MIS Quarterly. 36(4): 1165–1188.
Doyle, J. and R. Green. 1993. Data envelopment analysis and multiple criteria deci-
sion making. Omega. 21(6): 713–715.
Duarte, B. P. M. 2001. The expected utility theory applied to an industrial decision
problem—What technological alternative to implement to treat industrial solid
residuals. Computers & Operations Research. 28: 357–380.
Dyer, J. S. 1990. Remarks on the analytic hierarchy process. Management Science. 36/3:
249–258.
Dyer, J. S., P. C. Fishburn, R. E. Steuer, J. Wallenius, and S. Zionts. 1992. Multiple
criteria decision making, multiattribute utility theory: The next ten years.
Management Science. 38(5): 645–654.
Edwards, W. 1977. How to use multiattribute utility measurement for social decision
making. IEEE Transactions on Systems, Man, and Cybernetics. 7/5: 326–340.
Elkarni, F. and I. Mustafa. 1993. Increasing the utilization of solar energy
­technologies (SET) in Jordan: Analytic hierarchy process. Energy Policy. 21:
978–984.
Emrouznejad, A. 1995–2001. Ali Emrouznejad’s DEA HomePage. Warwick Business
School, Coventry CV4 7AL, UK.
Evans, G. W. 1984. An overview of techniques for solving multiobjective mathemati-
cal programs. Management Science. 30(11): 1268–1282.
Fare, R., S. Grosskopf, and D. Tyteca. 1996. An activity analysis model of the environ-
mental performance of firms—application to fossil-fuel-fired electric utilities.
Ecological Economics. 18: 161–175.
French, S. 1984. Interactive multi-objective programming: Its aims, applications and
demands. Journal of the Operational Research Society. 35: 827–834.
French, S., R. Hartley, L. C. Thomas, and D. J. White. 1983. Multi-Objective Decision
Making. Academic Press, London.
Ganley, J. A. and J. S. Cubbin. 1992. Public Sector Efficiency Measurement: Applications
of Data Envelopment Analysis. North-Holland, Amsterdam.
Geoffrion, A. 1968. Proper efficiency and theory of vector maximum. Journal of
Mathematical Analysis and Applications. 22: 618–630.
Multi-Criteria Decision Making 61

Gershon, M. and L. Duckstein. 1983. Reply. Water Resources Research. 19(1): 295–296.
Golden, B. L., E. A. Wasil, and D. E. Levy. 1989. Applications of the analytic hierarchy
process: A categorized, annotated bibliography, In: B. L. Golden, E. A. Wasil,
and P. T. Harker (Eds.), 37–58. The Analytic Hierarchy Process: Applications and
Studies. Springer-Verlag, Berlin.
Govindan, K., S. Rajendran, J. Sarkis, and P. Murugesan. 2015. Multi criteria decision
making approaches for green supplier evaluation and selection: A literature
review. Journal of Cleaner Production. 98: 66–83.
Graves, S. B., J. L. Ringuest, and J. F. Bard. 1992. Recent developments in screening
methods for nondominated solutions in multiobjective optimization. Computers
& Operations Research. 19(7): 683–694.
Guitouni, A. and J.-M. Martel. 1998. Tentative guidelines to help choosing an appro-
priate MCDA method. European Journal of Operational Research. 109: 501–521.
Halme, M., T. Joro, P. Korhonen, S. Salo, and J. Wallenius. 1999. A value efficiency
approach to incorporating preference information in data envelopment analy-
sis. Management Science. 45: 103–115.
Hanne, T. 1999. Meta decision problems in multiple criteria decision making, In:
T. Gal, T. J. Stewart, and T. Hanne (Eds.), Multicriteria Decision Making: Advances
in MCDM Models, Algorithms, Theory, and Applications, Kluwer Academic
Publishers, Boston, pp. 6.1–6.25.
Henig, M. I. and J. T. Buchanan. 1996. Solving MCDM problems: Process concepts.
Journal of Multi-Criteria Decision Analysis. 5: 3–21.
Ho, W., X. Xu, and P. K. Dey. 2010. Multi-criteria decision making approaches for
supplier evaluation and selection: A literature review. European Journal of
Operational Research. 202(1): 16–24.
Hokkanen, J. and P. Salminen. 1997. ELECTRE III and IV decision aids in an environ-
mental problem. Journal of Multi-Criteria Decision Analysis. 6: 215–226.
Howard, R. 1992. Heathens, heretics and cults: The religious spectrum of decision
aiding. Interfaces. 22(6): 15–27.
Hwang, C. L. and A. Masud. 1979. Multiple Objective Decision Making—Methods and
Applications: A State of the Art Survey. Lecture notes in economics and math-
ematical systems. 164, Springer-Verlag, Berlin.
Jacquet-Lagrèze, E. and J. Siskos. 1982. Assessing a set of additive utility functions for
multicriteria decision making: The UTA method. European Journal of Operational
Research. 10: 151–164.
Jacquet-Lagrèze, E. and Y. Siskos. 2001. Preference disaggregation: 20 years of MCDA
experience. European Journal of Operational Research. 130: 233–245.
Jato-Espino, D., E. Castillo-Lopez, J. Rodriguez-Hernandez, and J. C. Canteras-
Jordana. 2014. A review of application of multi-criteria decision making meth-
ods in construction. Automation in Construction. 45: 151–162.
Jones, M., C. Hope, and R. Hughes. 1990. A multi-attribute value model for the
study of UK energy policy. Journal of the Operational Research Society. 41(10):
919–929.
Joro, T., P. Korhonen, and J. Wallenius. 1998. Structural comparison of data envelop-
ment analysis and multiple objective linear programming. Management Science.
40: 962–970.
Kabir, G., R. Sadiq, and S. Tesfamariam. 2014. A review of multi-criteria decision-
making methods for infrastructure management. Structure and Infrastructure
Engineering. 10(9): 1176–1210.
62 Big Data Analytics Using Multiple Criteria Decision-Making Models

Karwan, M. H., J. Spronk, and J. Wallenius. (Eds.). 1997. Essays in Decision Making: A
Volume in Honour of Stanley Zionts. Springer-Verlag, Berlin.
Keeney, R. L. 1992. Value Focused Thinking: A Path to Creative Decision-making. Harvard
University Press, Cambridge, MA.
Keeney, R. L. 1999. Developing a foundation for strategy at Seagate software.
Interfaces. 29/6: 4–15.
Keeney, R. L. and T. L. McDaniels. 1999. Identifying and structuring values to guide
integrated resource planning at bc gas. Operations Research. 47/5: 651–662.
Keeney, R. L. and H. Raiffa. 1976. Decisions with Multiple Objectives: Preferences and
Value Tradeoffs. Wiley, New York. Another edition available in 1993 from
Cambridge University Press.
Kirkwood, C. W. 1992. An overview of methods for applied decision analysis.
Interfaces. 22(6): 28–39.
Korhonen, P. 1992. Multiple criteria decision support: The state of research and future
directions. Computers & Operations Research. 19(7): 549–551.
Koundinya, S., D. Chattopadhyay, and R. Ramanathan. 1995. Combining qualitative
objectives in integrated resource planning: A combined AHP—Compromise
programming model. Energy Sources. 17(5): 565–581.
Kuriger, G. and A. Ravindran. 2005. Intelligent search methods for nonlinear goal
programs. Information Systems and Operational Research. 43: 79–92.
Larichev, O. I. 1999. Normative and descriptive aspects of decision making, In: T.
Gal, T. J. Stewart, and T. Hanne (Eds.), Multicriteria Decision Making: Advances
in MCDM Models, Algorithms, Theory, and Applications, Kluwer Academic
Publishers, Boston, pp. 5.1–5.24.
Larichev, O. I. 2000. Qualitative comparison of multicriteria alternatives, In: Yong,
S. and M. Zeleny (Eds.), New Frontiers of Decision Making for the Information
Technology Era, World Scientific Publishing Co. Pte. Ltd., Singapore, pp. 207–224.
Larichev, O. I. and H. Moshkovich. 1995. ZAPROS-LM—A method and system for
ordering multiattribute alternatives. European Journal of Operational Research.
82(3): 503–521.
Lee, H., Y. Shi, S. M. Nazem, S. Y. Kang, T. H. Park, and M. H. Sohn. 2001. Multicriteria
hub decision making for rural area telecommunication networks. European
Journal of Operational Research. 133: 483–495.
Lee, S. M. and D. L. Olson. 1999. Goal programming, In: T. Gal, T. J. Stewart, and
T. Hanne (Eds.), Multicriteria Decision Making: Advances in MCDM Models,
Algorithms, Theory, and Applications, Kluwer Academic Publishers, Boston,
pp. 8.1–8.33.
Leung, P. S., K. Heen, and H. Bardarson. 2001. Regional economic impacts of fish
resources utilization from the Barents Sea: Trade-offs between economic
rent, employment, and income. European Journal of Operational Research. 133:
432–446.
Lofti, V., T. J. Stewart, and S. Zionts. 1992. An aspiration-level interactive model
for multiple criteria decision making. Computers & Operations Research. 19(7):
671–681.
Lootsma, F. A. 1999. Multi-Criteria Decision Analysis Via Ratio and Difference Judgement.
Dordrecht: Kluwer Academic Publishers.
Lv, W., X. Hong, and K. Fang. 2015. Chinese regional energy efficiency change and its
determinants analysis: Malmquist index and Tobit model. Annals of Operations
Research. 228(1): 9–22.
Multi-Criteria Decision Making 63

Malczewski, J. and M. Jackson. 2000. Multicriteria spatial allocation of educational

resources: An overview. Socio-Economic Planning Sciences. 34: 219–235.
Mareschal, B. 1988. Weight stability intervals in multicriteria decision aid. European
Journal of Operational Research. 33: 54–64.
Mareschal, B. and J. P. Brans. 1988. Geometric representations for MCDA. European
Journal of Operational Research. 34: 69–77.
Maxwell, D. T. 2000. Decision analysis: Aiding insight V—Fifth biennial survey.
ORMS Today. October: 28–35. Also available on the internet at the URL: http://
www.lionhrtpub.com/orms/surveys/das/das.html
Miettinen, K. 1999. Nonlinear Multiobjective Optimization. Kluwer Academic Publishers,
Boston.
Miettinen, K. and M. M. Mäkelä. 2000. Interactive multiobjective optimization
system WWW-NIMBUS on the Internet. Computers & Operations Research. 27:
709–723.
Millet, I. and P. T. Harker. 1990. Globally effective questioning in the analytic hierar-
chy process. European Journal of Operational Research. 48: 88–97.
Min, H. and J. Storbeck. 1991. On the origin and persistence of misconceptions in
goal programming. Journal of the Operational Research Society. 42(4): 301–312.
Özelkan, E. C. and L. Duckstein. 2000. Multi-objective fuzzy regression: A general
framework. Computers & Operations Research. 27: 635–652.
Paelnick, J. 1978. Qualiflex, a flexible multiple criteria method. Economics Letters. 3:
193–197.
Parra, M. A., A. B. Terol, and M. V. Rodríquez Uría. 2001. A fuzzy goal program-
ming approach to portfolio selection. European Journal of Operational Research.
133: 287–297.
Pohekar, S. D. and M. Ramachandran. 2004. Application of multi-criteria decision
making to sustainable energy planning—A review. Renewable and Sustainable
Energy Reviews. 8(4): 365–381.
Ramanathan, R. 1997. A note on the use of goal programming for Multiplicative AHP.
The Journal of Multi Criteria Decision Analysis. 6: 296–307.
Ramanathan, R. 1998. A multi-criteria methodology to the global negotiations on
climate change. IEEE Transactions on Systems, Man and Cybernetics—Part C:
Applications and Reviews. 28(4): 541–548.
Ramanathan, R. 1999. Selection of appropriate greenhouse gas mitigation options.
Global Environmental Change: Human and Policy Dimensions. 9(3): 203–210.
Ramanathan, R. 2000. A holistic approach to compare energy efficiencies of different
transport modes. Energy Policy. 28(11): 743–747.
Ramanathan, R. 2001a. A note on the use of analytic hierarchy process for environ-
mental impact assessment. Journal of Environmental Management. 63(1): 27–35.
Ramanathan, R. 2001b. A data envelopment analysis of comparative performance
of schools in the Netherlands. OPSEARCH—The Indian Journal of Operational
Research. 38(2): 160–182.
Ramanathan, R. 2001c. Comparative risk assessment of energy supply technologies: A
data envelopment analysis approach. Energy—The International Journal. 26: 197–203.
Ramanathan, R. 2002. Combining indicators of energy consumption and CO2 emis-
sions: A cross-country comparison. International Journal of Global Energy Issues.
17(3): 214–227.
Ramanathan, R. 2003. An Introduction to Data Envelopment Analysis: A Tool for
Performance Measurement. SAGE Publications, New Delhi.
64 Big Data Analytics Using Multiple Criteria Decision-Making Models

Ravindran, A. 2016. Multiple criteria decision making: An overview, Chapter 2, In:

A. R. Ravindran (Ed.), 15–50. Multiple Criteria Decision Making in Supply Chain
Management. CRC Press, Boca Raton, FL.
Ravindran, A., K. M. Ragsdell, and G. V. Reklaitis. 2006. Engineering Optimization:
Methods and Applications, 2nd ed. (Chapter 11). John Wiley, Hoboken, NJ.
Romero, C. 1986. A survey of generalized goal programming (1970–1982). European
Journal of Operational Research. 25: 183–191.
Romero, C. 1996. Multicriteria decision analysis and environmental economics: An
approximation. European Journal of Operational Research. 96: 81–89.
Roubens, M. 1982. Preference relations on actions and criteria in multicriteria deci-
sion making. European Journal of Operational Research. 10: 51–55.
Roy, B. 1996. Multicriteria Methodology for Decision Aiding. Kluwer Academic
Publishers, Dordrecht, The Netherlands.
Roy, B. 1999. Decision-Aiding today: What should we expect, In: T. Gal, T. J. Stewart,
and T. Hanne (Eds.), Multicriteria Decision Making: Advances in MCDM Models,
Algorithms, Theory, and Applications, Kluwer Academic Publishers, Boston, pp.
1.1–1.35.
Roy, B., M. Présent, and D. Silhol. 1986. A programming method for determining
which Paris metro stations should be renovated. European Journal of Operational
Research. 24: 318–334.
Roy, B. and D. Vanderpooten. 1996. The European school of MCDA: Emergence, basic
features and current works. Journal of Multi-Criteria Decision Analysis. 5: 22–38.
Saaty, T. L. 1980. The Analytic Hierarchy Process: Planning, Priority Setting and Resource
Allocation. McGraw-Hill, New York.
Saber, H. M. and A. Ravindran. 1996. A Partitioning Gradient Based (PGB) algorithm
for solving nonlinear goal programming problem. Computers and Operations
Research. 23: 141–152.
Salminen, P., J. Hokkanen, and R. Lahdelma. 1998. Comparing multicriteria meth-
ods in the context of environmental problems. European Journal of Operational
Research. 104: 485–496.
Salo, A. A. and R. P. Hamalainen. 2001. Preference ratios in multiattribute evaluation
(PRIME)-elicitation and decision procedures under incomplete information.
IEEE Transactions on Systems, Man, and Cybernetics-Part A: Systems and Humans.
31(6): 533–545.
Schniederjans, M. J. 1995. Goal Programming: Methodology and Applications. Kluwer
Academic Publishers, Boston.
Scott, J. A., W. Ho, and P. K. Dey. 2012. A review of multi-criteria decision-making
methods for bioenergy systems. Energy. 42(1): 146–156.
Seiford, L. M. 1996. Data envelopment analysis: The evolution of the State of the art
(1978–1995). The Journal of Productivity Analysis. 7: 99–137.
Shi, Y. and M. Zeleny. (Eds.). 2000. New Frontiers of Decision Making for the Information
Technology Era. World Scientific Publishing Co. Pte. Ltd., Singapore.
Shim, J. P. 1989. Bibliographical research on the Analytic Hierarchy Process (AHP).
Socio Economic Planning Sciences. 23(3): 161–167.
Shin, W. S. and A. Ravindran. 1991. Interactive multi objective optimization: Survey
I- Continuous Case. Computers and Operations Research. 18: 97–114.
Smithson, M. 1987. Fuzzy Set Analysis for Behavioral and Social Sciences. Springer,
New York.
Multi-Criteria Decision Making 65

Soltani, A., Hewage, K., Reza, B., and Sadiq, R. 2015. Multiple stakeholders in multi-
criteria decision-making in the context of municipal solid waste management:
A review. Waste Management. 35: 318–328.
Starr, M. K. and M. Zeleny. 1977. MCDM—State and future of the arts, In: M. K.
Starr and M. Zeleny (Eds.), 5–29. Multiple Criteria Decision Making: Studies
in the Management Sciences, Vol. 6. North-Holland Publishing Company,
Amsterdam.
Steuer, R. E. 1986. Multiple Criteria Optimization: Theory, Computation and Application.
John Wiley & Sons, New York.
Steuer, R. E., L. R. Gardiner, and J. Gray. 1996. A bibliographic survey of the activities
and international nature of multiple criteria decision making. Journal of Multi-
Criteria Decision Analysis. 5: 195–217.
Stewart, T. J. 1992. A critical survey on the status of multiple criteria decision mak-
ing theory and practice. Omega—The International Journal of Management Science.
20(5/6): 569–586.
Stewart, T. J. 1994. Data envelopment analysis and multiple criteria decision making:
A response. Omega. 22(2): 205–206.
Tamiz, M., D. F. Jones, and E. Eldarzi. 1995. A review of goal programming and its
applications. Annals of Operations Research. 58: 39–53.
Thanassoulis, E. and R. G. Dyson. 1992. Estimating preferred target input-output
levels using data envelopment analysis. European Journal of Operational Research.
56: 80–97.
Triantaphyllou, E. 2013. Multi-Criteria Decision Making Methods: A Comparative Study,
Vol. 44. Springer Science & Business Media, Dordrecht, Netherlands.
Urli, B. and R. Nadeau. 1999. Evolution of multi-criteria analysis: A scientometric
analysis. Journal of Multi-Criteria Decision Analysis. 8: 31–43.
Vansnick, J. C. 1986. On the problems of weights in MCDM: The noncompensatory
approach. European Journal of Operational Research. 24: 288–294.
Vargas, L. G. 1990. An overview of the analytic hierarchy process and its applica-
tions. European Journal of Operational Research. 48: 2–8.
Vincke, P. 1999. Outranking approach, In: T. Gal, T. J. Stewart, and T. Hanne (Eds.),
Multicriteria Decision Making: Advances in MCDM Models, Algorithms, Theory, and
Applications, Kluwer Academic Publishers, Boston, pp. 11.1–11.29.
von Winterfeldt, D. and Edwards, W. 1986. Decision Analysis and Behavioral Research.
Cambridge University Press, Cambridge.
Wang, J. J., Y. Y. Jing, C. F. Zhang, and J. H. Zhao. 2009. Review on multi-criteria
decision analysis aid in sustainable energy decision-making. Renewable and
Sustainable Energy Reviews. 13(9): 2263–2278.
White, D. J. 1990. A bibliography on the applications of mathematical programming
multiple-objective methods. Journal of the Operational Research Society. 41(8):
669–691.
Wierzbicki, A. P. 1998. Reference Point Methods in Vector Optimization and Decision
Support. Interim Report IR-98-017, International Institute for Applied Systems
Analysis, Laxenburg, Austria.
Wierzbicki, A. P. 1999. Reference point approaches, In: T. Gal, T. J. Stewart, and
T. Hanne (Eds.), Multicriteria Decision Making: Advances in MCDM Models,
Algorithms, Theory, and Applications, Kluwer Academic Publishers, Boston,
pp. 9.1–9.39.
66 Big Data Analytics Using Multiple Criteria Decision-Making Models

Yeh, Q. 1996. The application of data envelopment analysis in conjunction with finan-
cial ratios for bank performance evaluation. Journal of the Operational Research
Society. 47: 980–988.
Zadeh, L. 1965. Fuzzy sets. Information and Control. 8: 338–353.
Zanakis, S. H., A. Solomon, N. Wishart, and S. Dublish. 1998. Multi-attribute deci-
sion making: A simulation comparison of select methods. European Journal of
Operational Research. 107: 507–529.
Zeleny, M. 1980. The pros and cons of goal programming. Computers & Operations
Research. 8: 357–359.
Zeleny, M. 1982. Multiple Criteria Decision Making (McGraw-Hill Series In Quantitative
Methods For Management). McGraw-Hill Book Co., New York.
Zhao, L., Y. Zha, K. Wei, and L. Liang. 2016. A target-based method for energy saving
and carbon emissions reduction in China based on environmental data envel-
opment analysis. Annals of Operations Research. (in press).
Zimmermann, H. J. 1983. Using fuzzy sets in operational research. European Journal
of Operational Research. 13: 201–216.
Zionts, S. 1992. Some thoughts on research in multiple criteria decision making.
Computers & Operations Research. 19(7): 567–570.
Zionts, S. 2000. Some thoughts about multiple criteria decision making for ordinary
decisions, In: Yong, S. and M. Zeleny (Eds.), New Frontiers of Decision Making
for the Information Technology Era, World Scientific Publishing Co. Pte. Ltd.,
Singapore, pp. 17–28.
Zoupounidis, C. and A. I. Dimitras. 1998. Multicriteria Decision Aid Methods for the
Prediction of Business Failure. Kluwer Academic Publishers, Boston.
3
Basics of Analytics and Big Data

U. Dinesh Kumar, Manaranjan Pradhan, and

Ramakrishnan Ramanathan

CONTENTS
3.1 Introduction................................................................................................... 67
3.2 Analytics........................................................................................................ 69
3.2.1 Descriptive Analytics....................................................................... 71
3.2.2 Predictive Analytics......................................................................... 72
3.2.3 Prescriptive Analytics...................................................................... 73
3.3 Big Data: Volume, Variety, Velocity, and Veracity.................................... 74
3.4 Limitations of Traditional Technologies for Big Data.............................. 75
3.5 Analytics Life Cycle...................................................................................... 76
3.5.1 Data Capture, Store, Prepare, Analyze, and Share...................... 76
3.5.2 NoSql Systems................................................................................... 78
3.5.3 Data Store: Distributed File Systems.............................................. 79
3.6 Prepare and Analyze in Big Data...............................................................80
3.6.1 MapReduce Paradigm......................................................................80
3.6.2 Hadoop Ecosystem...........................................................................80
3.7 Big Data Analytics, Multicriteria Decision Making, and BI................... 82
3.8 Conclusions....................................................................................................84
Bibliography...........................................................................................................84

3.1 Introduction
Business analytics (BA) and big data have become essential components
that every organization should possess to compete effectively in the market.
Hopkins et al. (2010) claimed that analytics sophistication is one of the primary
differentiators between high-performing and low-performing organizations.
BA is a set of statistical, mathematical and machine-learning management tools,
and processes used for analyzing the past data, to understand hidden trends,
which can assist in problem solving and/or drive fact-based decision making
in an organization. In the 1980s, many organizations did not collect data or the
data was not in an appropriate form to derive insights. Organizations at that
point in time found decision making and/or problem solving arduous due to
the nonavailability of data; with the advent of enterprise resource planning

67
68 Big Data Analytics Using Multiple Criteria Decision-Making Models

(ERP) systems, most of the organizations have ensured the ­availability of data,
which could be called upon whenever needed. However, for effective and effi-
cient problem solving and ­decision making thereby, the data stored within the
ERP systems needed to be a­ nalyzed, and this gave birth to the use of analytics.
Analytics can be grouped into three categories: descriptive analytics, pre-
dictive analytics, and prescriptive analytics. Descriptive analytics deals with
describing past data using descriptive statistics and data visualization; useful
insights may be derived using descriptive analytics. Predictive analytics aims
to predict future events such as demand for a product/service, customer churn,
and loan default. Prescriptive analytics on the other hand provides an optimal
solution to a given problem or offers the best alternative among several alter-
natives. In other words, descriptive analytics captures what happened, pre-
dictive analytics predicts what is likely to happen, and prescriptive analytics
provides the best alternative to solve a problem. Although all three compo-
nents of analytics are important, the value-add and the usage of different ana-
lytics components are shown in Figure 3.1. For all the hype around analytics,
vast majority of organizations use descriptive analytics in the form of business
intelligence (BI). Significantly, a smaller group of organizations use predictive
analytics, mainly for forecasting; the number of organizations using prescrip-
tive analytics is minimal at this point in time in comparison with descriptive
and predictive analytics. However, it is interesting to note that the value-add
to a company increases many fold if organizations were to use predictive and
prescriptive analytics conjointly as compared to descriptive analytics alone.
Today, with the ever-growing use of the Internet, social media platforms,
smartphones, and Internet of things (IOTs), the amount of data that gets

Prescriptive
analytics
Value-add to organization

A very small proportion of

companies use predictive analytics
Predictive
analytics

A few companies use predictive

analytics
Descriptive
analytics Many companies use descriptive
analytics in the form of business
intelligence

Types of analytics solution

FIGURE 3.1
Types of analytics solution and the value-add.
Basics of Analytics and Big Data 69

generated everyday has increased several thousand fold over the past few
years. An estimate by www.vcloudnews.com claims that 2.5 Exabytes of data
gets generated every day and it will increase exponentially in the future;
and all these data provide greater advantage to enterprises or organizations
around the world that look to leverage these data to get diversified insights.
Enterprises/organizations today can better understand their dynamic busi-
ness environments, their customer’s behavior and preferences, predict accu-
rate market trends, weather forecasts, and thereby optimize resources at
granular levels to increase efficiency to an extent they never believed was
possible earlier.
One such wonderful case study was Google’s flu trends, which predicts flu
trends in real time even before public organizations like Centre for Disease
Control (CDC) know it. Google built a system that extracted search terms,
their search frequencies by regions and time from several billions of histori-
cal search requests it received over few years, and correlated with the actual
incident reported by CDC; and after correlating millions of search terms with
reported incidents, it found about 45 search terms which are highly corre-
lated. Since 2009, Google has been using those very search term frequencies
in real time to predict flu trends.
Unlike in the past, enterprises today do not only use transactional data for
getting insights into business and customers but also use other sources of data
like weblogs or clickstreams, social feeds, emails to get deeper understanding
of customers. For example, e-commerce companies (which have mushroomed
in recent times) do not wait until a new customer makes some purchase to
understand his or her preferences, but know that from browsing patterns (i.e.,
by checking on the different links that the customer may have visited or spent
time on). Thus, by this example one can possibly gage the data size that would
be generated every day for these e-commerce sites. To illustrate this further and
to put numbers, let us consider Amazon, which has about 188 million visitors
(Anon, 2016) to its site every month, and it stores information about every single
link they click, their wish lists, and purchases. With this humongous amount of
data, it is a real challenge for e-commerce companies first to capture, store, and
finally analyze the data for them to gain insights into their customers, under-
stand their preferences, and thereby make purchase recommendations.
In this chapter, we will be discussing in detail various aspects of analytics
and big data with few examples of real-life applications. Finally, we end with
an example of how analytics is used in multicriteria decision making.

3.2 Analytics
The primary objective of analytics is enabling to take informed decisions as
well as solve business problems. Organizations would like to understand the
70 Big Data Analytics Using Multiple Criteria Decision-Making Models

association between the key performance indicators (KPIs) and factors that
have significant impact on the KPIs for effective management. Knowledge
of relationship between KPIs and factors would then provide the decision
maker with appropriate actionable items. Analytics thus is a knowledge
repository consisting of statistical and mathematical tools, machine-­learning
algorithms, data management processes such as data extraction, transfor-
mation, and loading (ETL) and computing technologies such as Hadoop
that create value by developing actionable items from data. Devonport and
Harris (2009) reported that there was a high correlation between the use of
analytics and business performance. They reported that a majority of high
performers (measured in terms of profit, shareholder return, revenue, etc.)
strategically apply analytics in their daily operations as compared to low
performers.
The theory of bounded rationality proposed by Herbert Simon (1972) is
becoming very evident in the current context of managing organizations
and competing in the market. The increasing complexity of business prob-
lems, existence of several alternative solutions, and the paucity of time avail-
able for decision making demand a highly structured decision-making
process using past data for the effective management of organizations. There
are several reasons for the existence of bounded rationality such as uncer-
tainty, incomplete information about alternatives, and lack of knowledge
about cause and effect relationships between parameters of importance.
Although decisions are occasionally made using the “highest paid person’s
opinion” (HiPPO) algorithm especially in a group decision-making scenario
and Flipism (all decisions are made by flipping a coin), there is a significant
change in the form of “data-driven decision-making” among several compa-
nies. Many companies use analytics as competitive strategy and many more
are likely to use this in the near future, and here is why; a typical data-driven
decision-making process uses the steps as shown in Figure 3.2.

Stage 1 Stage 3
Stage 2
Identify problems/ Process the data for
Identify sources of data
improvement missing/incorrect data
required for problem
opportunities and prepare for analytics
identified in stage 1
model building

Stage 4
Stage 5 Stage 6
Build the best analytical
Communicate the Implement solution/
model and validate the
analytics output decision
model

FIGURE 3.2
Data-driven decision-making flow diagram.
Basics of Analytics and Big Data 71

1. Identify the problem or opportunity for value creation.

2. Identify sources of data (primary as well as secondary data).
3. Preprocess the data for missing and incorrect data. Prepare the data
for analytics model building, if necessary transform the data.
4. Build the analytical models and identify the best model using model
validation.
5. Communicate the data analysis output and decisions effectively.
6. Implement solution/decision.

BA is a set of statistical, mathematical management tools and processes

used for analyzing the past data that can assist in problem solving and/or
drive fact-based decision making in an organization. Increasing complexi-
ties associated with businesses demand an unparalleled understanding of
the customer expectations in order to serve better. According to the theory
of bounded rationality proposed by Herbert Simon (1972), the human mind
lacks the ability to choose and make the right decisions due to the complex-
ity of problems that organizations face and the limited time available for
decision making. In the 1980s, the culture of data collection was poor, several
organizations did not collect data or the data was not in a form, which could
be used for deriving insights, and this in turn resulted in organizations find-
ing it difficult especially for prompt decision making. With the introduction
of ERP systems, organizations today are ensured of the data availability that
can be called upon whenever needed. However, the data sitting in the ERP
systems need to be analyzed for problem solving and decision making; this
need is now met using analytics. BA helps organizations to derive BI that
helps organizations to manage the data-to-decision cycle. BA can be grouped
into three types: descriptive analytics, predictive analytics, and prescriptive
analytics. In the following sections, we discuss the three types of analytics
in detail.

3.2.1  Descriptive Analytics

Descriptive analytics is the simplest form of analytics that mainly uses sim-
ple descriptive statistics, data visualization techniques, and business-related
queries to understand the past data. As a utility tool, it captures the past and
present data and is not able to predict the future, which is generally done by
predictive analytics. Descriptive analytics is also known to be a power tool
to communicate hidden facts and trends in data. The following are a few
examples of descriptive analytics:

1. Most shoppers turn toward their right when they enter a retail store.
2. Men who kiss their wife before going to work earn more and live
longer than those who do not.
72 Big Data Analytics Using Multiple Criteria Decision-Making Models

3. Many marriages/relationships break up in January (known as rela-

tionship freeze).
4. Conversion rate among female shoppers is higher than male shop-
pers in consumer durable shops.

Simple descriptive statistics, charts (data visualization), or query can be

used to extract such information, and at the same time it also lends very use-
ful insights for businesses. For example, retailers keep products with higher
profit on the right side of the store since most people turn right. There are
many such strategies (or in the analytics nomenclature, “actionable items,”
used by decision makers).
Edward Tufte (2001), in his book titled The Visual Display of Quantitative
Information, demonstrated how innovative visuals can be used to commu-
nicate data effectively. For businesses today, dashboards form the core of
their BI and are an important element of analytics. Indian companies such
as Gramener* have used innovative data visualization tools to communicate
hidden facts in the data. Descriptive analytics may also be understood as the
initial stages of creating the analytics capability; and descriptive statistics
could help Small and Medium Enterprises (SME) uncover inefficiencies and
thereby eliminate and/or minimize them.
Simple data analysis can lead to business practices that result in financial
rewards. For instance, companies such as RadioShack and Best Buy found
a high correlation between the success of individual stores and the num-
ber of female employees in the sales team (Underhill, 2009). Underhill (2009)
also reported that the conversion rate (percentage of people who purchased
something) in consumer durable shops was higher among female shoppers
than among male shoppers. Sometimes, all you need is a simple query, which
could even lead to fraud detection. Consider the following example: recently,
China Eastern Airline found a man who had booked a first class ticket more
than 300 times within a year and cancelled it just before its expiry for full
refund so that he could eat free food at the airport’s VIP lounge.† (It is sur-
prising that the airline took so long to uncover this!)

3.2.2  Predictive Analytics

In the analytics capability maturity model (ACMM), predictive analytics
comes after descriptive analytics and is the most important analytics capa-
bility that aims to predict the probability of occurrence of a future event
such as customer churn, loan defaults, and stock market fluctuations. While
descriptive analytics is used for finding what has happened in the past,
predictive analytics is used for what is likely to happen in the future. The

* Source: https://gramener.com/
† Source: http://articles.timesofindia.indiatimes.com/2014-01-30/mad-mad-world/46827501_​

1_free-meals-single-ticket-issued-ticket
Basics of Analytics and Big Data 73

ability to predict a future event such as an economic slowdown, a sudden

surge or decline in a commodity’s price, which customer is likely to churn,
what will be the total claims from auto insurance customer, how long a
patient is likely to stay in the hospital, and so on will help organizations to
plan their future course of action. Anecdotal evidence suggests that predic-
tive analytics is the most frequently used type of analytics across several
industries. The reason for this is that almost every organization would like
to forecast demand for the products that they sell, the price of the material
used by them, and so on. Irrespective of the type of business, organizations
would like to forecast the demand for their products or services and under-
stand the causes of demand fluctuations. For example, when Hurricane
Charley struck the United States in 2004, Linda M. Dillman, Walmart’s
Chief Information Officer, wanted to understand the purchase behavior of
its customers (Hays, 2004). Using data-mining techniques, Walmart found
that the demand for strawberry pop-tarts went up over seven times during
the hurricane compared to its normal sales rate; the prehurricane top-sell-
ing item was found to be beer. These insights were used by Walmart when
the next hurricane—Hurricane Frances—hit the United States in August–
September 2004; most of the items predicted by Walmart sold quickly.
Although the high prehurricane demand for beer can be intuitively pre-
dicted, the demand for strawberry pop-tarts was a complete surprise. The
use of analytics can reveal relationships that were previously unknown and
are not intuitive.
The most popular example of the application of predictive analytics is
Target’s pregnancy prediction model. In 2002, Target hired statistician
Andrew Pole; one of his assignments was to predict whether a customer is
pregnant (Duhigg, 2012a). New parents are the holy grail of marketers since
they are price-insensitive customers who would like to buy the best things
for their new born. In 2010, it was reported that parents spent about USD
6800 on average on a child before his/her first birthday; the North American
new baby market was worth USD 36.3 billion (Duhigg, 2012b). At the outset,
the assignment put forward by the marketing department to Pole may look
bizarre, but it made great business sense.

3.2.3  Prescriptive Analytics

Prescriptive analytics is the highest level of capability of analytics today,
wherein firms decide what to do once they gain insights through descriptive
and predictive analytics. Prescriptive analytics assists users in finding the
optimal solution to a problem or in making the right choice/decision from
among several alternatives. Unlike predictive analytics (which, in many
cases, provides the probability of a future event), prescriptive analytics in
most cases provides an optimal solution/decision to a problem. Operations
research (OR) models form the core of prescriptive analytics. Ever since their
introduction during World War II, OR models have been used in every sector
74 Big Data Analytics Using Multiple Criteria Decision-Making Models

and in every industry. The potent prescriptive analytics “tools” comprising

of several applications have been widely used, and several companies across
the world have benefitted from them.
Coca-Cola Enterprises (CCE) is the largest distributor of Coca-Cola
products. In 2005, CCE distributed 2 billion physical cases containing
42 billion bottles and cans of Coca-Cola in the United States (Kant et al.,
2008). CCE developed an OR model that would meet several objectives
such as improved customer satisfaction and optimal asset utilization for
its distribution network of Coca-Cola products from 430 distribution cen-
ters to 2.4 million retail outlets. The optimization model resulted in cost
savings of USD 54 million and improved customer satisfaction. A simi-
lar distribution network problem (vehicle routing) was solved by the IIM
Bangalore team for Akshaya Patra. The Akshaya Patra Midday Meal Routing
and Transportation Algorithm (AMRUTA) was developed to solve the vehicle
routing problem; this was implemented at Akshaya Patra’s Vasanthapura
campus, resulting in savings of USD 75,000 per annum (Mahadevan et al.,
2013). A major challenge for any e-commerce company is to improve the
conversion of visits to transactions and order sizes. Hewlett Packard
(HP) established HPDirect.com in 2005 to build online sales. HP Global
Analytics developed predictive and prescriptive analytics techniques to
improve sales. The analytical solutions helped HP to increase conversion
rates and order sizes (Rohit et al., 2013).
Inventory management is one of the teething problems that are most
frequently addressed using prescriptive analytics. Samsung implemented
a set of methodologies under the title “Short Life and low Inventory in
Manufacturing” (SLIM) to manage all the manufacturing and supply
chain problems. Between 1996 and 1999, Samsung implemented SLIM in
all its manufacturing facilities, resulting in a reduction in the manufactur-
ing cycle time of random access memory devices from more than 80 days
to less than 30 days. SLIM enabled Samsung to capture additional markets
worth USD 1 billion (Leachman et al., 2002). In the next section, we will
discuss the more frequently used predictive and prescriptive analytics
tools.

3.3  Big Data: Volume, Variety, Velocity, and Veracity

Big data has four main characteristics: volume, variety, velocity, and veracity.
Volume specifies the need to deal with large amount of data; but how large
is large enough to be called big data? One definition is the data is so large
that it cannot be stored and processed in any of the traditional platforms that
enterprises were using so far. Thus, it is related to the existing technology
Basics of Analytics and Big Data 75

and its ability to store and process the data. There is a need for an alternate
solution or platform that has the margin to scale in order to accommodate
and process the exponentially increasing size of data.
Variety in big data refers to different types of data; all the data that are
captured nowadays cannot be arranged into rows and columns, or in other
words, not all data are structured. For example, data such as texts, images,
and machine-generated data in its original form cannot be arranged in the
traditional rows and columns. Such unstructured data can be in any form
such as XML, Json, free text, images, audios, or videos, and so on and there
is a need to analyze these data to get insights; for example, the sentiments
expressed by customers on a product or service provided by a company are
very important for improving the product and service overall.
Velocity is the rate of data growth. We have seen a little earlier that data
is growing exponentially, and thus it is imperative to analyze data faster,
almost in real time. As data starts to become backdated, their value dimin-
ishes hampering organizations to study and analyze data, which would have
enabled them to improve service delivery and decision-making. Veracity
refers to the quality and reliability of the data. Though we can leverage big
data to get new insights, the real challenge lies in how to capture, store, and
process these ever-increasing data size. Can the traditional platforms help or
we need to take a fresh approach to deal with big data?

3.4  Limitations of Traditional Technologies for Big Data

Most of the traditional IT platforms were built to be deployed on sin-
gle systems, and the amount of resources available on single systems
is always limited. The default approach to add more compute and stor-
age to a single system is called vertical scalability or scale-up approach.
However, this approach is not highly scalable as beyond a certain point,
no more resources can be added to a single system. The other approach
is to keep adding more systems: this is an infinitely scalable approach
and is called horizontal scalability or scale-out approach; but most of the
software frameworks or platforms are not built to leverage this approach,
mostly due to an increase in complexity at the software level. In a scale-out
approach, the software typically needs to deal with more failure points
due to the presence of multiple systems and complex coordination mecha-
nisms thereof. Thus, a number of traditional softwares were developed to
be a single system solution or scale-out in limited capacity, and therefore
most of these systems cannot accommodate the sheer size of the data that
need to be handled today.
76 Big Data Analytics Using Multiple Criteria Decision-Making Models

3.5  Analytics Life Cycle

Typically in analytics, the data follows a life cycle that has multiple stages,
such as:

• Data capture
• Data store
• Prepare
• Analyze
• Share

3.5.1  Data Capture, Store, Prepare, Analyze, and Share

Data Capture is the stage where data is received from multiple sources such
as transactional databases, customer surveys, financial reports, weblogs, mes-
sage queues, web services, or social feeds. It is to be noted here that at times the
incoming data rate could be very high, which in effect could mean millions
of messages in a second or minutes or billions of records per day. Let us look
at an example: an energy analytics provider in North America called OPower
receives about 200 readings from a single customer household every day and
it has about 25 million household customers, which effectively translates to 5
billion records per day (25 million × 20)—the challenge quite understandably
is to design a system that would capture those records within a single system.
This is the big data challenge at the capture stage itself, as there are hardly
any traditional platforms that can look to store several billion writes per day.
However, the primary problem is not just capturing the large volume of
data in this case. Enterprises want the data to be stored for months and years
altogether for future analysis. If the data received in a day is about 2 or 3 Tera
bytes (TBs), by end of the year it becomes almost 1 Peta byte (PB). Therefore,
in order to handle such huge data, enterprises need a platform that has
elasticity and can scale as the data grows to store it for a longer period of
time,  thereby giving a vital leverage to build complex analytical systems.
This is the big data problem at the data store stage.
Once the data is stored, the next herculean task is to navigate through
these data sets and extract only the relevant information for analytics. In
most cases, the captured data are not fit for analysis; lot of cleaning, aggrega-
tions, filtering, and data munging tasks need to be applied on the raw data
in order to make it ready for analytics. Not to forget the series of algorithms,
which has to scan 100 TBs and PBs of data, needing super computational
capacity; this is the data prepare stage.
Analyze is a stage where the prepared data sets are used for applying algo-
rithms like Structured Query Language (SQL) queries, statistical techniques,
or machine-learning algorithms. In the final stage, that is, share, the analytics
Basics of Analytics and Big Data 77

output is shared for visualization or fed to another system for usage. There
could be a big data problem in any one or more of the above stages. We will
discuss each stage in detail in later sections. The analytics life cycle is shown
in Figure 3.3.
Data from multiple sources come through various channels such as files,
web services, message queues, plain Transmission Control Protocol (TCP)
sockets, or Rich Site Summary (RSS) feeds. Data does also flow at different
rates, like a few dumps every hour or continuous streaming messages every
second like social feeds or transactional records or messages from devices
or online transactional processing (OLTP) systems. If the data inflow rate is
very high, then the systems which capture these data are in all likelihood to
run out of space and in the end may collapse; thus newer systems need to be
designed and ready to be used as a queue or buffer between source systems
and data capture systems. And finally, the data capture systems need to be
scaled out to support high volume writes. The entire architecture of data
capture in atypical big data environment is shown in Figure 3.4.
In summary, data capture poses three challenges:

• The ability to integrate multiple sources

• The ability to match the arrival rate of messages and capture rate
• The ability to capture/write data faster (millions of writes per second)

Capture Store Prepare Analyze Share

FIGURE 3.3
Analytics life cycle.

Transactions

Logs
• Integrates all
sources
• Provides a buffer
or queue
• Transforms or
enriches data
Large-scale distributed systems
optimized for high volume of reads
and writes

FIGURE 3.4
Data capture in big data environment.
78 Big Data Analytics Using Multiple Criteria Decision-Making Models

There are open-source tools like Apache Flume which acts as an intermediate
system listening to multiple sources and aggregates, transforms or enriches,
and writes to systems capturing the data. Tools like Apache Kafka works as a
queuing system between the source and destination with different rates of
arrival and capture of data, and has the ability to scale out to a great extent to
support millions of messages flowing through the system per second. Most
of the Not only SQL (NoSql) Systems are designed to read and write at very
high volume (millions of reads and writes per second) by scaling out systems.

3.5.2  NoSql Systems

Traditional relational database management systems (RDBMS) enforce
schema very strictly. All data that are captured need to conform to the
schema. Any minor changes to schema will end up altering all existing
records and prove to be very expensive. In today’s world, where data model
keeps evolving and changing very rapidly, using this approach is very hard
to manage. All we need is a system that would allow schema-less design and
allow defining schema dynamically where data is inserted. This in turn will
allow records to follow an overall schema, but is by and large free to add or
remove specific attributes or elements to it.
Another challenge with RDBMS is that they are not very conducive for stor-
ing real-world object which are more hierarchical in nature; RDBMS tend to
be more flat in structure and need an OR (Object to Relational) mapping for
storing data. This approach is expensive as OR mapping divides the objects
and stores into multiple tables. This makes all reads and writes of the objects
very costly and slow and cannot perform high volume reads and writes.
Finally, most of the RDBMS are not highly scalable systems. If we have to
store billions of records, we will need a system that can scale out to 100 s and
1000 s of servers. Traditional RDBMS systems do not support this architecture.
So, NoSql systems were developed to support these three features.

• Support schema-less design

• Support storing real-world object forms (deformalized forms)
• Highly scalable

If a table is very large and cannot be stored in a single system, NoSql system
splits the table into multiple shards and distributes these shards across sev-
eral servers. Now each server dealing with a set of shards manages all reads
and writes for those shards. By splitting the tables appropriately into sev-
eral hundred shards, we can load balance high volume incoming reads and
writes to different servers and hence can support large-scale data ingestions.
There are four types of NoSQL databases:

• Key-value store: These databases are designed for storing data in a key-
value fashion like a map. Each record within consists of an indexed
key and a value. Examples: DyanmoDB, Reddis, Riak, BerkeleyDB.
Basics of Analytics and Big Data 79

• Document database: Store data in key-value fashion where values

are stored as “documents,” which are designed to store complex
structure or objects. Each record is a document and assigned
a unique key, which is used to retrieve the document. A docu-
ment can be a JavaScript Object Notation (JSON, a lightweight
data-interchange format) messages; the document elements can
be indexed for advanced searches. Examples: MongoDB and
CouchDB.
• Column family store: Store records with grouping-related columns as
column families. Each record needs to have a set of column families,
but can have different column attributes in their respective column
families. These stores offer very high performance and a highly
­scalable architecture. Examples: HBase and HyperTable.
• Graph database: Based on graph theory, these databases are designed
for data whose relations are well represented as a graph with edges
and nodes.

3.5.3  Data Store: Distributed File Systems

Traditional file systems have failed to scale out sufficiently in order to store
the large influx of data in recent times, and therefore new distributed file
systems have been developed; some among them are:

• Google File System (GFS)

• GlusterFS
• Hadoop-Distributed File System (HDFS)
• OneFS
• XtremeFS

Among these, HDFS has been very popular as it is a part of Hadoop

Distribution, which has become a default big data platform for many enter-
prises. We will explain the HDFS architecture at a high level in this sec-
tion; other distributed file systems are pretty much similar in architecture
(Figure 3.5).
HDFS is an abstract level file system which splits the files and stores
across a cluster of machines called datanodes and maintains the metadata
in a master machine called namenode. The metadata provides informa-
tion about which file is split into how many chunks and stores in what all
datanodes, and it facilitates read and write operations by redirecting clients
to the appropriate datanodes. After splitting a file into multiple chunks, each
chunk is replicated multiple times and stored in different datanodes so that
it can withstand datanodes failures. As metadata is critical for all operations
on the file systems, it is a stand by for the master node (namenode), in case
the primary or active namenode fails.
80 Big Data Analytics Using Multiple Criteria Decision-Making Models

Namenode Namenode
(active) (stand by)

HA using shared storage/NFS

Datanode Datanode Datanode

1 2 3 1 3 2 2 3 1

Blocks

FIGURE 3.5
Hadoop-distributed system.

As data grows, more datanodes can be added to the cluster, thereby

increasing its storage capacity. Theoretically it can grow to infinite scale. The
largest deployed HDFS cluster in an enterprise today, at the time of this book
writing, stores around 512 PBs of data.

3.6  Prepare and Analyze in Big Data

3.6.1  MapReduce Paradigm
After storing large volume of data, the next challenge is to process and ana-
lyze them. Conventional way of reading data from filesystem and presenting
to a program or process cannot work anymore as the data is large and stored
across multiple systems. So, now the program or process needs to be moved
to the machines where data is stored. So, the programming paradigm of data
to the process needs to change to process to data.
MapReduce programming paradigm as shown in Figure 3.6 provides
the above feature of distributing the algorithms to process each split in
datanodes and then provides a mechanism to distribute the intermediate
outputs to different reduce nodes for final consolidation. The real challenge
is to decompose every algorithm to map and reduce stages and a mechanism
to distribute the data between maps and reduce stages.

3.6.2  Hadoop Ecosystem

Apache Hadoop is one of the most widely adopted platforms for storing
and processing big data. The Hadoop framework has three core layers or
Basics of Analytics and Big Data 81

Input Shuffle and sort Output

The, 1
Quick, 1
Brown, 1
The quick
Map Fox, 1
Brown fox The, (1,1,1) Brown, 2
The, 1 Brown, (1,1)
Fox, 1 Fox, (1,1) How, 1
The, 1 Mouse, (1)
Ate, 1 Now, 1
Mouse, 1 The, 3
The fox ate Reduce
Map
The mouse Ate, 1
How, 1 Cow, 1
Now, 1 Fox, 2
Brown, 1
Cow, 1 Mouse, 1
How now Quick, 1
Map
Brown cow

FIGURE 3.6
MapReduce approach to a simple word count task.

components; the bottom-most component is the distributed file system (as

explained in the previous section), which is designed to scale out and store
TBs and PBs of data. The next layer is Yet Another Resource Negotiator
(YARN), which is responsible for keeping track of how many systems there
are in a cluster, what resources are being currently consumed, what is
remaining, and how to schedule analytical task on different systems as well
as load balance the tasks. The final layer is MapReduce, which enables devel-
oping and executing tasks in parallel on the cluster nodes.
Hadoop is available as part of Apache Open Source license and being
developed by a very large community of developers who work for compa-
nies such as Yahoo, Facebook, Twitter, LinkedIn, and so on. The software is
free and designed to run on commodity hardware, which also brings down
the cost of the implementation. One of the largest clusters is deployed by
Yahoo and runs 40,000 nodes and stores, and processes around 512 PBs of
data.
The Hadoop ecosystem as shown in Figure 3.7 also provides a few more
vital components like Hive, Pig, and Mahout. As Hadoop is primarily devel-
oped in Java, the map-reducing algorithms need to be developed in Java,
which in itself is a constraint as most data analysts are not familiar with
programming languages. So, the Hadoop Ecosystem has developed a few
abstractions like Hive, which supports SQL syntax or Pig, which supports
a scripting interface for analyzing structure as well as unstructured data
like log files or social feeds. Mahout is a dedicated library for machine-
learning algorithms to be applied on big data. It supports algorithms such as
Regression, Classification, Clustering, and Collaborative filtering.
82 Big Data Analytics Using Multiple Criteria Decision-Making Models

Hive Pig Latin Mahout

DW system data analysis machine learning

Map reduce framework

YARN (cluster resource management)

HDFS 2.0—HA and federation

Sqoop
Flume/
Kafka
Import Import or export

Streaming data like logs or messages, and so on Structured data

FIGURE 3.7
Hadoop and its ecosystem.

After data is analyzed and an insight is derived, these can be integrated

with BI tools to visualize—like create graphs, charts, and so on. Or in some
cases, big data platforms can be used to build analytical models, which can
be then fed into real-time systems to make predictions in real time.

3.7 Big Data Analytics, Multicriteria Decision Making,

and BI
One of the primary objectives of analytics and big data is to assist organiza-
tions with decision making. Organizations are engaged in collecting big data
and analyzing the data using suitable analytics methods because they expect
to generate useful business insights by doing so. As highlighted in Section
3.2, a number of descriptive, predictive, and prescriptive analytic tools are
available in the literature. Mathematical modelling tools are a subset of these
analytic tools.
MCDM methods presented in the previous chapter are a class of math-
ematical modelling tools when decisions have to be made in the presence
of multiple criteria and sufficient data for developing such models can be
generated. Many decision-making problems encountered by organizations
have multiple criteria, and an optimal decision has to be arrived considering
Basics of Analytics and Big Data 83

all criteria. One such example is performance-based contracts (PBC) that are
becoming very popular among capital equipment users especially in the
defense and the aerospace industry. In PBC, the customers demand perfor-
mance as measured through multiple criteria such as reliability, availability,
total cost of ownership, and logistic foot print.
Predicting reliability and availability involves collecting historical failure
and maintenance data and finding the probability distribution of time to
failure and time to maintain distribution function. Estimating cost of owner-
ship will involve breaking down the total cost into different cost components
and predicting the future costs (such as operation and maintenance costs)
throughout the life of the systems. The original equipment manufacturer
will have to use prescriptive analytics techniques optimize the various KPIs
to provide equipment under PBC.
More examples are available in the remaining chapters of this book. The
role of MCDM in big data is illustrated in Figure 3.8. Big data is important
because the modern digital economy generates huge volumes of data every
day and this data can be used to understand customers and ultimately help
businesses. It has to be noted that data as such will not be sufficient for mak-
ing business decisions. There are a number of analytic tools and MCDM
models are one class of the whole range of analytics tools. However, big data
and MCDM models (and other analytics tools) will not be sufficient; one

Big data

MCDM tools

Business analytics tools

Analytics reports Human interpretation

Business intelligence

FIGURE 3.8
Big data analytics, MCDM, and business intelligence.
84 Big Data Analytics Using Multiple Criteria Decision-Making Models

needs a shrewd business mind to make sense of the results from these ana-
lytics tools to generate BI insights (Ramanathan et al., 2012).

3.8 Conclusions
As enterprises try to understand their respective businesses more deeply to
constantly deliver value to their customers, data is going to be the core focus
in the time to come. As Angela Ahrendts, CEO of Burberry, says “whoever
unlocks the reams of data and uses it strategically will win” (Andrew Gill,
2013). Hence, enterprises will have to deal with challenges of managing and
analyzing big data. As technologies evolve over a period of time, enterprises
need to adopt and learn quickly to benefit from it and remain competitive.
Similarly, governments are also adopting digital platforms to administer and
deliver services and quality life to its citizens. One such example is u
­ niversal
identification service (UID) implementation by Government of India. It has
reached a billion users and promises to be the core platform to manage and
deliver services under various social service programs.

Bibliography
Anon 2012. 2.5 Quintillion bytes created each day. Storage Newsletter.com, http://
www.vcloudnews.com/every-day-big-data-statistics-2-5-quintillion-bytes-of-
data-created-daily/, accessed on April 6, 2016.
Anon 2016. Most popular websites in the United States as of September 2015 ranked by
visitors. http://www.statista.com/statistics/271450/monthly-unique-visitors-
to-us-retail-websites/, accessed on April 6, 2016.
Brody, H., Rip, M.R., Johansen, P.V., Paneth, N., and Rachman, S. 2000. Map mak-
ing and myth making in broad street: The London Cholera epidemic 1954.
The Lancet, Vol. 356, pp. 64–69.
Bruhadeeswaran, R. 2012. Shubham Housing Development Finance raises $7.8M
from Elevar, Helion, Others. Vccircle. www.vccircle.com/news/2012/11/12/
shubham-housing-development-finance-raises-78m-elevar-helion-others.
Coles, P.A., Lakhani, K.R., and Mcafee, A.P. 2007. Prediction markets at Google.
Harvard Business School Case (No 9-607-088).
Devonport, T.H. 2006. Competing on analytics. Harvard Business Review, January,
pp. 1–10.
Devonport, T.H. and Harris, J.G. 2009. Competing on Analytics—The New Science of
Winning, Harvard Business School Press, Boston, MA.
Devonport, T.H., Iansiti, M., and Serels, A. 2013. Managing with analytics at Proctor
and Gamble. Harvard Business School Case (Case Number 9-613-045).
Basics of Analytics and Big Data 85

Devonport, T.H. and Patil, D.J. 2012. Data scientist: The sexiest job of 21st century.
Harvard Business Review, 2–8.
Duhigg, C. 2012a. How companies learn your secret. New York Times, February 16,
2012.
Duhigg, C. 2012b. The Power of Habit: Why We Do What We Do in Life and Business.
William Heinemann, London.
Grill, A. 2013. IBM CEO Ginni Rometty believes big data and social will change
everything—How about other CEOs. Leadership, Social Business and Social Media.
http://londoncalling.co/2013/03/ibm-ceo-ginni-rometty-believes-big-data-
and-social-will-change-everything-how-about-other-ceos/
Hayes, B. 2013. First links in the Markov chain. American Scientist, Vol. 101,
pp. 92–97.
Hays, C.L. 2004. What Wal-Mart knows about customers’ habits. New York Times,
November 14, 2004.
Hopkins, M.S., LaValle, S., Balboni, F., Kruschwitz, N., and Shockley, R. 2010.
10 insights: A first look at the new intelligence enterprise survey on winning
with data. MIT Sloan Management Review, Vol. 52, pp. 21–31.
Howard, R. 2002. Comments on the origin and applications of Markov decision
­processes. Operations Research, Vol. 50, No. 1, pp. 100–102.
Kant, G., Jacks, M., and Aantjes, C. 2008. Coca-Cola enterprises optimizes vehicle
routes for efficient product delivery. Interfaces, Vol. 38, pp. 40–50.
Leachman, R.C., Kang, J., and Lin, V. 2002. SLIM: Short cycle time and low inventory
in manufacturing at Samsung Electronics. Interfaces, Vol. 32, pp. 61–77.
Lewis, M. 2003. Moneyball: The Art of Winning an Unfair Game. W W Norton &
Company, London, UK.
Mahadevan, B., Sivakumar, S., Dinesh Kumar, D., and Ganeshram, K. 2013.
Redesigning mid-day meal logistics for the Akshaya Patra foundation: OR at
work in feeding hungry school children. Interfaces, Vol. 43, No. 6, pp. 530–546.
Mayer-Schonberger, V. and Cukier, K. 2013. Big Data: A Revolution That Will Transform
How We Live, Work and Think. John Murray, New York.
Ramanathan, R., Duan, Y., Cao, G., and Philpott, E. 2012. Diffusion and impact of
business analytics in UK retail: Theoretical underpinnings and a qualita-
tive study. EurOMA Service Operations Management Forum, Cambridge,
September 19–20, 2012.
Rohit, T., Chakraborty, A., Srinivasan, G., Shroff, M., Abdullah, A., Shamasundar,
B., Sinha, R., Subramanian, S., Hill, D., and Dhore, P. 2013. Hewlett Packard:
Delivering Profitable Growth for HPDirect.Com Using Operations Research.
Interfaces, Vol. 43(1), pp. 41–61.
Savant, M.V. 1990. Ask Marilyn. Parade Magazine, p. 16, September 9, 1990.
Siegel, E. 2013. Predictive Analytics: The Power to Predict Who Will Click, Buy, Lie or Die.
John Wiley and Sons, Hoboken, NJ.
Simon, H. 1972. Theories of bounded rationality. In Decisions and Organizations.
McGuire, C.B., and Radner, R. (Eds.). North-Holland Publishing Company,
New York, pp. 161–176.
Sirkin, H.L., Keenan, P., and Jackson, A. 2005. The hard side of change management.
Harvard Business Review, Vol. 83, No. 10, pp. 109–118.
Snow, S.J. 1999. Death by Water: John Snow and cholera in the 19th century. Liverpool
Medical Institution. http://www.lmi.org.uk/Data/10/Docs/11/11Snow.pdf
86 Big Data Analytics Using Multiple Criteria Decision-Making Models

Stelzner, M.A. 2013. 2013 Social Media Marketing Industry Report—How Marketers Are
Using Social Media to Grow Their Businesses. Social Media Examiner Report 2013.
Suhruta, K., Makija, K., and Dinesh Kumar, U. 2013. 1920 Evil Returns—Bollywood
and Social Media Marketing, IIMB Case Number IMB437.
Tandon, R., Chakraborty, A., Srinivasan, G., Shroff, M., Abdulla, A., Shamsundar,
B., Sinha, R., Subramaniam, S., Hill, D., and Dhore, P. 2013. Hewlett Packard:
Delivering profitable growth for HPDirect.Com using operations research.
Interfaces, Vol. 43, No. 1, pp. 48–61.
Tufte, E. 2001. Visual Display of Quantitative Information, Graphics Press, Connecticut.
Underhill, P. 2009. Why We Buy: The Science of Shopping. Simon & Schuster (paper-
back), New York.
4
Linear Programming (LP)-Based Two-
Phase Classifier for Solving a Classification
Problem with Multiple Objectives

Sakthivel Madankumar, Pusapati Navya, Chandrasekharan

Rajendran, N. Srinivasa Gupta, and B. Valarmathi

CONTENTS
4.1 Introduction................................................................................................... 88
4.2 Research Problem Description and Assumptions................................... 89
4.2.1 Proposed MILP Model..................................................................... 91
4.2.2 Proposed LP-Based Classifier (LP Model 1): Based on the
Conventional Approach of Considering All the Attributes
with Power Index Equal to 1........................................................... 92
4.2.3 Proposed LP-Based Classifier with a Crisp Boundary
(LP Model 2)....................................................................................... 93
4.2.4 Numerical Illustration for the Constraints in Proposed
LP-Based Classifier (LP Model 1).................................................... 96
4.2.5 Numerical Illustration for the Constraints in Proposed
LP-Based Classifier (LP Model 2).................................................... 97
4.3 Proposed LP-Based Two-Phase Classifier................................................. 98
4.3.1 Proposed LP Model for Two-Phase Classifier.............................. 98
4.3.2 Algorithm for the Proposed LP-Based Two-Phase
Classifier......................................................................................100
4.3.3 Numerical Illustration for the Constraints in Proposed
LP-Based Two-Phase Classifier..................................................... 102
4.4 Results and Discussion.............................................................................. 102
4.4.1 Comparison Study of the Proposed LP-Based Classifiers
with the LP Classifier with Fuzzy Measure and the
Choquet Integral by Yan et al. (2006)........................................... 103
4.4.2 Comparison Study of the Proposed LP-Based Classifiers
for the Recommendation Data Set with Artificial Neural
Networks.......................................................................................... 105
4.5 Summary...................................................................................................... 111
Acknowledgment................................................................................................. 112
References.............................................................................................................. 112

87
88 Big Data Analytics Using Multiple Criteria Decision-Making Models

4.1 Introduction
Classification models are used to predict the group/category, to which a
new observation belongs, based on the training data set or set of observa-
tions for which the categories are known in advance. Linear programming
(LP)-based classifiers for classification problem were discussed by Freed
and Glover (1986), and the objective function of the classifier/LP model
was either set to minimize the sum of errors, or set to minimize the maxi-
mum error for the training data set. Multiobjective LP model was used
by Shi et  al. (2001) for data mining in portfolio management. LP-based
classifiers provide good results in terms of accuracy when the data set is
linearly separable. To handle and to improve the accuracy of the classifier
when the data set is not linearly separable, researchers and practitioners
mostly consider logistic regression, support vector machines, and artifi-
cial neural networks (ANN) as classifiers. The problem of maximizing the
accuracy of classification is computationally hard, and the methods such
as LP-based classifier, logistic regression, support vector machines, and
ANN work on their specific/respective objective functions with respect
to each classifier, which indirectly improves the accuracy of the classifica-
tion. In order to improve the accuracy of the conventional LP model when
the data is not linearly separable, the concept of fuzzy measure was intro-
duced and utilized by Yan et al. (2006).
The big data is generally characterized by volume, velocity, variety, and
value of data. The technological advancements and the Internet of things
(IoT) enable the massive collection of data with high velocity from various
sources. The big data analytics is the process of understanding the hidden
pattern in this large volume of data to get better insights about the data.
The multiple criteria decision making (MCDM) is a topic in Operations
Research which deals with decision problems that involve multiple objec-
tives. In order to solve the decision problem with multiple objectives, the
researchers and practitioners consider goal programming or epsilon con-
straint-based models. The MCDM models can effectively be applied in the
field of big data analytics to solve the underlying decision problems.
This chapter first proposes a mixed integer linear programming (MILP)
model in order to maximize the accuracy of the classification. Since the prob-
lem is computationally hard, the MILP model can only solve the data set of
small samples (e.g., in order of 100 samples) with reasonable execution time.
In big data analytics, we often encounter the classification problem which
deals with large data sets. The study mainly concentrates on the develop-
ment of computationally efficient LP-based classifiers that can handle the
large volume of data, and also the ability to identify the non-dominated set
of solutions with respect to multiple objectives.
In order to handle the large volume of data, batch processing technologies
are used (e.g., Apache Hadoop) and to handle the high velocity of data, stream
Linear Programming (LP)-Based Two-Phase Classifier 89

processing technologies are used (e.g., Apache Spark or Apache Storm).

These technologies help in processing the big data to derive its meaning and
also to convert the unstructured data from various sources into structured
data. Once we have the structured data, we can apply the proposed LP-based
classifiers to derive the hidden pattern/decisions in this large volume of data
such that we can identify the non-dominated set of solutions with respect to
multiple objectives.
So, in the next part of the chapter, we consider the conventional LP-based
classifier with crisp boundary (with respect to the categories of classification)
which can only solve the data set that is linearly separable. Subsequently, to
further improve the accuracy, we propose an LP-based classifier with crisp
boundary and with the consideration of interaction among the attributes and
the contribution of attributes from their higher-order polynomial degrees.
Our approach (unlike the conventional LP-based approach for classification)
is able to capture the curvilinear/nonlinear boundary between the catego-
ries. In order to improve the accuracy, especially with respect to multiple
objectives, and to produce non-dominated set of solutions with respect to
multiple objectives, we propose a two-phase classifier for classification. Our
LP-based two-phase classifier considers the interaction among the attri-
butes and the contribution of attributes from their higher-order polynomial
degrees to solve the data set that is not linearly separable. In the first phase,
our proposed two-phase classifier runs the LP model by considering the
bandwidth of boundary, and in the second phase, the classifier transforms
the bandwidth of boundary into a crisp boundary, and in this process, the
proposed classifier identifies a non-dominated set of solutions with respect
to multiple objectives. Practitioners mostly consider the receiver operating
characteristics (ROC) curve to determine the best threshold settings for the
classifier, whereas the proposed LP-based two-phase classifier identifies the
non-dominated set of solutions with respect to multiple objectives.
To evaluate the performance of the proposed LP-based classifiers, we con-
sider two data sets that are already available in the literature (Yan et al., 2006;
Maas et al., 2011). We also compare the accuracy of all the proposed LP-based
classifiers with ANN, and the results indicate that the proposed LP-based
two-phase classifier is able to give good results even when the data is not
linearly separable, proving our claim that our LP-based two-phase classifier
can handle the data set that is not linearly separable.

4.2  Research Problem Description and Assumptions

A set of observations for which the categories (two predetermined catego-
ries: category 0 and category 1) are known in advance is referred to as train-
ing data set, and a set of observations for which the categories need to be
90 Big Data Analytics Using Multiple Criteria Decision-Making Models

predicted is referred to as test data set or validation data set. Observations

in both training data set as well as test data set have same number of fea-
tures/attributes, and for the training data set, the value for the target variable
(category) is known in advance and for the test and validation data set, the
classifier predicts the value of the target variable on the basis of the training
data set.
In Section 4.2.1, we propose a MILP model for solving the classification,
and the same model can be extended to solve multiple objectives with the
help of goal programming and epsilon-constraint method. However, this
MILP model, when implemented, can handle only small-sized data set (a
couple of hundred observations). In order to solve the classification prob-
lem that commonly arises in big data analytics and to overcome the MILP
model’s limitation to solve large data sets, in Section 4.2.2, we present the
conventional LP-based classifier with crisp boundary (with respect to the
classification of categories) that can solve the data set that is linearly sepa-
rable. In order to further improve the accuracy, in Section 4.2.3, we propose
an LP-based classifier (with crisp boundary) with the consideration of inter-
action among the attributes and the contribution of attributes from their
higher-order polynomial degrees. This approach to data transformation is
undertaken to capture the curvilinear/nonlinear boundary between the cat-
egories. Finally, in Section 4.3, we propose an LP-based two-phase classifier
for classification which considers the interaction among the attributes and
the contribution of attributes from their higher-order polynomial degrees to
solve the data set that is not linearly separable and also to improve the accu-
racy with respect to multiple objectives. In the first phase, the classifier runs
the LP model by considering the bandwidth of boundary, and in the second
phase, the proposed classifier transforms the bandwidth of boundary into
a crisp boundary, and in this process, the proposed classifier identifies a
non-dominated set of solutions with respect to multiple objectives.
The salient contributions of the proposed LP-based two-phase classifier
are in terms of treating the decision variables as unrestricted in sign, the con-
tribution of attributes from their interaction effects, the contribution of attri-
butes from their higher-order polynomial degrees, treating the classification
threshold/cut-off as a decision variable, and converting the bandwidth of
boundary of threshold to a crisp boundary, thereby behaving like a nonlin-
ear classifier to an extent. We also consider multiple objectives to produce
non-dominated set of solutions, when we determine the crisp boundary. The
multiple objectives include: maximize the total accuracy with respect to both
categories, maximize the accuracy with respect to category 1, and maximize
the accuracy with respect to category 0. We consider such multiple objectives
because in the application areas such as medical diagnosis, absence of an
alarm (failing to predict the category 1) is more serious than a false alarm.
The proposed LP-based two-phase classifier can find a non-dominated set
of solutions (with respect to multiple objectives) for the enhanced decision
support.
Linear Programming (LP)-Based Two-Phase Classifier 91

4.2.1  Proposed MILP Model

Parameters

A Total number of attributes in each observation

N Total number of observations in the training data set
j Index for the attribute
i Index for the observation
ai,j Value of attribute j for observation i
epsilon A constant with a very small value 0.00001
normalize Normalize is the function which converts a priori the value of the respective
expression to a value between 0 and 1. For example, while applying the
normalize function for the expression ai,j which corresponds to attribute j
for observation i, it is expressed as follows:

normalize ( ai , j ) =
(a i,j − min{ai′ , j }
i′∈N )
(max{a )
.
i′ , j } − min{ai′ , j }
i′∈N i′∈N

Decision Variables

x ′′j An unrestricted variable to capture the coefficient for the term ai,j
y An unrestricted variable which acts as a surrogate constant to improve the accuracy
successi A binary variable to indicate whether the prediction is correct or not

The objective is to maximize the accuracy/success of the prediction:

Maximize
N

Z= ∑ success i (4.1)
i =1
subject to the following constraints, for all i:
(
when the observation i is of category 0 (i.e., ∑ Aj =1(normalize( ai , j ) × x′′j + y ) should )
be less than or equal to 0.5 in the case of category-0 observation), we have
A

∑ (normalize(a
j =1
i, j ) × x′′j ) + y ≤ 0.5 + (1 − successi ) × M , (4.2)

(
when the observation i is of category 1 (i.e., ∑ Aj =1(normalize( ai , j ) × x′′j + y ) should )
be greater than or equal to (0.5 + epsilon) in the case of category-1 observa-
tion), we have
A

∑ (normalize (a
j =1
i, j ) × x′′j ) + y ≥ 0.5 + epsilon − (1 − successi ) × M. (4.3)

92 Big Data Analytics Using Multiple Criteria Decision-Making Models

In the above constraints (4.2) and (4.3), x′′j and y are unrestricted in sign,
and successi is a binary variable. The decision variable successi determines
whether the observation i is predicted correctly with respect to the corre-
sponding category, and objective function (4.1) maximizes the total accuracy
of the prediction.

4.2.2 Proposed LP-Based Classifier (LP Model 1): Based on the

Conventional Approach of Considering All the
Attributes with Power Index Equal to 1
The set of parameters and decision variables (except successi) given in Section
4.2.1 are also used in this section in addition to the following decision
variable.

Decision Variable

errori: A real variable to capture the error value in the expression

The objective is to minimize the total errors:

Minimize

Z= ∑ error i (4.4)
i =1

subject to the following constraints, for all i:

(
when the observation i is of category 0 (i.e., ∑ Aj =1(normalize( ai , j ) × x′′j + y ) )
should be less than or equal to 0.5 in the case of category-0 observation), we
have

 A

errori ≥ 
 ∑ (normalize (a
j =1
i, j ) × x′′j ) + y  − 0.5,

(4.5)

(
when the observation i is of category 1 (i.e., ∑ Aj=1 (normalize( ai , j ) × x ′′j + y ) )
should be greater than or equal to (0.5 + epsilon) in the case of category-1
observation), we have

 A

errori ≥ (0.5 + epsilon) − 
 ∑ (normalize (ai, j ) × x′′j ) + y .
j =1 
(4.6)

In the above constraints (4.5) and (4.6), x′′j and y are unrestricted in sign,
and errori is a real continuous variable. The decision variable errori captures
Linear Programming (LP)-Based Two-Phase Classifier 93

the amount of deviation from their respective threshold for each observation
(if present), and objective function (4.4) minimizes the total errors of all the
observations in the training data set, so as to improve the accuracy of the
prediction.
The proposed classifier (LP Model 1) splits the data set into training data
set and test data set, and then uses the training data set to train the model,
and uses the test data set to validate the model.
The proposed classifier (LP Model 1) runs the above LP model for the
training data set to train the model. Then the proposed classifier (LP Model
1) makes use of the trained model to predict the category for the test data set.
Note that the decision variable errori is used only in the above LP model (LP
Model 1) for the training data set to train the model and then the value for
( ( ) )
the expression ∑ Aj =1 normalize( ai , j ) × x′′j + y is calculated for each observa-
tion i in the test data set to predict the category. If the value of the expression
is less than or equal to 0.5, then the corresponding observation is predicted
as category 0; and if the value of the expression is greater than or equal to
(0.5 + epsilon), then the corresponding observation is predicted as category 1.
Then the accuracy with respect to test data set is calculated for the follow-
ing objectives:

maximize the total accuracy with respect to both categories,

(number of samples predicted correctly )

Objective 1 = × 100; (4.7)
(total number of samples)

maximize the accuracy with respect to category 1,

(number of category − 1 samples predicted correctly )

Objective 2 = × 100;
(ttotal number of category − 1 samples)
(4.8)

and maximize the accuracy with respect to category 0,

(number of category − 0 samples predicted correctly )

Objective 3 = × 100.
(ttotal number of category − 0 samples)

(4.9)

4.2.3  Proposed LP-Based Classifier with a Crisp Boundary (LP Model 2)

We consider the power index for each attribute in the range [1, 5] (range
being chosen in this study), and the incorporation of interaction effects of
attributes in the proposed LP-based classifier with a crisp boundary.
94 Big Data Analytics Using Multiple Criteria Decision-Making Models

Parameters

A Total number of attributes in each observation

N Total number of observations in the training data set
j, j′, j″ Indices for the attributes
i Index for the observation
Pj A set of power terms for attribute j, used with respect to the value of
attribute’s higher-order polynomial degrees
Qj A set of power terms for attribute j, used with respect to the interaction
effects of attributes
pj An element from set Pj/* Each attribute can have its own set of power terms:
For example, P1 = {1, 2, 3, 4,}; P2 = {1, 2, 3} and we have p1 ∈ P1; p2 ∈ P2 */
qj An element from set Qj /* Each attribute can have its own set of power terms
with respect to the interaction effects of attributes: For example, Q1 = {1, 2, 3};
Q2 = {1, 2, 3} and we have q1 ∈ Q1; q2 ∈ Q2 */
ai,j Value of attribute j for observation i
epsilon A constant with a very small value 0.00001
normalize Normalize is the function which converts a priori the value of the respective
expression to a value between 0 and 1. For example, while applying the
pj
normalize function for the term a i , j which corresponds to attribute j for
observation i, it is expressed as follows:

normalize ( a ) =
pj
(a
{ }) ; pj
i,j − min ai′ ,j j
i′∈N
p

(max {a } − min {a })
i,j
pj pj
i′ , j i′ , j
i′∈N i′∈N

and

( q j′ q j′′
×a )=
(aq j′
i , j′
q
{
× ai ,jj′′′′ − min ai′j,′j′ × ai′j,′′j′′
i′∈N
q q
})
(max {a })
normalize a .
} − miin {a
i , j′ i , j′′ q j′ q j′′ q j′ q
i′ , j′ ×a i′ , j′′ i′ , j′ × ai′j,′′j′′
i′∈N i′∈N

Decision Variables

xj,pj An unrestricted variable to capture the coefficient for the term/expression

p
ai , jj
x ′j′ , j′′ ,q j′ ,q j′′ An unrestricted variable to capture the coefficient for the interaction term
(
with respect to aiq,jj′′ × aiq,jj′′′′ )
y An unrestricted variable which acts as a surrogate constant to improve the
accuracy
errori A real variable to capture the error value in the expression

The objective is to minimize the total errors:

Minimize

Z= ∑ error i (4.10)
i =1
Linear Programming (LP)-Based Two-Phase Classifier 95

subject to the following constraints, for all i:

when the observation i is of category 0 (i.e.,

(∑ A
j =1 (
∑ p j ∈Pj normalize aip, jj × x j , p j( ) )
( ( )
+ ∑ Aj ′=−11 ∑ Aj ′′= j ′+ 1 ∑ q j′ ∈Q j′ ∑ q j′′ ∈Q j′′ normalize aiq,jj′ ′ × aiq,jj′′′′ × x′j ′ , j ′′ , q j′ , q j′′ + y ) )
should be less than or equal to 0.5 in the case of category-0 observation), we
have

 A

errori ≥ 
 ∑ ∑ (normalize (a ) × x )
j =1 p j ∈Pj
pj
i, j j,pj

A −1 A

+ ∑ ∑ ∑ ∑ (normalize (a
j ′=1 j ′′= j ′+1 q j′ ∈Q j′ q j′′ ∈Q j′′
q j′
i , j′
q
) )
× ai ,jj′′′′ × x′j′ , j′′ ,q j′ ,q j′′ + y − 0.5 . (4.11)


When the observation i is of category 1 (i.e.,


∑ ∑ (normalize (a ) × x )
A
pj
 i, j j,pj
 j =1 p j ∈Pj

A −1

∑ ∑ ∑ ∑ (normalize (a )
A

+ q j′
i , j′
q
)
× ai ,jj′′′′ × x ′j′ , j′′ ,q j′ ,q j′′ + y 

j′=1 j′′= j′ +1 q j ′ ∈Q j ′ q j ′′ ∈Q j ′′

should be greater than or equal to (0.5 + epsilon) in the case of category-1

observation), we have


∑ ∑ (normalize (a ) × x )
A
pj
errori ≥ (0.5 + epsilon) −  i, j j,pj
 j =1 p j ∈Pj

A −1

∑ ∑ ∑ ∑ (normalize (a )
A

+ q j′
i , j′
q
)
× ai ,jj′′′′ × x ′j′ , j′′ ,q j′ ,q j′′ + y  .

(4.12)
j′=1 j′′= j′+1 q j ′ ∈Q j ′ q j ′′ ∈Q j ′′

In the above constraints (4.11) and (4.12), x j , p j , x′j ′ , j ′′ , q j′ , q j′′ , and y are unre-
stricted in sign, and errori is a real variable. Constraints (4.11) and (4.12)
capture the contribution of attributes from their higher-order polynomial
degrees, and also capture the interaction effects among the attributes. The
decision variable errori captures the amount of deviation from their respec-
tive threshold for each observation (if present). Objective function (4.10)
96 Big Data Analytics Using Multiple Criteria Decision-Making Models

minimizes the total errors of all the observations in the training data set, so
as to improve the accuracy of the prediction.
The proposed classifier (LP Model 2) splits the data set into training data
set and test data set, and then uses the training data set to train the model,
and uses the test data set to validate the model.
The proposed classifier (LP Model 2) runs the above LP model for the train-
ing data set to train the model. Then this classifier makes use of the trained
model to predict the category for the test data set. Note that the decision vari-
able errori is used only in the above LP model (LP Model 2) for the training
data set to train the model, and then the value for the expression


∑ ∑ (normalize (a ) × x )
A
pj
 i, j j,pj
 j =1 p j ∈Pj

A −1

∑ ∑ ∑ ∑ (normalize (a )
A

+ q j′
i , j′
q
)
× ai ,jj′′′′ × x ′j′ , j′′ ,q j′ ,q j′′ + y 

j′=1 j′′ = j′+1 q j ′ ∈Q j ′ q j ′′ ∈Q j ′′

is calculated for each observation i in the test data set to predict the category.
If the value of the expression is less than or equal to 0.5, then the correspond-
ing observation is predicted as category 0; and if the value of the expression
is greater than or equal to (0.5 + epsilon), then the corresponding observation
is predicted as category 1. Then the accuracy with respect to test data set is
calculated for the objectives, Objective 1, Objective 2, and Objective 3.

4.2.4 Numerical Illustration for the Constraints in Proposed

LP-Based Classifier (LP Model 1)
In this section, we present the numerical example for the constraints in pro-
posed LP-based classifier (LP Model 1), and for the purpose of numerical
illustration, let the number of attributes be 2, and values of attributes be in
the range [0, 9]. Table 4.1 represents the samples with respect to each category
for the numerical illustration.
As the values of attributes are in the range [0, 9], for attribute 2 of sample 1,
the normalize function is expressed as follows:

( 5 − 0)
normalize (5) = = 0.56.
(9 − 0 )

For the samples in Table 4.1, the constraints of the proposed LP model are
expressed as follows:
With respect to sample 1, LP model Constraint (4.5) appears as follows:

errori ≥ (0.11 × x1′′+ 0.56 × x2′′ + y ) − 0.5. (4.13)

Linear Programming (LP)-Based Two-Phase Classifier 97

TABLE 4.1
Samples for the Numerical Illustration
Sample Attribute 1 Attribute 2 Category
1 1 5 0
2 4 9 1

With respect to sample 2, LP model Constraint (4.6) appears as follows:

errori ≥ (0.5 + epsilon) − (0.44 × x1′′+ 1.00 × x2′′ + y ) . (4.14)

4.2.5 Numerical Illustration for the Constraints in Proposed

LP-Based Classifier (LP Model 2)
In this section, we present the numerical example for the constraints in pro-
posed LP-based classifier (LP Model 2), and for the purpose of numerical
illustration, let the number of attributes be 2, and values of attributes be in
the range [0, 9]. Table 4.1 represents the samples with respect to each category
for the numerical illustration. Sets of power terms with respect to higher-
order polynomial degrees of attributes and with respect to interaction effects
of attributes are as follows:

P1 = {1, 2, 3}; (4.15)

P2 = {1, 2, 3}; (4.16)

Q1 = {1, 2}; and (4.17)

Q2 = {1, 2}. (4.18)

As the values of attributes are in the range [0, 9], for attribute 2 of sample 1
with power term 2, the normalize function is expressed as follows:

((5 × 5) − 0)
normalize (52 ) = = 0.31.
((9 × 9) − 0)

For the samples in Table 4.1, the constraints of the proposed LP model are
expressed as follows:
With respect to sample 1, LP model Constraint (4.11) appears as follows:

errori ≥ (0.11 × x1,1 + 0.01 × x1,2 + 0.00 × x1,3 + 0.56 × x2 ,1 + 0.31 × x2 ,2 + 0.17
× x2 ,3 + 0.06 × x1′ ,2 ,1,1 + 0.03 × x1′ ,2 ,1,2 + 0.01 × x1′ ,2 ,2 ,1 + 0.00 × x1′ ,2 ,2 ,2 + y ) − 0.5.
(4.19)
98 Big Data Analytics Using Multiple Criteria Decision-Making Models

Note: The coefficients are presented in this entire chapter in two decimal
points precision, and hence the expression 0.00 × x1,3 in Constraint (4.19) rep-
resents actually a small coefficient for the variable x1,3.
With respect to sample 2, LP model Constraint (4.12) appears as follows:

errori ≥ (0.5 + epsilon) − (0.44 × x1,1 + 0.20 × x1, 2 + 0.09 × x1, 3 + 1.00 × x2 ,1
+ 1.00 × x2 , 2 + 1.00 × x2 , 3 + 0.44 × x1′ , 2 ,1,1 + 0.44 × x1′ , 2 ,1, 2 + 0.20 × x1′ , 2 , 2 ,1
+ 0.20 × x1′ , 2 , 2 , 2 + y ).

(4.20)

4.3  Proposed LP-Based Two-Phase Classifier

In this section, we propose an LP-based two-phase classifier for classification
which considers the interaction among the attributes and the contribution
of attributes from their higher-order polynomial degrees to solve the data
set that is not linearly separable. In the first phase, the classifier runs the
LP model by considering the bandwidth of boundary (without invoking the
crisp boundary), and this bandwidth is treated as a decision variable, and
in the second phase, the proposed classifier transforms the bandwidth of
boundary into a crisp boundary. During this process, the proposed classi-
fier also identifies a non-dominated set of solutions with respect to multiple
objectives.

4.3.1  Proposed LP Model for Two-Phase Classifier

The set of parameters and decision variables given in Section 4.2.3 are also
used in this section in addition to following decision variable.
Decision Variable

b A real variable to capture the classification threshold/cut-off

Now, the LP model in the proposed two-phase classifier is presented.

Note that in the bandwidth of boundary [b, (b + 1)], b is treated as a decision
variable.
The objective is to minimize the total errors:
Minimize

Z= ∑ error
i =1
i (4.21)
Linear Programming (LP)-Based Two-Phase Classifier 99

subject to the following constraints, for all i: when the observation i is of

category 0 (i.e.,

(∑ A
j =1 ( ( )
∑ p j ∈Pj normalize aip, jj × x j , p j )
( ( )
+ ∑ Aj′=−11 ∑ Aj′′= j′+1 ∑ q j′ ∈Q j′ ∑ q j′′ ∈Q j′′ normalize aiq,jj′′ × aiq,jj′′′′ × x ′j′ , j′′ ,qq j′ ,q j′′ + y ) )
should be less than or equal to b in the case of category-0 observation), we have

 A

errori ≥ 
 ∑ ∑ (normalize (a ) × x )
j =1 p j ∈Pj
pj
i, j j,pj

A −1 A

+ ∑ ∑ ∑ ∑ (normalize (a
j’ =1 j’’ = j’ +1 q j’ ∈Q j’ q j’’ ∈Q j’’
q j′
i , j′
q
) )
× ai ,jj′′′′ × x′j′ , j′′ ,q j′ ,q j′′ + y  − b ,

(4.22)

when the observation i is of category 1 (i.e.,

(∑ A
j =1 ( ( )
∑ p j ∈Pj normalize aip, jj × x j , p j )
( ( )
+ ∑ Aj′=−11 ∑ Aj′′= j′+1 ∑ q j′ ∈Q j′ ∑ q j′′ ∈Q j′′ normalize aiq,jj′′ × aiq,jj′′′′ × x ′j′ , j′′ ,q j′ ,q j′′ + y ) )
should be greater than or equal to (b + 1) in the case of category-1 observa-
tion), we have


∑∑ (normalize (a ) × x )
A
pj
errori ≥ (b + 1) −  i, j j,pj
 j =1 p j ∈Pj

A −1

∑ ∑ ∑ ∑ (normalize (a )
A

+
q j′
i , j′
q
)
× ai ,jj′′′′ × x ′j′ , j′′ ,q j′ ,q j′′ + y  .

(4.23)
j′=1 j′′= j′+1 q j ′ ∈Q j ′ q j ′′ ∈Q j ′′

In the above constraints (4.22) and (4.23), x j , p j , x ′j′ , j′′ ,q j′ ,q j′′ , and y are unre-
stricted in sign, and errori and b are real variables. We have “1” in Constraint
(4.23), since we have dichotomous classification. Constraints (4.22) and (4.23)
capture the contribution of attributes from their higher-order polynomial
degrees, and also capture the interaction effects among the attributes. The
decision variable b captures the classification threshold/cut-off for the respec-
tive category. The decision variable errori captures the amount of deviation
from their respective threshold for each observation (if present). Objective
function (4.21) minimizes the total errors of all the observations in the train-
ing data set, so as to improve the accuracy of the prediction.
100 Big Data Analytics Using Multiple Criteria Decision-Making Models

4.3.2  Algorithm for the Proposed LP-Based Two-Phase Classifier

The proposed classifier runs in two phases. In the first phase, the classifier runs
the proposed LP model (LP model for the two-phase classifier) using the train-
ing data set to train the model. In the second phase, the proposed classifier uses
the test data set to transform the bandwidth of boundary [b, (b + 1)] into a crisp
boundary b + c, and while determining the crisp boundary multiple objectives
are considered. The proposed classifier iteratively increments the value of c
from 0 to 1 by 0.05 (step size), and identifies a non-dominated set of solutions
with respect to multiple objectives, while applying the trained model/expres-
sion on the test data set. Finally, for each of the solution in non-dominated set,
the classifier identifies the non-dominated set of solutions with respect to objec-
tives, Objective 1, Objective 2, and Objective 3, for the validation data set.

Phase 0:
• In this phase, the proposed classifier splits the data set into three:
training data set, test data set, and validation data set.
Step 1: Split the data set into training data set, test data set, and valida-
tion data set.
Phase 1:
• In this phase, the proposed classifier uses the training data set to train
the model (i.e., to get the values of the decision variables b, x j , p j , x′j ′ , j ′′ , q j′ , q j′′,
and y) and also identifies the bandwidth of boundary [b, (b + 1)].
Step 1: Run the proposed LP model (LP model in the proposed two-
phase classifier) for the training data set to get the LP solution
in terms of the value of the variable b, and also the values of the
variables x j , p j , x ′j′ , j′′ ,q j′ ,q j′′, and y, to compute the following:

 A


 ∑ ∑ (normalize (a ) × x )
j =1 p j ∈Pj
pj
i, j j,pj

A −1 A

+ ∑ ∑ ∑ ∑ (normalize (a
j′=1 j′′= j ′ +1 q j′ ∈Q j′ q j′′ ∈Q j′′
q j′
i , j′
q
) )
× ai ,jj′′′′ × x′j , j′′ ,q j′ ,q j′′ + y  .

(4.24)
  
Phase 2:
• The second phase of the proposed classifier comprises two parts.
Part 1:
• In this part, the proposed classifier uses the test data set to trans-
form the bandwidth of boundary into a crisp boundary b + c. In this
process, the proposed classifier identifies a non-dominated set of
solutions with respect to multiple objectives, and captures the cor-
responding set of c values.
Linear Programming (LP)-Based Two-Phase Classifier 101

Step 1: Iteratively increment the value of c from 0 to 1 by 0.05 (step size),

and for each value of c, do the following:
Step 1.1: For each observation in the test data set, calculate the value for the
expression (4.24) and assign the value to the variable val.
If val ≤ b + c
then declare the observation as category 0.
If val ≥ b + c + epsilon
then declare the observation as category 1.
Step 1.2: Calculate the objectives, Objective 1, Objective 2, and Objective 3.
Step 1.3: Form the non-dominated set of solutions with respect to Objective
1, Objective 2, and Objective 3 and store the corresponding value
of c with respect to every solution.
Part 2:
• In this part, the proposed classifier uses the validation data set to
validate the model. In this process, the proposed classifier identifies
a non-dominated set of solutions with respect to multiple objectives
for the validation data set by evaluating the validation data set for
the chosen set of c values.
Step 1: With the values of c that are obtained (from Part 1) correspond-
ing to the set of non-dominated solutions (see Step 1.3 of Part
1), consider the validation data set and hence obtain the set of
non-dominated solutions. This set of non-dominated solutions
constitutes the solutions with respect to objectives Objective 1,
Objective 2, and Objective 3 for the validation data set. This set of
solutions is used to benchmark/evaluate the performance of the
proposed LP-based two-phase classifier.

Note:

• In the case of a single-objective optimization problem, during

the second phase, the proposed classifier first identifies the best
solution with respect to the objective under consideration for the
test data set (see Part 1), and captures the corresponding c value;
and then the proposed classifier reports the single solution with
respect to the objective under consideration for the validation data
set by evaluating the validation data set for the chosen c value (see
Part 2).
( q q
• With the consideration of interaction among the attributes ai ,jj′′ × ai ,jj′′′′)
and the contribution of attributes from their higher-order polyno-
( )
p
mial degrees ai , jj , the proposed LP-based two-phase classifier is
able to capture the curvilinear/nonlinear boundary between the
categories.
102 Big Data Analytics Using Multiple Criteria Decision-Making Models

4.3.3 Numerical Illustration for the Constraints in Proposed

LP-Based Two-Phase Classifier
In this section, we present a numerical example for proposed LP-based two-
phase classifier, and for the purpose of numerical illustration, let the number
of attributes be 2, and values of attributes be in the range [0, 9]. Table 4.1 repre-
sents the samples with respect to each category for the numerical illustration.
We also have the power term settings according to Equations 4.15 through 4.18.
For the samples in Table 4.1, the constraints of the proposed LP model are
expressed as follows:
With respect to sample 1, LP model Constraint (4.22) appears as follows:

errori ≥ (0.11 × x1,1 + 0.01 × x1,2 + 0.00 × x1,3 + 0.56 × x2 ,1 + 0.31 × x2 ,2

+ 0.17 × x2 ,3 + 0.06 × x1′,2 ,1,1 + 0.03 × x1′,2 ,1,2 + 0.01 × x1′,2 ,2 ,1
+ 0.00 × x1′,2 ,2 ,2 + y ) − b.
(4.25)

With respect to sample 2, LP model Constraint (4.23) appears as follows:

errori ≥ (b + 1) − (0.44 × x1,1 + 0.20 × x1,2 + 0.09 × x1,3

+ 1.00 × x2 ,1 + 1.00 × x2 ,2 + 1.00 × x2 ,3 + 0.44 × x1′ ,2 ,1,1 + 0.44 × x1′ ,2 ,1,2
+ 0.20 × x1′ ,2 ,2 ,1 + 0.20 × x1′ ,2 ,2 ,2 + y ).
(4.26)

4.4  Results and Discussion

The proposed MILP model cannot be executed on large-sized data sets that
commonly arise in big data analytics. The proposed MILP model can run on
the training data set of size up to 100 or 200 observations, and in most cases it
overfits the training data, and it fails to generalize the underlying pattern in
the data set. Hence, in this section, we mainly concentrate on the performance
of the proposed LP-based classifiers (LP Model 1, LP Model 2, and LP-based
two-phase classifier). In Section 4.4.1, we present the comparison study of the
proposed LP-based classifiers (LP Model 1, LP Model 2, and LP-based two-
phase classifier) with LP classifier with fuzzy measure and the Choquet inte-
gral by Yan et al. (2006). In Section 4.4.2, we present the comparison study of
the proposed LP-based classifiers (LP Model 1, LP Model 2, and LP-based
two-phase classifier) with ANN. For the comparison study, we use the fol-
lowing settings which seem to be sufficient to produce reasonable accuracy:

Pj = {1, 2, 3, 4, 5} ∀j , (4.27)

Q j = {1} ∀j. (4.28)

Linear Programming (LP)-Based Two-Phase Classifier 103

4.4.1 Comparison Study of the Proposed LP-Based Classifiers

with the LP Classifier with Fuzzy Measure and
the Choquet Integral by Yan et al. (2006)
To evaluate the performance of the proposed LP-based classifiers (LP
Model 1, LP Model 2, and LP-based two-phase classifier), we use the same
data set presented in the paper by Yan et al. (2006). For the purpose of numer-
ical illustration, we take two samples from the data set, and present the con-
straints of the respective classifiers. This data set contains 200 observations
with two attributes and the range of values with respect to each attribute is
presented in Table 4.2, and the samples are presented in Table 4.3. We also
have the power term settings according to Equations 4.27 and 4.28.
For the samples in Table 4.3, the constraints of the proposed LP-based clas-
sifier (LP Model 1) are expressed as follows:
With respect to sample 1, LP model Constraint (4.5) appears as follows:

errori ≥ (0.06 × x1′′+ 0.43 × x2′′ + y ) − 0.5. (4.29)

With respect to sample 2, LP model Constraint (4.6) appears as follows:

errori ≥ (0.5 + epsilon) − (0.39 × x1′′+ 0.93 × x2′′ + y ) . (4.30)

For the samples in Table 4.3, the constraints of the proposed LP-based clas-
sifier (LP Model 2) are expressed as follows:
With respect to sample 1, LP model Constraint (4.11) appears as follows:

errori ≥ (0.06 × x1,1 + 0.00 × x1,2 + 0.00 × x1,3 + 0.00 × x1, 4 + 0.00 × x1,5
+ 0.43 × x2 ,1 + 0.19 × x2 ,2 + 0.08 × x2 ,3 + 0.04 × x2 , 4 + 0.02 × x2 ,5
+ 0.04 × x1′ ,2 ,1,1 + y ) − 0.5.
(4.31)
TABLE 4.2
Range of Values with Respect to Attributes
in Data Set Presented in the Paper by
Yan et al. (2006)
Attribute 1 Attribute 2
[0.01,0.99] [0.01,0.99]

TABLE 4.3
Two Samples from the Data Set Presented in the Paper
by Yan et al. (2006)
Sample Attribute 1 Attribute 2 Category
1 0.07 0.43 0
2 0.39 0.92 1
104 Big Data Analytics Using Multiple Criteria Decision-Making Models

With respect to sample 2, LP model Constraint (4.12) appears as follows:

errori ≥ (0.5 + epsilon) − (0.39 × x1,1 + 0.16 × x1,2 + 0.06 × x1,3

+ 0.02 × x1, 4 + 0.00 × x1,5 + 0.93 × x2 ,1 + 0.86 × x2 ,2 + 0.80 × x2 ,3
+ 0.75 × x2 , 4 + 0.69 × x2 ,5 + 0.45 × x1′ ,2 ,1,1 + y ). (4.32)

For the samples in Table 4.3, the constraints of the proposed LP-based two-
phase classifier are expressed as follows: With respect to sample 1, LP model
Constraint (4.22) appears as follows:

errori ≥ (0.06 × x1,1 + 0.00 × x1,2 + 0.00 × x1,3 + 0.00 × x1, 4 + 0.00 × x1,5
+ 0.43 × x2 ,1 + 0.19 × x2 ,2 + 0.08 × x2 ,3 + 0.04 × x2 , 4
+ 0.02 × x2 ,5 + 0.04 × x1′ ,2 ,1,1 + y ) − b. (4.33)

With respect to sample 2, LP model Constraint (4.23) appears as follows:

errori ≥ (b + 1) − (0.39 × x1,1 + 0.16 × x1,2 + 0.06 × x1,3 + 0.02 × x1, 4

+ 0.00 × x1,5 + 0.93 × x2 ,1 + 0.86 × x2 ,2 + 0.80 × x2 ,3 + 0.75 × x2 , 4
+ 0.69 × x2 ,5 + 0.45 × x1′ ,2 ,1,1 + y ). (4.34)

To calculate the objectives of the proposed LP-based classifiers (LP Model 1

and LP Model 2), we split the data into two sets, training data set (70%) and
test data set (30%), and the calculated objectives with respect to test data set
are listed in Table 4.4.
To calculate the objectives of the proposed LP-based two-phase classifier, we
split the data into three sets, training data set (60%), test data set (10%), and
validation data set (30%), and the calculated objectives of the validation data set
(see Part 2 of Phase 2) are listed in Table 4.4. The results indicate that the pro-
posed LP-based two-phase classifier is also able to give 100% accuracy for all
the objectives, and it is able to predict the outcome perfectly even when the data

TABLE 4.4
Performance of the Proposed LP-Based Classifiers (LP Model 1, LP Model 2, and the
LP-Based Two-Phase Classifier) for the Data Set by Yan et al. (2006)
Objective 1 (%) Objective 2 (%) Objective 3 (%)
LP Model 1 48.00 100.00 1.89
LP Model 2 98.00 98.94 97.17
Two-phase classifier 100.00 100.00 100.00
Note: LP Model 1 and LP Model 2 cannot address multiobjective optimization; A single
­non-dominated solution of the proposed two-phase classifier (for the validation data set).
Linear Programming (LP)-Based Two-Phase Classifier 105

is not linearly separable. Note that Yan et al. (2006) included all the samples in
the data set for training the model and the accuracy of their classifier was 100%.

4.4.2 Comparison Study of the Proposed LP-Based Classifiers for the

Recommendation Data Set with Artificial Neural Networks
To evaluate the performance of the proposed LP-based classifiers, we use
the data set with respect to movie recommendations (Maas et  al., 2011)
and note that while the data set presented by Maas et  al. (2011) contains
the movie reviews in text format, the last two authors of this chapter con-
verted the text reviews into a numerical data set in their earlier work. For
the purpose of numerical illustration, we take two samples from the data
set, and present the constraints of the respective classifiers. This large data
set contains 25,000 observations with eight attributes and the range of values
with respect to each attribute is presented in Table 4.5, and the samples are
presented in Table 4.6. We also have the power term settings according to
Equations 4.27 and 4.28.
For the samples in Table 4.6, the constraints of the proposed LP-based clas-
sifier (LP Model 1) are expressed as follows:
With respect to sample 1, LP model Constraint (4.5) appears as follows:

errori ≥ (0.25 × x1′′+ 0.09 × x2′′ + 0.21 × x3′′ + 0.11 × x 4′′ + 0.26 × x5′′
+ 0.05 × x6′′ + 0.42 × x7′′ + 0.31 × x8′′ + y ) − 0.5. (4.35)

With respect to sample 2, LP model Constraint (4.6) appears as follows:

errori ≥ (0.5 + epsilon) − (0.28 × x1′′+ 0.22 × x2′′ + 0.09 × x3′′ + 0.22 × x 4′′
+ 0.14 × x5′′ + 0.005 + 0.12 × x7′′ + 0.06 × x8′′ + y ). (4.36)

TABLE 4.5
Range of Values with Respect to Attributes in Recommendation Data Set
Attribute Attribute Attribute Attribute Attribute Attribute Attribute Attribute
1 2 3 4 5 6 7 8
[0,36] [0,23] [0,47] [0,18] [0,35] [0,21] [0,26] [0,16]

TABLE 4.6
Two Samples from the Recommendation Data Set
Attri­ Attri­ Attri­ Attri­ Attri­ Attri­ Attri­ Attri­
Sample bute 1 bute 2 bute 3 bute 4 bute 5 bute 6 bute 7 bute 8 Category
1 9 2 10 2 9 1 11 5 0
2 10 5 4 4 5 1 3 1 1
106 Big Data Analytics Using Multiple Criteria Decision-Making Models

For the samples in Table 4.6, the constraints of the proposed LP-based clas-
sifier (LP Model 2) are expressed as follows:
With respect to sample 1, LP model Constraint (4.11) appears as follows:

errori ≥ (0.25 × x1,1 + 0.062 × x1,2 + 0.02 × x1,3 + 0.00 × x1,4 + 0.00 × x1,5
+0.09 × x2 ,1 + 0.01 × x2 ,2 + 0.00 × x2 ,3 + 0.00 × x2 ,4 + 0.00 × x2 ,5
+0.21 × x3 ,1 + 0.05 × x3 ,2 + 0.01 × x3 ,3 + 0.00 × x3 ,4 + 0.00 × x3 ,5
+0.11 × x 4 ,1 + 0.01 × x 4 ,2 + 0.00 × x 4 ,3 + 0.00 × x 4 , 4 + 0.00 × x 4 ,5
+0.26 × x5 ,1 + 0.07 × x5 ,2 + 0.02 × x5 ,3 + 0.00 × x5 ,4 + 0.00 × x5 ,5
+0.05 × x6 ,1 + 0.00 × x6 ,2 + 0.00 × x6 ,3 + 0.00 × x6 , 4 + 0.00 × x6 ,5
+0.42 × x7 ,1 + 0.18 × x7 ,2 + 0.08 × x7 ,3 + 0.03 × x7 ,4 + 0.01 × x7 ,5
+0.31 × x8 ,1 + 0.10 × x8 ,2 + 0.03 × x8 ,3 + 0.01 × x8 ,4 + 0.00 × x8 ,5
+0.04 × x1′,2 ,1,1 + 0.17 × x1′,3 ,1,1 + 0.08 × x1′,4 ,1,1 + 0.12 × x1′,5 ,1,1 + 0.03 × x1′,6 ,1,1
+0.32 × x1′,7 ,1,1 + 0.24 × x1′,8 ,1,1 + 0.03 × x2′ ,3 ,1,1 + 0.02 × x2′ ,4 ,1,1 + 0.04 × x2′ ,5 ,1,1
+0.00 × x2′ ,6 ,1,1 + 0.13 × x2′ ,7 ,1,1 + 0.07 × x2′ ,8 ,1,1 + 0.03 × x3′ , 4 ,1,1 + 0.14 × x3′ ,5 ,1,1
+0.03 × x3′ ,6 ,1,1 + 0.29 × x3′ ,7 ,1,1 + 0.16 × x3′ ,8 ,1,1 + 0.06 × x 4′ ,5 ,1,1 + 0.02 × x 4′ ,6 ,1,1
+0.12 × x 4′ ,7 ,1,1 + 0.12 × x 4′ ,8 ,1,1 + 0.02 × x5′ ,6 ,1,1 + 0.20 × x5′ ,7 ,1,1 + 0.09 × x5′ ,8 ,1,1
+0.04 × x6′ ,7 ,1,1 + 0.02 × x6′ ,8 ,1,1 + 0.28 × x7′ ,8 ,1,1 + y ) − 0.5.
(4.37)
With respect to sample 2, LP model Constraint (4.12) appears as follows:

errori ≥ (0.5 + epsilon)

−( x1,1 × 0.28 + x1,2 × 0.08 + x1,3 × 0.02 + x1,4 × 0.01 + x1,5 × 0.00
+ x2 ,1 × 0.22 + x2 ,2 × 0.05 + x2 ,3 × 0.01 + x2 , 4 × 0.00 + x2 ,5 × 0.00
+ x3 ,1 × 0.09 + x3 ,2 × 0.01 + x3 ,3 × 0.00 + x3 ,4 × 0.00 + x3 ,5 × 0.00
+ x 4 ,1 × 0.22 + x 4 ,2 × 0.01 + x 4 ,3 × 0.01 + x 4 ,4 × 0.00 + x 4 ,5 × 0.00
+ x5 ,1 × 0.14 + x5 ,2 × 0.02 + x5 ,3 × 0.00 + x5 ,4 × 0.00 + x5 ,5 × 0.00
+ x6 ,1 × 0.05 + x6 ,2 × 0.00 + x6 ,3 × 0.00 + x6 ,4 × 0.00 + x6 ,5 × 0.00
+ x7 ,1 × 0.12 + x7 ,2 × 0.01 + x7 ,3 × 0.00 + x7 ,4 × 0.00 + x7 ,5 × 0.00
+ x8 ,1 × 0.06 + x8 ,2 × 0.00 + x8 ,3 × 0.00 + x8 ,4 × 0.00 + x8 ,5 × 0.00
+ x1′,2 ,1,1 × 0.10 + x1′,3 ,1,1 × 0.08 + x1′,4 ,1,1 × 0.18 + x1′,5 ,1,1 × 0.08 + x1′,6 ,1,1 × 0.03
+ x1′,7 ,1,1 × 0.10 + x1′,8 ,1,1 × 0.05 + x2′ ,3 ,1,1 × 0.03 + x2′ ,4 ,1,1 × 0.08 + x2′ ,5 ,1,1 × 0.05
+ x2′ ,6 ,1,1 × 0.01 + x2′ ,7 ,1,1 × 0.09 + x2′ ,8 ,1,1 × 0.03 + x3′ ,4 ,1,1 × 0.03 + x3′ ,5 ,1,1 × 0.03
+ x3′ ,6 ,1,1 × 0.01 + x3′ ,7 ,1,1 × 0.03 + x3′ ,8 ,1,1 × 0.01 + x 4′ ,5 ,1,1 × 0.07 + x 4′ ,6 ,1,1 × 0.04
+ x 4′ ,7 ,1,1 × 0.07 + x 4′ ,8 ,1,1 × 0.05 + x5′ ,6 ,1,1 × 0.01 + x5′ ,7 ,1,1 × 0.03 + x5′ ,8 ,1,1 × 0.01
+ x6′ ,7 ,1,1 × 0.01 + x6′ ,8 ,1,1 × 0.00 + x7′ ,8 ,1,1 × 0.02 + y ).

(4.38)
Linear Programming (LP)-Based Two-Phase Classifier 107

For the samples in Table 4.6, the constraints of the proposed LP-based two-
phase classifier are expressed as follows:
With respect to sample 1, LP model Constraint (4.22) appears as follows:

errori ≥ (0.25 × x1,1 + 0.062 × x1,2 + 0.02 × x1,3 + 0.00 × x1,4 + 0.00 × x1,5
+0.09 × x2 ,1 + 0.01 × x2 ,2 + 0.00 × x2 ,3 + 0.00 × x2 ,4 + 0.00 × x2 ,5
+0.21 × x3 ,1 + 0.05 × x3 ,2 + 0.01 × x3 ,3 + 0.00 × x3 ,4 + 0.00 × x3 ,5
+0.11 × x 4 ,1 + 0.01 × x 4 ,2 + 0.00 × x 4 ,3 + 0.00 × x 4 , 4 + 0.00 × x 4 ,5
+0.26 × x5 ,1 + 0.07 × x5 ,2 + 0.02 × x5 ,3 + 0.00 × x5 ,4 + 0.00 × x5 ,5
+0.05 × x6 ,1 + 0.00 × x6 ,2 + 0.00 × x6 ,3 + 0.00 × x6 , 4 + 0.00 × x6 ,5
+0.42 × x7 ,1 + 0.18 × x7 ,2 + 0.08 × x7 ,3 + 0.03 × x7 ,4 + 0.01 × x7 ,5
+0.31 × x8 ,1 + 0.10 × x8 ,2 + 0.03 × x8 ,3 + 0.01 × x8 ,4 + 0.00 × x8 ,5
+0.04 × x1′,2 ,1,1 + 0.17 × x1′,3 ,1,1 + 0.08 × x1′,4 ,1,1 + 0.12 × x1′,5 ,1,1 + 0.03 × x1′,6 ,1,1
+0.32 × x1′,7 ,1,1 + 0.24 × x1′,8 ,1,1 + 0.03 × x2′ ,3 ,1,1 + 0.02 × x2′ ,4 ,1,1 + 0.04 × x2′ ,5 ,1,1
+0.00 × x2′ ,6 ,1,1 + 0.13 × x2′ ,7 ,1,1 + 0.07 × x2′ ,8 ,1,1 + 0.03 × x3′ , 4 ,1,1 + 0.14 × x3′ ,5 ,1,1
+0.03 × x3′ ,6 ,1,1 + 0.29 × x3′ ,7 ,1,1 + 0.16 × x3′ ,8 ,1,1 + 0.06 × x 4′ ,5 ,1,1 + 0.02 × x 4′ ,6 ,1,1
+0.12 × x 4′ ,7 ,1,1 + 0.12 × x 4′ ,8 ,1,1 + 0.02 × x5′ ,6 ,1,1 + 0.20 × x5′ ,7 ,1,1 + 0.09 × x5′ ,8 ,1,1
+0.04 × x6′ ,7 ,1,1 + 0.02 × x6′ ,8 ,1,1 + 0.28 × x7′ ,8 ,1,1 + y ) − b.
(4.39)
With respect to sample 2, LP model Constraint (4.23) appears as follows:

errori ≥ (b + 1)
−( x1,1 × 0.28 + x1,2 × 0.08 + x1,3 × 0.02 + x1,4 × 0.01 + x1,5 × 0.00
+ x2 ,1 × 0.22 + x2 ,2 × 0.05 + x2 ,3 × 0.01 + x2 ,4 × 0.00 + x2 ,5 × 0.00
+ x3 ,1 × 0.09 + x3 ,2 × 0.01 + x3 ,3 × 0.00 + x3 ,4 × 0.00 + x3 ,5 × 0.00
+ x 4 ,1 × 0.22 + x 4 ,2 × 0.01 + x 4 ,3 × 0.01 + x 4 , 4 × 0.00 + x 4 ,5 × 0.00
+ x5 ,1 × 0.14 + x5 ,2 × 0.02 + x5 ,3 × 0.00 + x5 ,4 × 0.00 + x5 ,5 × 0.00
+ x6 ,1 × 0.05 + x6 ,2 × 0.00 + x6 ,3 × 0.00 + x6 , 4 × 0.00 + x6 ,5 × 0.00
+ x7 ,1 × 0.12 + x7 ,2 × 0.01 + x7 ,3 × 0.00 + x7 ,4 × 0.00 + x7 ,5 × 0.00
+ x8 ,1 × 0.06 + x8 ,2 × 0.00 + x8 ,3 × 0.00 + x8 ,4 × 0.00 + x8 ,5 × 0.00
+ x1′,2 ,1,1 × 0.10 + x1′,3 ,1,1 × 0.08 + x1′,4 ,1,1 × 0.18 + x1′,5 ,1,1 × 0.08 + x1′,6 ,1,1 × 0.03
+ x1′,7 ,1,1 × 0.10 + x1′,8 ,1,1 × 0.05 + x2′ ,3 ,1,1 × 0.03 + x2′ ,4 ,1,1 × 0.08 + x2′ ,5 ,1,1 × 0.05
+ x2′ ,6 ,1,1 × 0.01 + x2′ ,7 ,1,1 × 0.09 + x2′ ,8 ,1,1 × 0.03 + x3′ ,4 ,1,1 × 0.03 + x3′ ,5 ,1,1 × 0.03
+ x3′ ,6 ,1,1 × 0.01 + x3′ ,7 ,1,1 × 0.03 + x3′ ,8 ,1,1 × 0.01 + x 4′ ,5 ,1,1 × 0.07 + x 4′ ,6 ,1,1 × 0.04
+ x 4′ ,7 ,1,1 × 0.07 + x 4′ ,8 ,1,1 × 0.05 + x5′ ,6 ,1,1 × 0.01 + x5′ ,7 ,1,1 × 0.03 + x5′ ,8 ,1,1 × 0.01
+ x6′ ,7 ,1,1 × 0.01 + x6′ ,8 ,1,1 × 0.00 + x7′ ,8 ,1,1 × 0.02 + y ).

(4.40)
108 Big Data Analytics Using Multiple Criteria Decision-Making Models

TABLE 4.7
Performance of the Proposed LP-Based Classifiers (LP Model 1 and LP Model 2)
for the Recommendation Data Set
LP-Based Classifier (LP Model 1) LP-Based Classifier (LP Model 2)
Objective 1 49.35 Objective 1 52.22
Objective 2 50.39 Objective 2 52.11
Objective 3 48.30 Objective 3 52.32
Note: LP Model 1 and LP Model 2 cannot address multiobjective optimization.

To calculate the objectives of the proposed LP-based classifiers (LP Model 1

and LP Model 2), we split the data into two sets, training data set (70%: 17,500
observations) and test data set (30%: 7500 observations), and the calculated
objectives with respect to test data set are listed in Table 4.7.
To calculate the objectives of the proposed LP-based two-phase classifier,
we split the data into three sets, training data set (60%: 15,000 observations),
test data set (10%: 2500 observations), and validation data set (30%: 7500
observations), and the set of non-dominated solutions for the test data set (see
Part 1 of Phase 2) is listed in Table 4.8, and the same is shown in Figure 4.1,

TABLE 4.8
Performance of the Proposed LP-Based Two-Phase Classifier for the Test Data Set
(2500 Observations) with Respect to Three Objectives
Non-Dominated Set of Solutions
Solution c Value Objective 1 Objective 2 Objective 3
1 0.00 72.52 93.80 50.93
2 0.05 73.84 92.77 54.63
3 0.10 74.64 91.82 57.21
4 0.15 75.76 90.95 60.35
5 0.20 77.04 89.52 64.38
6 0.25 78.44 88.64 68.09
7 0.30 78.76 86.74 70.67
8 0.35 79.32 84.83 73.73
9 0.40 79.64 83.40 75.83
10 0.45 79.96 81.41 78.49
11 0.50 80.16 79.90 80.42
12 0.55 80.16 77.12 83.24
13 0.60 80.20 75.06 85.41
14 0.65 80.32 72.99 87.75
15 0.70 80.16 71.01 89.44
16 0.75 79.52 68.39 90.81
17 0.80 78.44 64.89 92.18
18 0.85 77.40 61.87 93.15
19 0.90 76.04 58.38 93.96
20 0.95 75.08 55.76 94.68
21 1.00 74.00 52.74 95.57
Linear Programming (LP)-Based Two-Phase Classifier 109

Objective 3

90.0000

80.0000

70.0000

60.0000

60.00
70.00
80.00 74.00
76.00
Objective 2 90.00 80.00 78.00
Objective 1

FIGURE 4.1
Non-dominated set of solutions with respect to test data set.

Objective 3

90.0000

80.0000

70.0000

60.0000

60.00
70.00 74.00
80.00 76.00
Objective 2 90.00 80.00 78.00
Objective 1

FIGURE 4.2
Non-dominated set of solutions with respect to validation data set.
110 Big Data Analytics Using Multiple Criteria Decision-Making Models

and the set of non-dominated solutions for the validation data set (see Part 2
of Phase 2) is listed in Table 4.9, and the same is shown in Figure 4.2.
We can also observe from Table 4.8 that (in the case of single objective) if
we choose Objective 1 as the primary objective under consideration, then the
proposed classifier (in Part 1 of Phase 2) selects 0.50 as the best c value; and if
we choose Objective 2 as the primary objective under consideration, then the
proposed classifier (in Part 1 of Phase 2) selects 0.00 as the best c value; and
if we choose Objective 3 as the primary objective under consideration, then
the proposed classifier (in Part 1 of Phase 2) selects 1.00 as the best c value.
We also compare the accuracy of the proposed LP-based classifiers with
the accuracy of ANN. We consider the ANN model with sigmoid function
as the activation function for artificial neurons, in which the number of hid-
den layers is 1, and the number of input neurons is set to 8 to match with the
number of attributes in the data set, and the number of output neurons is set
to 2 to match with the number of categories, and the number of neurons in
the hidden layer is 8. We train the ANN using training data set (70%: 17,500
observations), and we present the results of the ANN with respect to the test

TABLE 4.9
Performance of the Proposed LP-Based Two-Phase Classifier for the Validation Data
Set (7500 Observations) with Respect to Three Objectives
Non-Dominated Set of Solutions
Solution c Value Objective 1 Objective 2 Objective 3
1 0.00 72.85 94.40 51.54
2 0.05 73.88 93.22 54.75
3 0.10 74.99 92.23 57.93
4 0.15 76.01 90.97 61.22
5 0.20 77.00 89.87 64.27
6 0.25 77.79 88.58 67.11
7 0.30 78.40 87.02 69.87
8 0.35 79.08 85.55 72.68
9 0.40 79.43 83.51 75.38
10 0.45 79.96 81.55 78.38
11 0.50 80.01 79.30 80.72
12 0.55 79.96 77.00 82.89
13 0.60 79.83 74.88 84.72
14 0.65 79.57 72.49 86.58
15 0.70 78.97 69.62 88.22
16 0.75 78.31 66.81 89.68
17 0.80 77.57 63.78 91.22
18 0.85 76.68 60.78 92.41
19 0.90 76.07 58.39 93.55
20 0.95 75.15 55.58 94.51
21 1.00 74.27 52.90 95.41
Linear Programming (LP)-Based Two-Phase Classifier 111

TABLE 4.10
Performance of the Artificial Neural Networks for the Recommendation Data Set
Solution Objective 1 Objective 2 Objective 3
1 79.00% 81.47% 76.51%

data set (30%: 7500 observations), for the considered objectives Objective 1,
Objective 2, and Objective 3 in Table 4.10. The results indicate that the pro-
posed LP-based two-phase classifier performs well, and it is able to find a
non-dominated set of solutions with respect to multiple objectives (Objective
1, Objective 2, and Objective 3). We can also observe that solution 9 (in Table
4.9) by the proposed LP-based two-phase classifier clearly dominates the
solution by the ANN.

• Note that in Figures 4.1 and 4.2, Objective 1 maximizes the total
accuracy with respect to both categories, Objective 2 maximizes
the accuracy with respect to category 1, and Objective 3 maximizes
the accuracy with respect to category 0.

4.5 Summary
This chapter proposes MILP-based classifier and LP-based classifiers for
binary classification, and when we compare the accuracy of all the proposed
LP-based classifiers with ANN, the results indicate that the proposed LP-based
two-phase classifier is able to give better results. Consequently, the proposed
LP-based two-phase classifier is able to handle data that are not inherently lin-
early separable, unlike the conventional MILP-based and LP-based classifiers.
The salient contributions of the proposed LP-based two-phase classifier are
in terms of treating the decision variables as unrestricted in sign; accounting
for the contribution of attributes from their interaction effects and the con-
tribution of attributes from their higher-order polynomial degrees; treating
the classification threshold/cut-off as a decision variable; converting the band-
width of boundary of threshold to a crisp boundary with the consideration
of multiple objectives. The proposed LP-based two-phase classifier considers
such multiple objectives because in the application areas such as medical diag-
nosis, absence of an alarm (failing to predict the category 1) is more serious
than a false alarm. The proposed LP-based two-phase classifier is an efficient
method in terms of the ability to solve (linear programming model) the under-
lying classification problem, and thus it can be effectively used in conjunction
with more sophisticated and computationally demanding approaches such as
random forest, support vector machines, and ANN to improve the accuracy
further on the data that has high variety and high volume.
112 Big Data Analytics Using Multiple Criteria Decision-Making Models

Acknowledgment
We are thankful to the reviewers and the editors for their valuable comments
and suggestions to improve our chapter.

References
Freed, N. and Glover, F. 1986. Evaluating alternative linear programming models
to solve the two-group discriminant problem. Decision Sciences, 17(2), 151–162.
Maas, A. L., Daly, R. E., Pham, P. T., Huang, D., Ng, A. Y., and Potts, C. 2011. Learning
word vectors for sentiment analysis. In Proceedings of the 49th Annual Meeting
of the Association for Computational Linguistics: Human Language Technologies—
Volume 1, 142–150. Association for Computational Linguistics, Portland,
Oregon.
Shi, Y., Wise, M., Luo, M., and Lin, Y. 2001. Data mining in credit card portfolio man-
agement: A multiple criteria decision making approach. In Köksalan, M. and
Zionts, S. (Eds.), Multiple Criteria Decision Making in the New Millennium, 427–
436. Springer Berlin Heidelberg, Ankara, Turkey.
Yan, N., Wang, Z., Shi, Y., and Chen, Z. 2006. Nonlinear classification by linear pro-
gramming with signed fuzzy measures. In 2006 IEEE International Conference
on Fuzzy Systems, 408–413. IEEE, Vancouver, British Columbia, Canada.
5
Multicriteria Evaluation of
Predictive Analytics for Electric
Utility Service Management

Raghav Goyal, Vivek Ananthakrishnan,

Sharan Srinivas, and Vittaldas V. Prabhu

CONTENTS
5.1 Introduction................................................................................................. 114
5.2 Layout of Electric Supply Distribution Network................................... 115
5.3 Predictive Analysis..................................................................................... 119
5.3.1 Decision Trees................................................................................. 120
5.3.2 Logistic Regression......................................................................... 120
5.3.3 Boosting............................................................................................ 120
5.3.4 Random Forest................................................................................ 120
5.3.5 Support Vector Machines.............................................................. 121
5.3.6 Artificial Neural Networks........................................................... 121
5.4 Evaluating Prediction Models Using Multicriteria
Decision-Making Techniques................................................................... 122
5.4.1 Performance of Classification Model........................................... 122
5.4.2 Criteria for Model Evaluation....................................................... 123
5.4.3 Multicriteria Ranking Techniques............................................... 123
5.5 Data Description......................................................................................... 124
5.5.1 Weather Data................................................................................... 124
5.5.2 Power Outage Data......................................................................... 125
5.6 Experimental Results................................................................................. 125
5.7 Smart Staffing.............................................................................................. 130
5.7.1 Objective Functions........................................................................ 130
5.7.1.1 Objective 1: Minimize Staffing Costs............................ 130
5.7.1.2 Objective 2: Minimize Power Restoration Time.......... 131
5.7.2 Model Constraints.......................................................................... 131
5.7.2.1 Restriction on Workforce Capacity................................ 131
5.7.2.2 Worker Type Requirement.............................................. 131
5.7.2.3 Nonnegativity Restriction.............................................. 131
5.8 Conclusions.................................................................................................. 132
References.............................................................................................................. 132

113
114 Big Data Analytics Using Multiple Criteria Decision-Making Models

5.1 Introduction
Utility companies are responsible for the infrastructure of the power
­delivery system and power interruptions disrupt customers as well as cause
significant economic losses. In the United States, the estimated cost of power
interruptions is $79 billion per year (LaCommare and Eto, 2006). The out-
age costs are directly proportional to the customer’s dependence upon elec-
tricity during an outage. With annual electricity use in a typical U.S. home
increasing 61% since 1970, it is becoming increasingly important to reduce
and prevent outages (Swaminathan and Sen, 1998). Outage costs vary sig-
nificantly depending upon the outage attributes such as frequency, duration,
and intensity of the outage. In this chapter, an outage is considered to be
a complete or total loss of service, typically resulting from a distribution-
related cause or a transmission failure.
The priority of every organization is twofold: providing customers the best
possible product at the lowest cost while maintaining the quality. The situ-
ation is not very different in the case of a service industry. Electric utility
­companies aim to optimize every form of the electric service system. Every
company tries to provide the most reliable service in the form of consistent
and uninterrupted service to its customers while reducing the cost of supply
and maximizing profits. Companies often have to deal with problems such
as downtime or outage time due to a failure in the supply system. Therefore,
better recovery planning and forecasting of outages is necessary to provide
reliable service to customers.
There are several reasons attributed to an electric power outage, which
range from equipment failure and overloading of the line to weather-
related events. Power systems are most vulnerable to storms and extreme
weather events. Seasonal storms combined with wind, snow, rain, ice, etc.
can cause significant outages. Data on weather-related outages have been
used in the past to estimate the costs of an outage and the impact it has
on consumers. According to past weather-related outage data, 90% of cus-
tomer outage-­minutes are owing to events which affect the local distribu-
tion systems, while the remaining 10% are from generation and transmission
problems (Campbell, 2012). Electric utility companies can reduce the outages/­
damages resulting from severe weather conditions by enhancing the overall
­condition of the power delivery system and better prediction of the outage.
There has been a lot of research that focuses on predicting outages as shown
in Table 5.1.
However, most of the previous work does not use hourly weather fore-
cast to predict short-term outages. Increasingly, companies are looking
to tackle outages due to both local distribution systems and larger trans-
mission ­systems by developing strategies to reduce or prevent outages.
Short-term forecasts of an electric power outage and the cost param-
eters associated with the outage would help companies optimize their
Multicriteria Evaluation of Predictive Analytics 115

TABLE 5.1
Summary of Past Research Done
Predicting Predicting
Power Outage Power-Related
Cost Analysis for Extreme Damages for
of Power Weather Extreme Weather
Author Name Outages Conditions Conditions
Balijepalli et al. (2005) ✓
Cerruti and Decker (2012) ✓
Davidson et al. (2003) ✓
DeGaetano et al. (2008) ✓
Huang et al. (2001) ✓
J. Douglas (2000) ✓
LaCommare and Eto (2006) ✓
Li et al. (2010) ✓
Liu et al. (2007) ✓
Reed (2008) ✓ ✓
Reed et al. (2010) ✓ ✓
Sullivan et al. (1996) ✓
Winkler et al. (2010) ✓ ✓
Zhou et al. (2006) ✓
Zhu et al. (2007) ✓

manpower and resource planning. The advancement of science and tech-

nology has led to accurate short-term weather forecasts (up to 72 hours).
Data from National Digital Forecast Database for 5 years of U.S. weather
would be over 267 TB. The focus of this chapter is to leverage such short-
term forecasts by combining it with analytics for electric utility services.
These concerns have helped to stimulate research activity in the area of
electric power supply system’s m
­ odeling and analysis and are the motiva-
tion for this chapter.

5.2  Layout of Electric Supply Distribution Network

In a product supply chain, the product is manufactured using raw material
from suppliers, then passed on to the retailers and finally to the end cus-
tomer through a distribution network. A service supply chain in the form
of an electric distribution system involves customer–supplier service supply
relationships. Figure 5.1 shows the flow of the service, product, and informa-
tion through the supply chain.
116 Big Data Analytics Using Multiple Criteria Decision-Making Models

Supplier

Service Service
system Customer
provider

Information Service
flow product flow

FIGURE 5.1
Service supply chain.

The electric power supply industry is an integral part of the service system
industry and also important for the economy as all businesses rely h ­ eavily
on electric power to operate. The Energy Information Administration
­predicts that there would be an increase of 29% in the electricity demand in
the United States between 2012 and 2040 (Sieminski, 2014). The power gen-
erators own and operate electricity-generating facilities or power plants and
sell the power produced to the utility service providers. These service pro-
viders are the key players in the network as they are responsible for provid-
ing a reliable source of electric power while ensuring uninterrupted service
at an ­affordable cost.
The utility service provider is directly involved in the design of the ser-
vice and is answerable directly to the customer. For instance, the products
and services in an electric power utility service supply chain network are
limited to the electric power supplied and transmission services provided.
The ­service utility in a particular geographical area handles the transmis-
sion and distribution of electricity to the end users through a vertically
­integrated structure. The various decision makers in this system such as the
power generators, the power suppliers, the transmitters, and the customers
operate in a decentralized system. A depiction of the distribution network
for electric power is shown in Figure 5.2 (Nagurney and Matsypura, 2007;
Nagurney et al., 2007).
Understanding the outline of the electric power supply system is crucial
for identifying the critical points in the service supply network:

• Power generation: The electric power generating station could be

any power plant (gas, oil, nuclear, thermal, etc.) that converts fuel
sources into electricity. The power generated is then stepped up to
as high as 500,000 V and passed through transmission lines to the
distribution lines.
 Subtransmission
customer
Transmission lines 26 kV and 69 kV
765, 500, 345, 230, and 138 kV

Substation Primary customer

Generating station step-down 13 kV and 4 kV
transformer
Multicriteria Evaluation of Predictive Analytics

Secondary customer
Transmission
Generator step-up 120 kV and 240 kV
customer
transformer
138 kV or 20 kV

FIGURE 5.2
Basic electrical distribution system. (From U.S.–Canada Power System Outage Task Force, Final Report on the August 14, 2003, Blackout in the United States
and Canada: Causes and Recommendations, April 2004.)
117
118 Big Data Analytics Using Multiple Criteria Decision-Making Models

• Transmission: The transmission stage of the network involves trans-

mitting electricity over long distances at very high voltages since
energy transfer efficiency is high at very high voltages. Large players
in the transmission market operate thousands of miles of transmis-
sion lines.
• Distribution: The electricity transmitted via the transmission lines
is then stepped down using a step-down transformer. The dis-
tribution line networks, which are made up of the feeders and
laterals, deliver electricity to commercial, industrial, and residen-
tial customers. These feeders and laterals handle current with a
voltage range of 13–23 kV, while the service lines handle a volt-
age range of 120–480 V. The distribution costs for industrial and
commercial users are less due to the high-voltage power being
supplied.

Power outages could be broken down into three components: num-

ber of customers affected, total minutes of power outage, and number of
­interruptions. In addition, there are three customer types recognized by
electric u­ tilities—industrial, commercial, and residential. Depending on the
­geographic location, an electric utility company can serve a particular kind
of customer or could serve all three at the same time.
Earlier, utility companies primarily focused on reducing cost. However,
companies have shifted their focus to include supply assurance and risk
management. One of the key metrics used to evaluate the supply assur-
ance and risk management is system average interruption duration index,
measured in units of time. The median value for North American utilities is
approximately 1.50 hours (90 minutes). Typically, energy companies strive to
improve this measure by directly reducing the total power outage minutes
along with the number of customers affected by those outages. The predic-
tive analytic tools used in this chapter aim to predict power outage occur-
rences by using the hourly weather forecasts. This information could then be
used to reduce the interruption minutes by improving workforce planning
of the repair crew.
As discussed in Chapter 3, analytics can be grouped into three types:
descriptive analytics, predictive analytics, and prescriptive analytics.
Electric utilities typically use descriptive analytics such as Pareto charts
for descriptive statistics and data visualization to derive insights about the
most frequent causes of power interruptions. This chapter mainly focuses
on predictive analytics that uses weather forecasts to predict power inter-
ruptions. In this chapter, prescriptive analytics is limited to formulating
a multiple criteria mathematical programming (MCMP) model that mini-
mizes staffing the cost and duration of power interruption. The result-
ing data-driven decision p ­ rocess, as suggested in Chapter 3, would be as
follows:
Multicriteria Evaluation of Predictive Analytics 119

Chapter 3 Analytics for Electric Utility

1. Knowledge of relationship between KPIs and factors CMI reduction
would then provide the decision maker with
appropriate actionable items. Identify the problem or
opportunity for value creation
2. Identify sources of data (primary as well secondary Weather forecast
data) Outage data
3. Preprocess the data for missing and incorrect data. Combine the two data sources
Prepare the data for analytics model building, if to forecast outages
necessary transform the data
4. Build the analytical models and identify the best MCDM problem in selecting
model using model validation predictive analytic technique
5. Communicate the data analysis output and decisions Ranked list
effectively
6. Implement solution/decision Bicriteria optimization for
staffing decision

5.3  Predictive Analysis

Traditionally, electric utility companies used multiple-level Pareto charts to
identify the root cause of the outage and develop a strategy to minimize
­disruptions. However, in recent times, weather predictions have improved
substantially. For example, 96–120-hour predictions of a weather-related
event are nearly as accurate as the actual occurrence.
In addition to other attributes, companies could utilize hourly weather
forecasts to predict power outage occurrences by the hour and use them to
efficiently plan the usage of their resources.
The prediction of an outage using hourly weather forecasts can be
treated as a binary classification problem (0: no outage; 1: outage). In order
to predict the outage, it is necessary to have a data set that includes all the
input variables (weather attributes) and the target values (either 0: no out-
age or 1: outage). The initial data set is split into a training set and testing
set. Each instance in the training set contains one target value (either 0:
no outage or 1: outage) and several input parameters (weather attributes).
A machine-learning ­classifier uses the training data set and trains the
prediction model to learn the underlying relationship between the inputs
and the targets. A  learned classifier uses the inputs of the testing data
set to predict the ­outage occurrence. Finally, the classifier is evaluated by
comparing the actual o ­ utput and the predicted output. There are several
classification techniques and ­algorithms that are used in the world of pre-
dictive analytics. Several prominent algorithms and their results are dis-
cussed below.
120 Big Data Analytics Using Multiple Criteria Decision-Making Models

5.3.1  Decision Trees

Decision tree algorithms are one of the earliest classification algorithms that
have been used in predictive modeling (Quinlan, 1986). The algorithm works
on the philosophy of dividing the main problem (root node) into smaller
successive decisions based on a unique node until a particular class gains a
majority. This division is done until any further partitioning being done is
determined to be useless. The algorithm follows a top-down, greedy search
approach with the selection of the attribute that best classifies the set as the
root node. The advantages of using a decision tree algorithm are (1) it can be
applied to any type of data, (2) the final structure of the classifier is quite sim-
ple and can be easily interpreted by decision makers, the resulting trees can be
easily used to obtain a better understanding of the phenomenon in question.

5.3.2  Logistic Regression

The logistic regression model is used to predict a dichotomous outcome
using one or more independent variables (Menard, 2002). The outcome
or the response variable Y is the outage occurrence (i.e., 0 or 1), while the
­independent variables (X) are input parameters (i.e., weather attributes).
Therefore, in logistic regression, the probability of power outage and no
­outage are computed as shown below:

e(β0 + β⋅X ) 1
p(Y = 1) = = , (5.1)
1 + e(β0 + β⋅X ) 1 + e−(β0 + β⋅X )

p(Y = 0) = 1 − P(Y = 1). (5.2)

5.3.3 Boosting
Boosting is one of the most recent and important developments in the
domain of predictive classification techniques. It works on the principle of
sequentially applying a classification algorithm to reweighted versions of the
training data, and then taking a weighted majority vote of the sequence of
classifiers produced as a result of this sequential process. For the two-class
problem, boosting can be viewed as an approximation to additive modeling
on the logistic scale using maximum Bernoulli likelihood as a criterion. This
simple strategy is found to yield drastic improvements in results for many
classification algorithms due to statistic principles of additive modeling and
maximum likelihood estimation.

5.3.4  Random Forest

Random forest is a popular ensemble learning method for generating
­classification and regression models (Breiman, 2001; Ho, 1995). The method
Multicriteria Evaluation of Predictive Analytics 121

works by constructing a multitude of decision trees using the training set

and outputting the mode of the class and the mean prediction of the individ-
ual trees for a classification problem and regression problem, respectively.
In other words, random forests average multiple deep decision trees, trained
on d ­ ifferent parts of the same training set. The objective of reducing the
­variance comes at the expense of a small increase in the bias and some loss
of interpretability. However, this greatly boosts the performance of the final
model and it also corrects the problem of overfitting, which is a common
occurrence in decision trees. The general technique of bootstrap aggregating
or bagging is implemented in the training algorithm.

5.3.5  Support Vector Machines

Support vector machines (SVMs) are supervised learning methods used
for classification and regression tasks that generate nonoverlapping parti-
tions and usually employ all the attributes (Gunn, 1998; Gunn et  al., 1997;
Vapnik and Vapnik, 1998). The entity space is partitioned in a single pass and
is based on maximum margin linear discriminants, similar to probabilistic
approaches, but does not consider the dependencies among attributes. SVMs
have gained popularity as they are based on the structural risk minimiza-
tion (SRM) principle. This gives SVMs greater generalization ability, which
is the goal in statistical learning. SVMs rely on preprocessing the data to
­represent patterns in a high dimension. Data from two categories can always
be s­ eparated by a hyperplane when an appropriate nonlinear mapping is
used. One maximizes the distance between itself and the nearest target
value (optimal separating hyperplane). The basic idea behind SVM classifier
is to choose the hyperplane that has the maximum margin (distance from
itself to the nearest class).

5.3.6  Artificial Neural Networks

Artificial neural networks (ANNs) are used to estimate functions that can
depend on a large number of inputs and are generally unknown (Boger and
Guterman, 1997; Braspenning et al., 1995). ANNs assign numeric weights to
the connections between the input and output variables that can be tuned
based on experience. There are many different kinds of learning rules used
by neural networks, with the delta rule being the most popular. The delta
rule learns by updating the weights depending upon the error magnitude
(i.e., the difference between the predicted output and the actual output).
An initial guess and subsequent error corrections due to weight adjustments
lead to a final optimal weight. ANNs provide an analytical alternative to
conventional techniques which are often limited by strict assumptions of
normality, linearity, variable independence, etc. ANNs can capture many
kinds of relationships and thus allow the user to easily model phenomena
that are not easily explainable.
122 Big Data Analytics Using Multiple Criteria Decision-Making Models

5.4  Evaluating Prediction Models Using

Multicriteria Decision-Making Techniques
Multicriteria decision making (MCDM) has been heavily used in supplier
selection to improve the supply chain of any organization. Comprehensive
literature review on different MCDM techniques used for supplier selec-
tion and evaluation has been done in the past (Ho et al., 2010). In ­addition
to s­ upply chain management, MDCM techniques are used in several other
areas/fields such as healthcare, marketing, and financial management
(Zopounidis and Doumpos, 2002). For instance, MCDM techniques are used
in financial ­management for developing an efficient portfolio, credit-risk
assessment, etc. However, none of the previous research articles use MCDM
techniques to evaluate and select the best predictive model. This chapter
explores how different MCDM techniques such as L2 metric, Borda count,
and rating method could be used to pick the best predictive model, given
multiple selection c­ riteria such as accuracy, specificity, and sensitivity.

5.4.1  Performance of Classification Model

The classification methods discussed in Section 5.3 are used to predict
the occurrence of an interruption, given an hourly weather condition.
A  ­confusion matrix is generally used to illustrate the performance of the
classification model, and is generated by comparing the predicted outcome
and the actual outcome. Table 5.2 shows the layout of a confusion matrix.
The elements of the confusion matrix are as follows:

• True negatives (TN) represent the number of times the prediction

model accurately predicted 0 (no interruption).
• False positives (FP) represent the number of times the prediction
model  predicted 1 (interruptions) when in reality there were no
interruptions (represented by 0). This type of error is also called
type I error.
• True positives (TP) represent the number of times the model
­accurately predicted 1 (interruptions).

TABLE 5.2
Confusion Matrix
Actual Outcome
Predicted outcome 0 (no outage) 1 (outage)
0 (no outage) TN FN
1 (outage) FP TP
Multicriteria Evaluation of Predictive Analytics 123

• False negatives (FN) represent the number of times model pre-

dicted 0 (no interruptions) when in reality there were interruptions
(­represented by 1). This type of error is called type II error.

5.4.2  Criteria for Model Evaluation

The criteria considered for model evaluation are accuracy, area under the
receiver operating characteristic (ROC) curve, sensitivity, specificity, and
precision. Based on the confusion matrix, the values of the criteria for each
prediction model are determined, and are as follows:

• Accuracy {(TN + TP)/(TN + TP + FP + FN)}: This is a measure of

overall accuracy of the model. This is calculated by dividing
­
­accurate predictions by total number of instances.
• Area under the ROC curve (AUC): It is a standard measure of the
predictive accuracy. If it is below 0.50, then the model prediction is
worse than an unbiased flip of a coin.
• Sensitivity {TP/(FN + TP)}: It measures the proportion of outages that
were correctly classified.
• Specificity {TN/(TN + FP)}: It measures the proportion of no outages
that were accurately classified.
• Precision {(TP/(FP + TP)}: It measures the relevance of the positively
classified (i.e., 1: outage) instances.

5.4.3  Multicriteria Ranking Techniques

The following multicriteria ranking methods are used to determine the best
classification method:

• Rating method
• Borda count
• L2 metric method

Rating Method: In this method, the participants rate each parameter

on a scale of 1–10 (1 being least important and 10 being most important).
The  weights of each criterion are then calculated by normalizing the rat-
ings (Ravindran,  2016). The final score of each prediction model is the
weighted sum of the criteria values, and the model with the highest score is
ranked first.
Borda Count: In this method, the decision maker is asked to rank all the
criteria based on their importance. Therefore, the n criteria are ranked from
1 (most important) to n (least important). The most important criterion is
given n points, the next most important criterion gets n−1 points, and the
124 Big Data Analytics Using Multiple Criteria Decision-Making Models

least important criterion gets 1 point. In the presence of multiple decision

makers, the number of times a criterion is ranked at a ­particular rank on the
survey is multiplied with its corresponding points to determine the overall
points for each of the criteria. The weights are then c­ omputed by dividing
the points of each criterion by the sum of points of all the criteria. The final
score of each prediction model is the weighted sum of the criteria values, and
the model with the highest score is ranked first.
L2 Metric Method: It measures the distance between the vector of ideal
­solutions and the vector representing criteria values for each prediction
model as shown below (Ravindran, 2016).

1/2
 n

L2 (k ) = 

∑
j =1
( x jk − y j )

, (5.3)

where
k denotes the prediction model, j denotes the criterion, and n is the total
number of criteria.
xjk is the value of criterion j for prediction model k.
xj is the ideal value of criterion j.

Therefore, each prediction model will have a score, and the model with the
least L2 score is ranked first, followed by the next smallest L2 score, etc.

5.5  Data Description

The primary data set used in this chapter is the hourly recordings of weather
conditions for all the days in 2013, at the location of the electric utility.
The other data set contains time-stamped details of the complaints reported
by customers in the same year. Each instance of the complaint is captured
along with various other information such as equipment failure, causes of
failure, parts affected, and duration of the outage. The detailed variables of
these data files are described below.

5.5.1  Weather Data

The hourly weather information collected for the year 2013 is described in
Table 5.3.
The variable “Event” is further divided into four binary variables, namely,
fog, rain, tornado (hurricane), thunderstorm, to describe the type of event
that occurred. The variable “Event” takes a value 1 if at least one of the four
Multicriteria Evaluation of Predictive Analytics 125

TABLE 5.3
Weather Data Variables Description
Variable Name Variable Description Variable Type
Temperature Temperature of the location (°F) Numeric
Heat index Index combining effect of temperature and humidity (°F) Numeric
Dew point Dew point of the location (°F) Numeric
Humidity Humidity of the location (%) Numeric
Pressure Atmospheric pressure (in) Numeric
Visibility Visibility of the location (miles) Numeric
Wind speed Speed of the wind flowing at the location (mph) Numeric
Gust speed Gust speed of the wind at the location (mph) Numeric
Precipitation Amount of precipitation at the location (in) Numeric
Event Special weather condition at the location Binary

event types is 1. The final data set has over 8700 instances with 697 instances
experiencing a weather-related event. A sample of the final data set is shown
in Figure 5.3.

5.5.2  Power Outage Data

The power outage data set provided by an electric utility company contains
data on the outages experienced by the company in the year 2013. The data
set is time-stamped with each instance in this data set being referred to as a
ticket (outage).
The data set contains more than 42,000 such instances with outages
­affecting 1,989,032 customers. The total minutes of disruption (MI) for tickets
are totaled at 2,362,144 while the customer minutes interrupted (CMI) for
the entire year is 48,355,553 minutes. In this chapter, binary MI, which is
assigned a value of 0 when there is no outage and value of 1 when there is
an outage, is selected as the response variable (output). This information has
been rolled up hourly to integrate with previously mentioned weather data.
The predictive models used in this chapter aim to predict the binary MI
(response variable) with high accuracy. A sample of the final data table used
is shown in Figure 5.4.

5.6  Experimental Results

In this section, the criteria values for each of the prediction models are
obtained and the models are evaluated using multicriteria ranking
­techniques. The experimental results are performed by using the data set
obtained from the electric utility company. The prediction models are coded
 126

Dew Wind Gust Thunder

Date Hour Temp Heat Index Humidity Pressure Visibility Precipitation Event Rain Fog Tornado
Point Speed Speed storm

5/2/2013 7 73.90 0.00 69.10 0.85 29.89 10.00 11.50 0.00 0.04 1 1 0 0 0

5/2/2013 8 75.55 0.00 71.35 0.87 29.91 10.00 11.55 0.00 0.00 0 0 0 0 0

5/2/2013 9 73.75 0.00 71.03 0.91 29.91 10.00 10.95 5.48 0.00 1 1 1 0 0

5/2/2013 10 75.83 0.00 71.48 0.87 29.91 10.00 9.20 0.00 0.00 1 1 0 0 0

5/2/2013 11 74.43 0.00 71.03 0.89 29.91 6.50 15.28 11.50 0.00 1 1 0 0 0

5/2/2013 12 74.43 0.00 70.13 0.92 29.91 0.85 9.80 0.00 0.26 1 1 1 0 0

5/2/2013 13 71.68 0.00 69.84 0.94 29.90 2.60 8.30 0.00 0.82 1 1 1 1 0

5/2/2013 14 71.43 0.00 69.87 0.95 29.91 5.00 3.87 0.00 0.03 1 1 1 0 0

5/2/2013 15 69.57 0.00 66.43 0.90 29.90 3.33 8.47 6.13 0.16 1 1 1 0 0

FIGURE 5.3
Sample data table for historical weather.
Big Data Analytics Using Multiple Criteria Decision-Making Models
 Temp Heat Index Dew Point Humidity Pressure Visibility Wind Speed Gust Speed Precipitation Event Rain Thunderstorm Fog Tornado Binary MI
73.9 0 61 0.64 30.18 10 8.1 0 0 0 0 0 0 0 0
73.2 0 63.3 0.71 30.2 8.5 8.1 0 0 1 1 0 0 0 1
70 0 66 0.87 30.21 3 6.9 0 0.02 1 1 0 0 0 1
68 0 64.9 0.9 30.18 6 4.6 0 0.05 1 1 0 0 0 0
69.1 0 66 0.9 30.15 5 5.8 0 0.02 1 1 0 0 0 0
69.27 0 65.53 0.88 30.13 4.33 5.8 0 0.06 1 1 0 0 0 1
71.1 0 66 0.84 30.12 7 0 0 0 1 1 0 0 0 0
70 0 66.9 0.9 30.11 7 0 0 0.01 0 0 0 0 0 1
71.35 0 66.1 0.84 30.11 10 3.5 0 0 0 0 0 0 0 1
69.45 0 66.1 0.89 30.12 6.5 6.35 0 0 1 1 0 0 0 0
Multicriteria Evaluation of Predictive Analytics

68.9 0 66.1 0.91 30.18 10 5.8 0 0 0 0 0 0 0 1

68 0 66.13 0.94 30.14 8.67 4.23 0 0 1 1 0 0 0 1
66.9 0 64.9 0.93 30.13 10 5.8 0 0 0 0 0 0 0 0
66.9 0 66 0.97 30.13 10 4.6 0 0 0 0 0 0 0 1

FIGURE 5.4
Sample data table for prediction model.
127
128 Big Data Analytics Using Multiple Criteria Decision-Making Models

TABLE 5.4
Criteria Values for Prediction Models
Prediction Model Accuracy AUC Sensitivity Specificity Precision
Decision tree 0.698 0.661 0.923 0.211 0.718
Random forest 0.713 0.693 0.896 0.313 0.746
Boosting 0.710 0.720 0.906 0.288 0.734
Support vector machines 0.706 0.672 0.938 0.204 0.718
Logistic regression 0.683 0.710 0.910 0.219 0.691
Artificial neural networks 0.706 0.711 0.922 0.239 0.724

and executed in R statistical software using a computer with Intel Core i5

2.50 GHz processor with 8 GB of RAM.
The data set must be divided into training set and testing set. If there is
less training data, then the parameter estimates will have higher variance.
On the other hand, if there is less testing data, then the classifier perfor-
mance will have higher variance. Therefore, different data splits are used for
learning and validation and are as follows:

• Data used in training, 67%; data used for testing, 33%

• Data used in training, 70%; data used for testing, 30%
• Data used in training, 75%; data used for testing, 25%

The classification model is replicated 20 times with random sampling

for each  of the above splits, and the criteria values are recorded. It was
observed that the data split did not have any impact on the performance of
the p ­ rediction model because the criteria values did not have any substan-
tial deviation. Therefore, the average of the results is taken and compiled as
shown in Table 5.4.
In order to ensure easy comparison and avoid any bias in estimating
the overall score for each prediction model, the criteria values are scaled
using the ideal value method, where the criteria values are divided by their
ideal (best) value. For example, the ideal (best) value for accuracy is 0.713.
The ­accuracy value for each of the prediction models is divided by the ideal
value for accuracy. Therefore, the accuracy of decision tree will be scaled to
0.979 (i.e., 0.698/0.713). Note that the scaled criteria values will always be ≤1,
and the best value of each criterion is 1. Table 5.5 presents the scaled criteria
­values for all the prediction models.
It can be observed from Table 5.5 that none of the prediction models
­perform the best with respect to all the criteria. For instance, random forest is
the best prediction model with respect to accuracy, specificity, and precision.
However, boosting and SVMs perform the best with respect to AUC and
sensitivity, respectively. Therefore, the rating, Borda count, and L2 ­metric
method are used to determine the best model. The criteria weights for rating
method and Borda count method are shown in Table 5.6.
Multicriteria Evaluation of Predictive Analytics 129

TABLE 5.5
Summary of Scaled Results
Prediction Model Accuracy AUC Sensitivity Specificity Precision
Decision tree 0.979 0.918 0.984 0.676 0.963
Random forest 1.000 0.962 0.955 1.000 1.000
Boosting 0.997 1.000 0.966 0.921 0.983
Support vector machines 0.990 0.934 1.000 0.652 0.963
Logistic regression 0.958 0.986 0.970 0.699 0.927
Artificial neural networks 0.991 0.988 0.983 0.764 0.971

TABLE 5.6
Criteria Weights Using Rating and Borda Count Method
Weights Using Weights Using Borda
Criteria Rating Method Count Method
Accuracy 0.17 0.35
AUC 0.22 0.32
Sensitivity 0.22 0.07
Specificity 0.15 0.09
Precision 0.24 0.17

The overall score and the corresponding rank for each prediction model
obtained using the three MCDM ranking methods are presented in
Table 5.7.
Note that the final scores of each prediction model obtained using rating
method and Borda count method are calculated by multiplying the scaled
weights of the criterion with the corresponding criterion value.
Based on the analysis of data presented in Table 5.7, it is evident that r­ andom
forest, boosting, and ANNs are the best methods as they are consistently

TABLE 5.7
Results from L2 Metric
MCDM Ranking Methods
Rating Method Borda Count L2 Metric
Final Final Final
Prediction Model Score Rank Score Rank Score Rank
Decision tree 0.9174 6 0.9298 6 0.540 5
Random forest 0.9817 1 0.9847 2 0.8240 1
Boosting 0.9761 2 0.9866 1 0.7650 2
Support vector machines 0.9227 4 0.9378 5 0.537 6
Logistic regression 0.9205 5 0.9392 4 0.560 4
Artificial neural networks 0.9497 3 0.9657 3 0.637 3
130 Big Data Analytics Using Multiple Criteria Decision-Making Models

ranked in the top three. Since random forest is ranked first under the rating
method and L2 metric method, it is regarded as the best classifier to predict
outages for the given data set.

5.7  Smart Staffing

The machine-learning classifiers can predict future interruptions and their
intensity using predictors such as weather forecasts and history of main-
tenance. The advancements in science and technology have led to accurate
short-term weather forecasts (up to 72 hours) and, therefore, enable the
development of good prediction models. These predictions can be used to
plan the workforce (crew) for power restoration. Low staffing levels decrease
the ­staffing cost. However, it leads to high repair time resulting in higher
­restoration time. On the other hand, higher staffing levels ­substantially
decrease the restoration times at the expense of increased  staffing cost.
Therefore, the ­staffing problem involves conflicting criteria, and hence an
MCMP model is used to formulate the staffing problem.

Set and indices

w ∈ W Set of all worker types
f ∈ F Set of all failure types

Parameters
Rwf Total man-hours required by worker of type w to repair failure type f
Aw Total workers of type w available for repairs
Qwf 1 if failure type f requires worker of type w; 0 otherwise
M Large positive number

Decision variables
xwf Number of workers of type w hired to repair failure type f

5.7.1  Objective Functions

5.7.1.1  Objective 1: Minimize Staffing Costs
A stable repair crew with essential skills is necessary to restore power in
case of outages. However, the increase in the number of workers results in an
increase in cost for the electric utility company. Therefore, the first objective
seeks to minimize the total workers hired as shown below:

Minimize z1 = ∑∑x
f ∈F w ∈W
wf . (5.4)

Multicriteria Evaluation of Predictive Analytics 131

5.7.1.2  Objective 2: Minimize Power Restoration Time

The time taken to restore power depends on the intensity of failure and the
number of workers assigned to restore power. Assigning few workers for
a high-intensity failure would increase the restoration time. Therefore, the
second objective seeks to minimize the power restoration times as shown in

 Rwf 
Minimize z2 = ∑ ∑  x
f ∈F w ∈W
.
wf 
(5.5)

5.7.2  Model Constraints

5.7.2.1  Restriction on Workforce Capacity
The total number of workers of a specific type is finite (Aw). Therefore,
­constraint (5.6) ensures that the total number of workers hired is always
within the finite capacity restriction for all the worker types:

∑xf
wf ≤ Aw ∀ w ∈W . (5.6)

5.7.2.2  Worker Type Requirement

Each failure type requires a crew with different worker types. For instance, a
particular failure type may require two different worker types such as line-
man and foreman. It is essential to predict the type of failure and have the
crew prepared in advance. Constraints (5.7) and (5.8) ensure that the worker
type requirement is satisfied. If there is a requirement for a particular worker
type (i.e., Qwf = 1), then constraint (5.7) ensures that at least one worker of
that particular type is assigned to the failure, while constraint (5.8) becomes
redundant. However, if a failure does not require a particular worker type
(i.e., Qwf = 0), the constraints (5.7) and (5.8) ensure that no workers of that
type are assigned to the failure:

xwf ≥ Qwf ∀w ∈ W , f ∈ F , (5.7)

xwf ≤ MQwf ∀ w ∈ W , f ∈ F. (5.8)

5.7.2.3  Nonnegativity Restriction

Constraint (5.9) ensures that the number of workers hired is always a positive
integer:

xwf ≥ 0, integer ∀w ∈W , f ∈ F. (5.9)

132 Big Data Analytics Using Multiple Criteria Decision-Making Models

The model can be solved using solution techniques such as goal pro-
gramming and ε-constraint method to obtain a set of efficient solutions
that provide a trade-off between staffing costs and restoration time. The set
of efficient solutions can then be presented to a decision maker to obtain
the best compromise solution that meets the need of the electric utility
company.

5.8 Conclusions
In the United States, power outages cause billions of dollars in losses. This
chapter aims at predicting the power outage occurrences accurately for
an electric utility company that serves over 9 million people in the United
States. Several machine-learning classifiers are used to predict the outages
using the hourly weather forecasts. The machine-learning classifiers are
later ranked based on different metrics, namely, accuracy, AUC, sensitiv-
ity, specificity, and precision using MCDM techniques.
Based on the analysis using multicriteria ranking methods, it was evi-
dent that random forest was the best method to predict the power outage
occurrences as it was ranked first by two out of the three MCDM ranking
techniques. In addition, an MCMP model was presented to determine the
appropriate staffing levels using the outputs of the prediction models. In
future work, we plan to develop prediction models to measure the intensity
of the outage, and solve the MCMP model to obtain the set of efficient solu-
tions that presents the trade-off between staffing costs and restoration time.

References
Balijepalli, N., Venkata, S. S., Richter Jr, C. W., Christie, R. D., and Longo, V. J. 2005.
Distribution system reliability assessment due to lightning storms. IEEE
Transactions on Power Delivery, 20(3), 2153–2159.
Boger, Z. and Guterman, H. 1997. Knowledge extraction from artificial neural network
models (Vol. 4, pp. 3030–3035). IEEE International Conference on Systems, Man,
and Cybernetics. Computational Cybernetics and Simulation, IEEE, Orlando, FL.
Braspenning, P. J., Thuijsman, F., and Weijters, A. J. M. M. 1995. Artificial Neural
Networks: An Introduction to ANN Theory and Practice (Vol. 931). Springer Science
& Business Media, Berlin, Germany.
Breiman, L. 2001. Random forests. Machine Learning, 45(1), 5–32.
Campbell, R. J. 2012. Weather-related power outages and electric system resiliency.
Congressional Research Service, Library of Congress, Washington, DC.
Multicriteria Evaluation of Predictive Analytics 133

Cerruti, B. J. and Decker, S. G. 2012. A statistical forecast model of weather related

damage to a major electric utility. Appl Meteor Climatol, 51(2), 191–204.
Clemmensen, J. 1993. Estimating the cost of power quality. IEEE Spectr, 30(6), 40–41.
Davidson, R. A., Liu, H., Sarpong, I. K., Sparks, P., and Rosowsky, D. V. 2003. Electric
power distribution system performance in Carolina hurricanes. Natural Hazards
Review, 4(1), 36–45.
DeGaetano, A. T., Belcher, B. N., and Spier, P. L. 2008. Short-term ice accretion fore-
casts for electric utilities using the weather research and forecasting model and
a modified precipitation-type algorithm. Weather and Forecasting, 23(5), 838–853.
Douglas, J. 2000. Research Analyst to EPRI. Personal Communication to J. Eto. Lawrence
Berkeley National Laboratory, Berkeley, California.
Gunn, S. R. 1998. Support Vector Machines for Classification and Regression. ISIS Technical
Report, 14.
Gunn, S. R., Brown, M., and Bossley, K. M. 1997. Network performance assessment
for neurofuzzy data modelling. In Liu, X., Cohen, P., Berthold, M. R. (Eds.),
Advances in Intelligent Data Analysis Reasoning about Data (pp. 313–323). Springer,
London, UK.
Ho, T. K. 1995. Random decision forests (Vol. 1, pp. 278–282). Proceedings of the Third
International Conference on Document Analysis and Recognition, IEEE, Montreal,
Canada.
Ho, W., Xu, X., and Dey, P. K. 2010. Multi-criteria decision making approaches for
supplier evaluation and selection: A literature review. European Journal of
Operational Research, 202(1), 16–24.
Huang, Z., Rosowsky, D., and Sparks, P. 2001. Hurricane simulation techniques for
the evaluation of wind-speeds and expected insurance losses. Journal of Wind
Engineering and Industrial Aerodynamics, 89(7), 605–617.
LaCommare, K. H. and Eto, J. H. 2006. Cost of power interruptions to electricity con-
sumers in the United States (US). Energy, 31(12), 1845–1855.
Li, H., Treinish, L. A., and Hosking, J. R. 2010. A statistical model for risk manage-
ment  of electric outage forecasts. IBM Journal of Research and Development,
54(3), 8–1.
Liu, H., Davidson, R. A., and Apanasovich, T. V. 2007. Statistical forecasting of elec-
tric power restoration times in hurricanes and ice storms. Power Systems, IEEE
Transactions on Power Systems, 22(4), 2270–2279.
McNulty, S. and Solutions, E. 2001. The Cost of Power Disturbances through Industrial
and Digital Economy Companies. EPRI’s Consortium for Electric Infrastructure
for a Digital Society (CEIDS), USA, Tech. Rep, 108829.
Menard, S. 2002. Applied Logistic Regression Analysis (Vol. 106). Sage, Thousand Oaks,
California.
Nagurney, A., Liu, Z., Cojocaru, M.-G., and Daniele, P. 2007. Dynamic electric power
supply chains and transportation networks: An evolutionary variational
inequality formulation. Transportation Research Part E: Logistics and Transportation
Review, 43(5), 624–646.
Nagurney, A. and Matsypura, D. 2007. A supply chain network perspective
for ­electric  power generation, supply, transmission, and consumption. In
Kontoghiorghes, E. and Gatu, C. (Eds.), Optimisation, Econometric and Financial
Analysis (pp. 3–27). Springer, Germany.
Quinlan, J. R. 1986. Induction of decision trees. Machine Learning, 1(1), 81–106.
134 Big Data Analytics Using Multiple Criteria Decision-Making Models

Ravindran, A. R. 2016. Managing supply chains: An introduction. Multiple Criteria

Decision Making in Supply Chain Management, 1, 1–14.
Reed, D. A. 2008. Electric utility distribution analysis for extreme winds. Journal of
Wind Engineering and Industrial Aerodynamics, 96(1), 123–140.
Reed, D. A., Powell, M. D., and Westerman, J. M. 2010. Energy supply system perfor-
mance for Hurricane Katrina. Journal of Energy Engineering, 136(4), 95–102.
Sieminski, A. 2014. International Energy Outlook. Energy Information Administration
(EIA), https://www.eia.gov/pressroom/presentations/sieminski_11182014.pdf.
Sullivan, M. J., Suddeth, B. N., Vardell, T., and Vojdani, A. 1996. Interruption costs,
customer satisfaction and expectations for service reliability. IEEE Transactions
on Power Systems, 11(2), 989–995.
Swaminathan, S. and Sen, R. K. 1998. Review of Power Quality Applications of Energy
Storage Systems. Sandia National Laboratories, Livermore, California.
Vapnik, V. N. and Vapnik, V. 1998. Statistical Learning Theory (Vol. 1). Wiley, New York.
Winkler, J., Duenas-Osorio, L., Stein, R., and Subramanian, D. 2010. Performance
assessment of topologically diverse power systems subjected to hurricane
events. Reliability Engineering & System Safety, 95(4), 323–336.
Zhou, Y., Pahwa, A., and Yang, S.-S. 2006. Modeling weather-related failures of
­overhead distribution lines. IEEE Transactions on Power Systems, 21(4), 1683–1690.
Zhu, D., Cheng, D., Broadwater, R. P., and Scirbona, C. 2007. Storm modeling for
­prediction of power distribution system outages. Electric Power Systems Research,
77(8), 973–979.
Zopounidis, C. and Doumpos, M. 2002. Multicriteria classification and sorting meth-
ods: A literature review. European Journal of Operational Research, 138(2), 229–246.
6
Multiobjective Forecasting: Time
Series Models Using a Deterministic
Pseudo-Evolutionary Algorithm

Nagulapally Venkat Ramarao, P. Y. Yeshwanth Babu,

Sankaralingam Ganesh, and Chandrasekharan Rajendran

CONTENTS
6.1 Introduction................................................................................................. 136
6.1.1 Seasonality....................................................................................... 137
6.1.2 Stationarity and Nonstationarity................................................. 137
6.1.3 Overview of ARMA Models......................................................... 137
6.1.4 Brief Literature Review on ARIMA Models............................... 138
6.2 Proposed Multiobjective Deterministic Pseudo-Evolutionary
Algorithm..................................................................................................... 139
6.2.1 Step-by-Step Procedure of the Proposed MDPEA..................... 139
6.2.1.1 Phase 1: MAPE Being the Primary Objective.............. 140
6.2.1.2 Phase 2: MaxAPE Being the Primary Objective.......... 142
6.2.1.3 Phase 3: Generating the Combined Netfront with
Respect to the Training Period....................................... 144
6.2.1.4 Phase 4: Generating the Combined Netfront
with Respect to the Test Data Set (of Size I″
Time Periods) from the Models Which Form the
Netfront Corresponding to the Training Period......... 145
6.2.1.5 Phase 5: Stop..................................................................... 147
6.3 Computational Evaluation of the Proposed MDPEA............................ 148
6.3.1 Data Set............................................................................................. 148
6.3.2 Multiobjective Netfront for a Retail Segment Sales Data
(90:10 with Respect to the Split of Training Data Set:
Test Data Set)................................................................................... 148
6.4 Summary and Conclusions....................................................................... 152
Acknowledgments............................................................................................... 152
References.............................................................................................................. 152

135
136 Big Data Analytics Using Multiple Criteria Decision-Making Models

6.1 Introduction
Autoregressive integrated moving average (ARIMA) method is a generaliza-
tion of autoregressive moving average (ARMA) models developed by Box
and Jenkins (1970). ARIMA method provides a parsimonious description of
the stationary data in terms of two polynomials, one for the autoregression
and the other for the moving average. In case of nonstationary data, an ini-
tial differencing step (corresponding to the integrated part of the model) is
applied to reduce nonstationarity. Nonseasonal ARIMA models are denoted
by ARIMA (p, d, q) where parameters p, d, and q are nonnegative integers,
p  is the order of the autoregressive part, d is the degree of differencing,
and q is the order of the moving average part of ARIMA model. Seasonal
ARIMA models are denoted as ARIMA (p, d, q) (P, D, Q)m, where m refers to
the ­number of periods in each season, and P, D, Q refer to the autoregres-
sive, differencing, and the moving average terms for the seasonal part of the
ARIMA model.
In big data analytics, we encounter the forecasting problem that deals with
large data sets. In most business scenarios, the older data might be less use-
ful in building the forecast models and more weightage has to be given to
the immediate past; this is where ARIMA, seasonal ARIMA, and hybrid
ARIMA models come into play. Using batch-processing technologies such
as Apache’s Hadoop to handle the volume component of big data and using
stream-processing technologies such as Apache’s Spark to handle the veloc-
ity component of big data, we could convert the unstructured data into struc-
tured data. Once the data is structured, a forecasting model can be used to
arrive at forecasts, either with a single objective or with respect to multiple
objectives.
The usual industry requirement for time series forecasting has objectives
such as reducing the average error in predictions for a time period, as well
as restricting the maximum error that can pop up in any given period. A
typical example would be inventory management where the objective of the
time series model would be that the average inventory of products in any
quarter should be minimum and at any point of time, the inventory should
not go beyond a threshold based on the capacity of the warehouse. At times,
such objectives are conflicting since a model, which has the least mean abso-
lute percentage error (MAPE), does not necessarily give the least maximum
absolute percentage error (MaxAPE; true in most of cases). This chapter
addresses the multiple criteria decision analysis involved in the time series
forecasting, by using the ARIMA model. The algorithm detailed in Section
6.2 is scalable to any number of time series, and it can cater to present indus-
trial requirement of forecasting the sales at stock-keeping unit level (which
is very large in fast moving consumer goods [FMCG] or any other consumer
goods sector).
Multiobjective Forecasting 137

6.1.1 Seasonality
Seasonal components consist of effects that are reasonably stable with respect
to timing, direction, and magnitude. Seasonality in a time series can be iden-
tified by regularity in crests and troughs, which have consistent direction
and magnitude, relative to the trend. Commonly employed approaches to
modeling seasonal patterns include the Holt–Winters exponential smooth-
ing model and ARIMA model.

6.1.2  Stationarity and Nonstationarity

A stationary time series is the one whose statistical properties such as mean,
variance, and autocorrelation remain constant over time. In other words, the
joint probability distribution of the series does not change when shifted in
time. If a time series does not hold the above property, then it is said to be
nonstationary. Seasonality is a characteristic of a time series in which the
data experiences regular and predictable changes recurring at certain inter-
vals. The data can be checked for stationarity by plotting the time series and
by Dickey–Fuller test. Sometimes, differencing of the data is needed to arrive
at a stationary series.
We could arrive at the values for the p, d, q and P, D, Q parameters via the
autocorrelation factor (ACF) and the Partial ACF (PACF) of the time series.
The ACF explains the correlation between values of the series at different
points in time. Given measurements Y1, Y2, Y3, …, YN at time X1,X2,X3, …, XN,
respectively, the lag k autocorrelation function is defined as

∑
N −k
(Yi − Y )(Yi + k − Y )
rk = i =1
. (6.1)
∑
N
(Yi − Y )2
i =1

The PACF gives the partial correlation of a time series with its own lagged
values, controlling for the values of the time series at all shorter lags.

6.1.3  Overview of ARMA Models

The autoregressive or AR model can be written in the form:

yt = c + α 1 yt −1 +  + α p yt − p + et ; or yt = c + ∑α y i t−i + et , (6.2)
i =1

where the terms in α are autocorrelation coefficients with respect to lags

1, 2, …, p, and et is a residual error term. Note that this error term specifically
relates to the current time period t.
138 Big Data Analytics Using Multiple Criteria Decision-Making Models

The moving average or MA part of the model can be written as follows:

q

yt = c + β0 et + β1et −1 +  + β q et − q ; or yt = c + ∑β e i t−i , (6.3)

i=0
where the βi terms are the weights applied to prior error values in the time
series, and it is usual to define β0 = 1 without loss of generality.
ARMA(p, q) model can be expressed as below:

y t = α 0 + α 1 y t − 1 +  + α p y t − p − β 0 e t − β 1e t − 1 −  − β q e t − q . (6.4)

Seasonal ARIMA (that is, SARIMA(p, d, q)(P, D, Q)m) model can be
expressed as below, where p, d, q, P, D, and Q are nonnegative integers and
m is periodicity:

ϕ(B)Φ(Bm )(1 − B)d (1 − Bm )D (Zt − µ) = θ(B)Θ(Bm )at , (6.5)

where
φ(B) = 1 − φ1B − φ2B2 − … − φpBp, φ1, φ2, …, φp are coefficients;
Φ(Bm) = 1 − Φ1Bm − Φ2B2*m − … − ΦpBP*m, Φ1, Φ2, …, ΦP are coefficients;
θ(B) = 1 + θ1B + θ2B2 + … + θqBq, θ1, θ2, …, θq are coefficients; and
Θ(Bm) = 1 + Θ1Bm + Θ2B2*m + … + ΘQBQ*m, Θ1, Θ2, …, ΘQ are coefficients.

As shown above, the four functions are polynomials in B of degrees p, q,

P, and Q, where B is the backward shift operator, that is, Byt = yt−1, B2yt = yt−2,
B3yt = yt−3, etc., d is the order of regular differences, D is the order of seasonal
differences, and Zt denotes the observed value of time series data. For detailed
explanation of these equations, refer to Hyndman and Athanasopoulos (2013).

6.1.4  Brief Literature Review on ARIMA Models

Zou and Yang (2004) proposed an algorithm to assign weights and combine
ARIMA models for a better performance of prediction. Cools et al. (2009) used
ARIMAX and SARIMAX models to forecast the daily traffic counts. Lee and
Hamzah (2010) developed an ARIMAX model to forecast monthly sales of
Muslim boys’ clothes in Indonesia. This model combines ARIMA model and
calendar variation effect during Eid holidays using linear regression, and had
better forecast results than decomposition method, SARIMA, and Artificial
Neural Network (ANN). The regression–SARIMA modeling framework cap-
tures important drivers of electricity demand. Nie et al. (2012) introduced a
hybrid model which integrates ARIMA with support vector machines (SVM)
to forecast short-term load forecasting for energy management. Chikobvu
and Sigauke (2012) predicted the daily peak electricity demand in South
Africa using SARIMA and regression–SARIMA. The performance of the
developed models was evaluated by comparing them with Winter’s triple
Multiobjective Forecasting 139

exponential smoothing model. Empirical results from the study showed that
the SARIMA model produced more accurate short-term forecasts.
In this chapter, we attempt to improve on the basic ARIMA model with
the consideration of multiple objectives such as the minimization of MAPE
and MaxAPE. We develop offspring time series from the best parent ARIMA
models by considering the fitness values of parent ARIMA models. These
fitness values are computed using the appropriate primary objective func-
tion, namely, MAPE/MaxAPE; and considering the parent ARIMA models,
coupled with their relative fitness values, we deterministically generate the
offspring time series.

6.2 Proposed Multiobjective Deterministic

Pseudo-Evolutionary Algorithm
In this chapter, we have considered sales data of two retail segments. The
real-life sales data of a beverage company and the real-life sales data of an
online fashion store are being used here. Since the data has granularity
(Section 6.3) at day level, we observe seasonality for every 7 days. The details
of the data sets are not furnished here for the purpose of confidentiality, and
for making the chapter concise we present the application of the algorithm
considering one data set only.
We fit 288 parent SARIMA models to the time series with possible com-
binations of the parameters p ∈ {0, 1, 2, 3, 4, 5}, d ∈ {0, 1}, q ∈ {0, 1, 2, 3, 4, 5},
(P,D,Q) ∈ {(0,0,0),(0,1,1),(1,1,0),(1,1,1)}, and seasonality m as 7. Once these models
are obtained, we then proceed to calculate MAPE and MaxAPE for each of
these parent SARIMA models and arrange them in nondecreasing order of
the respective objective function. Fitness Values are calculated for each time
series generated from the parent SARIMA model, as explained in step 3 in
phases 1 and 2 of the step-by-step procedure of the proposed multiobjective
deterministic pseudo-evolutionary algorithm (MDPEA). Once the relative
fitness values are obtained for each set of parent time series, we obtain deter-
ministically the offspring time series corresponding to the consideration of
these parent time series. Subsequently, we compare offspring and parent
time series to get the netfront in the training period, followed by their use in
the generation of time series in the test period.

6.2.1  Step-by-Step Procedure of the Proposed MDPEA

MAPE and MaxAPE are chosen as the criteria or objectives for the following
reason. The MaxAPE provides the scenario of the worst-case model error
and the MAPE provides the scenario of the average model error. Also, MAPE
and MaxAPE are better metrics to judge the forecast model’s performance
140 Big Data Analytics Using Multiple Criteria Decision-Making Models

in real life, compared to other theoretical likelihood metrics such as Akaike

information criteria (AIC) or Bayesian information criteria (BIC).
The algorithm has four phases. In the first phase, we fit SARIMA models
to the time series, and find the top parent models with MAPE as the primary
objective. The parent time series from the respective SARIMA models are
assigned relative fitness values, according to their corresponding MAPE val-
ues. Thereafter we generate offspring time series by using respective parent
time series.
In the second phase, we choose the top parent SARIMA models with
MaxAPE as the primary objective and generate offspring time series; here,
the relative fitness values are assigned to parent time series according to the
MaxAPE of the corresponding SARIMA models, and thereafter we generate
offspring time series by using respective parent time series.
In the third phase, we find the nondominated netfront with respect to
MAPE and MaxAPE by considering parent time series and offspring time
series, with respect to the training period.
In the fourth phase, we find the nondominated netfront with respect to
MAPE and MaxAPE by considering those parent SARIMA models and off-
spring SARIMA time series (that have entered the netfront corresponding to the
training period) in the test period. Note that while generating these offspring
time series in the test period, we make use of the corresponding and respective
parent-series models (include their fitness values) that have been obtained in
the training period. Let I′ be the size of training set and I″ be that of test set.

6.2.1.1  Phase 1: MAPE Being the Primary Objective

Step 1: Set

p ∈ {0,1, 2, 3, 4, 5},
q ∈ {0,1, 2, 3, 4, 5},
d ∈ {0,1},
(P , D,Q) ∈ {(0, 0, 0), (0,1,1), (1,1, 0), (1,1,1)}, an
nd
m = 7.

Step 2: Run the SARIMA model for every combination of (p, d, q,)(P, D, Q)m,
and choose the best N parent models with respect to MAPE and arrange
them in the nondecreasing order of their respective MAPE.
Denote the following:

• Actual data point of the time series as yi, where i = 1, 2, 3, …, I′, cor-
responding to the training set.
• Predicted value of nth parent model, with respect to data point i (i.e.,
p,I ′
time period i) corresponding to training set, as y1n , i for n = 1, 2, 3, …,
N, and i = 1, 2, 3, …, I′.
Multiobjective Forecasting 141

• The MAPE with respect to nth parent time series considering the
training set is given as follows:

 I′
p ,I ′
y1n ,i − yi 
E1 p ,I ′
n

=

∑
i =1
yi
 1
× 100 × .

I′
(6.6)

• The MaxAPE with respect to nth parent time series (chosen with the
primary objective of minimizing the MAPE) considering the train-
ing set is given as follows:

 y1p , I ′ − y 
i
p,I ′  n,i  (6.7)
ξ1 n = max i  × 100 .
yi
 

Step 3: Do the following to generate offspring time series with the primary
objective of minimizing the MAPE, by considering ( n′ − 1) parent time series
at a time to generate a corresponding offspring time series, and by consider-
ing the relative fitness of these chosen parents:

Step 3.1: Set n′ = 3.

Step 3.2: Calculate the fitness of nth parent model (i.e., parent time
series), where 1 ≤ n ≤ n′ − 1, as follows, given n′:

f 1n ,n′ = E1np′,I ′ − E1np,I ′. (6.8)

Step 3.3: Calculate the relative fitness of nth parent time series, where
1 ≤ n ≤ n′ − 1:

f1 n, n′
f 1′n , n ′ = . (6.9)
∑
n ′− 1
f 1n ′′ , n ′
n ′′= 1

Step 3.4: Generate the predicted values with respect to offspring, that
is, the offspring time series n″ with respect to training set, where
n″ = n′ − 2:

∑ ( y1 )
n ′− 1
o,I ′ p,I ′
y1n ′′ , i = n,i
× f 1′n , n ′ , for i = 1, 2, 3, … , I ′. (6.10)
n =1
142 Big Data Analytics Using Multiple Criteria Decision-Making Models

Step 3.5: Set n′ = n′ + 1 and repeat step 3.2 through 3.5 up to n′ = N.

Step 4: Calculate MAPE and MaxAPE values for n″th offspring time series
with respect to the training set:

 I′
o ,I ′
y1n′′ ,i − yi 
E1 o ,I ′
n′′

=

∑
i =1
yi
 1
× 100 × ,

I′
(6.11)

 y1o , I ′ − y 
i
 n ′′ , i 
ξ1on,′′I ′ = max i  × 100 for n′′ = 1, 2, 3, … ,( N − 2). (6.12)
yi
 

6.2.1.2  Phase 2: MaxAPE Being the Primary Objective

Step 1: Set

p ∈ {0, 1, 2, 3, 4, 5},
q ∈ {0, 1, 2, 3, 4, 5},
d ∈ {0, 1},
(P , D, Q) ∈ {(0, 0, 0), (0, 1, 1), (1, 1, 0), (1, 1, 1)}, and
m = 7.

Step 2: Run SARIMA models for every combination of (p, d, q,)(P, D, Q)m and
choose the best N parent models with respect to MaxAPE and arrange them
in nondecreasing order of their respective MaxAPE.
Denote the following:

• Actual data point of time series as yi, where i = 1, 2, 3, …, I′ corre-

sponds to the training set.
• Predicted value of nth parent model with respect to data point i cor-
p,I ′
responding to training set as y2n , i for n = 1, 2, 3, …, N, and i = 1, 2,
3, …, I′ (note: these parents are obtained with the consideration of
MaxAPE).
• The MAPE with respect to nth parent time series considering the
training set (chosen with the primary objective of minimizing
MaxAPE) is given as follows:
Multiobjective Forecasting 143

 I′
p ,I ′
y2n ,i − yi 
E2 p ,I ′
n

=

∑i =1
yi
 1
× 100 × .

I′
(6.13)

• The MaxAPE with respect to nth parent time series considering the
training period is:

  p,I ′ 
y 2 − yi
 n,i 
ξ2np, I ′ = max i  × 100 . (6.14)
yi
 

Step 3: Do the following to generate offspring time series with the primary
objective of minimizing the MaxAPE, by considering (n′ − 1) parent time
series:

Step 3.1: Set n′ = 3.

Step 3.2: Calculate the fitness of nth parent time series, 1 ≤ n ≤ n′ − 1, as
follows, given n′:

f 2n , n ′ = ξ 2np,′ I ′ − ξ 2np, I ′. (6.15)

Step 3.3: Calculate the relative fitness of nth parent model, 1 ≤ n ≤ n′ − 1,
as follows, given n′:

f2 n, n′
f 2′n , n ′ = . (6.16)
∑
n ′− 1
f 2n ′′ , n ′
n ′′= 1

Step 3.4: Generate the predicted values with respect to offspring, that
is, the offspring time series n″ with respect to training set, where
n″ = n′ − 2:

∑ ( y2 )
n ′− 1
o,I ′ p,I ′
y2n ′′ , i = n,i
× f 2′n , n ′ , for i = 1, 2, 3, … , I ′. (6.17)
n =1

Step 3.5: Set n′ = n′ + 1 and repeat step 3.2 through 3.5 up to n′ = N.
144 Big Data Analytics Using Multiple Criteria Decision-Making Models

Step 4: Calculate MAPE and MaxAPE values for n″th offspring time series
with respect to training set:

 o,I ′ 
I′ y2n ′′ , i − yi
E2 o,I ′
n ′′

=

∑
i =1
yi
 1
× 100 × ,

I′
(6.18)

  o,I ′ 
y 2 − yi
 n ′′ , i 
ξ2on,′′I ′ = max i  × 100 for n′′ = 1, 2, 3, … ,( N − 2). (6.19)
yi
 

6.2.1.3 Phase 3: Generating the Combined Netfront with

Respect to the Training Period
Dominated model: A model is referred to as dominated if both MAPE and
MaxAPE are bettered by or equal to at least one other model in the lot (con-
sisting of the parents and offspring).
Nondominated model: A model is referred to as nondominated if its MAPE
or MaxAPE is better than any other model in the lot. Eliminating all the
dominated models in the lot will result in the nondominated netfront.
We compare the performance of parent and offspring time series with
respect to both MAPE and MaxAPE in the training period, and identify the
nondominated sets of parent and offspring time series. We proceed as fol-
lows considering this nondominated netfront:

• Let the performance of top N parent SARIMA time series with

( )
respect to MAPE be denoted by E1np, I ′ , ξ1np, I ′ , n = 1, 2, 3, …, N; From
these time series, the nondominated parents, which enter the net-
front corresponding to training period, are denoted by the set Ψ1P.
• Let the performance of top parent SARIMA time series with respect
( )
to MaxAPE denoted by E2np, I ′ , ξ 2np, I ′ , n = 1, 2, 3, …, N. From these
time series, the nondominated parents, which enter the netfront cor-
responding to training period, are denoted by the set Ψ2P.
• Let the performance of offspring time series derived in phase 1 be
( )
denoted by E1on,′′I ′ , ξ1on,′′I ′ , n″ = 1, 2, 3, …, (N − 2); From these offspring,
the nondominated offspring, which enter the netfront correspond-
ing to training period, are denoted by the set Ψ1o.  Note that these
offspring are generated with the consideration of the corresponding
f 1′n ,n′ associated with the respective parents.
• Let the performance of offspring time series derived in phase 2 be
( )
denoted by E2on,′′I ′ , ξ 2on,′′I ′ , n″ = 1, 2, 3, …, (N − 2). From these offspring,
Multiobjective Forecasting 145

the nondominated offspring, which enter the netfront correspond-

ing to training period, are denoted by Ψ2o. Note that these offspring
are generated with the consideration of the corresponding f 2′n , n ′
associated with the respective parents.
• Note that when we carry over offspring in the sets Ψ1o and Ψ2o, we
also carry over the respective parent’s parameters (p, d, q) (P, D, Q)m
and their relative fitness values that have led to the generation of the
offspring, along with the offspring’s n″ and n′. This information is
used, while evaluating the offspring in the test period.

6.2.1.4 Phase 4: Generating the Combined Netfront with Respect to the

Test Data Set (of Size I″ Time Periods) from the Models Which
Form the Netfront Corresponding to the Training Period

• The performance of the parent time series models in the test data
set (with MAPE as primary objective), which have earlier entered
the netfront corresponding to the training data set, is denoted by
( )
E1np, I ′′ , ξ1np, I ′′ , n∈Ψ1p. Note that using the parent time series models
and their corresponding (p, d, q) (P, D, Q)m values obtained from the
training data set, we now generate the parent time series (forecasts)
for the test data set, and their performance is denoted as above. The
generated parent time series are checked for possible entry into the
nondominated netfront corresponding to the test period. The non-
dominated parents which enter the netfront corresponding to test
(
period are denoted by their respective performance E1np, I ′′ , ξ1np, I ′′ , )
n ∈ Ω1p.
• The performance of the parent time series models in the test data
set (with MaxAPE as primary objective), which have earlier entered
the netfront corresponding to the training data set, is denoted by
( )
E2np, I ′′ , ξ 2np, I ′′ , n ∈ Ψ2p. Note that using the parent time series mod-
els and their corresponding (p, d, q) (P, D, Q)m values obtained from
the training data set, we generate the parent time series (forecasts)
for the test data set. The generated parent time series are checked
for possible entry into the nondominated netfront correspond-
ing to the test period. The nondominated parents which enter the
( )
­netfront corresponding to test period are denoted by E2np, I ′′ , ξ 2np, I ′′ ,
n ∈ Ω2p.

The following steps give the procedure to obtain the offspring time series
with respect to the test data set.
Step 1: MAPE as the primary objective:
Considering the offspring time series in Ψ1o one by one, and using the
corresponding set of parent time series models and their fitness values
146 Big Data Analytics Using Multiple Criteria Decision-Making Models

that have led to the generation of this offspring time series, we do the
following.

Step 1.1: Generate the forecasted time series for the test data set by
considering the parent time series model in the set of parent time
series models that have led to the generation of this offspring time
series; denote such a forecasted time series given by every such
p , I ′′
parent model as y1n , i with respect to the time period i in the test
data set.
Step 1.2: Using the above and the parents’ corresponding fitness val-
ues (that have been obtained in the training period) with respect to
this offspring in Ψ1o and this offspring’s corresponding n″ and n′, we
obtain the corresponding forecasted offspring time series for the test
data set as follows:

∑ ( y1 )
n ′− 1
o , I ′′ p , I ′′
y1n ′′ , i = n,i
× f 1′n , n ′ , for i = 1, 2, 3, … , I ′′. (6.20)
n =1

Note that f 1′n , n ′ is inherited from the training data set (Equation 6.9).
Step 1.3: Calculate:

 I ′′
o , I ′′
y1n ′′ , i − yi 
E1 o , I ′′
n ′′

=

∑ i =1
yi
 1
× 100 × ;

I ′′
(6.21)

 y1o , I ′′ − y 
i
o , I ′′  n ′′ , i 
ξ1 n ′′ = max i  × 100 . (6.22)
yi
 

By repeating the above for every offspring in Ψ1o, we get the correspond-
ing set of offspring time series with respect to the test period. This set of
offspring time series with respect to the test period is checked for pos-
sible entry into the set of nondominated solutions with respect to the test
period. Let the set of nondominated offspring time series in the test period
be denoted by Ω1o. Let the performance of the nondominated offspring
which enter the netfront corresponding to the test period be denoted by
( )
E1on,′′I ′′ , ξ1on,′′I ′′ , n″ ∈ Ω1o.
Step 2: MaxAPE as the primary objective:
Considering the offspring time series in Ψ2o one by one, and using the
corresponding set of parent time series models and their fitness values
Multiobjective Forecasting 147

that have led to the generation of this offspring time series, we do the
following.

Step 2.1: Generate the forecasted time series for the test data set by con-
sidering every parent time series model in the set of parent time series
models that have led to the generation of this offspring time series,
denote such a forecasted time series given by every such parent model
p , I ′′
as y2n , i with respect to the time period i in the test data set.
Step 2.2: Using the above and the parents’ fitness values (that have been
obtained in the training period) with respect to this offspring from
Ψ2o and this offspring’s corresponding n″ and n′, we obtain the cor-
responding forecasted offspring time series for the test data set as
follows:

∑ ( y2 )
n ′− 1
o , I ′′ p , I ′′
y2n ′′ , i = n,i
× f 2′n , n ′ , for i = 1, 2, 3, … , I ′′. (6.23)
n =1

Note that f 2′n , n ′ is inherited from the training data set (Equation 6.16).
Step 2.3: Calculate:

 o , I ′′ 
I ′′ y2n ′′ , i − yi
E2 o , I ′′
n ′′

=

∑ i =1
yi
 1
× 100 × ;

I ′′
(6.24)

  o , I ′′ 
y 2 − yi
 n ′′ , i 
ξ2on,′′I ′′ = max i  × 100 . (6.25)
yi
 

By repeating the above for every offspring in Ψ2 , we get the correspond- o

ing set of offspring time series with respect to the test period. This set of
offspring time series with respect to the test period is checked for possi-
ble entry into the set of nondominated solutions with respect to the test
period. Let the set of nondominated offspring time series in the test period
be denoted by Ω2o. Let the performance of the nondominated offspring
which enter the netfront corresponding to the test period be denoted by
( )
E2on,′′I ′′ , ξ 2on,′′I ′′ , n″ ∈ Ω2o.

6.2.1.5  Phase 5: Stop

The algorithm is terminated and the time series denoted by Ω1p, Ω2p, Ω1o,
and Ω2o form the nondominated time series with respect to the test period.
148 Big Data Analytics Using Multiple Criteria Decision-Making Models

6.3  Computational Evaluation of the Proposed MDPEA

We consider one data set (Figure 6.1) from a retailer segment. The details
are masked for confidentiality. In a generic sense, the time series can be con-
sidered as sales of consumer products, which are durable and conveniently
available in market, and have a weekly seasonality since the data is at daily
granularity.

6.3.1  Data Set

Figure 6.1 details the real-life sales data of a beverage company.

6.3.2 Multiobjective Netfront for a Retail Segment Sales Data (90:10

with Respect to the Split of Training Data Set: Test Data Set)
Multiple data set SARIMA models are fitted to the sales time series and
arranged in the increasing order of MAPE and MaxAPE separately. These
two sets are considered as the parent models to derive the offspring models.
Table 6.1 shows the parent and offspring nomenclature, and their respective
description. Note that we report 20 top parent time series with respect to
MAPE and 20 top parents time series with respect to MaxAPE, because the
next 10 parent time series with respect to MAPE and the next 10 parent time
series with respect to MaxAPE have not entered the nondominated front in
the training data set. However, we consider the top 32 parent time series
with respect to MAPE and top 32 parent time series with respect to MaxAPE
to generate 30 offspring time series with respect to MAPE and 30 offspring
time series with respect to MaxAPE, respectively.

14.0K
12.0K
10.0K
8.0K
Sales

6.0K
4.0K x axis 1 cm = 1 day
2.0K y axis 1 cm = 2000 units
0.0K
1
12
23
34
45
56
67
78
89
100
111
122
133
144
155
166
177
188
199
210
221
232
243
254
265
276
287
298
309
320
331
342
353
364

Days

FIGURE 6.1
Real-life sales data of a retail segment; time period: 1 year at a day-level granularity.
Multiobjective Forecasting 149

TABLE 6.1
Labels for Parent and Offspring Time Series
Top 20 Parents with MAPE Top 20 Parents with MaxAPE
Label as Primary Objective Label as Primary Objective
P-1 ARIMA(5, 0, 3)(1, 1, 1) P-21 ARIMA(0, 0, 0)(0, 1, 1)
P-2 ARIMA(4, 1, 5)(0, 1, 1) P-22 ARIMA(0, 0, 0)(1, 1, 1)
P-3 ARIMA(5, 1, 4)(0, 1, 1) P-23 ARIMA(0, 0, 4)(0, 1, 1)
P-4 ARIMA(5, 0, 2)(0, 1, 1) P-24 ARIMA(0, 0, 4)(1, 1, 1)
P-5 ARIMA(5, 0, 5)(1, 1, 1) P-25 ARIMA(1, 0, 4)(0, 1, 1)
P-6 ARIMA(5, 1, 5)(1, 1, 1) P-26 ARIMA(2, 0, 3)(1, 1, 1)
P-7 ARIMA(4, 0, 4)(1, 1, 1) P-27 ARIMA(2, 0, 4)(0, 1, 1)
P-8 ARIMA(5, 0, 2)(1, 1, 1) P-28 ARIMA(2, 0, 4)(1, 1, 1)
P-9 ARIMA(4, 0, 5)(0, 1, 1) P-29 ARIMA(2, 0, 5)(1, 1, 1)
P-10 ARIMA(4, 0, 3)(0, 1, 1) P-30 ARIMA(3, 0, 4)(0, 1, 1)
P-11 ARIMA(5, 0, 5)(0, 1, 1) P-31 ARIMA(4, 0, 2)(0, 1, 1)
P-12 ARIMA(5, 1, 3)(0, 1, 1) P-32 ARIMA(4, 0, 2)(1, 1, 1)
P-13 ARIMA(5, 1, 5)(0, 1, 1) P-33 ARIMA(4, 0, 3)(1, 1, 1)
P-14 ARIMA(5, 1, 4)(1, 1, 1) P-34 ARIMA(4, 1, 5)(1, 1, 0)
P-15 ARIMA(4, 0, 4)(0, 1, 1) P-35 ARIMA(4, 1, 5)(1, 1, 1)
P-16 ARIMA(5, 1, 3)(1, 1, 1) P-36 ARIMA(5, 0, 4)(0, 1, 1)
P-17 ARIMA(3, 1, 4)(1, 1, 1) P-37 ARIMA(5, 0, 4)(1, 1, 0)
P-18 ARIMA(5, 1, 1)(1, 1, 1) P-38 ARIMA(5, 0, 4)(1, 1, 1)
P-19 ARIMA(5, 1, 2)(1, 1, 1) P-39 ARIMA(5, 0, 5)(1, 1, 0)
P-20 ARIMA(4, 1, 3)(1, 1, 1) P-40 ARIMA(5, 0, 5)(1, 1, 1)
Label Parent description (MAPE as primary objective)
P-1 to P-20 Top 20 parent SARIMA time series with respect to MAPE as primary
objective
Label Parent description (MaxAPE as primary objective)
P-21 to P-40 Top 20 parent SARIMA models with respect to MaxAPE as primary
objective
Label Offspring time series description (MAPE as primary objective)
O-41 Offspring obtained from top 3 parents with MAPE as primary objective
O-42 Offspring obtained from top 4 parents with MAPE as primary objective
… …
O-70 Offspring obtained from top 32 parents with MAPE as primary objective
Label Offspring time series description (MaxAPE as primary objective)
O-71 Offspring obtained from top 3 parents with MaxAPE as primary objective
O-72 Offspring obtained from top 4 parents with MaxAPE as primary objective
… …
O-100 Offspring obtained from top 32 parents with MaxAPE as primary objective
150 Big Data Analytics Using Multiple Criteria Decision-Making Models

Scenario 1:
In this scenario, we see the formation of nondominated solutions when we
consider only 20 offspring, with respect to MAPE and MaxAPE. We compare
O-41 to O-60 and O-71 to O-90 offspring time series with P-1 to P-40 par-
ent time series considering the training period and we create a netfront of
nondominated time series (Figure 6.2). The parent and offspring time series,
which enter this netfront, are evaluated in the test period and the final net-
front is created as shown in netfront test period chart (Figure 6.3). We refrain
from labeling all the data points due to congestion of data points.
As we can see from Figures 6.2 and 6.3, both parent and offspring time
series have entered the netfront of nondominated solutions in the training
set; however, of those time series that have entered the netfront in the training
set, only one offspring time series has entered the netfront of nondominated

20.60%

O-42 P-1
O-45
P-4
MaxAPE

O-48
O-49 P-33
O-50
O-90
O-89
O-77
O-76
O-71
O-72
18.60%
2.95% 3.45%
MAPE

FIGURE 6.2
Netfront corresponding to the training period with consideration of (O-41 to O-60) and (O-71
to O-90) offspring time series, (P-1 to P-20) and (P-21 to P-40) parent time series.

30.60%

28.60%

26.60% O-72
MaxAPE

24.60%

22.60%

20.60%

18.60%
2.95% 3.45% 3.95% 4.45% 4.95% 5.45% 5.95%
MAPE

FIGURE 6.3
Netfront corresponding to the test period.
Multiobjective Forecasting 151

20.60%

O-70
O-69
O-42 O-68
O-45 P-1
MaxAPE

O-48 P-4
O-49
O-50 P-33

O-100
O-99 O-77
O-76
O-71
O-72
18.60%
2.80% 3.20% 3.60%
MAPE

FIGURE 6.4
Netfront corresponding to the training period with consideration of (O-41 to O-70) and (O-71 to
O-100) offspring time series, and (P-1 to P-20) and (P-21 to P-40) parent time series.

solutions in the test set, when dealing with multiple objectives, that is, MAPE
and MaxAPE.

Scenario 2:
In this scenario, we see the formation of nondominated solutions when
we consider a set of 30 offspring, with respect to MAPE and MaxAPE. We
compare O-41 to O-70 and O-71 to O-100 offspring time series with P-1 to
P-40 parent time series considering the training period and create a netfront
of nondominated time series as shown in netfront training period chart
(Figure 6.4). The parent and offspring time series which enter this netfront
are ­evaluated in the test period and the final netfront is created as shown in
netfront test period chart (Figure 6.5).

30.60%

28.60%

26.60% O-72
MaxAPE

24.60%

22.60%

20.60%

18.60%
2.80% 3.20% 3.60% 4.00% 4.40% 4.80% 5.20% 5.60% 6.00%
MAPE

FIGURE 6.5
Netfront corresponding to test period for offspring.
152 Big Data Analytics Using Multiple Criteria Decision-Making Models

As we can see from Figures 6.4 and 6.5, both parent and offspring time
series have entered the netfront of nondominated solutions in the training set;
­however, of those time series that have entered the netfront in the t­ raining set,
only one offspring time series has entered to the netfront of ­nondominated
solutions in the test set, when dealing with multiple o
­ bjectives, that is, MAPE
and MaxAPE.

6.4  Summary and Conclusions

The proposed MDPEA offers the development of offspring time series from
the parent time series, deterministically by considering the fitness values of
parent SARIMA models. These fitness values are computed using the appro-
priate primary objective function, namely, MAPE/MaxAPE. Considering
the parent SARIMA models deterministically, we generate these offspring
time series by considering the respective parents whose fitness values are
­separately calculated with respect to the MAPE and MaxAPE.
The salient contributions of the proposed MDPEA are as follows:

1.
Accuracy: It is evident from the computational experiments that the
MDPEA is able to produce the offspring time series that perform
better than the parent time series with respect to the multiple objec-
tives of minimizing the MAPE and MaxAPE.
2.
Speed: MDPEA gets the best of both worlds, that is, it exploits the
simplicity of a deterministic model and uses the inheritance feature
from the evolutionary algorithms.
3.
Compatibility: The offspring forecasts from this algorithm can be
used as inputs to ANN to improve the performance.

Acknowledgments
The authors are grateful to the editors and reviewers for their suggestions
and comments to improve the earlier version of the chapter.

References
Box, G.E.P. and Jenkins, G.M. 1970. Time Series Analysis Forecasting and Control.
Holden-Day, San Francisco, CA.
Multiobjective Forecasting 153

Chikobvu, D. and Sigauke, C. 2012. Regression-SARIMA modelling of daily peak

electricity demand in South Africa. Journal of Energy in Southern Africa 23(3):
23–30.
Cools, M., Moons, E., and Wets, G. 2009. Investigating variability in daily traffic
counts using ARIMA X and SARIMA X models: Assessing impact of holidays
on two divergent site locations. In Transportation Research Board (TRB) 88th
Annual Meeting Compendium of Papers DVD, Washington, DC, January 11–15,
2009.
Hyndman, R.J. and Athanasopoulos, G. 2013. Forecasting principles and practice.
Available via: www.OTexts.com/fpp, accessed December 26, 2015.
Lee, M. and Hamzah, N. 2010. Calendar variation model based on ARIMAX for fore-
casting sales data with Ramadhan effect. In Proceedings of the Regional Conference
on Statistical Sciences, Kuala Lumpur, Malaysia, June 2010, pp. 349–361.
Nie, H., Liu, G., Liu, X., and Wang, Y. (2012). Hybrid of ARIMA and SVMs for short-
term load forecasting. Energy Procedia 16: 1455–1460.
Zou, H. and Yang, Y. 2004. Combining time series models for forecasting. International
Journal of Forecasting 20(1): 69–84.
7
A Class of Models for Microgrid
Optimization

Shaya Sheikh, Mohammad Komaki, Camelia Al-Najjar,

Abdulaziz Altowijri, and Behnam Malakooti

CONTENTS
7.1 Introduction................................................................................................. 155
7.2 Energy Models for Microgrids.................................................................. 158
7.2.1 Energy Operations Model............................................................. 159
7.2.1.1 Microgrid Operations Optimization Example............ 161
7.2.2 Microgrids Design Optimization................................................. 164
7.2.2.1 Example for Microgrids Design..................................... 165
7.2.3 Multiperiod Energy Model............................................................ 166
7.2.3.1 Multiperiod Energy Planning Optimization
Example............................................................................. 167
7.2.4 Bicriteria Optimization of Microgrids......................................... 169
7.2.4.1 Bicriteria Microgrid Operations Example.................... 169
7.3 Conclusion................................................................................................... 171
References.............................................................................................................. 171

7.1 Introduction
Microgrid is defined as a set of local energy generators, energy transmitters,
energy storages, and users that can work either independent (islanded mode)
from the main grid or in connection with the main grid. In practice, main
grids could function as a backup system for microgrids. Microgrid concept
was originally presented as a solution for creating a sustainable, green, and
efficient energy model; see Zamora and Srivastava (2010).
Renewable energy resources have had increasing and accelerating pen-
etration rate in microgrids since the turn of the century. However, the inter-
mittent nature of renewables such as solar and wind has always been a
challenge for their integration in microgrids. For example, electricity gener-
ated from solar panels is impacted by factors such as weather conditions and
time of day. Fortunately, energy storages can decrease this impact by add-
ing more flexibility to the system and by alleviating the imbalance between

155
156 Big Data Analytics Using Multiple Criteria Decision-Making Models

energy supply and demand. There has been promising progress toward
building and deploying energy storages in microgrid scale and with reason-
able costs.
Microgrids are becoming more and more intelligent and connected to each
other. Generators, storages, and users are equipped with the sensors that can
measure and monitor parameters of the system at every spot and in every
second. Entities are also connected to each other through internet or intranet
where they can communicate to each other. A controller (such as heating,
ventilation, and air conditioning [HVAC] in buildings) can then receive these
data and adjust the parameters of the system (e.g., energy usage) in real time.
These set of sensors generate huge amount of data that can grow exponen-
tially within a short span of time. Energy big data and renewable resources
amplify the urgency for designing more efficient and flexible microgrids.
Energy big data contains variety of data sets, including but not lim-
ited to weather data, energy market data, geographical data (e.g., global
positioning system [GPS] information), and field measurement data (e.g.,
device status, electricity consumption, storage level, etc.). These sets of
data are blended together and then classified in order to generate use-
ful information and insights about the whole system. For example, GPS
data help to visualize locational marginal price in a geographic context.
Common sources of big data in microgrids are smart appliances and
metering points throughout the grid. Metering points can generate new
set of data as frequent as every 5–15 minutes. These points generate large
volume of data which makes communication, complexity, and data storage
an inefficient and costly process; see Diamantoulakis et al. (2015). Data is
also being generated from generators, transformers, and local distribu-
tors. In addition, information from grid monitoring and maintenance are
generated on a regular basis. The volume of collected data from a grid
can easily reach to terabytes over a year. As a result, data-mining and
machine-learning tools are required to refine the data, discover meaning-
ful patterns in data, and generate useful information from it. Information
and insights are then applied to analyze energy price fluctuations in real
time, to analyze energy generators status, and to plan and monitor energy
system’s status.
According to Zhou et al. (2016), seven steps for managing big energy data
are: data collection, data cleaning, data integration, data mining, visualiza-
tion, intelligent decision making, and smart energy management. Given the
large volume of generated data and its dynamic nature, the data processing
would be a daunting and challenging task.
Data-mining methods such as regression, clustering, neural network, and
support vector machines are important tools for extracting useful and rel-
evant information from a large pool of data. Regression models, time series
models, and state–space models are among the most popular short-term
forecasting methods; see Kyriakides and Polycarpou (2007). Also, forecast-
ing can predict the demand accurately and optimization tool can achieve
A Class of Models for Microgrid Optimization 157

the optimal energy generation and distribution. Implementing these tools

together can result in effective dynamic pricing, planning, and operations
policies in energy systems.
Dynamic nature of data in microgrids highlights the importance of
short-term, very short-term, and ultrashort-term forecasting methods; see
Diamantoulakis et  al. (2015). Very short-term and ultrashort-term forecast-
ing are mainly used for immediate forecasting. However, there are fewer
­methods (such as artificial neural network) for these two categories. All
aforementioned data-mining tools are mainly used for averaging out
the measures in energy systems and are not useful for individual meters.
Therefore, ­customized methods such as empirical mode decomposition and
extreme learning with kernel are suggested for analyzing data from single
meters; see Yoo et al. (2011).
In this chapter, we concentrate on the modeling of microgrids in their sim-
plest form. We assume that data and energy are allowed to flow between
users, generators, and storages at any time. Therefore, all entities of a
microgrid communicate to each other and a controller is constantly collect-
ing and processing data from all entities in the system. One of the presented
models, called multiperiod energy model, utilizes the temporal data such as
energy prices as the input and generates the best energy generation sched-
ule among renewable generators, and best distribution among storages, and
consumers.
We present a number of operational and design energy models for solv-
ing microgrids problems. The first model, called energy operation model, is
used as a base to represent a microgrid. The second model, energy design
optimization, offers the best design for a microgrid with specified demand
range. The extension of first and second model that considers multiperiods
is also presented. Finally, the last model combines two important criteria of
cost and environmental impact in energy operation model. For further infor-
mation about presented models, see Malakooti (1986, 2014), Malakooti et al.
(2013), and Ravindran and Warsing (2012). A multicriteria energy operation
model is also explained in Sheikh et al. (2014).
Factories, residential buildings, office buildings, and hybrid cars are some
examples of microgrids. A basic microgrid is composed of energy genera-
tors, energy users, and energy storages; see Carmona and Ludkovski (2010).
These entities are linked together in order to supply, store, and consume
energy; see Figure 7.1. A system of interconnected microgrids is presented
in Figure 7.2.
There are always different alternatives for generating or procurement of
energy, but there are trade-offs among these alternatives. The multicrite-
ria microgrid approach distinguishes the most preferred different sources
with their associated generation and transportation costs. For further infor-
mation about multicriteria energy system models, see Sheikh et al. (Sheikh
and Malakooti, 2011, 2012; Sheikh, 2013; Sheikh et al., 2014, 2015), Ren et al.
(2010), Liu et  al. (2010), Fazlollahi et  al. (2012), Fadaee and Radzi (2012),
158 Big Data Analytics Using Multiple Criteria Decision-Making Models

FIGURE 7.1
A simple microgrid.

Microgrid 1 Microgrid 5

Microgrid 6

Microgrid 2
Microgrid 4

Microgrid 3

FIGURE 7.2
System of microgrids.

Alarcon-Rodriguez et  al. (2010), Shabanpour and Seifi (2015), and Sharafi
and ELMekkawy (2014).

7.2  Energy Models for Microgrids

Microgrid consists of three connected entities: users, energy storages, and
generators. A user such as a factory machinery receives energy from either
a generator or a storage. Storage, such as a battery, is a physical entity which
stores energy for later use. If the rate of energy production is higher than
the rate of energy consumption, the excess energy can be stored in the stor-
age for later use or be sold to the main grid. A generator such as a wind
A Class of Models for Microgrid Optimization 159

turbine generates energy. Demand of each user needs to be satisfied such

that it reduces energy costs and stores some energy for future use. Major
costs c­ onsist of generating and transportation (transmission and distribu-
tion) costs.
Users, generators, and storages are coordinating to minimize the total cost
and to satisfy the needed demands of the users. The energy stored by the
storages is paid for by the users of the system for future use. Stored energy
at the end of the period is left for the next period’s use. The amount in the
storage is set depending on the prediction of the next period’s energy prices.
If the prediction shows that generator’s prices will be higher, then more
energy is stored in storages for future use and vice versa. Δsp is defined as
the ­predicted change in the future price of energy. The storage can also act as
a hub in which all input energy from the generator is transported to the user
without storing energy in the storage.
Supply and demand behavior of storages: Depending on the price of the energy
in the future, it may be economical to store or release energy from storages;
see Escribano et al. (2011), Marckhoff and Wimschulte (2009). If the future
price of energy will be higher, then more energy should be stored in the
­storage now. On the other hand, if the future price of energy will be lower,
then the storage should sell as much energy as it can now, and the energy
level of the storage should be set to the minimum; see Conejo et al. (2005).

7.2.1  Energy Operations Model

A linear programming optimization approach for the operation of microgrids
is presented in this section. This model finds the optimal amount of energy
generated by each energy generator, stored in each storage for future use,
and transported on each link. Decision variables and costs for formulating
the energy operations model are presented in Table 7.1. Net flow is defined
as the difference between all incoming energy flows and all outgoing energy
flows:

Net flow = All outgoing energy flows − All incoming energy flows

The sign of net flow for each entity signifies the nature of that entity.
Specifically,

Net flow > 0 Generator

Net flow < 0 User
Net flow = 0 Storage

Nomenclature: By using the notation provided in Table 7.1, the microgrids

model for a given period is formulated as follows. For further explanation
160 Big Data Analytics Using Multiple Criteria Decision-Making Models

TABLE 7.1
Notations and Symbols Used for Microgrids
Parameter Index
Ki: Maximum capacity of gen. i i: Origin entity
D: Set of energy demand for users j: Destination entity
gi: Cost per unit of energy for gen. i
Decision Variable
qp, max: Maximum capacity of storage p
xij: Amount of energy from
qp, min: Minimum capacity of storage p
entity i to entity j
Δsp: Projected gain in dollars of energy price for next period
qp: Energy level of storage p
cij: Energy transportation cost from entity i to entity j

and applications of Model 1, see Malakooti et  al. (2013), Malakooti (2014),
Sheikh and Malakooti (2011, 2012), Sheikh (2013), and Sheikh et al. (2014, 2015).
Model 1: Energy operation optimization

Minimize f1 = ∑ ∑cx
for all i for all j
ij ij (7.1)

∑g ∑x
for all i
i
for all j
ij (7.2)

G ∑ s q .
for all p
p p (7.3)

Subject to:

∑x
for all i
ij ≥ Dj for all j demand constraint for users (7.4)

∑ x ≤k
for all j
ij i for all i capacity constraint for generators (7.5)

q p ,0 + ∑x −∑x
for all j
ij
for all i
ij = qp for all p (energy level of storagees) (7.6)

q p ≤ q p ,max for all (maximum allowable level for storages) (7.7)

q p ≥ q p ,min for all (minimum required level for storages) (7.8)

x ij ≥ 0, q p ≥ 0 for all i , j, and p. (7.9)

A Class of Models for Microgrid Optimization 161

Equation 7.1 calculates transportation cost from generators to users,

from generators to storages, and from storages to users. Cost of purchas-
ing energy from generators to users and storages are shown with Equation
7.2. G in Equation 7.3 is a given large positive number (e.g., set G = 100,000).
This equation finds total monetary gain of the system by storing energy at
current time. The objective function minimizes the total cost of generat-
ing, transferring, and storing energy. The marginal gain or loss for energy
storages is defined such that if Δsp > 0, then the energy price of the next
period will be higher. Therefore, store (or buy) energy for future use (i.e.,
q p is maximized) and if Δsp < 0, then the energy price of the next period
will be lower. Therefore, sell the existing energy in the storage (i.e., q p is
minimized).
The demand constraints (7.4) ensure that the total amount of energy
being transferred to the users from the generators and the storages is
greater than or equal to the energy demand by the users. The production
constraints (7.5) show that the amount of energy being transferred from
each generator to the users and the storages is not more than the maxi-
mum capacity of the given generator. Constraints (7.6) show the amount of
energy that is present in the storages at the end of the period. This amount
is equal to the amount of energy that is initially in the storage plus the
amount that is transferred from the generators to the storage minus the
amount that is transferred from the storage to users. Constraints (7.7) show
that the amount of energy in storage p at the end of the period should
be less than or equal to the maximum allowable capacity of the storage.
Constraints (7.8) ensure that the amount of energy in storage p at the end
of the period is more than or equal to the required minimum level of the
storage for emergency purposes. Based on the prediction of energy prices,
either more energy is stored in the storage or it is sold to result in less
energy in the storage. The nonnegativity constraints (7.9) ensure that all
amounts of energy generated, transferred, or stored will be nonnegative
values. See microgrid operations example.

7.2.1.1  Microgrid Operations Optimization Example

There exist three users, three generators, and two storages shown in Figure 7.3.
The demands for users 1, 2, and 3 and 4 are 400, 450, and 350 units of energy,
respectively. The generating energy cost per unit of energy for generators 1,
2, and 3 are $44.5, $44, and $43.5, respectively. Maximum capacities of gen-
erators 1, 2, and 3 are 1000, 950, 1200 units of energy, respectively. The initial
energy levels in storages 1 and 2 are 150 and 150 and the maximum capacity
of storages 1 and 2 are 250 and 240 units of energy, respectively. The mini-
mum capacity of storages 1 and 2 are 50 and 20 units of energy, respectively.
The data for this example is presented in Table 7.2. Suppose that Δs1 = +1 and
162 Big Data Analytics Using Multiple Criteria Decision-Making Models

Gen. 1 User 1
K1= 1000 Saver 1 D1= 400

User 2
D2= 450
Gen. 2
K2 = 950
Saver 2
User 3
Gent 3 D3=350
K2 = 120

FIGURE 7.3
Illustration for systems operations example with Δs1 = +1 and Δs2 = +2.

TABLE 7.2
Data for Microgrid Operations

User j User j
cij 1 2 3 cij 1 2 3
Gen. i 1 6 5 8 Storage i 1 1 2 1
2 7 4 5 2 2 0.5 1.5
3 5 4 6 Demand Dj 400 450 350

Storage Generator g i ($) Ki Storage q p,0 q p,min q p,max

j
1 44.5 1000 1 150 50 250
cij 1 2
2 44 950 2 150 20 240
Gen.i 1 4.5 5
3 43.5 1200
2 6 4
3 5 6

Δs2 = +2. The next period’s energy price is predicted to be higher for storages
1 and 2 (i.e., buy energy or Δsp > 0 for all storages).
Formulation for the microgrids operation is shown below. This formula-
tion is used to find the optimal solution that minimizes the total cost for the
energy operation example.
Optimal distribution of energy for above microgrid is as follows. Lingo 13
was used to find the optimal solution for this problem.
A Class of Models for Microgrid Optimization 163

Energy User j Energy Storage p Energy User j

Trans. x ij 1 2 3 Tot. Trans. y ip 1 2 Tot. Trans. zpj 1 2 3 Tot.
Gen. 1 0 0 0 0 Gen. 1 0 0 0 Storage 1 0 0 0 0
i 2 0 0 350 350 i 2 0 90 90 p 2 0 0 0 0
3 400 450 0 850 3 100 0 100 Tot. 0 0 0 0
Tot. 400 450 350 1200 Tot. 100 90 190

According to the solution, generator 2 should generate 350 and 90 units

of energy for user 3 and storage 2, respectively, and generator 3 should
generate 400, 450, and 100 units of energy for users 1, 2, and storage 1,
respectively. In total, generator 2 generates (350 + 90) = 440 units of energy
and generator 3 generates (400 + 450 + 100) = 950 units. At the end of
the period, the energy levels of storages 1 and 2 are q1 = 150 + 100 = 250
and  q2 = 150 + 90 =  240, respectively. The solution is presented in
Figure 7.4. The objective function value, f1, and total benefit of this exam-
ple is $66,365.

Energy User j Energy Storage p Energy User j

Trans. x ij 1 2 3 Tot. Trans. y ip 1 2 Tot. Trans. zpj 1 2 3 Tot.
Gen. 1 0 0 0 0 Gen. 1 0 0 0 Storage 1 0 0 100 100
i 2 0 0 120 120 i 2 0 0 0 p 2 0 0 130 130
3 400 450 0 850 3 0 0 0 Tot. 0 0 230 230
Tot. 400 450 120 970 Tot. 0 0 0

400 units

100 units

450 units

90 units

FIGURE 7.4
Solution for systems operations example.
164 Big Data Analytics Using Multiple Criteria Decision-Making Models

400 units

100 units

120 units 450 units

100 units

130 units

FIGURE 7.5
Solution for systems operations example with Δs1 = −1 and Δs2 = −2.

If the future prices were predicted to be lower than the current prices,
Δs1 = −1 and Δs2 = −2, the solution as presented in Figure 7.5 would be: q1 = 50,
q2 = 20, and f1 = $47,040.

7.2.2  Microgrids Design Optimization

In Section 2.1, an optimization model for planning and operations of existing
microgrids was discussed. This section presents the design of microgrids.
As an example, in a community, there are choices to buy different energy
generators and storages. In this model, both operational and design costs are
taken into consideration. The formulation of this model is based on energy
operation model discussed earlier.
Generator- and storage-purchasing costs need to be considered in the
objective function (7.1 through 7.3). The formulations for purchasing genera-
tors or storage are similar. Expression (7.10) considers the total purchasing
cost for generators which needs to be added to the objective function. ui is a
binary variable and ci is the cost of purchasing generator i.

+ ∑c u . i i (7.10)
i =1

The constraints of microgrids design are the same as energy operation

model except that binary variables for each generator are added, where 1
means the item is purchased, and 0 means it is not purchased. Generator
constraints (7.5) need to be replaced with constraints (7.11).

∑x
for all j
ij ≤ K iu i for i = 1, … , I (generator investment constraints),, (7.11)

A Class of Models for Microgrid Optimization 165

where

1 If generator i is chosen to be purchased 

ui :  .
0 otherwise 

For further explanation regarding Models 1 and 2, see Malakooti et  al.
(2013, 2014).
Model 2: Microgrid design optimization

Minimize objective function (7.1 through 7.3) + Expression (7.10)

Constraints (7.4), (7.6) through (7.9) and (7.11)

7.2.2.1  Example for Microgrids Design

The microgrid is illustrated in Figure 7.6. For simplicity, the arrows between
entities are not shown in the figure. Suppose that there is no generator and it
is possible to buy one or more of three possible generators. The users and
storages are existing entities as given in energy operation example. The input
parameters are as follows (Table 7.3).
The optimal design and operation plan for this example are found by
CPLEX 13. The solution is shown in Table 7.4.
In the final solution, u1 = 0, u2 = 1, u3 = 1, and u4 = 1 which means that the
best decision is to purchase generators 2, 3, and 4. The total cost is $249,192.5.
The solution is shown in Figure 7.7.

Gen. 1
K1= 500 Saver 1 1300 units

Gen. 2 1400 units

K2 = 1000

Saver 2
Gen. 3
K2= 900

Gen. 4
K3= 950

FIGURE 7.6
Energy systems design.
166 Big Data Analytics Using Multiple Criteria Decision-Making Models

TABLE 7.3
Information for Microgrids Design

User j User j
cij 1 2 cpj 1 2
Gen. i 1 4 4 Storage p 1 3.5 3
2 5 3 2 4 3.5
3 5 4 Demand Dj 1300 1400
4 4 6

cip,Trans. Storage Purchase Gen. Init. Min. Max.

Cost p Cost in Costs, Gen. Enr Enr Enr
from i to Gen $1000 g i ($) Cap., K i Storage q p,0 q p,min q p,max
p 1 2
1 25 60.5 500 1 50 50 300
Gen. i 1 1 2
2 25.5 60.75 1000 2 50 50 350
2 2 0.75
3 25.3 60.25 900
3 2 0.8
4 24.9 60.5 950
4 1 1

TABLE 7.4
Solution for Microgrids Design

User j Storage p User j

Energy Energy Energy
Trans. x ij 1 2 Tot Trans. y ip 1 2 Tot. Trans. zpj 1 2 Tot.
Gen. 1 0 0 0 Gen. 1 0 0 0 Storage 1 0 0 0
i i 2 0 0 0 p 2 350 0 350
2 0 1000 1000
3 0 400 400 3 0 350 350 Tot. 350 0 350
4 950 0 950 4 0 0 0
Tot 950 1400 2350 Tot. 0 350 350

7.2.3  Multiperiod Energy Model

Energy price in many microgrids are set as a variable rate where this rate is
a function of supply and demand. As a result, the market value of energy
changes during a given day. Variable energy price is addressed by energy
operations optimization for multiple periods. Suppose that there are T peri-
ods, where t = 1, 2, … , T. For multiperiod models, the formulation is sim-
ilar to Model 1. However, index t is added to all decision variables. Also,
Equation 7.6 changes to:
A Class of Models for Microgrid Optimization 167

950 units

1000 units

400 units

350 units
350 units

FIGURE 7.7
Solution for energy design example.

q p , t −1 + ∑x −∑x
for all j
ijt
for all i
ijt = q p ,t for all p (energy level of storages), (7.12)

where q p,t−1 and q p,t represent energy level in storage p at the end of period
t − 1 (or start of period t) and the end of period t, respectively. The total gain
for future savings is not used in this formulation since future prices of given
periods are known.
We introduce two approaches for solving multiperiod energy model: (1)
Period-by-period approach which is used when energy price and demands per
period are only known at the beginning of that period. In this case, the energy
model is solved at the beginning of each period; (2) Aggregate multiperiod
approach which is used when energy price and demands of a given number of
future periods are known at the beginning of the first period. If the data for all
upcoming periods are known at the beginning, multiperiod energy operation
can be solved in one step. In this case, energy stored in storages at the end of
period t − 1 will be considered as the initial energy level for period t.

7.2.3.1  Multiperiod Energy Planning Optimization Example

Consider energy operations example with a day and night period. Suppose
that the information for both periods is known at the beginning of the first
period. Different costs of buying and selling energy are considered. The
demands and cost information for day period are shown in Table 7.5. The
initial energy levels in storages 1 and 2 are 100 and 150, respectively.
Data for day period is provided in Table 7.6.
The formulation is similar to energy operation model except that
Constraints (7.13) show input/output of the different periods and their con-
nection together. The solution would be as shown in Table 7.7.
168 Big Data Analytics Using Multiple Criteria Decision-Making Models

TABLE 7.5
Information for Night-Period Energy Operation
Predicted Predicted
Change Change
in Energy in Energy Gen- Gen- Gen-
Price for Price for erator 1 erator 2 erator 3
Demand Demand Demand Storage 1, Storage 2, Costs Costs Costs
Period User 1 User 2 User 3 Δs1,t Δs2,t (g1,t) (g 2,t) (g3,t)
1, Night 200 250 100 4 3 44.5 44 43.5

TABLE 7.6
Information for Day-Period Energy Operation
Predicted Predicted
Change Change
in Energy in Energy Gen- Gen- Gen-
Price for Price for erator 1 erator 2 erator 3
Demand Demand Demand Storage 1, Storage 2, Costs Costs Costs
Period User 1 User 2 User 3 Δs1,t Δs2,t (g1,t) (g 2,t) (g3,t)
2, Day 500 600 300 −4 −3 44.5 44 43.5

Energy levels for storages in period 1 are q1,2,1 = 60, q2,2,1 = 150. In period

2, energy level for storage 2 will remain unchanged. However, storage 2
decreases its energy to 24. These energy levels are in accordance with the
expected behavior from storages.

TABLE 7.7
Solution for Multiperiod Energy Problem

User j Storage p User j

Energy Energy Energy
Trans. x ij1 1 2 3 Tot. Trans. y ip1 1 2 Tot. Trans. zpj1 1 2 3 Tot.
Gen. 1 0 0 0 0 Gen. 1 0 0 0 Storage 1 0 0 40 40
i i p 2 0 0 0 0
2 0 0 60 60 2 0 0 0
3 200 250 0 450 3 0 0 0 Tot. 0 0 40 40
Tot. 200 250 60 510 Tot. 0 0 0

User j Storage p User j

Energy Energy Energy
Trans. x ij2 1 2 3 Tot. Trans. y ip2 1 2 Tot. Trans. zpj2 1 2 3 Tot.
Gen. 1 0 0 0 0 Gen. 1 0 0 0 Storage 1 0 0 0 0
i 2 0 0 174 174 i 2 0 0 0 p 2 0 0 126 126
3 500 600 0 1100 3 0 0 0 Tot. 0 0 126 126
Tot. 500 600 174 1274 Tot. 0 0 0
A Class of Models for Microgrid Optimization 169

7.2.4  Bicriteria Optimization of Microgrids

Energy cost is considered an important criterion in microgrids. Another
important criterion is the environmental impact of the microgrid. Different
energy generators have different environmental impacts. This section intro-
duces bicriteria microgrids operations optimization with two criteria func-
tions: energy cost and environmental impact. The environmental impact
factor per unit of energy for each energy generator is presented by mi for
i = 1, … , I. The bicriteria model is as follows:
Model 3: Bicriteria of energy operations

Minimize total cost f1 = Use (7.1 through 7.3)

Minimize total environmental impact f2 = ∑m ∑x.
for all i
i
for all j
ij (7.13)

Subject to Constraints (7.4) through (7.9)
Σ for all j xij in (7.13) is the generated energy by generator i. This amount is
multiplied by environmental factor per unit of energy (mi) to find the total
environmental impact of generator i. The constraints of Model 3 are the same
as the constraints of energy operation model. To find the best alternative for
a given system, use the following Multi-Criteria Decision Making (MCDM)
process. In the first step, an upper and lower bound is found for each criteria
using the following:

Solution
Problem I: Minimize f1 subject to Constraints of energy operation model f1,min f2,max
Problem II: Minimize f2 subject to Constraints of energy operation model f1,max f2,min

Finally, an evenly distributed set of values for f2 is selected and used as a

constraint in the following bicriteria operations model.

Minimize f1 = Equations 7.1 through 7.3

Subject to: f2 = Equation 7.13
f2 ≤ the selected f2 value and Constraints of Model 1

Varying f2 over its range will result in many alternatives with a wide range
of solutions. The decision maker will be able to select the best alternative by
specifying a value for f2.

7.2.4.1  Bicriteria Microgrid Operations Example

Consider the following microgrid where the first generator is operated by
wind energy and the second generator is operated by coal with environ-
mental factors of m1 = 0.1 and m2 = 0.9, respectively. A set of alternatives are
170 Big Data Analytics Using Multiple Criteria Decision-Making Models

TABLE 7.8
Information for Bicriteria Microgrid Example

User j User j
cij, Trans. Cost cpj, Trans. Cost
from i to j 1 2 3 4 from p to j 1 2 3 4
Gen. i 1 6 5 8 4 Storage p 1 1 2 1 2
2 7 4 5 7 2 2 0.5 1.5 1
Demand Dj 600 800 400 300

cip, Trans. Storage p Gen. Gen. Env. Init. Min. Max.

Cost from Costs, Cap. Imp. Enr Enr Enr
i to p 1 2 Gen. g i ($) Ki mi Storage q p,0 q p,min q p,max
Gen. 1 4.5 5 1 40 1900 0.1 1 50 50 300
i 2 6 4 2 34 1800 0.9 2 50 50 350

presented for this example. Note that environmental impacts of storages are
not considered in this example (Table 7.8).
First, the range of the criteria must be found. Then, the criteria must be
calculated when minimizing only f1, the total cost. This model is as follows:

Minimize f1 Subject to : Constraints of Model 3

The solution to this example is shown as Problem I in Table 7.9.
Then, the criteria must be calculated when minimizing only f2, the
­environmental impact. This problem is as follows:

Minimize f 2 Subject to : Constraints of Model 3

The solution to this problem is shown as Problem II in Table 7.9.
Minimize f1
Subject to: f2 ≤ 1651.6
Constraints of Model 1

In the next step, an evenly distributed range of values for f2 is found. The
following five evenly distributed values of f2 are used: 384.4, 637.5, 1021.5,
1405.5, and 1651.6.
TABLE 7.9
Optimal Solution Considering Single Objectives
Solution
Problem I: Minimize f1 subject to Constraints of Energy Operation Model 83,886 1651.6
Problem II: Minimize f2 subject to Constraints of Energy Operation Model 94,806 384.4
A Class of Models for Microgrid Optimization 171

TABLE 7.10
Solution for Bicriteria Microgrid Example

User j Storage p User j

Enr. Enr. Enr Trans.
Trans. x ij 1 2 3 4 Tot. Trans. y ip 1 2 Tot. zpj 1 2 3 4 Tot.
Gen. 1 0 0 0 300 300 Gen. 1 16 0 16 Storage 1 26 0 0 0 26
i 2 0 800 400 0 1200 i 2 0 600 600 p 2 574 0 0 0 574
Tot 0 800 400 300 1500 Tot. 16 600 616 Tot 600 0 0 0 600

Storage Energy Level, qp 1 2

50 50

For example, for f2 = 1651.6, the following model is solved to find the opti-
mal solution for bicriteria energy operations example.
The solution to this bicriteria microgrids example is shown in Table 7.10.
The total cost (f1) and total environmental impact (f2) for this solution are
f1 = $83,886, and f2 = 1651.6, respectively. This solution shows that generator
1 generates a total of 316 units of energy while generator 2 generates 1800
units of energy.

7.3 Conclusion
Integrating renewable energy resources with microgrids requires develop-
ing flexible and efficient energy systems. In this chapter, energy operation
optimization problem was used as base for developing other applicable mod-
els such as multiple period microgrid operation model and bicriteria opera-
tion model. Linear structure of these models enables us to solve medium- to
large-scale problems in a matter of few seconds to few minutes. All presented
models can be applied as a base for more complex microgrid operation and
design problems where parameter values are subject to change every few
minutes. Bicriteria energy model can also be expanded to multicriteria model
by incorporating additional objectives such as reliability of energy system or
thermal comfort of energy users; see Sheikh (2013); Sheikh et al. (2014).

References
Alarcon-Rodriguez, A., Ault, G., and Galloway, S. 2010. Multi-objective planning of
distributed energy resources: A review of the state-of-the-art. Renewable and
Sustainable Energy Reviews, 14(5), 1353–1366.
172 Big Data Analytics Using Multiple Criteria Decision-Making Models

Carmona, R. and Ludkovski, M. 2010. Valuation of energy storage: An optimal

switching approach. Quantitative Finance, 10(4), 359–374.
Conejo, A. J., Fernandez-Gonzalez, J. J., and Alguacil, N. 2005. Energy procure-
ment for large consumers in electricity markets. IEE Proceedings-Generation,
Transmission and Distribution, 152(3), 357–364.
Diamantoulakis, P. D., Kapinas, V. M., and Karagiannidis, G. K. 2015. Big data
­analytics for dynamic energy management in smart grids. Big Data Research,
2(3), 94–101.
Escribano, A., Peña, J. I., and Villaplana, P. 2011. Modelling electricity prices:
International evidence. Oxford Bulletin of Economics and Statistics, 73(5), 622–650.
Fadaee, M. and Radzi, M. A. M. 2012. Multi-objective optimization of a stand-alone
hybrid renewable energy system by using evolutionary algorithms: A review.
Renewable and Sustainable Energy Reviews, 16(5), 3364–3369.
Fazlollahi, S., Mandel, P., Becker, G., and Maréchal, F. 2012. Methods for multi-
objective investment and operating optimization of complex energy systems.
Energy, 45(1), 12–22.
Kyriakides, E. and Polycarpou, M. (eds.) 2007. Short term electric load forecasting:
A ­tutorial. In Trends in Neural Computation (pp. 391–418). Springer, Berlin.
Liu, P., Pistikopoulos, E. N., and Li, Z. 2010. A multi-objective optimization approach
to polygeneration energy systems design. AIChE Journal, 56(5), 1218–1234.
Malakooti, B. 1986. Implementation of MCDM for the glass industry energy system.
IIE Transactions, 18(4), 374–379.
Malakooti, B. 2014. Production and Operations Systems with Multi-Objectives. John
Wiley.
Malakooti, B., Sheikh, S., Al-Najjar, C., and Kim, H. 2013. Multi-objective energy
aware multiprocessor scheduling using bat intelligence. Journal of Intelligent
Manufacturing, 24, 805–819.
Marckhoff, J. and Wimschulte, J. 2009. Locational price spreads and the pricing of
contracts for difference: Evidence from the Nordic market. Energy Economics,
31(2), 257–268.
Ravindran, A. R. and D. P. Warsing, Jr. 2012. Supply Chain Engineering: Models and
Applications. Boca Raton, FL: CRC Press.
Ren, H., Zhou, W., Nakagami, K. I., Gao, W., and Wu, Q. 2010. Multi-objective
optimization for the operation of distributed energy systems considering
­
­economic and environmental aspects. Applied Energy, 87(12), 3642–3651.
Shabanpour-Haghighi, A. and Seifi, A. R. 2015. Multi-objective operation manage-
ment of a multi-carrier energy system. Energy, 88, 430–442.
Sharafi, M. and ELMekkawy, T. Y. 2014. Multi-objective optimal design of hybrid
renewable energy systems using PSO-simulation based approach. Renewable
Energy, 68, 67–79.
Sheikh, S. 2013. Optimization and Risk Scenario Analysis of Energy Planning
Problem. PhD dissertation. Advisor: Professor Behnam Malakooti, Case
Western Reserve University.
Sheikh, S., Komaki, M., and Malakooti, B. 2014. Multi-objective energy operation
problem using Z utility theory. International Journal of Advanced Manufacturing
Technology, 74(9), 1303–1321.
Sheikh, S., Komaki, M., and Malakooti, B. 2015. Integrated risk and multi-objective
optimization of energy systems. Computers & Industrial Engineering, 90, 1–11.
A Class of Models for Microgrid Optimization 173

Sheikh, S. and Malakooti, B. 2011. Integrated energy systems with multi-objective.

IEEE Energy Tech Conference, Cleveland, OH, May 25–26, 2011, pp. 1–5.
Sheikh, S. and Malakooti, B. 2012. Design, operation, and efficiency of energy
­systems. IEEE Energy Tech Conference, Cleveland, OH.
Yoo, P. D., Ng, J. W., and Zomaya, A. Y. 2011. An energy-efficient kernel framework for
large-scale data modeling and classification. In IEEE International Symposium
on Parallel and Distributed Processing Workshops and PhD Forum (IPDPSW),
pp. 404–408.
Zamora, R. and Srivastava, A. K. 2010. Controls for microgrids with storage: Review,
challenges, and research needs. Renewable and Sustainable Energy Reviews, 14(7),
2009–2018.
Zhou, K., Fu, Chao, and Yang, Shanlin. 2016. Big data driven smart energy manage-
ment: From big data to big insights. Renewable and Sustainable Energy Reviews,
56, 215–225.
8
A Data-Driven Approach for Multiobjective
Loan Portfolio Optimization Using
Machine-Learning Algorithms and
Mathematical Programming

Sharan Srinivas and Suchithra Rajendran

CONTENTS
8.1 Introduction................................................................................................. 176
8.1.1 Impact of Banks on the Society..................................................... 176
8.1.2 Risks Associated with Lending.................................................... 176
8.1.3 Loan Approval Process.................................................................. 177
8.1.4 Problem Statement and Objective................................................ 177
8.2 Literature Review....................................................................................... 179
8.2.1 Predictive Analytics Approach for Loan Approvals................. 180
8.2.2 Motivation........................................................................................ 181
8.3 Methodology............................................................................................... 181
8.3.1 Predicting the Risk Associated with Each Applicant................ 182
8.3.1.1 Data Description.............................................................. 182
8.3.1.2 Data Cleansing................................................................. 183
8.3.2 Training and Testing Using Machine-Learning Algorithms......184
8.3.2.1 Logistic Regression.......................................................... 184
8.3.2.2 Artificial Neural Network.............................................. 184
8.3.2.3 Random Forests................................................................ 186
8.3.2.4 Stochastic Gradient-Boosted Decision Trees................ 188
8.3.2.5 Stacking............................................................................. 189
8.3.3 Evaluating a Learning Algorithm................................................ 189
8.3.4 MCMP Model.................................................................................. 191
8.3.4.1 Model Objectives.............................................................. 192
8.3.4.2 Model Constraints........................................................... 193
8.3.4.3 Solution Approach: Multiobjective Optimization
Using ε-Constraint Method............................................ 195
8.3.4.4 Demand Uncertainty in Loan Application Requests....196
8.4 Experimental Results................................................................................. 197
8.4.1 Input Parameters of Machine-Learning Algorithms................. 197
8.4.2 Analysis............................................................................................ 198

175
176 Big Data Analytics Using Multiple Criteria Decision-Making Models

8.4.3 Input Parameters of MCMP Model.............................................. 200

8.4.4 MCMP Model Analysis.................................................................. 201
8.4.4.1 Scenario Analysis............................................................. 201
8.5 Conclusions.................................................................................................. 204
References.............................................................................................................. 206

8.1 Introduction
In the United States, one in three Americans experiences financial prob-
lems and may not be able to completely fulfill his/her needs using his/
her savings or income (Soergel, 2015). Financial institutions, such as banks,
lend money to qualified borrowers to help them meet their credit needs and
the borrowers are expected to repay the loan in installments over a certain
time period. Irrespective of the size of the bank, loans contribute a major
portion of the bank’s total asset. For large commercial banks, such as Wells
Fargo, JP Morgan, Citibank, and Bank of America, loans as a percentage of
asset size is nearly 45%. For medium-sized banks, such as Capital One and
PNC, loans contribute about half of their total asset (Perez, 2015). Therefore,
banks use a major portion of deposited funds to issue different types of
loans (e.g., credit card loans, mortgage loans, auto loans) and earn a profit
by collecting interest on the loaned amount. The profit earned from a loan
is the difference between the total amount of money collected from the bor-
rower throughout the loan repayment period and the amount lent to the
borrower.

8.1.1  Impact of Banks on the Society

Banks and societies are mutual constitutions. Banks generate value by pro-
moting the well-being and fulfillment of the global population. In other
words, banks generate value by supporting the global population in financ-
ing, savings, electronic payment systems, and asset management. For exam-
ple, customers can safely deposit their savings in banks and avoid the major
risks involved in protecting them. Therefore, the society’s well-being ensures
the well-being of the banks and hence, banks must work in the best interest
of the community.

8.1.2  Risks Associated with Lending

As discussed earlier, revenue generated from loans is one of the largest assets
for any bank. However, there is also a huge risk involved because a percent-
age of the total approved loans may result in bad loans (debts that are not
recovered in time). Bad loans are one of the most common causes for bank
revenue losses and sometimes bankruptcy. Many national and international
A Data-Driven Approach for Multiobjective Loan Portfolio Optimization 177

banks suffered huge losses due to late payment or loan default resulting in
financial instability. For instance, the net loss for Bank of India in the sec-
ond quarter of 2015 due to bad loans was about $170 million (Tripathy, 2015).
In late 2000s, banks issued a large sum of money toward home and personal
loans without considering the risk of bad loans. Due to this, banks could
not derive profits from all the approved loans leading to a financial crisis
and a significant decrease in the loan approval rate. The decrease in the loan
approval rate led to a fall in house prices and as a result, the borrowers had
to sell more assets to repay their loans. This was one of the major causes of
the recession in 2009 and could have been avoided if the banks had carefully
chosen their borrowers.

8.1.3  Loan Approval Process

The sequence of events involved in approving a loan starting from the
submission of loan applications by the potential borrowers to the final
­
­outcome is referred to as the loan approval process. Figure 8.1 illustrates the
steps involved in the loan approval process as described by Power (2002).
The  loan application process begins when the applicant initiates a loan
request by first submitting the preapproval documents (e.g., proof of employ-
ment, driver’s license) requested by the bank. These documents are then ver-
ified by the lender in the preapproval process. If the documents are valid,
then the loan preapproval statement is sent to the applicant and additional
supporting d
­ ­ ocuments (e.g., number of dependents, number of existing
loans) are requested to make a final decision. The lender’s team of under-
writers ­processes these documents and inputs the borrower’s attributes
into the loan approval model (a decision support system for loan approv-
als). Some of the criteria that are given as inputs to the loan approval model
are income, open line of credits and age, and these criteria may vary from
one bank to another. Based on the outcome of the decision support system
and the ­strategic goals of the company (e.g., limit on the number of loans
approved), the ­underwriter decides whether to issue a loan to the applicant
and he/she is notified of the bank’s decision.

8.1.4  Problem Statement and Objective

Even though the largest asset of banks is the revenue generated from loans, it
is necessary to consider the huge risk of approving bad loans to avoid bank-
ruptcy. Therefore, it is evident that a robust loan approval system is essential
for efficient operation of the banks. Banks have certain strategic goals (e.g.,
policies, targets) for their loan portfolio. In addition, the loan portfolio must
be efficient to achieve the optimal risk–return tradeoff (i.e., a portfolio that
offers the highest expected return for a specified risk level or lowest risk
level for a specified level of expected return). Hence, the objectives of this
chapter are to:
178 Big Data Analytics Using Multiple Criteria Decision-Making Models

Applicant requests loan

Applicant provides preapproval documents

Preapproval
documents pass No Deny the loan application
initial verification?

Yes

Financial institution requests additional supporting

documents from the applicant

Documents are sent to underwriter

Underwriter derives the attributes (features) associated

with the applicant using the supporting documents

Attributes associated with the applicant are given as

inputs to the loan approval model

Underwriter decides
No
to approve loan?

Yes

Loan approved

FIGURE 8.1
Steps involved in loan approval process.
A Data-Driven Approach for Multiobjective Loan Portfolio Optimization 179

• Analyze existing loan applicant data to predict the risk type (high
risk or low risk) of forthcoming applicants using machine-learning
algorithms.
• Develop a set of efficient loan portfolio that presents a risk–return
trade-off and meets the strategic goals of the financial institutions
using multicriteria mathematical programming (MCMP).

8.2  Literature Review

The problem of loan portfolio optimization has been studied for many
decades (e.g., Altman, 1980; Clemente, 1980; Taylor, 1980). Banks may lose
their revenue and customer base if there is a delay in the loan approval
process. Therefore, it is necessary for the banks to be responsive in process-
ing the loan requests and notifying the decision to the applicants (Lovati,
1975; Bennett, 1984; Van Leuvensteijn, 2007). An efficient decision support
system enables the banks to quickly make decisions on loan approvals.
Most type of loans, such as mortgage loans, have a long repayment period
leading to a long-term relationship with the borrowers. Due to this, banks
have to choose borrowers who are less likely to be delinquent on their
loan payment (Burke and Hanley, 2003; Greenbaum and Thakor, 2007).
In any portfolio management, the two key objectives are maximizing the
expected return and minimizing the expected risk (Sharpe, 1970; Fama and
Miller, 1972; Greenbaum and Thakor, 2007). An ideal loan portfolio would
have the highest return and lowest risk. However, in reality, it is not pos-
sible to achieve the ideal portfolio and the decision makers must, therefore,
resort to a portfolio that is efficient. A portfolio is considered to be efficient if
the improvement to an objective is only possible at the expense of the other
objective.
Saunders et  al. (2007) proposed an optimization model for loan portfo-
lio selection. However, Sirignano et  al. (2015) considered the optimization
model to be static and hence proposed a dynamic model. According to the
authors, the loan selection problem is a binary decision and the mean–vari-
ance ­models can be extended for loan portfolio selection. The authors proved
that for large-scale problems, the integer programming model becomes
extremely complex, and overcame this issue by using an approximate
­optimization approach.
Gerlach and Peng (2005) studied the relationship between residential
prices and bank lending in Hong Kong, and concluded that the bank lend-
ing is correlated with residential prices in both developing and developed
economies. Also, borrowing capacity and credit limit depend on the resi-
dential prices. Therefore, residential prices affect the banks’ capital posi-
tion both directly and indirectly. Moreover, the factors that affect the loan
180 Big Data Analytics Using Multiple Criteria Decision-Making Models

approvals may vary depending on the type of loan. For instance, factors
such as loan amount, total income, number of dependents, and real estate
securities are major factors that impact the approval of auto loans (Hill,
2014), and factors such as history of late payments, credit scores, payment
for existing debt, and loan tenure impact the approval of mortgage loans
(Choi, 2011).

8.2.1  Predictive Analytics Approach for Loan Approvals

Zurada and Barker (2011) suggested that the accuracy of predicting the risk
type of the borrower (good or bad borrower) plays a significant role in the
bank’s financial stability. Due to this, machine-learning techniques, such
as logistic regression and discriminant analysis, are studied (Bell et  al.,
1990; Khandani et al., 2010; Neelankavil, 2015). Ince and Aktan (2009) com-
pared different techniques such as discriminant analysis, logistic regres-
sion, neural networks, and classification and regression trees (CART) for
credit scoring in banking. Historical data from a financial institution were
used to evaluate the techniques, and the authors observed that CART
and neural networks outperformed other techniques. The authors also
considered the probability of making an error of granting loan for a bad
­c ustomer and the probability of not granting loan for a good customer.
A detailed review of the methods used to classify the “good” and “bad”
risk classes in the literature is given by Hand and Henley (1997). The
methods ­discussed in their paper classified the applicants based on their
repayment behavior.
Banks are faced with the challenge of using good classification models
to make sound decisions and gain competitive advantage. Several methods
used for classifying good and bad risks are given below:

• Discriminant analysis (Durand, 1941; Meyers and Forgy, 1963; Lane,

1972; Apilado et al., 1974)
• Regression (Orgler, 1970; Fitzpatrick, 1976; Wiginton, 1980; Srinivasan
and Kim, 1987; Leonard, 1993; Henley, 1994; Lucas, 2004)
• Mathematical programming methods (Hand, 1981; Showers and
Chakrin, 1981; Kolesar and Showers, 1985)
• Artificial neural network (ANN; Davis et  al., 1992; Ripley, 1994;
Rosenberg and Gleit, 1994)
• Time-varying models (Bierman and Hausman, 1970; Dirickx and
Wakeman, 1976; Srinivasan and Kim, 1987)
• Random forest (RF; Brown and Mues, 2012; Wang et al., 2012; Kruppa
et al., 2013)
• Stacking (Wang and Ma, 2012; Koutanaei et al., 2015)
A Data-Driven Approach for Multiobjective Loan Portfolio Optimization 181

8.2.2 Motivation
In reality, financial institutions offer different types of loans (such as
mortgage loans, commercial loans) to serve the varying customer needs.
The return varies depending on the type of loan and the risk type of the
­borrower. However, most of the existing researches primarily focus on
developing a recommender system for loan approvals and do not take into
account the different types of loans. Moreover, the strategic goals of the
financial institution (e.g., restriction on the total approvals for a particular
type of loan) are seldom taken into consideration. Also, most of the previ-
ous researches consider only single loan applications and ignore the pos-
sibility of joint loan applications (application with coborrower). If the
applicant has a bad credit score, then the probability of loan acceptance can
be increased by having a cosigner with a good credit score because most
banks use the better of the two credit scores to determine the eligibility for
loan approval (Somers and Hollis, 1996). Therefore, the present work aims
to address these gaps in the literature. The proposed two-stage decision
­support system involves the integration of data analytics with multiple
criteria decision making (MCDM). In the first stage, we evaluate different
machine-learning algorithms to classify the risk type of the applicants. In
the second stage, we propose an integer programming model with mul-
tiple objectives to construct the loan portfolio. The MCMP model uses the
risk type of the borrower as one of the inputs and constructs the portfolio
by considering the different types of loans, strategic goals of the banks,
and both single and joint applications.

8.3 Methodology
The proposed methodology is illustrated using Figure 8.2.
In the first stage, the borrowers are categorized as high risk or low risk by
different machine-learning classifiers using factors such as borrower’s age
and late payment history. The best machine-learning classifier is selected
based on the output performance measures discussed in Section 8.3.3.
It  is  important for banks to differentiate between high- and low-risk bor-
rowers because the interest rate for all borrowers depends on their risk type.
High-risk borrowers are often charged more interest rate to compensate for
the high probability of loan defaulting (Diette, 2000). In the  second stage,
a diversified loan portfolio is developed with the objective of achieving
higher interest rate returns and lower risk using an MCMP model. Since the
risk type of the potential borrowers impacts the return and the risk of the
portfolio, the solutions obtained in the first stage using machine-learning
­techniques are used as an input to the second stage.
182 Big Data Analytics Using Multiple Criteria Decision-Making Models

Stage 1: Machine learning to select best classifier

Raw data

Data cleansing

Stage 2: Multicriteria mathematical

programming
Training data Testing data
Objective
• Maximize portfolio return
Using training data Using testing
inputs and target outputs data inputs • Minimize portfolio risk

Subject to
Train the classifier Predict the output of
• Achieve diversification on different loan types
the trained classifier
• Achieve age diversification
Using testing data target • Limit the average
outputs and predicted output • Number of dependents in the portfolio
Choose another Evaluate the • Debt-to-income ratio
classifier classifier • Number of open credits of the portfolio

Yes
Evaluate another
classifier?

No
Compare the performance of all
evaluated classifiers

Select the best classifier

FIGURE 8.2
Overview of research methodology.

8.3.1  Predicting the Risk Associated with Each Applicant

In this section, the tasks involved in determining the risk associated with
applicants are discussed.

8.3.1.1  Data Description

The data used for determining the risk type of an applicant (high risk or
low risk) were downloaded from the Kaggle competition site (Kaggle, 2011).
A detailed description of the downloaded data is presented in Table 8.1.
The variable “Risk” is the output variable and all the other variables are
treated as input variables or features or predictors. Therefore, the data has
ten predictors and one response variable. The input variables are not highly
correlated, and hence all the variables can be used to predict the response
variable. The downloaded raw data contains 150,000 samples. As in any
real-world data, these data also suffer from user input errors and missing
values. The user input errors are values that appear to be meaningless for
a variable under consideration. For example, the age of some of the borrow-
ers was entered as 0 and the monthly income of some of the borrowers was
A Data-Driven Approach for Multiobjective Loan Portfolio Optimization 183

TABLE 8.1
Description of Data
Variable Name Description Type
Risk Risk associated with the borrower Binary (high risk or low risk)
Age Age of the borrower (in years) Integer
Debt ratio Ratio of monthly debt payments to Continuous between 0 and 1
monthly gross income
LOC Number of open loans and lines of credit Integer
Income Monthly income of the borrower Continuous
MREL Number of mortgage and real estate loans Integer
Dependents Number of dependents of the borrower Integer
Utilization Ratio of total balance on lines of credit to Continuous between 0 and 1
the total credit limits
30 days Number of times the borrower has been Integer
30–59 days past the due date in the last
2 years
60 days Number of times the borrower has been Integer
60–89 days past the due date in the last
2 years
90 days Number of times the borrower has been Integer
equal to or more than 90 days past the
due date

entered as $1. In addition, there were 29,731 instances in which the monthly
income was missing, and 3924 instances in which the number of dependents
was missing. The presence of erroneous or missing data can significantly
impact the results of the analyses, and therefore, it is important to cleanse the
data before using them as inputs to the machine-learning algorithm. Section
8.3.1.2 describes the various data-cleansing techniques used to handle the
user input errors and missing values.

8.3.1.2  Data Cleansing

The process of detecting any inconsistencies in the data and replacing them
with suitable values and making them usable is called data cleansing or
­scrubbing. In order to handle the user input errors and missing values in the
data, several alternatives are considered, and are as follows:
Substitute with a Unique Value: The data entry error and missing data are
coded with a value that never occurs in the data set. For example, all the
missing values and the user input errors can be substituted with −1.
Substitute with Median: The data entry error and missing data are substi-
tuted with the median value of that feature.
Discard Variable and Substitute with a Unique Value: The features with too
many missing values are removed and the data entry error and missing
­values for the remaining features are then coded as −1.
184 Big Data Analytics Using Multiple Criteria Decision-Making Models

Discard Variable and Substitute with Median: The features with too many
missing values are removed and the data entry error and missing data for
the remaining features are then substituted with the median value of that
feature.
Discard Incomplete Rows: The rows containing error or missing values are
removed from the data set.

8.3.2  Training and Testing Using Machine-Learning Algorithms

Several classification techniques are evaluated using the data obtained from
Kaggle (2011). The data is split into training data and testing data. Each algo-
rithm or classifier uses the training data to learn the underlying relationship
between the input features and the response variable, and the testing data to
evaluate the strength of the trained classifier. Therefore, the classifier auto-
matically learns to classify the risk type of the potential borrowers using the
features in the training phase, and the trained classifier uses the features of
the testing data to predict the output (i.e., risk type of the potential borrower).
Five different machine-learning techniques are used and the best method
for classifying the risk type is identified. The five methods considered are:
(1) logistic regression, (2) RFs, (3) neural networks, (4) gradient boosting, and
(5) stacking. A supervised learning procedure is used in the five machine-
learning algorithms, and hence the training of the methods involves the use
of known inputs and outputs.

8.3.2.1  Logistic Regression

The probability that the borrower will be of high risk or low risk is com-
puted using the input features and is shown in Equations 8.1 and 8.2.
b0 is a constant and bis are the regression coefficients of the input param-
eters. The training data set is used to estimate the regression constant and
­regression ­coefficients that best fit the observed data.

1
P(Risk Type = High) = − ( b0 + b1 * Age + b2 * Debt Ratio + b3 * LOC + b4 * Income ++ b10 * 90 Day ) (8.1)
1+ e

P(Risk Type = Low) = 1 − P (Risk Type = High). (8.2)

A schematic representation of the relationship between the input features

and the response variable with a logistic function is shown in Figure 8.3.

8.3.2.2  Artificial Neural Network

ANN is a machine-learning algorithm that is inspired by the biological neu-
ral network. As shown in Figure 8.4, ANN includes three different types
A Data-Driven Approach for Multiobjective Loan Portfolio Optimization 185

Age
b0

b1
Debt
ratio b2 Logistic
Training data

Final prediction
function
b3

LOC
b10

90 days

FIGURE 8.3
Representation of the logistic regression algorithm.

of layers, namely, an input layer, hidden layer, and output layer and each
layer has a certain number of nodes. Each node (i) in a given layer (l) is con-
nected to each node (j) in the next layer (l + 1) by a connection weight (wij).
In order to train the classifier, the features are given as inputs to the input
layer. Each input value is multiplied by a weight at the input layer, and the
weighted input is relayed to each node in the hidden layer. Each node in
the hidden layer will combine the weighted inputs that it receives, use it
with the activation function (e.g., sigmoid activation), and relay the value
to the nodes in the output layer. The output layer then determines the net-
work output (risk type) by performing a weighted sum of the outputs of the
hidden layer. The process of using the training inputs, hidden layers, and
activation function to compute the response variable (risk type) is called
feed-forwarding or forward pass. Initially, the training process begins with ran-
dom weights. At the end of each feedforward step, the predicted output is
compared with the actual output. If the predicted risk type is same as the
actual risk type, then the neural network’s weights are reinforced. On the
other hand, if the predicted risk type is incorrect, then the neural network’s
weights are adjusted based on a feedback. This process is called backward
pass. The forward pass and the backward pass are repeated for different
186 Big Data Analytics Using Multiple Criteria Decision-Making Models

1
W0a
W0b
W0n

Age W1a
W1b
a
W1n Wah

Debt W2a Wal High risk

ratio W2b

Final prediction
W2n Wbh
Training data

b
W3a
Wbl
LOC W3b
W3n

Wnh Low risk

W10a W10b Wnl
n
W10n
90 day

Input layer Hidden layer Output layer

FIGURE 8.4
Representation of a three-layered neural network algorithm.

training samples until the ANN classifier is fully trained. ANN can be use-
ful to uncover complex relationship between the features and the output.
However, it requires more parameters to be estimated, and therefore, may
require more time for training.

8.3.2.3  Random Forests

RF is an ensemble of decision trees proposed by Breiman (2001). Decision
trees use a tree-like structure to obtain the output class and have nodes
at each level of the tree. Each node splits into two nodes in the next level
and the data set is divided among the two nodes based on a test (e.g., is the
­feature “age” greater than 50?). This process is repeated until the output class
(risk type) is reached. At each level, it is necessary to select the feature that is
most useful for classifying the response variable, and information gain is a
metric to measure the usefulness of a feature at that level. Information gain
is the expected reduction in entropy due to sorting on a given node. Entropy
is the measure of impurity and a higher entropy indicates more information
A Data-Driven Approach for Multiobjective Loan Portfolio Optimization 187

content. If pi indicates the probability of class i, then Equations 8.3 and 8.4
give the entropy and information gain, respectively.

Entropy = ∑ − p log i 2 pi (8.3)

i

Information gain = Entropy (parent) − Weighted average

[Entropy (children)]. (8.4)

Each decision tree provides an output class and the final output class is the
plurality voting of all the decision trees. Figure 8.5 is a schematic representa-
tion of the RF algorithm.

Training data

Debt
Age LOC 90 days
ratio

Bootstrapped Bootstrapped Bootstrapped

data data data

Decision tree 1 Decision tree 2 Decision tree m

Prediction of tree 1 Prediction of tree 2 Prediction of tree m

Plurality voting

Final prediction

FIGURE 8.5
Representation of the random forests algorithm.
188 Big Data Analytics Using Multiple Criteria Decision-Making Models

8.3.2.4  Stochastic Gradient-Boosted Decision Trees

Stochastic gradient boosting is an ensemble method introduced by Friedman
(2002). It iteratively trains shallow decision trees with the objective of
­minimizing a loss function (e.g., negative log-likelihood) and sequentially
learns from the errors of the previous trees. Therefore, the trees are trained
one at a time and cannot be trained in parallel. A schematic representation
of stochastic gradient-boosted decision trees (SGBDT) algorithm is shown in
Figure 8.6.
The samples that are wrongly classified by a decision tree are upweighted,
and the samples that are correctly classified are downweighted. This
­procedure is continued iteratively for all the decision trees leading to a
higher weight for observations with correctly classified outputs and a lower
weight for observations with wrongly classified outputs. It is to be noted that

Training data

Debt
Age LOC 90 days
ratio

W1 Bootstrapped W2 Bootstrapped Bootstrapped

data data Wm data

Decision tree 1 Decision tree 2 Decision tree m

Prediction of Prediction of Prediction of

tree 1 tree 2 tree m

Weighted voting

Final prediction

FIGURE 8.6
Representation of the SGBDT algorithm.
A Data-Driven Approach for Multiobjective Loan Portfolio Optimization 189

the SGBDT randomly samples a subset of the training data to train each deci-
sion tree. The final output would be a weighted voting of the decision trees.

8.3.2.5 Stacking
Stacking is the idea of combining the predictions of multiple classifiers and
it involves two phases. In the first phase, the features are independently
trained using different classifiers (e.g., logistic regression, RFs). The classi-
fiers used to train the features are called base-level classifiers. In the second
phase, the predicted outputs obtained from the base-level classifiers are
used as inputs to the second-phase classifier (e.g., ANN, RFs) called meta-
level ­classifier. In  other words, the second phase combines the individual
­predictions of the base-level classifiers. The final class is the output obtained
from the meta-level classifier. A schematic representation of the stacking
­algorithm is shown in Figure 8.7.

8.3.3  Evaluating a Learning Algorithm

Given the set of input values, the trained classifier provides the probability
value for the risk type of borrowers. A threshold is then required to convert

Training data

Debt
Age LOC 90 days
ratio

Base classifier 1 Base classifier 2 Base classifier m

Prediction of Prediction of Prediction of

classifier 1 classifier 2 classifier m

Meta-level classifier

Final prediction

FIGURE 8.7
Representation of stacking.
190 Big Data Analytics Using Multiple Criteria Decision-Making Models

the probability value to either high risk or low risk. For instance, a threshold of
0.50 indicates that any value below 0.50 is low risk and above 0.50 is high risk.
A classifier may have any threshold value (e.g., 0.70) depending on the data.
Therefore, the area under the curve (AUC) value for receiver operating charac-
teristics (ROC) is used to obtain the best threshold value for a given classifica-
tion problem. The AUC value quantifies the overall ability of the classifier to
discriminate between the risk types of the applicant. A ­random classifier has
an AUC value of 0.50 and a perfectly accurate classifier has an AUC value of
1.0. The ROC curve plots the true-positive rate (TPR) versus the false-positive
rate (FPR) for various threshold settings as shown in Figure  8.8. Therefore,
each point on the ROC curve represents the TPR/FPR value c­ orresponding
to a particular threshold. The dotted line has an AUC value of 0.50 and is the
performance of a random classifier. A perfect classifier would yield a point in
the upper left corner or the coordinate (0,1) of the ROC curve. Hence, using the
ROC curve, a decision maker can choose a threshold value that is closest to the
(0,1) coordinate. In other words, the threshold should be chosen in such a way
that the TPR is high and the FPR is low. Once the threshold value is obtained,
the probability values below the threshold are categorized as low-risk bor-
rowers and values above the threshold are categorized as high-risk borrowers.
The accuracy of the classifier is determined by using the actual output and
the predicted output by constructing a confusion matrix. A model has high
accuracy if the actual outputs are same as the predicted outputs for many
instances. As shown in Figure 8.9, the confusion matrix has four categories,
namely, true positive, true negative, false positive, and false negative.
If a classifier predicts the risk type of a potential borrower as high risk
and if the actual risk type of the potential borrower is low risk, then it is a
false positive or Type I error. Type I error may result in loss of customers to
the financial institution because the borrower may go to a competitor who

ROC curve
1.0

0.8
True-positive rate

0.6

0.4

0.2

0.0
0.0 0.2 0.4 0.6 0.8 1.0
False-positive rate

FIGURE 8.8
A sample diagram of the ROC curve.
A Data-Driven Approach for Multiobjective Loan Portfolio Optimization 191

Actual output
High risk Low risk
False positive
Predicted High risk True positive (type I error)
output
False negative
Low risk (type II error) True negative

FIGURE 8.9
Elements of a confusion matrix.

offers a lower interest rate. If a classifier predicts the risk type of a potential
­borrower as low risk and if the actual risk type of the borrower is high risk,
then it is a false negative or Type II error. Type II error may result in loss of
revenue for the bank since the borrower has a higher chance to default on
his loan. It is important to avoid both the Type I and Type II errors because
the loan portfolio model seeks to reduce the risk and increase the return.
If a ­classifier predicts the risk type of a potential borrower as high risk and
if the actual borrower is high risk, then it is a true positive. If a c­ lassifier
­predicts the risk type of a potential borrower as low risk and if the actual
borrower is low risk, then it is a true negative. True positive and true nega-
tive both improve the accuracy of the classifier. Therefore, three metrics,
TPR (­sensitivity), true ­negative rate (specificity) and accuracy, are derived
from the confusion matrix and are used in this present work to evaluate a
­classifier, and are given below:

TPR (Sensitivity ) =
∑ True Positive (8.5)
∑ True Positive + ∑ False Negative
True Negative Rate (Specificity ) =
∑ True Negative
∑ True Negative + ∑ False Positive
(8.6)

Accuracy =
∑ True Positive + ∑ True Negative . (8.7)
∑ True Positive + ∑ False Negative + ∑ False Positive
+ ∑ True Negative

8.3.4  MCMP Model

In this section, the proposed multiobjective integer programming model for
constructing the loan portfolio is discussed. The notations used in the model
are given below:
192 Big Data Analytics Using Multiple Criteria Decision-Making Models

Sets and Indices

c ∈ C Set of all age categories
t ∈ T Set of all loan types
a ∈ A Set of all loan applications
a ∈ N(t) Set of all loan applications of loan type t
b ∈ B(a) Set of all applicants for application a B(a) ∈{1,2}
Note: B(a) = 1 if single loan application and 2 if joint loan application

Input Parameters Associated with Loan Applicant

Ea,b,c 1 if age of bth borrower of application a falls within age category c, 0 otherwise
Da,b Total number of dependents for bth borrower of application a
DRa,b Debt ratio for bth borrower of application a
Ra,b Risk associated with bth borrower of application a (output of Stage 1)
La,b Number of late payments made by bth borrower of application a
Oa,b Number of open credits for bth borrower of application a

Input Parameters Associated with Financial Institution

TcE Lower bound on the percentage of total borrowers for age category c
TD Upper bound on number of dependents
TDR Upper bound on debt ratio
TL Upper bound on late payment
TO Upper bound on open credit
LPt Lower bound on percentage of loan issued to loan type t
UPt Upper bound on percentage of loan issued to loan type t
Ia Interest rate of loan issued to application a

Decision Variables
Δa 1 if application a is approved; 0 otherwise
δa,b 1 if b is the primary borrower for application a; 0 otherwise

8.3.4.1  Model Objectives

Objective 1: Maximize the total return due to loan interest
Bank’s stability is essential for a bank to function effectively. The increase
in revenue results in the improved capital base of banks, thereby increasing
the stability. The main source of revenue for banks is from interest generated
from loans. Therefore, it is necessary to maximize the ­revenue generated
from loans and hence first objective considered in the loan portfolio opti-
mization model is to maximize the return generated from the loan as given
below:

Max Z1 = ∑(I × ∆ ).
a a (8.8)
a ∈A

Objective 2: Minimize the portfolio’s risk

A Data-Driven Approach for Multiobjective Loan Portfolio Optimization 193

If many borrowers fail to meet the financial obligations mentioned in the

loan agreement, then the asset of the bank becomes less than its liability
and may even reach a state of insolvency. Therefore, the bank must ­evaluate
the risk of the applicant defaulting on a loan and hence, objective 2 is to
­minimize the total risk incurred by the bank as given below:

Min Z2 = ∑ ∑ (R
a ∈A b ∈B( a )
a,b × δ a , b ). (8.9)

8.3.4.2  Model Constraints

8.3.4.2.1  Limit the Average Debt-to-Income Ratio Associated with the Borrowers
Debt-to-income ratio is a measure of the proportion of gross income being
paid toward debt. A high debt-to-income ratio indicates that the borrower
either has a high debt or does not have sufficient income to repay the debt.
Evidence from the literature suggests that a borrower with a high debt-to-
income ratio is more likely to have trouble in repaying the loan (Lanza, 2014).
Hence, financial institutions consider debt-to-income ratio as an important
factor when deciding to lend money or approve credit. Therefore, constraint
(8.10) ensures that, on an average, the debt ratio of the borrowers in the
­portfolio must be within the bank-specified debt ratio value (TDR).

∑ ∑ (DR × δ
a ∈A b ∈B( a )
a,b a,b )
≤ T DR . (8.10)

∑ ∑ δ a ∈A b ∈B( a )
a,b

8.3.4.2.2  Diversify the Types of Loan

Borrowers have a variety of purposes for requesting loans, such as p ­ urchasing
a new house or funding for a new business (Curtis, 2007). Bank loans can
be categorized into low interest generating loan types, such as c­ ommercial
loans or commercial real estate loans, and high interest generating loan
types, such as consumer loans and credit card loans. A diversified portfolio
helps to maintain a steady revenue when the types of loans are not highly
correlated. Therefore, it is necessary to consider different types of loans to
achieve loan diversification ensuring customer satisfaction. Constraints
(8.11) and (8.12) achieve the loan diversity across the different types of loans.

∑∆
a ∈N ( t )
a ≥ LPt × (| A|) ∀t ∈ T (8.11)

∑∆
a ∈N ( t )
a ≤ UPt × (| A|) ∀t ∈ T . (8.12)

194 Big Data Analytics Using Multiple Criteria Decision-Making Models

8.3.4.2.3  Limit the Average Late Payment Frequency of the Borrowers

Avoiding a payment, missing a payment, or late payment may result in the
reduction of the net revenue generated by the bank. In addition to the lost
revenue, there are other costs, such as the cost involved in contacting the
borrower to reclaim the loan through debt collection agency (Sullivan, 2014).
Therefore, it is necessary to choose borrowers who are less likely to make
late payments, and financial institutions use the borrower’s late ­payment
history to determine the likelihood of making late payments in the future.
Constraint  (8.13) ensures that, on an average, the late payments of the
­borrowers in the portfolio must be within the bank-specified late payment
value (TL).

∑ ∑ (L × δ
a ∈A b ∈B( a )
a,b a,b )
≤ T L. (8.13)

∑ ∑ δ a ∈A b ∈B( a )
a,b

8.3.4.2.4 Limit the Average Number of Dependents Associated

with the Borrowers
Studies have proven that the number of dependents is a significant fac-
tor to be considered when approving loans (Riungu, 2014). Borrower with
many dependents may experience increased financial burden and, therefore,
increases his/her likelihood of not paying the loan on time. Constraint (8.14)
ensures that, on an average, the number of dependents of the b ­ orrowers in
the portfolio must be within the bank-specified number of dependents
value (TD).

∑ ∑ (D × δ
a ∈A b ∈B( a )
a,b a,b )
≤ T D. (8.14)

∑ ∑ δ a ∈A b ∈B( a )
a,b

8.3.4.2.5  Limit the Average Open Credits Associated with the Borrowers
Opening a new line of credit indicates that the borrower is unable to repay
his/her debts and too many open credits may be riskier for borrowers who
do not have an established credit history. Hence, financial institutions do
not prefer borrowers with many open credits (Dykstra and Wade, 1997).
Constraint (8.15) ensures that, on an average, the number of open credits
of the borrowers in the portfolio must be within the bank-specified open
­credits value (TO).

∑ ∑ (O × δ
a ∈A b ∈B( a )
a,b a,b )
≤ TO. (8.15)

∑ ∑ δ a ∈A b ∈B( a )
a,b

A Data-Driven Approach for Multiobjective Loan Portfolio Optimization 195

8.3.4.2.6  Achieve Age Diversification among the Borrowers

According to Chapman (1940), bad loan experience for a certain age category
is significantly different from the other age categories. Also, older borrowers
are more likely to make wise decisions and do not run into bankruptcy (Carr,
2013). Therefore, in the optimal loan portfolio, it is necessary to consider at
least a certain percentage of borrowers to fall within that bank-specified
age category to ensure risk reduction and loan diversification. Therefore,
­constraint (8.16) is introduced to have at least a certain percentage of the loan
( )
application TcE to be accepted with borrowers in that age category.

∑ ∑ (E
a ∈A b ∈B( a )
a,b ,c × δ a,b )
≥ TcE ∀ c ∈ C. (8.16)

∑ ∑ a ∈A
δ a,b
b ∈B( a )

8.3.4.2.7  Hard Constraints

Constraint (8.17) ensures that in a joint loan application, only one of the two
borrowers can be a primary borrower, and Constraint (8.18) forces Δa to be
1 if application a is approved. Constraint (8.19) forces binary restrictions on
the model.

∑δ
b ∈B( a )
a,b ≤ 1 ∀a ∈ A (8.17)

∑δ
b ∈B( a )
a,b = ∆a ∀a ∈ A (8.18)

δ a , b ∈ {0, 1} ∀a ∈ A, b ∈ B( a). (8.19)

8.3.4.3 Solution Approach: Multiobjective Optimization Using

ε-Constraint Method
In a multicriteria problem, all objectives cannot be optimized simultane-
ously. Therefore, the main focus is to find the set of efficient solutions. A solu-
tion is said to be efficient (Pareto-optimal or nondominated) if it c­ annot be
improved any further without sacrificing the performance of at least one of
the other objective function value (Masud and Ravindran, 2008). The MCMP
model  described in Sections 8.3.4.1 and 8.3.4.2 can be solved using the
ε-constraint method.
The multicriteria problem under consideration has two objectives:

• Objective 1: maximize return (Max f1(x))

• Objective 2: minimize risk (Min f2(x))
196 Big Data Analytics Using Multiple Criteria Decision-Making Models

In the ε-constraint method, objective 1 is maximized (Equation 8.20) and

objective 2 is given as a constraint (Equation 8.21) to the model along with the
other regular constraints discussed in Section 8.3.4.2 (Equation 8.22).

Max f1( x) (8.20)

subject to

f 2 ( x) ≤ ε (8.21)

x ∈ S, where S is the feasible region. (8.22)

The RHS of the constrained objective function (ε) in Equation 8.21 is varied
to obtain the efficient frontier.
First, the model is solved as a single objective problem considering only
maximizing return and ignoring the risk constraint. The resulting total risk
and return are the upper bounds on the value of risk and return, ­respectively.
Next, the model is again solved as a single objective problem considering only
minimizing risk and ignoring the return. The resulting total risk and return
are the lower bounds on the value of risk and return, respectively. Finally,
the return is maximized as the risk is varied (the value of ε is ­varied) from
its lower bound to its upper bound. The model using ε-constraint method is
given below:
Objective

Max Z1 = ∑ ∑ (I × δ
a ∈A b ∈B( a )
a a,b ).

Subject to:
Constraints (8.10) through (8.19), and

∑ ∑ (R
a ∈A b ∈B( a )
a,b × δ a , b ) ≤ ε.

8.3.4.4  Demand Uncertainty in Loan Application Requests

The total number of loan application requests is unknown for any given time
period. Therefore, to study the effect of the variability of the total number of
loan applications on the portfolio’s return and risk, n ­different scenarios are
considered. For each scenario, the total number of loan a­ pplications is sam-
pled from a known distribution and therefore known a  ­priori, and then the
mathematical formulation is solved by varying the ε values. Finally, the mean
and the standard deviation of the output parameters across all s­ cenarios for
each ε value are obtained and analyzed.
A Data-Driven Approach for Multiobjective Loan Portfolio Optimization 197

8.4  Experimental Results

In this section, the input parameters are defined and the best classifier for
the given input parameters is selected. Model accuracy, specificity, sensitiv-
ity, and training time defined in Section 8.3.3 are the performance measures
used to identify the best classifier. The output of the best trained classifier is
given as inputs to the MCMP model.

8.4.1  Input Parameters of Machine-Learning Algorithms

As discussed in Section 8.3.1.1, the data set used for classifying the risk
type of the potential borrowers is downloaded from Kaggle (Kaggle, 2011),
and it has historical data for 150,000 borrowers. The historical data have
both the feature values (inputs) and the corresponding risk type (outputs).
In addition to the historical data with both inputs and outputs, Kaggle
also provided data for 100,000 borrowers that have only the feature values
(inputs), and in this work, they are assumed to represent the population of
the potential ­borrowers. The best classifier is used to predict the risk type
of these ­potential borrowers. However, the loan portfolio cannot be gener-
ated only based on the risk type of potential borrowers. It is essential to
consider v ­ arious ­strategic plans of banks, such as the diversification of the
loan types, achieving higher returns, maintaining low risks, and clearly
these goals are conflicting in nature. Therefore, the predicted risk type
of the potential b ­ orrower is used in the MCMP model to develop a loan
portfolio that aims to achieve the bank-specified goals. The summary of
the data parameters for machine-learning algorithms is given in Table 8.2.

TABLE 8.2
Summary of Data Parameters for Machine-Learning Algorithms
Data Value
Training data 100,000
Testing data 50,000
Population of potential borrowers 100,000
Number of features 10
Number of outputs 1
Number of classifiers evaluated 5
Number of base-level classifiers used in stacking 2
Base-level classifiers in stacking RF and SGBDT
Meta-level classifier in stacking ANN
Types of loans considered 5
Percentage of individual loan applications 80%
Percentage of joint loan applications 20%
198 Big Data Analytics Using Multiple Criteria Decision-Making Models

8.4.2 Analysis
The five classifiers are trained and tested using the caret package in R and the
MCMP model was coded in Visual C++ and executed in IBM CPLEX®12.4.0.0
optimizer. The experiments were conducted on a computer with 8 GB RAM,
Intel i5 2.50 GHz processor running Windows 10.
Section 8.3.1.2 discusses five different alternatives for data cleansing:
substitute with −1 (A1), substitute with median (A2), discard variable and
substitute with unique value (A3), discard variable and substitute with
­
median value (A4), and discard incomplete rows (A5). Each alternative is
then used to train and test each of the five classifiers. Finally, the AUC ­values
for each alternative for the five classifiers are obtained and are shown in
Table 8.3. The best approach for data cleansing is the alternative with the
highest AUC value.
Based on the values in Table 8.3, A1 has the highest AUC value for all the
classifiers except neural network. Neural network has the highest AUC value
for the alternative A3. However, the difference between the AUC value for
A3 and A1 obtained using the neural network is very small. Therefore, A1
is chosen as the best alternative for data cleansing and error and missing
values are substituted with −1. Two-thirds of the cleansed data is used for
training and the remaining is used for testing. The training data was resam-
pled using fourfold cross validation technique in which the training data is
internally split to train/test runs to determine the optimal parameters of the
classifier. The AUC value is then obtained for each classifier for the testing
set. It is important to note that the entries in the training set and testing set
are randomly sampled and the AUC value changes depending on the entries
in the training and testing set. Therefore, to estimate the accuracy value of
each classifier, the procedure of estimating the AUC values is repeated 50
times. Figure 8.10 shows the average AUC value over 50 replications along
with its standard deviation for the five classifiers.
It is observed that the standard deviation of the AUC values of the classi-
fiers is very low and therefore, all the classifiers are robust in their classifica-
tion. The average AUC value for logistic regression is very low compared
with the other classifiers. Hence, logistic regression is no longer considered

TABLE 8.3
AUC Values for Different Alternatives
Logistic Neural Random
Regression Network Forest SGBDT Stacking
A1 0.69522 0.85519 0.84529 0.86210 0.86351
A2 0.69319 0.84344 0.84526 0.86148 0.86192
A3 0.68583 0.85856 0.84084 0.86082 0.86126
A4 0.69343 0.84344 0.84527 0.86148 0.86188
A5 0.68190 0.83376 0.83534 0.85182 0.85332
A Data-Driven Approach for Multiobjective Loan Portfolio Optimization 199

1.0

0.75
Average AUC values

0.50

0.25

0.00
Logistic Neural Random SGBDT Stacking
regression network forests
Learning method

FIGURE 8.10
Average AUC values for each learning method.

in choosing the best classifier. In order to choose the best classifier among
the other four, the class probabilities should be converted to class labels
by using a specified threshold value. The ROC curve for the classifiers is
shown in Figure 8.11 and can be used to select an appropriate threshold
value. The threshold value impacts the sensitivity, specificity, and accuracy,
and hence, different threshold values are evaluated to obtain a trade-off
between ­sensitivity, specificity, and accuracy.
Using the ROC curve and by experimenting different threshold values, a
threshold of 0.82 is chosen for RF and a threshold of 0.80 is chosen for SGBDT
and Neural Network to obtain the respective class labels. Since stacking uses
the neural network as its meta-level classifier, its threshold is also set to 0.80.
The class label is low risk if the class probability is less than the threshold
and is high risk if it is greater than the threshold.
The sensitivity, specificity, and accuracy values for each classifier are
obtained using the confusion matrix and the time required to train each
classifier is recorded as shown in Table 8.4. It is to be noted that the c­ lassifiers
must be trained periodically to learn any new patterns emerged during
­environmental changes. An ideal method will have a sensitivity, specificity,
and accuracy values close to 1 and will require low-training time.
All the methods perform well in classifying the risk type and have
­accuracy values close to each other. It can be observed that stacking yields
better results in terms of sensitivity, specificity, and accuracy when com-
pared with neural network, RFs, and SGBDT. However, the total training
200 Big Data Analytics Using Multiple Criteria Decision-Making Models

ROC curve
Random Gradient Neural Stacking
forest boosting network
1.0

0.8
True-positive rate

0.6

0.4

0.2

0.0

0.0 0.2 0.4 0.6 0.8 1.0

False-positive rate

FIGURE 8.11
ROC curve for comparing the learning algorithms.

TABLE 8.4
Performance of Classification Algorithms
Algorithm Sensitivity Specificity Accuracy Training Time (in sec)
Neural network 0.7154 0.8308 0.8200 172
Random Forests 0.7212 0.8313 0.8200 229
SGBDT 0.7431 0.8231 0.8278 130
Stacking 0.7433 0.8314 0.8296 648

time (training time of base-level classifier + training time of meta-level

­classifier) is very high, compared with the other three classifiers. SGBDT
has better sensitivity, accuracy, and training time values, and has slightly
worse ­specificity value when compared with neural network and RFs.
Therefore, for this p ­ articular data set, SGBDT is preferable over neural
network and RFs. Further, ­sensitivity, specificity, and accuracy values of
SGBDT are close to those of the ­stacking method, and its training time
is almost five times less than that of the stacking method. Hence, SGBDT
is chosen as the best ­classifier to ­classify the risk type of the potential
borrowers.

8.4.3  Input Parameters of MCMP Model

Table 8.5 gives the summary of input parameters used in MCMP model.
Five different types of loans are considered: mortgage loans, credit
A Data-Driven Approach for Multiobjective Loan Portfolio Optimization 201

TABLE 8.5
Summary of Data Parameters for Mathematical Programming Model
Data Value
Bound on percentage of time the borrowers must be above 50 years of age 25%
Bound on number of dependents 2
Bound on debt ratio 0.4
Bound on late payments 5
Bound on open credits 10
Percentage of loan issued to loan type 1 29%
Percentage of loan issued to loan type 2 16%
Percentage of loan issued to loan type 3 13%
Percentage of loan issued to loan type 4 12%
Percentage of loan issued to loan type 5 30%
Interest rate of loan issued to loan type 1 5%
Interest rate of loan issued to loan type 2 3%
Interest rate of loan issued to loan type 3 2.5%
Interest rate of loan issued to loan type 4 2.5%
Interest rate of loan issued to loan type 5 5%

card loans, ­consumer loans, commercial real estate loans, and commer-
cial loans. The  interest rate and percentage of these five different types
of  loans  ­considered are  obtained from the literature (Dilworth, 2015;
Issa, 2015).

8.4.4  MCMP Model Analysis

Figure 8.12 illustrates risk–return curve for different scenarios. The bounds
for the risk and return vary for each scenario depending on the total num-
ber of loan applications. The risk objective is treated as ε-constraint and
the value of ε is varied from the lower bound to the upper bound with
a step size of 100. Across all scenarios, the return increases steeply when
the underwriter is willing to accept up to 100 high-risk borrowers (i.e.,
ε = 100) and the slope of the line gradually decreases because the maxi-
mum revenue ­generating high-risk borrowers are already granted loan
when ε = 100. Using the ­efficient frontier, the underwriter can estimate
the maximum return ­generated for each scenario for a bank-specified risk
value.

8.4.4.1  Scenario Analysis

Figures 8.13a through e illustrate the impact of portfolio risk on the percent-
age of different types of loans. From the error bars, it can be observed that the
percentage of each type of loan issued varies as the portfolio risk increases;
202 Big Data Analytics Using Multiple Criteria Decision-Making Models

Efficient frontier for different scenarios

6,500,000

6,000,000

5,500,000
Total portfolio return

5,000,000
Scenario 1
4,500,000
Scenario 2
4,000,000 Scenario 3
Scenario 4
3,500,000 Scenario 5
Scenario 6
3,000,000 Scenario 7
Scenario 8
2,500,000 Scenario 9
Scenario 10
2,000,000
0 100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400
Total portfolio risk

FIGURE 8.12
Efficient frontiers for different scenarios.

however, the variation is not statistically significant. Hence, the portfolio

risk does not significantly impact the percentage of different types of loans.
Also, within scenarios, the variation in the percentage of a particular type of
loan issued is not significant. Therefore, uncertainties in the total number of
applications received in a planning horizon do not impact the loan diversifi-
cation with respect to the type of loans.
Figure 8.14a shows the impact of the portfolio risk on the loan approval
rate. It can be observed that the loan approval rate increases as the portfo-
lio risk increases. This is so because, if the financial institution is willing
to t­olerate  more risk, then more high-risk borrowers are accepted since
the objective of the model is to maximize the return. This results in an
increase in the loan approval rate. Due to the increase in the loan approval
rate, the portfolio’s return also increases as risk increases as shown in
Figure 8.14b.
Also, based on the various risk settings tested, the underwriter can decide
on the portfolio return and loan approval rate. For example, consider a
­setting in which the bank is willing to accept up to 500 high-risk ­borrowers.
From Figure 8.14a and b, it can be concluded that the return generated is
on average around 4.75 million and the average loan approval rate is about
70%. Hence, for this multiobjective problem, the model serves as a decision
­support tool for determining the portfolio return and loan approval rate for
a bank-specified risk value.
A Data-Driven Approach for Multiobjective Loan Portfolio Optimization 203

(a) 40

Percentage of mortgage loans

30

20

10

0
0 100 200 300 400 500 600 700 800 900 1000
Portfolio risk (no. of high-risk borrowers)
(b)
40
Percentage of credit card loans

30

20

10

0
0 100 200 300 400 500 600 700 800 900 1000
Portfolio risk (no. of high-risk borrowers)
(c)
40
Percentage of consumer loans

30

20

10

0
0 100 200 300 400 500 600 700 800 900 1000
Portfolio risk (no. of high-risk borrowers)

FIGURE 8.13
Impact of portfolio risk on percentage of different types of loans. (Continued )
204 Big Data Analytics Using Multiple Criteria Decision-Making Models

(d)
40
Percentage of commercial loans

30

20

10

0
0 100 200 300 400 500 600 700 800 900 1000
Portfolio risk (no. of high-risk borrowers)
(e)
40
Percentage of other loans

30

20

10

0
0 100 200 300 400 500 600 700 800 900 1000
Portfolio risk (no. of high-risk borrowers)

FIGURE 8.13 (Continued )

Impact of portfolio risk on percentage of different types of loans.

8.5 Conclusions
Financial institutions, such as banks, cater to the needs of various customers
by offering a variety of loans, and the revenue generated by loans is one of
the largest assets for any bank. However, there is also a large risk associ-
ated with the approval of bad loans and may sometimes even lead to bank-
ruptcy. Therefore, in this work, a decision support recommender system is
­developed for loan approvals taking into account the different types of loans.
The developed recommender system is a two-stage system. In the first
stage, the best machine-learning classifier is selected among the five c­ lassifiers
A Data-Driven Approach for Multiobjective Loan Portfolio Optimization 205

(a) 100
Loan approval rate (in %)

75

50

25

0
0 100 200 300 400 500 600 700 800 900 1000
Portfolio risk (no. of high-risk borrowers)

(b)

5
Portfolio return (in millions)

0
0 100 200 300 400 500 600 700 800 900 1000
Portfolio risk (no. of high-risk borrowers)

FIGURE 8.14
Impact of the portfolio risk on the loan approval rate and portfolio return.

c­ onsidered to classify the potential borrowers as high risk or low risk. Several
factors of the applicant, such as age, late payment history, and number of open
credits, are given as inputs to the classifiers. The five different classifiers used
are logistic regression, RFs, neural networks, gradient b ­ oosting, and stacking.
The sensitivity, specificity, and accuracy values for each classifier are obtained
using the confusion matrix and the time required to train each c­ lassifier is
recorded. The method which has a s­ ensitivity, ­specificity, and accuracy val-
ues close to 1 and least training time is selected as the best classifier. The
output of the best classifier in the first stage along with other ­attributes of the
applicants is given as input to the second stage of the model. In the second
stage, a diversified loan portfolio is developed c­ onsidering the types of loans,
206 Big Data Analytics Using Multiple Criteria Decision-Making Models

return, and risk associated with the borrower. The strategic goals and con-
straints of the bank are also given as input to the second stage.
The data of about 150,000 samples with ten predictors and one response
variable were downloaded from the Kaggle competition site (Kaggle, 2011).
SGBDT performs the best for the given data set. The findings indicate that
the portfolio risk does not significantly impact the percentage of different
types of loans as well as the uncertainties in the total number of applications
received in a planning horizon. The loan approval rate and return increase
as the portfolio risk increases since the financial institution is willing to
­tolerate more high-risk borrowers resulting in an increased loan approval
rate and portfolio return. Finally, the efficient frontier is developed with the
objective of maximizing the return and minimizing the risk, using which
the underwriter can estimate the maximum return possible generated for
each scenario for a bank-specified risk value.

References
Altman, E. I. 1980. Commercial bank lending: Process, credit scoring, and costs of errors
in lending. Journal of Financial and Quantitative Analysis, Vol. 15(4), pp. 813–832.
Apilado, V. P., Warner, D. C., and Dauten, J. J. 1974. Evaluative techniques in ­consumer
finance—Experimental results and policy implications for financial institu-
tions. Journal of Financial & Quantitative Analysis, Vol. 9(2), pp. 275–283.
Bell, T. B., Ribar, G. S., and Verchio, J. 1990. Neural nets versus logistic regression:
A comparison of each model’s ability to predict commercial bank f­ailures.
In Proceedings of the 1990 Deloitte and Touche/University of Kansas Symposium of
Auditing Problems, Lawrence, KS, pp. 29–58.
Bierman, Jr, H. and Hausman, W. H. 1970. The credit granting decision. Management
Science, Vol. 16(8), pp. B-519–B532.
Breiman, L. 2001. Random forests. Machine Learning, Vol. 45(1), pp. 5–32.
Brown, I. and Mues, C. 2012. An experimental comparison of classification ­algorithms
for imbalanced credit scoring data sets. Expert Systems with Applications,
Vol. 39(3), pp. 3446–3453.
Burke, A. E. and Hanley, A. 2003. How do banks pick safer ventures? A theory
­relating the importance of risk aversion and collateral to interest margins and
credit rationing. The Journal of Entrepreneurial Finance, Vol. 8(2), pp. 13–24.
Carr, D. 2013. Why older minds make better decisions. http://www.forbes.com/sites/
nextavenue/2013/04/29/why-older-minds-make-better-decisions/, accessed on
December 6, 2015.
Chapman, J. M. 1940. Factors affecting credit risk in personal lending. In Commercial
Banks and Consumer Installment Credit. New York: National Bureau of Economic
Research, pp. 109–139.
Choi, C. 2011. Mortgage approved: 5 factors that lenders consider on home loan
applications in tighter financial market. http://www.postandcourier.com/
article/20110703/PC05/307039941/, accessed on December 2, 2015.
A Data-Driven Approach for Multiobjective Loan Portfolio Optimization 207

Clemente, Jr, D. A. 1980. Prediction of agricultural loan repayment performance.

Philippine Review of Economics, Vol. 17(1 & 2), pp.31–60.
Curtis, G. 2007. Different needs, different loans | Investopedia. http://www.
investopedia.com/articles/pf/07/loan_types.asp, accessed on October 10,
2015.
Davis, R. H., Edelman, D. B., and Gammerman, A. J. 1992. Machine-learning
algorithms for credit-card applications. IMA Journal of Management Mathematics,
Vol. 4(1), pp. 43–51.
Diette, M. D. 2000. How do lenders set interest rates on loans? https://www.
minneapolisfed.org/publications/community-dividend/how-do-lenders-set-
interest-rates-on-loans, accessed on December 6, 2015.
Dilworth, K. 2015. Credit card interest rates slide to 14.89%. http://www.creditcards.
com/credit-card-news/interest-rate-report-10715-down-2121.php, accessed on
August 6, 2016.
Dirickx, Y. M. and Wakeman, L. 1976. An extension of the Bierman-Hausman model
for credit granting. Management Science, Vol. 22(11), pp. 1229–1237.
Durand, D. 1941. Risk elements in consumer installment financing. In: NBER Books.
National Bureau of Economic Research, Inc, Cambridge, MA, number dura41-2,
June.
Dykstra, D. R. and Wade, P. M. 1997. U.S. Patent No. 5,611,052. Washington,
DC: U.S. Patent and Trademark Office.
Fama, E. F. and Miller, M. H. 1972. The Theory of Finance. New York: Holt Rinehart &
Winston.
Fitzpatrick, D. 1976. An analysis of bank credit card profit. Journal of Bank Research,
Vol. 7, pp. 199–205.
Friedman, J. H. 2002. Stochastic gradient boosting. Computational Statistics and Data
Analysis, Vol. 38(4), pp. 367–378.
Gerlach, S. and Peng, W. 2005. Bank lending and property prices in Hong Kong.
Journal of Banking & Finance, Vol. 29(2), pp. 461–481.
Greenbaum, S. I. and Thakor, A. V. 2007. Contemporary Financial Intermediation. Boston:
Academic Press, an imprint of Elsevier.
Hand, D. J. 1981. Discrimination and Classification. Wiley Series in Probability and
Mathematical Statistics, Chichester: Wiley.
Hand, D. J. and Henley, W. E. 1997. Statistical classification methods in consumer
credit scoring: A review. Journal of the Royal Statistical Society. Series A (Statistics
in Society), Vol. 160(3), pp. 523–541.
Henley, W. E. 1994. Statistical aspects of credit scoring. Doctoral dissertation.
Open  University. http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.241084,
accessed on August 6, 2016.
Hill, T. 2014. How lenders decide your auto loan rate. http://www.dallasnews.
com/business/personal-finance/headlines/20140815-how-lenders-decide-
your-auto-loan-rate.ece, accessed on December 15, 2015.
Ince, H. and Aktan, B. 2009. A comparison of data mining techniques for credit
­scoring in banking: A managerial perspective. Journal of Business Economics and
Management, Vol. (3), pp. 233–240.
Issa, E. 2015. American household credit card debt statistics: 2015—NerdWallet.
https://www.nerdwallet.com/blog/credit-card-data/average-credit-card-
debt-household/, accessed on December 15, 2015.
208 Big Data Analytics Using Multiple Criteria Decision-Making Models

Kaggle. 2011. Data—Give me some credit. https://www.kaggle.com/c/GiveMe​

SomeCredit/data, accessed on September 15, 2015.
Khandani, A. E., Kim, A. J., and Lo, A. W. 2010. Consumer credit-risk models
via machine-learning algorithms. Journal of Banking & Finance, Vol. 34(11),
pp. 2767–2787.
Kolesar, P. and Showers, J. L. 1985. A robust credit screening model using categorical
data. Management Science, Vol. 31(2), pp. 123–133.
Koutanaei, F. N., Sajedi, H., and Khanbabaei, M. 2015. A hybrid data mining model
of feature selection algorithms and ensemble learning classifiers for credit
­scoring. Journal of Retailing and Consumer Services, Vol. 27, pp. 11–23.
Kruppa, J., Schwarz, A., Arminger, G., and Ziegler, A. 2013. Consumer credit risk:
Individual probability estimates using machine learning. Expert Systems with
Applications, Vol. 40(13), pp. 5125–5131.
Lane, S. 1972. Submarginal credit risk classification. Journal of Financial and Quantitative
Analysis, Vol. 7(1), pp. 1379–1385.
Lanza, A. 2014. Afford a mortgage with student loan debt. http://www.usnews.com/
education/blogs/student-loan-ranger/2014/10/29/afford-a-mortgage-with-
student-loan-debt, accessed on December 15, 2015.
Leonard, K. J. 1993. Empirical Bayes analysis of the commercial loan evaluation
process. Statistics & Probability Letters, Vol. 18(4), pp. 289–296.
Lovati, J. M. 1975. The changing competition between commercial banks and thrift
institutions for deposits. Review July, pp. 2–8.
Lucas, A. 2004. Updating scorecards: Removing the mystique. In: Readings in Credit
Scoring: Foundations, Developments, and Aims. Thomas, L. C., Edelman, D. B., and
Crook, J. N. (eds.). New York: Oxford University Press, pp. 93–109.
Masud, A. and Ravindran, A. 2008. Multiple criteria decision making, Chapter 5.
In Operations Research and Management Science Handbook, ed. A. Ravindran. Boca
Raton, FL: CRC Press Taylor and Francis Group, pp. 1–35.
Meyers, J. H. and Forgy, E. W. 1963. The development of a numerical credit eval-
uation  system. Journal of the American Statistical Association, Vol. 58(303),
pp. 799–806.
Neelankavil, J. P. 2015. International Business Research. New York: Routledge.
Orgler, Y. E. 1970. A credit scoring model for commercial loans. Journal of Money,
Credit and Banking, Vol. 2(4), pp. 435–445.
Perez, S. 2015. Welcome to market realist. http://marketrealist.com/2015/03/loan-
assets-gaining-importance-banking-sector/, accessed on December 15, 2015.
Power, D. J. 2002. Decision Support Systems: Concepts and Resources for Managers.
Westport, CT: Greenwood Publishing Group.
Ripley, B. D. 1994. Neural networks and related methods for classification. Journal of
the Royal Statistical Society. Series B (Methodological), Vol. 56(3), pp. 409–456.
Riungu, M. K. 2014. Factors influencing loan repayment in micro-finance
­institutions: A case of South Imenti District. Doctoral dissertation.
Rosenberg, E. and Gleit, A. 1994. Quantitative methods in credit management: A sur-
vey. Operations Research, Vol. 42(4), pp. 589–613.
Saunders, D., Xiouros, C., and Zenios, S. 2007. Credit risk optimization using factor
models. Annals of Operations Research, Vol. 152(1), pp. 49–77.
Sharpe, W. F. 1970. Portfolio Theory and Capital Markets. McGraw-Hill, New York.
Showers, J. L. and Chakrin, L. M. 1981. Reducing uncollectible revenue from
­residential telephone customers. Interfaces, Vol. 11(6), pp. 21–34.
A Data-Driven Approach for Multiobjective Loan Portfolio Optimization 209

Sirignano, J., Tsoukalas, G., and Giesecke, K. 2015. Large-scale loan portfolio ­selection.
Available at SSRN: http://dx.doi.org/10.2139/ssrn.2641301
Soergel, A. 2015. 1 in 3 Americans near financial disaster. http://www.usnews.com/
news/blogs/data-mine/2015/02/23/study-suggests-1-in-3-americans-flirting-
with-financial-disaster/, accessed on December 15, 2015.
Somers, P. and Hollis, J. M. 1996. Student loan discharge through bankruptcy.
American Bankruptcy Institute Law Review, Vol. 4, pp. 457.
Srinivasan, V. and Kim, Y. H. 1987. Credit granting: A comparative analysis of
­classification procedures. Journal of Finance, Vol. 42(3), pp. 665–681.
Sullivan, B. 2014. 4 student loan debt collection tricks. http://www.cbsnews.com/
media/4-student-loan-debt-collection-tricks/, accessed on December 15, 2015.
Taylor, W. F. 1980. Meeting the equal credit opportunity act’s specificity requirement:
Judgmental and statistical scoring systems. Buffalo Law Review, Vol. 29, pp. 73.
Tripathy, D. 2015. Bank of India sinks to Q2 loss as bad debts jump. http://in.reuters.
com/article/bank-of-india-q2-results-idINKCN0SY1A020151109, accessed on
December 15, 2015.
Van Leuvensteijn, M., Bikker, J. A., Van Rixtel, A. A., and Kok Sorensen, C. 2007.
A new approach to measuring competition in the loan markets of the euro area.
Available at SSRN: http://ssrn.com/abstract=991604
Wang, G. and Ma, J. 2012. A hybrid ensemble approach for enterprise credit risk
assessment based on Support Vector Machine. Expert Systems with Applications,
Vol. 39(5), pp. 5325–5331.
Wang, G., Ma, J., Huang, L., and Xu, K. 2012. Two credit scoring models based on dual
strategy ensemble trees. Knowledge-Based Systems, Vol. 26, pp. 61–68.
Wiginton, J. C. 1980. A note on the comparison of logit and discriminant models
of consumer credit behavior. Journal of Financial and Quantitative Analysis,
Vol. 15(3), pp. 757–770.
Zurada, J. and Barker, R. M. 2011. Using memory-based reasoning for predicting
default rates on consumer loans. Review of Business Information Systems (RBIS),
Vol. 11(1), pp. 1–16.
9
Multiobjective Routing in a Metropolitan
City with Deterministic and Dynamic
Travel and Waiting Times, and
One-Way Traffic Regulation

Swaminathan Vignesh Raja, Chandrasekharan Rajendran,

Ramaswamy Sivanandan, and Rainer Leisten

CONTENTS
9.1 Introduction................................................................................................. 212
9.2 Literature Review....................................................................................... 213
9.3 Problem Definition: Multiobjective Shortest Path Problem
with Time Dependency.............................................................................. 215
9.4 Mathematical Model for the Time-Dependent Shortest Path
Problem When the Travel Times and Waiting Times, and One-
Way Traffic Regulation Are Dynamic and Time-Dependent in a
City Network............................................................................................... 217
9.4.1 Terminology.................................................................................... 217
9.4.2 Mathematical Model....................................................................... 219
9.4.3 Multiobjective Optimization Algorithm to Obtain
the Strictly Nondominated Solutions.......................................... 223
9.4.3.1 Terminology......................................................................223
9.4.3.2 Step-by-Step Procedure to Generate a Set of
Nondominated (Pareto-Optimal) Solutions,
Given εt............................................................................. 223
9.5 Experimental Settings, Results, and Discussion.................................... 226
9.5.1 Data with Respect to Distance and Travel Times along the
Roads and Waiting Times at Junctions........................................ 226
9.5.2 Sample Network Topology............................................................ 232
9.6 Implementation of the Proposed Multiobjective Model
to the Complete Chennai Network..........................................................234
9.7 Summary...................................................................................................... 236
Acknowledgment................................................................................................. 238
References.............................................................................................................. 241

211
212 Big Data Analytics Using Multiple Criteria Decision-Making Models

9.1 Introduction
A transportation network is a network of roads and streets, which permits
vehicles to move from one place to another. A directed or undirected graph
can be used to represent a transportation network, where edges of the graph
have a capacity and receive a flow. The vertices of the graph are called nodes
and the edges of the graph are called arcs. In any transportation network,
the vehicle starts from the source node and moves toward the sink node.
There are several types of network problems/models available in litera-
ture, for example, transportation problem, shortest path problem, minimum
spanning tree problem, maximum flow problem, and minimum flow prob-
lem. The shortest path problem has tremendous importance in network flow
models because it is applicable to any type of transportation network seen
in real life. One of the most important issues affecting the performance of a
transportation network is routing. The goal of routing between two points
in a network is to reach the destination as quickly as possible (shortest time
path problem) or as cheaply as possible (minimum cost or distance path
problem).
The shortest path problem is one of the most studied problems among net-
work optimization problems (Bertesekas, 1991; Ahuja et al., 1993; Schrijver,
2003). Two kinds of labelling approaches, namely, label-setting algorithm
and label-correcting algorithm, are in use to solve the shortest path problem.
Label-setting algorithm is in use only for acyclic network with nonnegative
costs, whereas label-correcting algorithm is applicable for acyclic network
with negative and nonnegative costs.
The shortest path problem with time dependency is dealt as a single-objec-
tive problem by several researchers, and one can classify the single-objective
problem into minimum cost path and fastest path problem. The minimum
cost path problem is solved to find the path having the minimum length with
respect to the cost by considering the travel time, while in the fastest path prob-
lem, the objective is to find the paths having the minimum length with respect
to time-dependent travel time. Bellman’s optimality principle (Bellman, 1958)
is a modification of the single-objective shortest path problem where arc travel
times are nonnegative integers for every time period, and achieve all-to-one
fastest paths to the single destination node from all other nodes. Dreyfus (1969)
proposed the modification of Dijkstra’s static shortest path algorithm to obtain
the fastest path between two vertices for a given departure time. The algo-
rithm of Dreyfus (1969) is suitable to solve first-in-first-out network problems.
Ziliaskopoulos and Mahmassani (1993) and Wardell and Ziliaskopoulos (2000)
proposed a label-correcting algorithm to obtain all-to-one minimum cost paths
and all-to-one fastest paths for all departure time. Chabini (1998) proposed
a label-setting algorithm running backward in the set of time parameters to
obtain all-to-one minimum cost and all-to-one fastest paths for all departure
time without the first-in-first-out assumption.
Multiobjective Routing in a Metropolitan City with Deterministic 213

The big data is characterized by volume, velocity, variety, and value of

data. With the help of technological advancements, we are now able to col-
lect the data about the transportation network of the entire city. The network
representation of roads and streets allows us to explore the topological and
geographical properties of the transportation network.
The multiple criteria decision making (MCDM) is an important area of
research in Operations Research which deals with decision problems that
involve multiple objectives. Real-life shortest path problem involves multi-
objectives; for example, minimizing cost, time, distance, maximizing reli-
ability, etc.
Typically, there does not exist a unique optimal solution for MCDM prob-
lems, and based on the decision maker’s preference, a solution is chosen
from the set of alternative solutions. The shortest path problem with mul-
tiobjectives is an nondeterministic polynomial time (NP)-hard problem
(Hamacher et al., 2006; Mohamed et al., 2010). In multiobjective optimization
problems, there may not exist a single solution that satisfies all the objec-
tives simultaneously, and the objective functions in most of the cases are
conflicting. A solution set is called nondominated set or Pareto set if none
of the objective function values can improve without degrading any of the
other objective function values. Ruzika and Wiecek (2005) presented a com-
prehensive survey of the literature (from 1975 to 2005) for the multiobjective
shortest path problem. Ehrgott and Wiecek (2005) presented a survey on
multiobjective integer programming, and discussed the scalarization tech-
niques for general continuous multiobjective optimization.
The organization of this chapter is as follows: In Section 9.2, we review the
literature; in Section 9.3, we define the multiobjective shortest path problem
with time dependency of travel times along roads and of waiting times at
junctions; we present our mathematical model and the multiobjective opti-
mization approach in Section 9.4; we explain the model with a numerical
illustration in Section 9.5; and in Section 9.6, we present the results with
respect to the complete travel in the city of Chennai from a given origin to a
given destination, followed by summary in Section 9.7.

9.2  Literature Review

In this section, we discuss the literature on multiobjective shortest path
­problems. Real-world shortest path problems are often time-dependent,
with more than one set of time-dependent parameters (Müller-Hannemann
and Schnee, 2007). Hamacher et  al. (2006) studied the application of the
shortest path problem for evacuation modeling where shortest paths rep-
resent evacuation routes. Several attributes associated with a route, such
as its length or reliability, are of particular interest for an evacuee in case
214 Big Data Analytics Using Multiple Criteria Decision-Making Models

of emergency. A ­collection of possible evacuation routes are used to model

a complete evacuation plan. They presented the interrelation of the time-
dependent network optimization problem and evacuation modeling con-
sidering two objectives. They proposed a heuristic algorithm to solve the
problem and compared its performance with respect to the existing algo-
rithms. Sung et al. (2000) considered a flow speed model of shortest path
problem with time-dependent networks. In their problem, as the interval
changes, the flow speed on the link changes, but not the travel time. They
modified the Dijkstra’s label-setting algorithm and proved that the flow
speed model is better than the link travel-time model. Dell’Amico et  al.
(2008) considered a shortest path problem, where the travel time on the
arcs may vary with time, and allow for the waiting time at the nodes.
Since the simple Dijkstra’s algorithm adaptation may fail to solve the dis-
continuities that exist on the routes, they proposed a new Dijkstra-like
algorithm to solve the problem. Mohamed et al. (2010) proposed a genetic
algorithm to solve the shortest path problem with the two objectives,
namely, minimizing cost and minimizing travel time. They compared the
performance of their genetic algorithm with the algorithm of Brumbaugh-
Smith and Shier (1989). Chitra and Subbaraj (2010) proposed an elitist
multiobjective ­evolutionary algorithm on the basis of the nondominated
sorting genetic algorithm, for solving the dynamic shortest path routing
problem in ­computer networks. They addressed the problem by consid-
ering delay and cost as a weighted sum of objectives, and generated the
Pareto-optimal solutions for the dynamic routing problem in a single run,
and compared the result with the result of a single-objective weighting
factor method.
Seshadri and Srinivasan (2010) proposed a new bound-based optimality
criterion for the optimal reliability path problem. On the basis of the bounds,
the authors proposed an algorithm to evaluate the path having the maxi-
mum travel-time reliability between given source and destination on the
network with stochastic, normal, and correlated link travel times specified
by the multivariate normal distribution. Seshadri and Srinivasan (2012) pro-
posed an algorithm to compute the minimum robust cost path on the cor-
related and stochastic link travel-time network.
They transformed the robust cost objective to a link separation or sum
of squares form. The level robust cost measure of the path is quantified
by using a weighted combination of squared mean and variance, and the
weights represent the importance of the travel-time variability of the user.
Prakash and Srinivasan (2014) presented a sample-based algorithm for
the minimum robust cost path on a network with link travel-time correla-
tions, and formulated it as a separable multiobjective problem. The authors
adopted a ­sample-based approach to represent the link travel-time distribu-
tions, implicitly capture the path correlations, and thus obviate the explicit
estimation of the link travel-time correlation matrix.
Multiobjective Routing in a Metropolitan City with Deterministic 215

The following gaps are identified from the review of literature.

• Several criteria exist in real-life situations. Hence, solving the prob-

lem as multiobjective considering two or more conflicting objectives
takes the problem closer to reality. Our model includes two conflict-
ing objectives, and provides the decision maker with a possible set of
nondominated solutions allowing to choose her/his preferred solu-
tion, given εt.
• Many authors do not consider real-life situations such as one-way
traffic in the rush period, time-dependent dynamic and determin-
istic travel times along the roads, and time-dependent dynamic
and deterministic waiting times at junctions; we consider these
real-life aspects in our mathematical model. It is possibly for the
first time such a multiobjective mixed-integer linear program-
ming (MILP) model is developed by the explicit consideration of
time-dependent travel times and time-dependent waiting times,
and time-dependent one-way traffic regulation in a major city
road network.

9.3 Problem Definition: Multiobjective Shortest

Path Problem with Time Dependency
In this study, we mainly focus on minimizing the distance, time, or cost to
traverse between two nodes. The objective of this work is to find a short-
est path with the minimum total travel time and distance between a given
pair of origin and destination (O–D). The time taken to travel a particular
link is called traverse time. Most of the authors consider the traverse time as
deterministic and static; however, in reality, it is not static because it depends
on various factors such as one-way traffic in rush periods, dynamic wait-
ing time at signals, varying traffic conditions, etc. In this work, we consider
multiobjective shortest path problem with time-dependent dynamic and
deterministic traverse times and waiting times, and time-dependent one-
way traffic regulation.
Let G = (V,E) be a directed graph, where V is a set of nodes, and E is a
set of edges. The departure time at the source node, source node s and des­
tination node d are given as inputs. The solution to the time-dependent
shortest path problem is to find an (s,d)-path that leaves a source node at a
given time and minimizes the total travel distance as well as total travel time
to reach the destination node which satisfies all the constraints. The objective
of the model is to arrive at an optimal route to traverse over a given network
216 Big Data Analytics Using Multiple Criteria Decision-Making Models

with consideration of minimizing the total travel time and the total distance
simultaneously.
In this work, we consider multiobjective shortest path problem with real-
life constraints such as time-dependent dynamic and deterministic travel
times, time-dependent dynamic and deterministic waiting times, and time-
dependent one-way traffic, and we propose a mathematical model to solve
the same. The biggest advantage of developing a mathematical model is the
flexibility of the resulting model; many cost functions can be chosen, and
many constraints can be added that otherwise would be difficult to satisfy
with a Dijkstra-like approach. For example, we address the one-way traffic
regulation in the proposed mathematical model. Dynamic programming
becomes time consuming and is not very computationally efficient due to
curse of dimensionality. Another motivation to go for MILP model is that the
same model can be extended for developing a multiobjective optimization
algorithm.
The technological advancements (Internet of Things) in big data enable
us to collect the data about the entire transportation network of a city. We
evaluate the proposed model with multiple objectives using the real-world
(in Chennai city) network of major roads consisting of 1658 nodes and 4224
links, mostly undirected graph, except the roads involving one-way traffic.
We consider the following aspects in our work that are taken into consider-
ation while collecting data from the travel and incorporated into the MILP
model with multiple objectives:

• Consider the source node as n′ and the destination node as n″, and
the arrival time at node n′ be denoted by An′.
• As travelers do not wait at the origin, we assume the waiting time at
the source node as 0.
• A day consists of a given number of travel-time intervals.
• Unit of the distance is 1 km and unit of the time is 1 min.
• Traffic corresponds to a given number of congestion levels, thereby
influencing the travel time along a road and waiting time at a
junction.
• Waiting time at a junction or node i depends on the time interval
during which the actual arrival at node i takes place; waiting time is
dynamic and deterministic.
• Travel time along the road (i,j) depends on the time interval in the
day during which the actual travel takes place; travel time is dynamic
and deterministic.
• Entry into the road (i,j) should be avoided inherently in the MILP
model during the one-way traffic intervals. Most roads allow two-
way traffic; however, there can be some roads that allow two-way
traffic for most periods in a day, except for some periods when
Multiobjective Routing in a Metropolitan City with Deterministic 217

the roads will allow traffic in one-way; for example, traffic flow is
allowed along a given road (i,j) during 8 a.m. to 10 a.m., but no traf-
fic along arc (j,i) during the same period, whereas traffic is allowed
along the road (j,i) during 6 p.m. to 8 p.m., but no traffic is allowed
along arc (i,j) during the same period.
• Two objectives, namely, the minimization of total travel time
(including waiting times at junctions) and the minimization of total
­distance traveled are considered.

9.4 Mathematical Model for the Time-Dependent

Shortest Path Problem When the Travel Times and
Waiting Times, and One-Way Traffic Regulation Are
Dynamic and Time-Dependent in a City Network
The mathematical model for the time-dependent shortest path problem is
presented in this section. We explain the terminologies associated with the
mathematical model in Section 9.4.1.

9.4.1 Terminology

Parameters Description
n Number of nodes in the network
n′ Origin node of travel
n″ Destination node of travel
i,j A pair of nodes
dij Distance from node i to node j /*note: It is not necessarily symmetric*/
k Index for time interval
∇ iw Number of waiting-time intervals with respect to node i
ϕ tij Number of travel-time intervals with respect to road (i,j)
Ωijo Number of one-way and two-way traffic intervals with respect to road (i,j)
tijk Travel time from node i to node j in the time interval k
Wik Waiting time at node i during the time interval k
LLwik Lower limit for the time interval k with respect to node i, to define the
time-dependent waiting time
ULwik Upper limit for the time interval k with respect to node i, to define the
time-dependent waiting time
LLtijk Lower limit for the time interval k with respect to road (i,j), to define the
time-dependent travel time
(Continued)
218 Big Data Analytics Using Multiple Criteria Decision-Making Models

Parameters Description
ULtijk Upper limit for the time interval k with respect to road (i,j), to define the
time-dependent travel time
/*note: For a given road (i,j), the actual travel time can vary depending upon
the actual time of arrival at the head of the road (i,j); for example, in a day of
1440 min and ϕ tij = 4, if interval 1 corresponds to 7.30 a.m., 11 a.m., we set
LLtij1 = 0, ULtij1 = 210, and tij1 = 10; interval 2 corresponds to 11 a.m., 5 p.m., we
have LLtij 2 = 211, ULtij 2 = 570, and tij2 = 8; interval 3 corresponds to 5 p.m., 9
p.m., we have LLtij 3 = 571, ULtij 3 = 810, and tij3 = 9; and interval 4 corresponds
to 9 p.m., 7.30 a.m., we have LLtij 4 = 811, ULtij 4 = 1440, and tij4 = 7 */
LLoijk Lower limit for the allowed traffic regulation during the time interval k with
respect to road (i,j), to define the time-dependent one-way traffic time/
regulation
ULoijk Upper limit for the allowed traffic regulation during the time interval k with
respect to road (i,j), to define the time-dependent one-way traffic time/
regulation
/*note: For a given road (i,j) that has no traffic for interval between 7.30 a.m.
and 11 a.m., two-way traffic between 11 a.m. and 5 p.m., one-way traffic
between 5 p.m. and 9 p.m., and two-way traffic between 9 p.m. and 7.30
a.m., we have Ωijo = 4; LLoij1 = 0 (i.e., 7.30 a.m.), ULoij1 = 210 (i.e., 11 a.m.), and
set ∆ ijo1 = 0; we have LLoij 2 = 211, ULoij 2 = 570, and ∆ ijo 2 ∈{0, 1}; we have
LLoij 3 = 571, ULoij 3 = 810, and set ∆ ijo 3 = 0; and we have LLoij 4 = 811, ULoij 4 = 1440
, and ∆ ijo 4 ∈{0, 1} */
J(i) Set of nodes or junctions to which there exists a direct connectivity from/to
node i
M A large value; set to 10,000 in this study

Decision
Variables Description
Yij A binary variable that takes a value 1 if the road (i,j) is chosen in the travel route;
0 otherwise
/* note: If there exists no direct connectivity between nodes/junctions i and j,
then we set dij = ∞ and/or set Yij = 0 */
δi An indicator (binary variable) that takes value 1 if node i is visited in the
travel route;
0 otherwise
∆ ikw An indicator (binary variable) that takes value 1 if node i is reached during
the interval k in the travel route;
0 otherwise
∆ tijk An indicator (binary variable) that takes value 1 if road (i,j) is traversed
during the interval k in the travel route;
0 otherwise
∆ ijk
o
An indicator (binary) that takes value 1 if the one-way is allowed during the
interval k in the travel route;
0 otherwise
ωi Waiting time at node i that takes a value if δi = 1. /*note: ωn′ = ωn″ = 0 */
Ai Arrival time at node i that takes a value if δi = 1./*note: An′ is given as input*/
(Continued )
Multiobjective Routing in a Metropolitan City with Deterministic 219

Decision
Variables Description
τij Travel time along the road (i,j) that takes a value if Yij = 1.
Aikw Arrival time at node i in the interval k, and it takes a value if δi = 1 and other
Aikw ’s are equal to 0
t
Aijk Arrival time on the road (i,j) in the interval k, and it takes a value if Yij = 1, and
t
other Aijk ’s are equal to 0
o
Aijk Arrival time on the road (i,j) (with the possible one-way traffic regulation) in
o
the interval k, and it takes a value if Yij = 1, and other Aijk ’s are equal to 0

9.4.2  Mathematical Model

Objective function:

Minimize Z1 = dist = ∑∑d Y .

i = 1 j ∈J ( i )
ij ij (9.1)

n n

Minimize Z2 = time = ∑∑
i = 1 j ∈J ( i )
τ ij + ∑ω .
i =1
i (9.2)

subject to the following:

/* Constraints (9.3) through (9.7) represent the conditions for at most one
inflow and one outflow */

∑Y
j∈J ( i )
ij ≤1 ∀ i = 1,…, n, i ≠ n′ and i ≠ n′′. (9.3)

∑Y
j∈J ( i )
ji ≤1 ∀ i = 1,…, n, i ≠ n′ and i ≠ n′′. (9.4)

∑Y =∑Y
j∈J ( i )
ij
j∈J ( i )
ji ∀ i = 1,…, n, i ≠ n′ and i ≠ n′′. (9.5)

∑Y
j∈J ( i )
ji = δi ∀ i = 1,…, n, i ≠ n′ and i ≠ n′′. (9.6)

Yij + Yji ≤ 1 ∀ i = 1, … , n, i ≠ n′ , i ≠ n′′ , and ∀j ∈ J (i). (9.7)

220 Big Data Analytics Using Multiple Criteria Decision-Making Models

/* Constraints (9.8) through (9.12) ensure that there is only one outflow and
no inflow for the source node n′, and only one inflow and no outflow for the
destination node n″ */

∑Y
j ∈J ( n ′ )
n ′j = 1. (9.8)

∑Y
j ∈J ( n ′′ )
jn ′′ = 1. (9.9)

∑Y
j ∈J ( n ′ )
jn ′ = 0. (9.10)

∑Y
j ∈J ( n ′′ )
n ′′ j = 0. (9.11)

δ n ′ = δ n ′′ = 1. (9.12)

/* Constraints (9.13) through (9.21) capture the waiting time that is depen-
dent on the arrival time at node i with respect to corresponding interval ∆ ikw
*/

ω n ′ = ω n ′′ = 0. (9.13)

An′ = arrival _ time / *given as an input to the model*/. (9.14)

Ai ≤ Mδ i ∀i = 1, … , n. (9.15)

A j ≤ Ai + ω i + τ ij + M(1 − Yij ) ∀i = 1, … , n, i ≠ n′′ , and ∀j ∈ J (i). (9.16)

A j ≥ Ai + ω i + τ ij − M(1 − Yij ) ∀i = 1, … , n, i ≠ n′′ , and ∀j ∈ J (i). (9.17)

LLwik ∆ ikw ≤ Aikw ≤ ULwik ∆ ikw ∀i = 1,…, n, i ≠ n′ and i ≠ n′′ , and k = 1, 2,…, ∇ iw .

(9.18)

∇iw

∑∆ w
ik = δi ∀ i = 1, … , n, i ≠ n′ , and i ≠ n′′. (9.19)
k =1
Multiobjective Routing in a Metropolitan City with Deterministic 221

∇iw

∑A w
ik = Ai ∀ i = 1, … , n, i ≠ n′ , and i ≠ n′′. (9.20)
k =1

∇iw

ωi = ∑(∆ W ) w
ik ik ∀ i = 1, … , n, i ≠ n′ , and i ≠ n′′. (9.21)
k =1

/* Constraints (9.22) through (9.29) capture the arrival time and travel time
along the road (i,j) based on the departure time at node i, that is, the actual
arrival time at the road (i,j), defined with respect to ϕ tij time intervals*/

LLtijk ∆ tijk ≤ Aijk

t
≤ ULtijk ∆ tijk ∀i = 1, … , n, i ≠ n′′ , ∀j ∈ J (i), and k = 1, … , ϕ tij .

(9.22)

ϕtij

∑∆ t
ijk = Yij ∀i = 1, … , n, i ≠ n′′ , and ∀j ∈ J (i). (9.23)
k =1

ϕtij

∑A t
ijk ≤ Ai + ω i + M(1 − Yij ) ∀i = 1, … , n, i ≠ n′′ , and ∀j ∈ J (i). (9.24)
k =1

ϕtij

∑A t
ijk ≥ Ai + ω i − M(1 − Yij ) ∀i = 1, … , n, i ≠ n′′ , and ∀j ∈ J (i). (9.25)
k =1

ϕtij

∑A t
ijk ≤ MYij ∀i = 1, … , n, i ≠ n′′ , and ∀j ∈ J (i). (9.26)
k =1

τ ij ≤ MYij ∀i = 1,…, n, and ∀j ∈ J (i). (9.27)

ϕtij

τ ij ≤ ∑ (∆ t ) + M(1 − Y )
t
ijk ijk ij ∀i = 1, … , n, i ≠ n′′ , and ∀j ∈ J (i). (9.28)
k =1

ϕtij

τ ij ≥ ∑ (∆ t ) − M(1 − Y )
t
ijk ijk ij ∀i = 1, … , n, i ≠ n′′ , and ∀j ∈ J (i). (9.29)
k =1
222 Big Data Analytics Using Multiple Criteria Decision-Making Models

/* Constraints (9.30) through (9.34) ensure the travel along the road (i,j)
only during the allowed traffic-time intervals when we arrive at node i. The
mathematical model should inherently avoid the entry into the roads having
one-way traffic regulation in a particular interval (Section 9.4.1 for the set-
tings of ∆ ijk
o
).*/

LLoijk ∆ ijk
o
≤ Aijk
o
≤ ULoijk ∆ ijk
o
∀i = 1, … , n, i ≠ n′′ , ∀j ∈ J (i), and k = 1, … , Ωijo .

(9.30)

Ωijo

∑∆ o
ijk = Yij ∀i = 1, … , n, i ≠ n′′ , and ∀j ∈ J (i). (9.31)
k =1

Ωijo

∑A o
ijk ≤ MYij ∀i = 1, … , n, i ≠ n′′ , and ∀j ∈ J (i). (9.32)
k =1

Ωijo

∑A o
ijk ≤ Ai + ω i + M(1 − Yij ) ∀i = 1, … , n, i ≠ n′′ , and ∀j ∈ J (i). (9.33)
k =1

Ωijo

∑A o
ijk ≥ Ai + ω i − M(1 − Yij ) ∀i = 1, … , n, i ≠ n′′ , and ∀j ∈ J (i). (9.34)
k =1

Yij , δ i , ∆ ikw , ∆ tijk , and ∆ ijk

o
are binary variables, and all other variables are
nonnegative.
/* note: Ωijo is the number of one-way and two-way traffic intervals opera-
tional along the road (i,j). If traffic is not allowed along the road (i,j) during
interval k, then set ∆ ijk o
= 0. For example, one-way traffic along the road (13,11)
in Figure 9.4 is being operational during interval 3, whereas (11,13) traffic is
not allowed. Then we set ∆ 11 o
,13 , 3 = 0 for the road (11,13) initially itself in the
model*/
Equation 9.1 is the objective function to minimize the total distance, and
Equation 9.2 is the objective function to minimize the total travel time.
Constraint (9.3) ensures a maximum of one outflow from the node i, except
the source node and the destination node. Constraint (9.4) ensures a maxi-
mum of one inflow to node i, except the source node and the destination
node. Constraint (9.5) assures that the size of inflow equals the size of out-
flow at all nodes, except source node and destination node. Constraint (9.6)
ensures every node in the path has an inflow. On the selection of an arc,
Constraint (9.7) ensures unidirectional flow in the arc. Constraints (9.8)
and (9.10) make sure that there is only outflow and no inflow at the source
Multiobjective Routing in a Metropolitan City with Deterministic 223

node. Constraints (9.9) and (9.11) ensure that there is only inflow and no
outflow at the destination node. Constraint (9.12) ensures the selection of
the source and destination nodes as part of the path. Constraints (9.15)
through (9.17) determine the arrival time at node j from node i. Constraints
(9.18) through (9.21) capture the time-dependent dynamic waiting time on
the basis of the arrival time interval k at node i. Constraints (9.22) through
(9.29) capture the time-dependent dynamic travel time along the arc (i,j)
on the basis of the departure time from node i. Constraints (9.30) through
(9.34) ensure that the travel takes place during the allowed traffic periods
along road (i,j) during the day.

9.4.3 Multiobjective Optimization Algorithm to Obtain

the Strictly Nondominated Solutions
Scalarization method or weighted sum method, ε-constraint method, goal
programming, and multilevel programming are in use to solve multiobjec-
tive optimization problems. In this study, we use ε-constraint method or
ε-approach to solve the bicriteria MILP model. In this method, the decision
maker has to choose an objective out of two objectives, while the remainder
of the objectives is constrained to satisfy the given target value (defined by
εt). Below is the step-by-step procedure of the ε-constraint method applied
to multiobjective routing in the metropolitan city (also see Tiwari, 2016 for a
related work).

9.4.3.1 Terminology

Variables Description
iter Iteration number
1
distiter Distance obtained in the iteration iter
1
timeiter 1
Minimum travel time obtained in the iteration iter, corresponding to distiter .
time* Optimal travel time
{δel } iter
ij Solution from the MILP, associated with distiter 1 1
and timeiter
γ 1
iter { }
Solution set with δelij , distiter , and timeiter
iter 1 1

/* note: Rest of the terminologies are introduced in Section 9.4.1*/

9.4.3.2 Step-by-Step Procedure to Generate a Set of

Nondominated (Pareto-Optimal) Solutions, Given εt
/* Do Step 1 to get a solution with the minimum total distance and the
­corresponding minimum total travel time; also this step gives us in Step 1.4
the optimal total travel time */
224 Big Data Analytics Using Multiple Criteria Decision-Making Models

Step 1: Set iter = 1.

Step 1.1: Execute the following MILP, called MILP-1:

Minimize
n
1
Z1 = distiter = ∑∑d Y
i = 1 j ∈J ( i )
ij ij

with the MILP given in Section 9.4.2.

Step 1.2: Execute the following MILP, called MILP-2:
Minimize

n n

Z2 = time 1
iter = ∑ ∑ τ + ∑ω
i = 1 j ∈J ( i )
ij
i =1
i

and subject to all constraints in the MILP given in Section 9.4.2, and
with the following add-on constraint:

∑ ∑ d Y = dist
i = 1 j ∈J ( i )
ij ij
1
iter . (9.35)

Denote the resultant solution for this MILP as δelijiter . { }

{ }
Step 1.3: Set γ 1iter = δelijiter , associated with distiter
1 1
, timeiter , and their
respective Yij’s.

/* note: We have the Pareto-optimal solution with minimum total distance

and corresponding minimum total travel time */
Step 1.4: Execute the MILP given in Section 9.4.2 with the following
objective function:
Minimize

n n

Z2* = time * = ∑∑
i = 1 j ∈J ( i )
τ ij + ∑ω i =1
i

subject to all constraints given in the MILP presented in Section 9.4.2.

Step 2:
/* Do this step to get further Pareto-optimal solutions, given εt (with
respect to time decrement)*/
Multiobjective Routing in a Metropolitan City with Deterministic 225

/* note: Skip Step 2 if time11 = time * */

Step 2.1: Set iter = iter + 1;

With respect to the original MILP given in Section 9.4.2, do the following:

Set:
Minimize

Z1 = ∑∑d Y = dist
i = 1 j ∈J ( i )
ij ij
1
iter

subject to all constraints given in the MILP presented in Section 9.4.2,

and with the following additional constraint:

( 1
if timeiter − 1 − ε t > time
*
)
then

n n

add : ∑∑
i = 1 j ∈J ( i )
τ ij + ∑ω ≤ time
i =1
i
1
iter − 1 − εt (9.36)

else

n n

add : ∑∑
i = 1 j ∈J ( i )
τ ij + ∑ω = time
i =1
i
*
(9.37)

and execute the resultant MILP.

Step 2.2: Execute the following MILP:
Minimize

n n
1
Z2 = timeiter = ∑∑
i = 1 j ∈J ( i )
τ ij + ∑ω i =1
i

and subject to all constraints in the MILP given in Section 9.4.2, and
with the following additional constraint

∑ ∑ d Y = dist
i = 1 j ∈J ( i )
ij ij
1
iter . (9.38)

226 Big Data Analytics Using Multiple Criteria Decision-Making Models

{
Denote the solution from the above MILP as δelijiter . }
{ } 1
Step 2.3: Set γ 1iter = δelijiter , associated with distiter 1
, timeiter , and their
respective Yij’s.

Step 3: If
1
timeiter = time *
then proceed to Step 4,
else return to Step 2.
Step 4: STOP: the set of strictly Pareto-optimal solutions is obtained,
{ } 1
denoted by γ 1iter , ∀iter with the corresponding distiter 1
, timeiter , and
their respective Yij’s, for the given εt.

9.5  Experimental Settings, Results, and Discussion

We now present the application of the proposed mathematical model to find
the shortest path for the Chennai city network with the objectives of mini-
mizing the total travel time and total distance travelled.

9.5.1 Data with Respect to Distance and Travel Times

along the Roads and Waiting Times at Junctions
The data used in this study consists of two sets: Chennai city network data
and a sample network topology of Chennai city network. City network
data is obtained from the Centre of Excellence in Urban Transport at the
Indian Institute of Technology (IIT) Madras. Data set includes the map of
Chennai city with parameters—from node ID., to node ID., the correspond-
ing distance and free-flow travel time (FFT) along the road. To illustrate
the working of the mathematical model, we consider a sample network
topology of 33 nodes as shown in Figure 9.1. We use this illustrative exam-
ple to show the solutions with respect to different traffic conditions. In
this example, node 1 is the source node and node 33 is the destination
node. The distance and the FFT are given in Tables 9.1 and 9.2, respec-
tively. Data for the other traffic levels (low-traffic and high-traffic travel
times) are generated from FFT. For the purpose of illustrating the working
of the proposed MILP model, we consider four travel-time intervals dur-
ing a day: high-traffic interval 1 (7.30  a.m. to 11 a.m.), free-flow traffic or
normal-traffic interval 2 (11 a.m. to 5 p.m.), high-traffic interval 3 (5 p.m. to
Multiobjective Routing in a Metropolitan City with Deterministic 227

2 1 4
3

7
8
5 6

9
10
11

12
13

14 15 16 17

18
19 20 21 22

23 24
25 26 27

30 29
28

33 32 31

FIGURE 9.1
Chennai road network topology: a sample (not drawn to scale). Note: origin and destination
nodes are denoted by node 1 and node 33, respectively.

9 p.m.), and low-traffic interval 4 (9 p.m. to 7.30 a.m.) in this study. We use
0.6 (i.e., 0.6 × FFT) as the multiplication factor for the low traffic and 1.8
(i.e., 1.8 × FFT) as the multiplication factor for the high traffic. The wait-
ing times with respect to normal-traffic interval at nodes are tabulated in
Table  9.3, and for other traffic levels, waiting times are generated using
these multiplication factors. Details of one-way traffic are shown in Table
9.4. Interested readers can obtain the time data from the authors at the
Indian Institute of Technology Madras.
 228

TABLE 9.1
Distance between Nodes (in km): Chennai Road Network Topology (Sample)
To/From 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33

1 – 1.6 – 0.28 – – – 1.6 – – – – – – – – – – – – – – – – – – – – – – – – –

2 1.7 – 0.08 – – – 0.28 – – – – – – – – – – – – – – – – – – – – – – – – – –
3 – 0.1 – – 0.24 – – – – – – – – – – – – – – – – – – – – – – – – – – – –
4 4.3 – – – – – – – – 0.3 – – – – – – – – – – – – – – – – – – – – – – –
5 – – 0.85 – – 0.25 – – – – – 0.62 – – – – – – – – – – – – – – – – – – – – –
6 – – – – 0.39 – 0.26 – – – 0.14 – – – – – – – – – – – – – – – – – – – – – –
7 – 0.26 – – – 0.09 – 0.09 – – – – – – – – – – – – – – – – – – – – – – – – –
8 1.7 – – – – – 0.26 – 0.11 – – – – – – – – – – – – – – – – – – – – – – – –
9 – – – – – – – 0.26 – 0.28 0.39 – – – – – – – – – – – – – – – – – – – – – –
10 – – – 2 – – – – 0.28 – – – – – – – 4.3 – – – – – – – – – – – – – – – –
11 – – – – – 0.14 – – 0.25 – – – 0.91 – – – – – – – – – – – – – – – – – – – –
12 – – – – 0.85 – – – – – – – – 0.62 – – – – – – – – – – – – – – – – – – –
13 – – – – – – – – – – 1.9 – – – 1.1 – – 0.68 0.68 – – – – – – – – – – – – – –
14 – – – – – – – – – – – 2.2 – – – – – 0.49 – – – – – – – – – – – – – – –
15 – – – – – – – – – – – – 4.3 – – 4.8 – – – – – – – – – – – – – – – – –
16 – – – – – – – – – – – – – – 3.7 – 1.1 – – – 3.7 – – – – – – – – – – – –
17 – – – – – – – – – 4.3 – – – – – 4.4 – – – – – 4.4 – – – – – – – – – – –
18 – – – – – – – – – – – – 4.8 0.73 – – – – 0.96 – – – – – – – – – – – – – –
19 – – – – – – – – – – – – 1.6 – – – – 0.16 – 0.55 – – 1.9 – – – – – – – – – –
20 – – – – – – – – – – – – – – 6 – – – 1.9 – 0.91 – – – 0.11 – – – – – – – –
(Continued)
Big Data Analytics Using Multiple Criteria Decision-Making Models
TABLE 9.1  (Continued)
Distance between Nodes (in km): Chennai Road Network Topology (Sample)
To/From 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33

21 – – – – – – – – – – – – – – – 6 – – – 0.95 – 2.4 – – – 2.4 – – – – – – –

22 – – – – – – – – – – – – – – – – 2 – – – 1.6 – – – – – 0.32 – – – – – –
23 – – – – – – – – – – – – – – – – – – 0.96 – – – – 0.73 – – – – – – – – 0.69
24 – – – – – – – – – – – – – – – – – – – – – – 0.21 – 0.49 – – – – 0.69 – – –
25 – – – – – – – – – – – – – – – – – – – 0.55 – – – 0.17 – 0.31 – – – – – – –
26 – – – – – – – – – – – – – – – – – – – – – – – – 0.35 – 0.67 – 0.88 – – – –
27 – – – – – – – – – – – – – – – – – – – – – 1.5 – – – 1.5 – 1.5 – – – – –
28 – – – – – – – – – – – – – – – – – – – – – – – – – – 0.32 – 0.62 – – – –
29 – – – – – – – – – – – – – – – – – – – – – – – – – 1.5 – 2.2 – – 0.95 – –
30 – – – – – – – – – – – – – – – – – – – – – – – 0.88 – – – – – – – 0.21 –
31 – – – – – – – – – – – – – – – – – – – – – – – – – – – – 0.35 – – 0.67 –
32 – – – – – – – – – – – – – – – – – – – – – – – – – – – – – 0.31 0.3 – 3.3
33 – – – – – – – – – – – – – – – – – – – – – – 1.9 – – – – – – – – 3.3 –
Multiobjective Routing in a Metropolitan City with Deterministic
229
 230

TABLE 9.2
Free-Flow Travel Time (FFT in Minutes): Chennai Road Network Topology (Sample)
To/
From 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33

1 – 4.364 – 0.764 – – – 4.364 – – – – – – – – – – – – – – – – – – – – – – – – –

2 4.636 – 0.184 – – – 0.764 – – – – – – – – – – – – – – – – – – – – – – – – – –
3 – 0.218 – – 0.542 – – – – – – – – – – – – – – – – – – – – – – – – – – – –
4 11.727 – – – – – – – – 0.818 – – – – – – – – – – – – – – – – – – – – – – –
5 – – 1.894 – – 0.682 – – – – – 1.378 – – – – – – – – – – – – – – – – – – – – –
6 – – – – 1.064 – 0.709 – – – 0.382 – – – – – – – – – – – – – – – – – – – – – –
7 – 0.709 – – – 0.245 – 0.245 – – – – – – – – – – – – – – – – – – – – – – – – –
8 4.636 – – – – – 0.709 – 0.3 – – – – – – – – – – – – – – – – – – – – – – – –
9 – – – – – – – 0.709 – 0.764 1.064 – – – – – – – – – – – – – – – – – – – – – –
10 – – – 5.455 – – – – 0.764 – – – – – – – 11.727 – – – – – – – – – – – – – – – –
11 – – – – – 0.382 – – 0.682 – – – 2.482 – – – – – – – – – – – – – – – – – – – –
12 – – – – 1.894 – – – – – – – – 1.378 – – – – – – – – – – – – – – – – – – –
13 – – – – – – – – – – 5.182 – – – 3 – – 1.855 1.855 – – – – – – – – – – – – – –
14 – – – – – – – – – – – 4.889 – – – – – 1.336 – – – – – – – – – – – – – – –
15 – – – – – – – – – – – – 11.727 – – 13.091 – – – – – – – – – – – – – – – – –
16 – – – – – – – – – – – – – – 10.091 – 3 – – – 10.091 – – – – – – – – – – – –
17 – – – – – – – – – 11.727 – – – – – 12 – – – – – 12 – – – – – – – – – – –
18 – – – – – – – – – – – – 13.091 1.991 – – – – 2.618 – – – – – – – – – – – – – –
19 – – – – – – – – – – – – 4.364 – – – – 0.436 – 1.5 – – 5.182 – – – – – – – – – –

(Continued)
Big Data Analytics Using Multiple Criteria Decision-Making Models
TABLE 9.2  (Continued)
Free-Flow Travel Time (FFT in Minutes): Chennai Road Network Topology (Sample)
To/
From 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33

20 – – – – – – – – – – – – – – 16.364 – – – 5.182 – 2.482 – – – 0.3 – – – – – – – –

21 – – – – – – – – – – – – – – – 16.364 – – – 2.591 – 6.545 – – – 6.545 – – – – – – –
22 – – – – – – – – – – – – – – – – 5.455 – – – 4.364 – – – – – 0.873 – – – – – –
23 – – – – – – – – – – – – – – – – – – 2.618 – – – – 1.991 – – – – – – – – 1.882
24 – – – – – – – – – – – – – – – – – – – – – – 0.573 – 1.336 – – – – 1.882 – – –
25 – – – – – – – – – – – – – – – – – – – 1.5 – – – 0.464 – 0.845 – – – – – – –
26 – – – – – – – – – – – – – – – – – – – – – – – – 0.955 – 1.827 – 2.4 – – – –
27 – – – – – – – – – – – – – – – – – – – – – 4.091 – – – 4.091 – 4.091 – – – – –
28 – – – – – – – – – – – – – – – – – – – – – – – – – – 0.873 – 1.378 – – – –
29 – – – – – – – – – – – – – – – – – – – – – – – – – 4.091 – 4.889 – – 2.591 – –
30 – – – – – – – – – – – – – – – – – – – – – – – 2.4 – – – – – – – 0.573 –
31 – – – – – – – – – – – – – – – – – – – – – – – – – – – – 0.955 – – 1.827 –
32 – – – – – – – – – – – – – – – – – – – – – – – – – – – – – 0.845 0.818 – 9
33 – – – – – – – – – – – – – – – – – – – – – – 5.182 – – – – – – – – 9 –
Multiobjective Routing in a Metropolitan City with Deterministic
231
232 Big Data Analytics Using Multiple Criteria Decision-Making Models

TABLE 9.3
Waiting Time (at Nodes/Traffic Junctions) of Chennai Road Network
Topology
Node i 1 2 3 4 5 6 7 8 9
Waiting 1.700 1.700 1.700 1.700 1.700 1.020 1.700 1.700 1.700
Time (min)

Node i 10 11 12 13 14 15 16 17 18
Waiting 1.700 1.700 1.700 2.380 2.380 2.380 1.020 1.700 1.700
Time (min)

Node i 19 20 21 22 23 24 25 26 27
Waiting 2.380 2.380 1.700 1.020 1.700 1.700 1.700 1.020 1.700
Time (min)

Node i 28 29 30 31 32 33
Waiting 1.020 1.700 1.700 1.020 1.700 1.020
Time (min)

9.5.2  Sample Network Topology

We consider the following scenarios.

• Scenario 1: In this scenario, the entire travel is completed in one

single traffic condition. For example, consider a traveler who wants
to travel from node 1 to node 33 under normal traffic conditions.
Suppose the traveler starts from node 1 at 4 p.m.; then the estimated
arrival time at the destination and the total distance traversed, based
on the optimization model, are 4.31 p.m. and 4.57 km (Table  9.5).
Figure 9.2 gives an optimal route (1–4–10–9–11–13–19–20–25–24–23–
33) with the minimum total distance and the corresponding total
time taken.

TABLE 9.4
One-Way Traffic along the Road (i,j) of Chennai Road Network Topology
One-Way Traffic along the Road (i,j) (with No Traffic
Road (i,j) Allowed along the Road (j,i)) during the Period(s)
(2,3) (7.30 a.m. to 11 a.m.); (11 a.m. to 5 p.m.); (5 p.m. to 9 p.m.).
(11,13) (7.30 a.m. to 11 a.m.); (5 p.m. to 9 p.m.).
(17,22) (7.30 a.m. to 11 a.m.); (5 p.m. to 9 p.m.).
(31,32) (7.30 a.m. to 11 a.m.); (5 p.m. to 9 p.m.).
Legend: Road (i,j) is blocked in the particular traffic interval.
Multiobjective Routing in a Metropolitan City with Deterministic 233

TABLE 9.5
Optimal Solution with Respect to Minimum Total Distance, and the
Corresponding Travel Times, and Waiting Times; Travel Starts from Node 1
at 4 p.m.
Distance along the Time along the Road Waiting Time at
Road (i,j) Road (km) (min) the Junction (min)
(1,4) 0.28 0.764 1.020
(4,10) 0.30 0.818 1.700
(10,9) 0.28 0.764 1.700
(9,11) 0.39 1.064 1.700
(11,13) 0.91 2.482 2.380
(13,19) 0.68 1.855 2.380
(19,20) 0.55 1.500 2.380
(20,25) 0.11 0.300 1.700
(25,24) 0.17 0.464 1.700
(24,23) 0.21 0.573 1.700
(23,33) 0.69 1.882 0.000
Total distance = 4.57 km Total travel time = 30.826 min

• Scenario 2: In this scenario, let us assume that there is a change in

traffic condition (say normal to heavy) during the travel from the
source to the destination. Suppose a traveler decides to start the
travel at 4.50 p.m. instead of 4 p.m., then based on our assumptions,
the travel is in the normal level from 4.50 p.m. to 5 p.m. and at 5
p.m., it changes to heavy traffic. From Table 9.6, we see that the traf-
fic condition changes from normal to heavy at node 11 resulting in
a total travel time of 45.865 min (compared with 30.826 min in sce-
nario 1) for the same route. Figure 9.3 displays the given route start-
ing from node 1 at 4.50 p.m. and ending in node 33 at 5.36 p.m. It
also displays an optimal route (1–4–10–9–11–13–19–20–25–24–23–33)
with the minimum total distance and the corresponding total time
taken.
• Scenario 3: In this scenario, we assume that there is one-way traffic
in addition to changeover of traffic condition in the middle of the
travel. Suppose the traveler encounters the unavailability of the arc
from node 11 to node 13 due to one-way traffic, especially in high-
traffic condition at 5 p.m. Here, our model considers that node 11 to
node 13 is unavailable during this interval and finds the optimal
route as (1–4–10–9–11–6–5–12–14–18–19–20–25–24–23–33) with a total
travel time of 65.744 min and distance 6.20 km (Table 9.7). Figure 9.4
displays the given route starting from node 1 at 4.50 p.m. It also dis-
plays an optimal route with the minimum total distance and the
corresponding total time taken.
234 Big Data Analytics Using Multiple Criteria Decision-Making Models

2 1 4
3 (0.28, 0.764)

(0.30, 0.818)
7
8
5 6

9
) 10
11 64 (0.28, 0.764)
1.0
) . 3 9,
2 (0
12 48
, 2.
13 91
(0.
(0.68, 1.855)

14 15 16 17

18
19 20 21 22
(0.55, 1.500)
(0.
11
,0
.30
0)

23 24
(0.21, 0.573) (0.17, 25 26 27
0.464)
2)

30 29
.88

28
9, 1
(0.6

33 32 31

FIGURE 9.2
Travel starts from node 1 at 4 p.m. without changeover (with respect to traffic conditions) of
travel time and waiting time (total distance = 4.57 km and total time covered = 30.826 min).
Note: optimal route (with respect to minimum total distance and the corresponding total time
covered/taken): {1–4–10–9–11–13–19–20–25–24–23–33}. Legend (distance, travel time): for exam-
ple (0.28, 0.764) in the above figure corresponds to distance and travel time from node 1 to node
4, and rest of the values follow suit.

9.6 Implementation of the Proposed Multiobjective

Model to the Complete Chennai Network
We implement the mathematical model for the shortest path problem with
multiobjectives for the Chennai city network. The high-traffic travel time
Multiobjective Routing in a Metropolitan City with Deterministic 235

TABLE 9.6
Travel Starts from Node 1 at 4.50 p.m.; Traffic Level Changes from 5 p.m. Onwards
Distance along the Time along the Road Waiting Time at the
Road (i,j) Road (km) (min) Junction (min)
(1,4) 0.28 0.764 1.020
(4,10) 0.30 0.818 1.700
(10,9) 0.28 0.764 1.700
(9,11) 0.39 1.064 1.700
(11,13) 0.91 2.482 4.280
(13,19) 0.68 3.339 4.280
(19,20) 0.55 2.700 4.280
(20,25) 0.11 0.540 3.060
(25,24) 0.17 0.835 3.060
(24,23) 0.21 1.031 3.060
(23,33) 0.69 3.388 0.000
Total distance = 4.57 km Total travel time = 45.865 min

is calculated and preprocessed using Bureau of Public Roads (BPR) vol-

ume-delay function (Ng and Waller, 2010) to make it closer to reality. In
order to address real-life scenarios, we consider time-dependent dynamic
and deterministic travel times, time-dependent dynamic and determin-
istic waiting times, and one-way traffic during high-traffic conditions, in
our proposed mathematical model. In Section 9.5, we explain the scenarios
such as changeover of traffic conditions and the one-way traffic during rush
period with the help of a topology and the corresponding optimal solution.
We also consider multiple objectives (total distance travelled and time taken
to reach the destination) simultaneously in our mathematical model. When
we implement the proposed approach to the city network, we have a set of
Pareto-optimal solutions for the given εt which is assumed as 1 min. Results
are presented by considering two levels of traffic. For the purpose of illus-
tration, we choose two extreme solutions from the nondominated solution
set. The routes are represented with the help of road map of Chennai city
network. Two figures are shown here to illustrate the paths over the road
map. An optimal solution with respect to total distance travelled (10.54 km)
and an optimal solution with respect to the total time taken to reach the
destination (97 min) during rush period are highlighted with thick line,
respectively, in Figures 9.5 and 9.6. We tabulate the nondominated set of
solutions with the total distance traversed and the total time taken for the
travel from IIT Madras (Node ID: 746) to Chennai Central station (Node ID:
496), and present them in Table 9.8 and Figure 9.7. The Chennai city network
data set with 1658 nodes and 4224 links as well as the results obtained from
the model are available upon request.
236 Big Data Analytics Using Multiple Criteria Decision-Making Models

2 1 4
3 (0.28, 0.764)

(0.30, 0.818)
7
8
5 6

9
) 10
11 054 (0.28, 0.764)
3 9 , 1.
.
12 .4 82) (0
13 91, 2
(0.
(0.68, 3.339)

14 15 16 17

18
19 20 21 22
(0.55, 2.700)
(0.
11
,0
.54
0)

23 24
(0.17, 25 26 27
(0.21, 1.031) 0.835)
38)

30 29
3.3

28
69,
(0.

33 32 31

FIGURE 9.3
Travel starts from node 1 at 4.50 p.m. with changeover of travel time and waiting time from
normal traffic period (up to 5 p.m.) to heavy traffic period (from 5 p.m. onwards) (total dis-
tance = 4.57 km and total time covered = 45.865 min). Note: optimal route (with respect to
minimum total distance and the corresponding total time covered/taken): {1–4–10–9–11–13–
19–20–25–24–23–33}. Legend (distance, travel time): For example (0.28, 0.764) in the above figure
corresponds to distance and travel time from node 1 to node 4, and rest of the values follow suit.

9.7 Summary
We propose a mathematical model for the shortest path problem considering
multiple conflicting objectives, namely, minimizing the total distance covered
and total time taken, thereby considering the problem closer to reality. We also
consider real-life aspects such as time-dependent dynamic and deterministic
travel times, time-dependent dynamic and deterministic waiting times, and
Multiobjective Routing in a Metropolitan City with Deterministic 237

TABLE 9.7
Travel Starts from Node 1 at 4.50 p.m., with the One-Way Traffic Regulation
from 5 p.m. Onwards
Distance along the Time along the Waiting Time at the
Road (i,j) Road (km) Road (min) Junction (min)
(1,4) 0.28 0.764 1.020
(4,10) 0.30 0.818 1.700
(10,9) 0.28 0.764 1.700
(9,11) 0.39 1.064 1.700
(11,6) 0.14 0.688 1.840
(6,5) 0.39 1.915 3.060
(5,12) 0.62 2.480 3.060
(12,14) 0.62 2.480 4.280
(14,18) 0.73 2.405 3.060
(18,19) 0.96 4.712 4.280
(19,20) 0.55 2.700 4.280
(20,25) 0.11 0.540 3.060
(25,24) 0.17 0.835 3.060
(24,23) 0.21 1.031 3.060
(23,33) 0.69 3.388 0.000
Total distance = 6.20 km Total travel time = 65.744 min
Note: Due to the one-way traffic being operational from 5 p.m. onwards with respect to road
(11,13), a new route starting from (11,6) is generated.

time-dependent one-way traffic regulation as constraints in our multiobjec-

tive mathematical model. A nondominated set of solutions are obtained using
the ε-constraint method, thereby allowing the user to select the preferred
solution. Our model inherently computes and determines the optimal route
which accounts for time-dependent dynamic and deterministic travel times,
time-dependent dynamic and deterministic waiting times, and time-depen-
dent one-way traffic regulation along the roads. The model is validated using
a sample topology for Chennai city network and also with a real-life large
metropolitan city network (Chennai city network) consisting of 1658 nodes
and 4224 arcs. Besides the quality of the solutions, an optimization model for
Chennai city network aids the decision maker to compare and choose a pref-
erable route from different alternatives. From the study, we observe that the
MCDM models can be effectively applied to large transportation network to
solve the multiple objectives that commonly arise in the shortest path prob-
lem. Our model has important real-life applications in intelligent transporta-
tion systems in relation to the route guidance and congestion mitigation on
networks with time-dependent dynamic and deterministic travel times, and
time-dependent dynamic and deterministic waiting times. As for future work,
we propose to consider the stochastic travel times along the roads in the net-
work which increase the available data size as well as the problem complexity.
238 Big Data Analytics Using Multiple Criteria Decision-Making Models

2 1 4
3 (0.28, 0.764)

(0.30, 0.818)
7
(0.14, 0.6 8
5 6 88)
(0.62, 2.480) (0.62, 2.480)

(0.39, 1.915) 9
) 10
11
1. 064 (0.28, 0.764)
39,
12 (0.
13

14 15 16 17
05)
2.4
73,
(0.

18
(0.96, 4.712) 19 20 21 22
(0.55, 2.700)
(0.
11
,0
.54
0)

23 24
(0.21, 1.031) (0.17, 25 26 27
0.835)
8)

30 29
.33

28
,3
69
(0.

33 32 31

FIGURE 9.4
Travel starts from node 1 at 4.50 p.m. with changeover of travel time and waiting time from
normal traffic to high-traffic period; one-way traffic along (13,11) being operational, but no traf-
fic between (11,13) being operational between 5 p.m. and 9 p.m. (total distance = 6.20 km and
total time covered = 65.744 min). Note: optimal route (with respect to minimum total distance
and the corresponding total time covered/taken): {1–4–10–9–11–6–5–12–14–18–19–20–25–24–23–33}.
Legend (distance, travel time): For example (0.28, 0.764) in the above figure corresponds to dis-
tance and travel time from node 1 to node 4, and rest of the values follow suit.

Acknowledgment
The authors gratefully acknowledge the support from the Centre of Excellence
in Urban Transport at the Indian Institute of Technology Madras, which is
funded by the Ministry of Urban Development, Government of India. The
first author acknowledges the support from the DAAD for carrying out a
portion of this work at the University of Duisburg-Essen. The authors are
thankful to the reviewers and the editors for their valuable comments and
suggestions to improve our chapter.
Multiobjective Routing in a Metropolitan City with Deterministic 239

FIGURE 9.5
An optimal solution with respect to the objective function (minimizing total distance) for the
travel from IIT Madras to Chennai Central Station.
240 Big Data Analytics Using Multiple Criteria Decision-Making Models

FIGURE 9.6
An optimal solution with respect to the objective function (minimizing total travel time) for
the travel from IIT Madras to Chennai Central Station.
Multiobjective Routing in a Metropolitan City with Deterministic 241

TABLE 9.8
A Set of Strictly Pareto-Optimal Solutions for Travel from IIT Madras to Chennai
Central Station, Given εt = 1 min
Distance (km) 10.54 10.77 11.42 11.72 11.95 12.60 12.83
Time (min) 114.6 111.4 107.6 105.0 101.8 100.2 97.0

116
(10.54, 114.6)
114

112
(10.77, 111.4)
110

108
(11.42, 107.6)
Time (min)

106
(11.72, 105.0)
104

102
(11.95, 101.8)
(12.60, 100.2)
100

98
(12.83, 97.0)
96
10.4 10.9 11.4 11.9 12.4 12.9 13.4
Distance (km)

FIGURE 9.7
Nondominated solutions (heuristically) for the travel from IIT Madras to Chennai Central
Station, given εt = 1 min.

References
Ahuja, R., Magnanti, T. and Orlin, V. 1993. Network Flows: Theory, Algorithms, and
Applications. Prentice Hall, Englewood Cliffs, NJ.
Bellman, R.E. 1958. On a routing problem. Quarterly of Applied Mathematics 16: 87–90.
Bertesekas, D. 1991. Linear Network Optimization: Algorithms and Codes. M.I.T. Press,
Cambridge, MA.
Brumbaugh-Smith, J. and Shier, D. 1989. An empirical investigation of some bicri-
terion shortest path algorithms. European Journal of Operational Research 43:
216–224.
Chabini, I. 1998. Discrete dynamic shortest path problems in transportation applica-
tions. Transportation Research Record 1645: 8170–8175.
Chitra, C. and Subbaraj, P. 2010. A non-dominated sorting genetic algorithm for
shortest path routing problem. International Journal of Electrical and Computer
Engineering 5: 53–63.
242 Big Data Analytics Using Multiple Criteria Decision-Making Models

Dell’Amico, M., Lori, M. and Pretolani, D. 2008. Shortest paths in piecewise continu-
ous time-dependent networks. Operations Research Letters 36: 688–691.
Dreyfus, S. 1969. An appraisal of some shortest-path algorithms. Operations Research
17: 395–412.
Ehrgott, M. and Wiecek, M. 2005. Multiobjective programming. In: J. Figueira, S.
Greco, and M. Ehrgott, editors, Multicriteria Decision Analysis: State of the Art
Survey, Springer Verlag, Boston, Dordrecht, London: 667–722.
Hamacher, H.W., Ruzika, S. and Tjandra, S.A. 2006. Algorithms for time-dependent
bicriteria shortest path problems. Discrete Optimization 3: 238–254.
Mohamed, C., Bassem, J. and Taicir, L. 2010. A genetic algorithm to solve the bicrite-
ria Shortest Path Problem. Electronic Notes in Discrete Mathematics 36: 851–858.
Müller-Hannemann, M. and Schnee, M. 2007. Finding all attractive train connec-
tions by multi-criteria Pareto search. In: F. Geraets, L. Kroon, A. Schoebel,
D.  Wagner, and C. Zaroliagiis, editors, Algorithmic Methods and Models for
Railways Optimization, Springer, Berlin, Heidelberg, 4359: 246–263.
Ng, M.W. and Waller, S.T. 2010. A computationally efficient methodology to char-
acterize travel time reliability using the fast Fourier transform. Transportation
Research Part B 44: 1202–1219.
Prakash, A. and Srinivasan, K.K. 2014. Sample-based algorithm to determine mini-
mum Robust cost path with correlated link travel Times. Transportation Research
Record: Journal of the Transportation Research Board 2467: 110–119.
Ruzika, S. and Wiecek, M.M. 2005. Approximation methods in multi-objective
­programming. Journal of Optimization Theory and Applications 126: 473–501.
Schrijver, A. 2003. Combinatorial Optimization: Polyhedra and Efficiency, Springer
Verlag, Berlin.
Seshadri, R. and Srinivasan, K.K. 2010. Algorithm for determining most reliable
travel time path on network with normally distributed and correlated link
travel times. Transportation Research Record: Journal of the Transportation Research
Board 2196: 83–92.
Seshadri, R. and Srinivasan, K.K. 2012. An algorithm for the minimum robust cost
path on networks with random and correlated link travel times. In: D. M.
Levinson, H. X. Liu, and M. Bell, editors, Network Reliability in Practice, Springer,
New York: 171–208.
Sung, K., Bell, M.G.H., Seong, M. and Park, S. 2000. Theory and Methodology-
Shortest paths in a network with time-dependent flow speeds. European Journal
of Operational Research 121: 32–39.
Tiwari, A. 2016. Multi-objective convoy routing problem: Exact and heuristic
­methods, Unpublished MS thesis, IIT Madras, Chennai, India.
Wardell, W. and Ziliaskopoulos, A. 2000. A intermodal optimum path algorithm for
dynamic multimodal networks. European Journal of Operational Research 125:
486–502.
Ziliaskopoulos, A. and Mahmassani, H. 1993. Time-dependent shortest path algo-
rithm for real-time intelligent vehicle/highway system. Transportation Research
Record 1408: 94–104.
10
Designing Resilient Global Supply Chain
Networks over Multiple Time Periods within
Complex International Environments

Rodolfo C. Portillo

CONTENTS
10.1 Introduction................................................................................................. 243
10.2 Literature Review....................................................................................... 246
10.3 Problem Description................................................................................... 247
10.4 Model Features............................................................................................ 248
10.4.1 Multiperiod Model Features......................................................... 250
10.4.1.1 Notation............................................................................. 250
10.4.1.2 Variables............................................................................ 253
10.4.2 Objective Function..........................................................................254
10.4.3 Model Scaling..................................................................................254
10.4.4 Set of Goal Constraints.................................................................. 255
10.4.5 Set of Constraints............................................................................ 257
10.4.6 Objective Function Considering Currency Exchange Rates....... 260
10.5 Data Collection............................................................................................ 262
10.6 Case Study................................................................................................... 262
10.7 Conclusions and Future Research............................................................ 265
References.............................................................................................................. 266

10.1 Introduction
With increased globalization, as stated by Friedman (2005), “In this world, a
smart and fast global supply chain is becoming one of the most important
ways for a company to distinguish itself from its competitors.” As tradition-
ally, the objective of every supply chain continues to be the maximization of
the overall value by reducing procurement cost, increasing responsiveness
to customers, and decreasing risk. The big change now is that global sup-
ply chain management involves a myriad of company’s worldwide interests,
customers, and suppliers rather than just a domestic perspective. Besides the

243
244 Big Data Analytics Using Multiple Criteria Decision-Making Models

financial aspects, companies now deal with a plethora of other factors when
doing business abroad. Within this environment, as part of the company’s
strategy to manage its global supply chain, it must make decisions such as its
overall sourcing plan, supplier selection, capacity and location of facilities,
modes of transportation, etc. The emphasis of this research is on develop-
ing mathematical models to determine optimal supply chain designs that
best support competitive strategies. Multicriteria mixed-­integer linear pro-
gramming models were developed to aid in a multiple echelon supply chain
design. This work also includes the definition of a set of design selection
criteria integrating financial, customer service, risk, and strategic factors.
A supply chain consists of (1) a series of physical entities (e.g., suppliers,
plants, warehouses, and retailers) and (2) a coordinated set of activities
concerned with the procurement of raw material and parts, production of
intermediate and final products, and their distribution to the customers
(Ravindran and Warsing 2013). The various decisions involved in manag-
ing a supply chain can be grouped into three types: strategic, tactical, and
operational. Strategic decisions deal primarily with the design of the sup-
ply chain network, namely, the number and location of plants and ware-
houses and their respective capacities. They are made over a longer time
horizon and have a significant impact with respect to the company’s assets
and resources, such as opening, expanding, closing, and downsizing facili-
ties. Tactical decisions are primarily of a planning nature and made over a
horizon of one or two years. They involve purchasing, aggregate production
planning, inventory management, and distribution decisions. Finally, opera-
tional decisions are short term and made on a daily or weekly basis, such as
setting customer delivery and weekly production schedules as well as inven-
tory replenishment.
Optimal supply chain design needs to balance among multiple conflicting
objectives, such as efficiency in terms of costs and profitability as well as speed
to source, produce, and distribute products to customers. Resiliency is also an
important objective and is measured in terms of the reliability of the supply
chain network when there are disruptions to the supply chain. The case study
presented in this chapter addresses strategic and tactical decisions in design-
ing and managing an agile global supply chain considering the effect of mul-
tiple foreign currency exchange rates over multiple periods of time.
Decision makers often need to consider multiple criteria in order to deter-
mine the best course of action to solve a particular problem. The relationship
among these decision criteria can be conflicting, which implies that trade-
offs need to be considered and carefully evaluated. The search for an opti-
mal solution for a multiobjective problem becomes a simultaneous process of
optimizing two or more conflicting objectives. Refer to Ravindran (2008) and
Ravindran et al. (2006) for further reading on multiple objective optimization
methods.
As described by Masud and Ravindran (2008), a multicriteria decision-
making problem in general can be represented as follows:
Designing Resilient Global Supply Chain Networks 245

Maximize

C1 ( x), C2 ( x),… , Ck ( x),

where

x ∈X

x is any specific alternative,

X is a set representing the feasible region or available alternatives, and
C1 is the lth evaluation criterion.

According to the authors, multicriteria decision-making (MCDM) prob-

lems can be classified in two types: (1) the multicriteria mathematical pro-
gramming problems (MCMP) with an infinite number of feasible alternatives
implicitly determined by a finite number of explicitly stated constraints, and
(2) the discrete multicriteria selection problems (MCSP) that consist of a finite
number of alternatives stated explicitly. Refer to Chapter 2 for more details
on MCDM models.
The material in this chapter is based on the doctoral dissertation of the
author (Portillo 2009). With increased globalization, global supply chain
management has become strategically important for many companies. The
objectives of every supply chain continue to be maximizing the overall value
generated by reducing the costs of procurement, increasing the responsive-
ness to customers, and decreasing the risks due to disruptions affecting the
supply chain network. The big change now is that global supply chain man-
agement involves a company’s worldwide interests: manufacturing facili-
ties, distribution centers (DCs), customers, and suppliers located in several
countries. Besides the conventional financial aspects, companies are now
required to deal with a plethora of other factors for doing business abroad,
such as duties, transfer prices, taxes, multiple exchange rates, and disrup-
tion risk. Transfer prices are charges among enterprise entities on goods
and s­ ervices. Disruption risk can be due to natural disasters, supplier qual-
ity issues, political tensions among countries, civil unrest, economic issues,
government controls, and strikes from unions, among others. Within this
environment, as part of their global supply chain management strategy,
a company must make decisions on its overall outsourcing plan, supplier
selection, the number of production plants and DCs that are needed, as well
as the locations of those facilities, modes of transportation, and customer
allocations to the DCs.
The emphasis of this chapter is on developing mathematical models to
determine optimal supply chain design to support specific competitive strat-
egies within the complex multinational environment across multiple peri-
ods of time. A multicriteria mixed-integer linear programming model has
been developed to aid in a multiple echelon supply chain design, including
246 Big Data Analytics Using Multiple Criteria Decision-Making Models

manufacturing and distribution facilities’ location/allocation, capacity and

expansion requirements, production and distribution variables, interna-
tional issues, exchange rates, lead times, transfer prices, and time periods.
This work also considers a variety of semifinished goods of a health and
hygiene consumer products company (e.g., tissue hard rolls, unpacked dia-
pers, and oily soap solution) and finished products (e.g., facial tissues, toilet
paper, toiletries, and gloves). Moreover, it includes the definition of a set of
supply chain design criteria that integrates financial aspects, customer ser-
vice, disruption risk, and strategic factors in the process. Strategic factors
may include decisions to open new markets, increase market share, and to
strengthen relationships with particular customers.
These methods for designing a resilient and responsive supply chain were
applied to a leading global health and hygiene company listed in the Fortune
500 that operates in 37 countries; its global brands are sold in more than 150
countries and used by approximately 1.3 billion people, holding first or second
market positions in the majority of markets. This case study focuses on the
company’s largest international division selling products across one continent
with several offices, distribution and manufacturing facilities in 22 countries.
As part of its competitive strategy, the firm serves from multinational chains
to thousands of small “mom and pop” stores. A complex supply chain struc-
ture resulted after a series of mergers and acquisitions having 45 distribution
facilities sourced by 21 plants spread out 10 countries within the division and
other facilities located at different continents. In order to enhance the divi-
sion’s competitive strategy, a robust, flexible, and efficient global supply chain
design was required to assure exceptional achievement of customer service
and financial goals while considering related risk and strategic factors.
This application represented a major big data effort (refer to Chapter 3 on Big
Data). An aggregated database consisting of approximately 75,000 data records,
including marketing, sales, production, distribution, and purchasing informa-
tion, was extracted from the company’s enterprise resource planning (ERP)
system, data warehouse, and other platforms by examining large sets contain-
ing structured and unstructured data of different types and sizes. Data quality
assessment methods such as data profiling and standardization, cross tabula-
tion, matching and linking, outlier detection, among others. In addition, query
programming was critical to automate data extraction. Refer to an overview
of the state of the art and focus on emerging trends to highlight the hardware,
software, and application landscape of big data analytics (Kambatla et al. 2014).

10.2  Literature Review

Over 30 years ago, researchers recognized that systematic optimization
should be used instead of common sense. Early models tended to treat only
Designing Resilient Global Supply Chain Networks 247

logistics aspects, while most developed thus far have focus primarily on
single-criterion financial measures. A vast majority has addressed portions
of the supply chain. Moreover, only a few have incorporated multinational
and global criteria. Research for supply chain design started early on with
a model developed for Hunt-Wesson Foods (Geoffrion and Graves 1974).
After a decade, a system was implemented for Nabisco Foods, Inc. (Brown
et al. 1987). Several applications followed for a petrochemical company (Van
Roy and Wolsey 1985), Libbey–Owens–Ford (Martin et  al. 1993), and Auli
Foods (Pooley 1994). A comprehensive global supply chain model (GSCM)
was applied to Digital Equipment Corporation (Arntzen et  al. 1995). Later,
a supply chain restructuring was supported at Procter & Gamble using
mathematical optimization models (Camm et  al. 1997). Other large-scale
comprehensive models were implemented at Caterpillar (Rao et al. 2000), an
agrochemicals company (Sousa et  al. 2008), and to a global chemicals firm
(Tsiakis and Papageorgiou 2008). All, except for the one at Digital Equipment
Corporation (1995), have been established either for optimizing costs or
profitability alone. This led to multiobjective approaches and surprisingly
enough, despite the nature of the problem, very little work has been devoted
using multicriteria techniques. The relationship among criteria can be con-
flicting, implying trade-offs, so the search for optimality becomes a simulta-
neous process. An early multicriteria approach was presented at Netherlands
Car BV (Ashayeri and Rongen 1987). Analytic hierarchy process (AHP) and
ranking methods for facility relocation were also proposed for solving sup-
ply chain design problems (Melachrinoudis and Min 1999). These authors
later included two weighted objectives within a single nonpreemptive Goal
Programing (GP) objective (Melachrinoudis et al. 2000). An AHP for optimiz-
ing the strategic importance of customers and their related risks was also
proposed (Korpela et al. 2002). More recently, a bicriteria nonlinear stochastic
optimization model was presented to determine the best supply chain (Gaur
and Ravindran 2006). Afterwards, a multiobjective model was presented for
solving a vendor selection problem (Wadhwa and Ravindran 2007). Arguably,
most attention has been paid to methodologies that break the problem into
pieces and simplify the inherent complexity of the supply chain.

10.3  Problem Description

In the last 15 years, a series of mergers and acquisitions across the conti-
nent led to a supply chain with a highly complex internal structure of many
production and distribution facilities without standard assets and product
technologies, together with redundancy of operational strategies and orga-
nizational structures. Most of the products sold in the region are supplied
by these 21 manufacturing plants, and the rest are imported from other
248 Big Data Analytics Using Multiple Criteria Decision-Making Models

company facilities located all around the world. In addition, cross-sourcing

activity within the continent has increased significantly in the last few years,
by blending the advantages of single and multiple sourcing strategies. Today,
more than 60% of the production facilities manufacture finished and semi-
finished products that are distributed to different countries in addition to
the local market. At least three facilities are continent-wide facilities sourc-
ing all markets within the division, and they also export products to other
company divisions worldwide. Until now, asset rationalization to improve
operational efficiencies and structural reorganization efforts have focused
on supply chain distribution designs for specific business units or division-
wide designs considering manufacturing facilities for particular products
only. In order to better support and enhance the division’s competitive strat-
egy, a robust, flexible, and efficient global supply chain design was required
to ensure exceptional achievement of customer service levels and financial
goals while considering related risk factors.
For a complete strategic and tactical optimization of the manufacturing
and distribution network, the model needed to manage more than 100 cus-
tomer zones or markets, as well as dozens of products manufactured in more
than 250 production lines. Because of the diverse international nature of the
problem, many global factors needed to be considered, such as domestic and
international freights, transfer prices, taxes and duties, among others. Some
products have a multistage production process and are processed in mul-
tiple echelons. In other words, the production process involves more than
one stage before the final product is ready, where inventory may need to
be managed and production technology and manufacturing lead times may
differ. For example, in tissue production, you may have multiple processes
to manufacture the paper itself in the form of semifinished hard rolls that
are later converted to finished products such as napkins, facial tissue, toi-
let paper, and kitchen towels. Multiple production rates are considered for
multiproduct machines and products. The costs in the optimization process
include facility overheads, fixed and variable costs of production lines, and
raw materials consumption costs.

10.4  Model Features

A base model for designing resilient global supply chain networks was pre-
sented by Portillo (2016). This chapter extends the application of this model
to consider supply chain network design tactical decisions across multiple
time periods as well as the impact of incorporating multiple foreign cur-
rency exchange rates given the global geographic scope of the analysis.
The multicriteria model integrates customer service levels, strategic factors,
and disruption risk criteria along with the financial measure of performance.
Designing Resilient Global Supply Chain Networks 249

Customer service level is measured using two factors: (1) demand fulfillment
and (2) speed of delivery. Demand fulfillment is defined as the portion of the
customer demand that is satisfied, namely the quantity that is effectively deliv-
ered to the customers. The ability to completely fulfill customer demand is
modeled as a goal constraint by specifying demand fulfillment targets for all
the combinations of products and customer zones. Speed of delivery is mea-
sured in terms of the lead time to deliver the products to the customers. This is
also modeled as a goal, by minimizing the quantity weighted lead time, based
on volume and the respective delivery lead times.* Weighted lead time targets,
for each customer zone, are explicitly considered in the GP model. In addition,
the multicriteria model considers the minimization of risk associated with
supply chain disruptions. Different measures of risk for domestic and global
sourcing are estimated for each manufacturing, converting, and distribution
location. These measures incorporate facility- and c­ ountry-specific risk fac-
tors. Facility-specific risk factors are determined based on assessments per-
formed by the decision makers. Country-specific risk factors are obtained by
considering the weighted average cost of capital rates for each country. The
objective of minimizing the risk measure is also modeled as a goal constraint
by setting the overall risk target value for the entire supply chain. Decisions
related to supply chain network design may also require the modeler to con-
sider strategic factors to open new markets, to increase market share, and to
strengthen relationships with customers. This model includes measures for
strategic factors for each facility, based on the ratings provided by the deci-
sion makers. A goal constraint is set to achieve the maximum possible overall
strategic measure for the entire supply chain network.
Among other features, the model allows the evaluation of outsourcing
decisions as well as the consideration of different product mix and corre-
sponding productivity rates on different production lines and at different
locations. The model supports both strategic and tactical decisions.
On the strategic side, the focus is on the design of the supply chain net-
work, in which the optimization model determines the facilities that need
to be opened and their locations, as well as the facilities that negatively
affect profitability and therefore need to be closed. In the case where cur-
rent network capacity (measured in product units) is not sufficient to fulfill
customers’ demand, the model provides for manufacturing and distribution
decisions, evaluating where and how capacity should be expanded or out-
sourced. Note that, when combining products with significantly different
specifications, a common standard unit of measure may be defined within the
enterprise, such as weight (i.e., tons) and volume (i.e., cubic meters) for mea-
suring capacity. Also, the ability to perform analysis at the production line
level facilitates decisions associated with the transfer of equipment among
facilities. Moreover, strategic decisions related to technological changes are

* Corresponding to each arc of the supply chain network that links a facility (plant/DC) to a
customer zone.
250 Big Data Analytics Using Multiple Criteria Decision-Making Models

supported by the model, such as what technologies are more convenient for
the required expansions, or what specific equipment should be considered
for write-off and replacement. The model also assists in tactical decisions,
such as customer zone assignments to the DCs, the development of high-level
production and distribution plans, product allocation to specific equipment,
and cross-sourcing among production facilities. The objective then becomes
the minimization of the deviations from the specified criteria targets: profit,
demand fulfillment, lead time, disruption risk, and strategic factors.
A detailed description of the additions to the multicriteria base model
(Portillo 2016) is included in the following sections, including some addi-
tional mathematical notation.

10.4.1  Multiperiod Model Features

Balancing supply and demand has been an art disguised as science, and the
ability to precisely match production and sales volume has become more
important than ever for highly globalized businesses—for both operational
and financial efficiencies.
With this objective, the model presented in this chapter allows on one
side the evaluation of production and sourcing capabilities compared with
expected demand volume at different time periods. By optimizing multiple
criteria, the model determines first the optimal configuration of manufac-
turing and distribution facilities for each period of time as well as the cor-
responding optimal production mix and cross-sourcing strategies to meet
the desired objectives. It also provides for the evaluation of different ­supply
chain network configuration alternatives over time, perhaps determin-
ing the optimal time period when a production line should be opened. In
addition, different values for each time period can be set for most of the
parameters included in the mathematical formulation, for example, allow-
ing for the inclusion of price and cost trends, productivity learning curves
and changes from continuous improvement, modifications in transfer prices
and international commercial terms over time, among others. Moreover, the
optimization model is flexible enough to handle different time horizons and
periodicities, allowing both long-term strategic analyses as well as short-term
tactical plans. Since the focus of this work is on strategic and some ­tactical
types of decision, more operational decisions that involve the analysis of
inventory levels have not been considered. However, including inventory
variables and modifying the mathematical formulation allows this feature
to be easily extended in the future.

10.4.1.1 Notation
10.4.1.1.1  Index Sets
Index m is included to represent time periods that can be defined depend-
ing on the level of analysis required as months, quarters, years, or others.
Designing Resilient Global Supply Chain Networks 251

Index o is added to represent the starting period for specific design alterna-
tives. By using the latter, it is possible, for example, to set different start-up
periods for a production line for which the model will determine the optimal
solution by considering the trade-offs from productivity learning curves,
different sourcing options, and machine installation and operating costs. In
this case, a particular production line under evaluation will have multiple
production capacity and operating costs in a specific time period depending
on when it started operations.

10.4.1.1.2 Parameters
This model considers a breakdown of customer demand as well as produc-
tion and distribution capacities, balancing them at each period and optimiz-
ing the supply chain design and flows accordingly. Different values may be
applied to sales prices, raw material and machine variable costs, transfer
prices, duties, and freights over time, allowing for the analysis of the impact
that changes with respect to these parameters in the network, creating cross
sourcing and improving gross profit. These parameters are:

Dkpm = customer demand volume in market k for category p at time

period m
Pkpm = sales price for category p in market k at time period m
MRChpm = raw material variable cost for category p in manufacturing
plant h at time period m
MUChpm = CIF (cost, insurance, and freight) unit cost for category p for
manufacturing facility h at period m
( 2 )  = cross-sourcing cost percentage for category p from manufactur-
I hipm
ing facility h to converting facility i at time period m
( 2 )  = freight in $/ton for category p from manufacturing facility h to
Thipm
converting facility i at time period m
( 2 )  = freight in $/ton for category p from manufacturing facility h to
TI hipm
entry port/customs when sending goods to converting facility i at
time period m
( 3 )  = freight in $/ton for category p from manufacturing facility h to
Thkpm
market k at time period m
CRCipm = raw material variable cost per ton for category p in converting
plant i at time period m
CMCtm = machine cost per hour for converting line t at time period m
CUCipm = CIF unit cost for category p for converting site i at time period m
( 4)
I ijpm  = cross-sourcing cost percentage for category p from converting
facility i to distribution facility j at time period m
( 4 )  = freight in $/standard unit for category p from converting facility
Tijpm
i to distribution center j at time period m
252 Big Data Analytics Using Multiple Criteria Decision-Making Models

( 4 )  = transportation cost for product p from converting facility i to

TI ijpm
entry port/customs when sending goods to distribution facility j in
$/standard unit at m
( 5 )  = freight in $/standard unit for category p from converting facility
Tikpm
i to market k at time period m
( 7 )  = freight in $/standard unit for category p from distribution facil-
Tjkpm
ity j to market k at time period m
( 6 )  = freight in $/standard unit for category p from distribution facil-
Tjlpm
ity j to distribution center l at time period m
( 6 )  = transportation cost for product p from DC j to entry port/cus-
TI jlpm
toms when sending goods to DC l in $ per standard unit at time
period m
DUCjlm = CIF unit cost for category p for distribution center j at time
period m
( 6 )  = cross-sourcing cost percentage for category p from distribution
I jlpm
facility j to distribution facility l at time period m
VOCj = variable operational cost of DC j in $/cubic meters at time
period m

In addition, parameters representing fixed operational costs for manu-

facturing, converting and distribution facilities and production lines, pro-
duction and shipping capacities, and productivity rates consider not only
m but also index o. Multiple values can be considered by the optimiza-
tion model for one period of time depending on the start date when a
given facility or production line started operations. These parameters are
defined next:

FMhmo = facility fixed costs of manufacturing plant h at time period m if

opened at time period o
FMMhtmo = fixed costs of production line t in manufacturing site h at
time period m if opened at time period o
MPTptmo = manufacturing time hours per ton for category p at produc-
tion line t at time period m if opened at time period o
CPTptmo = conversion time hours per ton for category p at production
line t at m for alternative o
MMChtmo = machine cost per hour for manufacturing line t at time
period m if opened at time period o
FCimo = facility fixed costs of converting plant i at time period m if
opened at o
FCMitmo = fixed costs of production line t in converting site i at time
period m if opened at o
Designing Resilient Global Supply Chain Networks 253

CMCitmo = machine cost per hour for production line t in converting

facility i if opened at o
FDjmo = fixed operational cost of DC j at time period m if opened at o
MChmo = manufacturing capacity of h at time period m if opened at o
CCimo = converting capacity of i at time period m if opened at o
SCjmo = shipping capacity of DC j at time period m if opened at o

10.4.1.2 Variables
Similarly, the variables corresponding to the manufacturing and converting
supply chain echelons include the m and o indices indicating the correspond-
ing flow between a pair of nodes at period m if the facility or production line
is opened at period o.

x(hiptmo
1)  = production of category p at production line t at manufacturing
facility h sent to converting facility i at time period m if production
line is opened at o
x hkptmo  = production of category p at production line t at manufactur-
( 2)

ing facility h sent to market k at time period m if production line is

opened at o
(1)
yijptmo  = production of category p at production line t at converting facil-
ity i sent to distribution facility j at time period m if production line
is opened at o
( 2)
yikptmo  = production of category p at production line t at converting facil-
ity i sent to market k at time period m if production line is opened at o

However, since it is assumed that the shipping capacity of a distribution

center will not be significantly affected by when it started operations, for
simplicity index o is not added to the distribution center outflow variables.
These variables are:

1)  = distribution of category p from DC j to DC l at time period m

z(jlpm
2 )  = distribution of category p from DC j to market k at time period m
z(jkpm

The binary variables only include index o, so multiple binary variables will
exist for a particular facility or production line depending on different start-
up period options. Note that index m is not included since the objective is
to decide to open/close an asset or not and if so then determine when is the
optimal time to perform the corresponding action.

δ (ho1)  = 1 if manufacturing facility h is opened at time period o, 0 otherwise

γ (1)
hto
 = 1 if production line t in manufacturing facility h is opened at time
period o, 0 otherwise
254 Big Data Analytics Using Multiple Criteria Decision-Making Models

δ (io2)  = 1 if converting facility i is opened at time period o, 0 otherwise

γ (ito2)  = 1 if production line t in converting facility i is opened at time
period o, 0 otherwise
( 3 )  = 1 if distribution facility j is opened at time period o, 0 otherwise
δ jo

The modified model formulation is presented in Equations 10.1 through 10.39.

10.4.2  Objective Function

In this model, the objective function includes the time period index m in
the demand negative deviation variable term, which now adds the demand
deviations for each of the market, product, and time period combinations.
All other terms in the objective function are the same as in previous ­models.
Similarly, the parameters wg and Pg are used to incorporate the cardinal
weights and ordinal priorities for a goal g, which are determined based on
the decision maker preferences. Weights or preferences are used depend-
ing on the type of the objective function, nonpreemptive, or preemptive,
respectively.
Note that in the model presented below the demand fulfillment deviation
variable is indexed on the time period m as well, such that d5−kpm . This means
that specific target values for demand fulfillment are specified for each
­product category p sold in a given market k at a particular time period m. The
sum of all the corresponding deviation variables is then averaged arithmeti-
cally by the total number of market, product, and time period combinations
with customer demand greater than zero, which is given by the product of
the cardinality of the sets MKS, PRS, and PDS such that Dkp > 0∀k , p . This
result is weighted then as a component of the goal-programming objective
function. See below:
Nonpreemptive


z = w d + w2 
− ∑ d2+k 
k ∈MKS

 + w3 d3+ + w4 d4− + w5 
∑ m∈PDS ∑ k ∈MKS ∑
d5−kpm 
.
p∈PRS
(10.1)
1 1
 MKS   MKS * PRS * PDS 
   

Preemptive


z = P1d1− + P2 
∑ d2+k 
k ∈MKS

 + P3 d3+ + P4 d4− + P5 
∑ m∈PDS ∑ k ∈MKS ∑
d5−kpm 
.
p∈PRS
(10.2)
 MKS   MKS * PRS * PDS 
   

10.4.3  Model Scaling

For the nonpreemptive version of the multicriteria objective function, scaling
is necessary for proper optimization. Note that the goal constraints explained
Designing Resilient Global Supply Chain Networks 255

below can be given in different units of measure and can significantly vary
in magnitude, for example, gross profit can be given in millions of currency
units, lead time from single to no more than double-digit quantity of days,
demand fulfillment as a percentage amount, and risk and strategy as single-
digit measures. The formulation incorporates each goal target for scaling
purposes assuring that the deviation variable values are such that 0 ≤ d ≤ 1.
In the nonpreemptive model, the objective function minimizes the weighted
result of the scaled deviation variables value. For the preemptive version
of the model, scaling is not necessary and consequently the goal constraint
should incorporate the target values on the right hand side of the equation,
allowing for the deviation variables to take a value according to the base unit
of measure of each goal constraint (currency, days, percentage points, and
risk and strategy units). In this case, the model will sequentially optimize
each of the goals based on their priority.

10.4.4  Set of Goal Constraints

Similar to the objective function, some terms in the goal constraints include
changes based on the modified parameters, including indices m and o.
These are presented below:
f1 = Max. gross profit

1 
∑ ∑ ∑ ∑ z(jkpm
2)
Pkpm
T µ(1)  k∈MKS p∈PRSKk j∈DCSKK m∈PDS
+ ∑ ∑ ∑ ∑ ∑ ( 2)
∑ y ikptmo Pkpm (10.3)
k ∈MKS p∈PRSK k i∈CFSK k t∈CTSCi m∈PDS o∈ODS
( 2)
+ ∑ ∑ ∑ ∑ ∑ ∑ x P
hkptmo kpm
k ∈MKS p∈PRSK k h∈MFSK k t∈CTSMh m∈PDS o∈ODS

− ∑ ∑ ∑ FMhmoδ (ho1) − ∑ ∑ ∑ FCimoδ (io2) − ∑ ∑ ∑ FDjmoδ (jo3 ) (10.4)

h∈MFS m∈PDS o∈ODS i∈CFS m∈PDS o∈ODS j∈DCS m∈PDS o∈ODS

− ∑ ∑ ∑ ∑ FMMhtm γ (hto
1)
− ∑ ∑ ∑ ∑ FCMitm γ (ito2) (10.5)
h∈MFS t∈CTSMh m∈PDS o∈ODS i∈CFS t∈CTSCi m o

 
− ∑ ∑ ∑ ∑ MRChpm  ∑ ∑ x(1)) + ∑ ∑ x( 2)
 i∈CFSMh t∈CTSMh hiptmo k∈MKSMh t∈CTSMh hkptmo 
(10.6)
m∈PDS h∈MFS p∈PRSMh o∈ODS

 
− ∑ ∑ ∑ ∑ ∑ MMCtmo MPTptmo  ∑ x(h1iptmo
 i∈CFSMh
)
+ ∑ x(hkptmo
2)

(10.7)
m∈PDS o∈ODS h∈MFS t∈CTSMh p∈PRSTt k ∈MKSMh

 
− ∑ ∑ ∑ CRCipm
i∈CFS p∈PRSCi m∈PDS
∑ ∑  ∑
t∈CTSCi o∈ODS
∑ y ij(1p)tmo + ∑
 j∈DCSCi m∈PDS
( 2)
∑ y ikptmo
k ∈MKSCi m∈PDS
 (10.8)

 
− ∑ ∑ ∑ ∑
m∈PDS o∈ODS i∈CFS t∈CTSCi p∈PRSTt
∑ CMCtmoCPTptmo  ∑ y i(j1ptmo
 j∈DCSCi
) ( 2)
+ ∑ y ikptmo
k ∈MKSCi
 (10.9)

256 Big Data Analytics Using Multiple Criteria Decision-Making Models

 ( 2) 
− ∑ ∑ VOC j  ∑ ∑ z(jlpm
1) ( 2)
DFjkp + ∑ ∑ z(jkpm
2)
DFjkp  (10.10)
m∈PDS j∈DCS  l∈DCSC j p∈PRSDt k ∈MKSD j p∈PRSK k

( 2)
− ∑ ∑ ∑ ∑ ∑ ∑ Thipm x(hiptmo
1)
m∈PDS h∈MFS i∈CFSMh p∈PRSCi t∈CTSCh o∈ODS
( 4 ) (1)
− ∑ ∑ ∑ ∑ ∑ ∑ Tijpm y ijptmo (10.11)
m∈PDS i∈CFS j∈DCSCi p∈PRSDj t∈CTSCi o∈ODS
( 6 ) (1)
− ∑ ∑ ∑ ∑ T z
jlpm jlpm
m∈PDS j∈DCS l∈DCSD j p∈PRSDl

(3)
− ∑ ∑ ∑ ∑ ∑ ∑ Thkpm x(hkptmo
2)
m∈PDS h∈MFS k ∈MKSMh p∈PRSK k t∈CTSCh o∈ODS
( 5) ( 2)
− ∑ ∑ ∑ ∑ ∑ ∑ Tikpm y ikptmo (10.12)
m∈PDS i∈CFS k ∈MKSCi p∈PRSK k t∈CTSCi o∈ODS
(7 ) ( 2)
− ∑ ∑ ∑ ∑ T z
ijpm jkpm
m∈PDS j∈DCS k ∈MKSD j p∈PRSK k

( 2)
− ∑ ∑ ∑ ∑ ∑ ∑ MUChpm I hipm x(h1iptmo
)
m∈PDS h∈MFS p∈PRSMh i∈CFSMh t∈CTSMh o∈ODS
( 4 ) (1)
− ∑ ∑ ∑ ∑ ∑ ∑ CUCipm I ijpm y ijptmo (10.13)
m∈PDS i∈CFS p∈PRSCi j∈DCSDi t∈CTSCi o∈ODS
( 6 ) (1)
− ∑ ∑ ∑ ∑ DUC I
jpm jlpm jlpm z
m∈PDS j∈DCS p∈PRSD j l∈DCSD j

( 2) ( 2)
− ∑ ∑ ∑ ∑ ∑ ∑ TI hipm I hipm x(hiptmo
1)
m∈PDS h∈MFS i∈CFSMh p∈PRSCi t∈CTSMh o∈ODS
(10.14)
− ∑ ∑ ∑ ∑ ∑ ∑ TI i(jpm
4 ) ( 4 ) (1)
I ijpm y ijptmo
m∈PDS i∈CFS j∈DCSCi p∈PRSD j t∈CTSCi o∈ODS

6 ) ( 6 ) (1) 
− ∑ ∑ ∑ ∑ TI (jlpm I jlpm z jlpm  + d1− − d1+ = 1. (10.15)
m∈PDS j∈DCS l∈DCSD j p∈PRSDl 

f2 = Min. lead time to markets

∑ L(ik5) ∑ ∑ ∑ ( 2)
∑ y ikptmo + ∑ L(j7k) ∑ z(jkpm
2)
i∈CFS o∈ODS m∈PDS p∈PRSK k t∈CTSCi j∈DCS p∈PRSK k

∑ ∑ Dkpm
m∈PDS p∈PRSK k
Dkpm >0

 2) 
T (k2)  ∑ ∑ ∑ ( 2)
∑ y ikptmo + ∑ ∑ z(jkpm 
 m∈PDS o∈ODS p∈PRSKk t∈CTSCi m∈PDS p∈PRSK k
−
+d − d =
2k
+
2k ∀k ∈ MKS. (10.16)
∑ ∑ Dkpm
m∈PDS p∈PRSK k
Dkpm >0

f3 = Min. risk of supply chain disruptions

1 
∑ ∑ R(h1)δ (ho1) + ∑ ∑ R(i2)δ (io2) + ∑ ∑ R(j3 )δ (jo3 )
T ( 3 )  h∈MFS o∈ODS i∈CFS o∈ODS j∈DCS o∈ODS

+ ∑
h∈MFS o∈ODS
(
∑ Rc(h1) 1 − δ (ho1) + ∑ ) i∈CFS o∈ODS
(
∑ Rci( 2) 1 − δ (io2) + ∑ ) j∈DCS o∈ODS
(
∑ Rc(j 3 ) 1 − δ (jo3 ) 

) (10.17)
+ d3− − d3+ = 1.

Designing Resilient Global Supply Chain Networks 257

f4 = Max. strategic factors

∑ ∑ Sth(1)δ (ho1) + ∑ ∑ Sti( 2)δ (io2) + ∑ ∑ St(j 3 )δ (jo3 )

h∈MFS o∈ODS i∈CFS o∈ODS j∈DCS o∈ODS
+ d4− − d4+ = 1. (10.18)
Tµ(4)

f5 = Max. demand fulfillment at markets

 2) 
∑  ∑ ∑ ∑ x(hkptmo
2)
+ ∑ ( 2)
∑ y ikpm o + ∑ z(jkpm 
 h∈MKSMh t∈CTSMh o∈ODS
m∈PDS i∈CFSK k o∈ODS j∈DCSK k

Dkpm
(10.19)
+ d5−kpm − d5+kpm = 1 ∀m ∈ PDS, k ∈ MKS, p ∈ PRSK k .

10.4.5  Set of Constraints

The set of constraints containing parameters or variables with index m are
now defined for each time period m. In addition, the constraints including
index o are presented in one of two forms depending on the constraint type.
On one side, in the capacity and binary constraints the variables are added
based on index o, and on the other the balance constraints are defined for
each alternative o, except the binary equation for the distribution to custom-
ers’ echelon. Modifications to the constraints presented in the base model
(Portillo 2016) are as follows:
Distribution capacity utilization

 ( 2) 
∑  ∑ z(jlpm1) (3)
DFjlp + ∑ ∑ z(jkpm
2)
DFjkp  ≤ o∈∑ SC jmoδ (jo3 ) ∀m, j ∈ DCS. (10.20)
 p∈PRSDl
l∈DCSD j k ∈MKSD j p∈PRSK k ODS

Balance at 2-stage product manufacturing

∑ ∑ ∑ ∑ DFip(1) x(hiptmo
1)
=
h∈MFSPp i∈CFSCi t∈CTSMh o∈ODS

∑ ∑ ∑ (1)
∑ y ijptmo + ∑ ∑ ∑ ( 2)
∑ y ikptmo , ∀m, p ∈ PRSS. (10.21)
i∈CFSFp j∈DCSCi t∈CTSCi o∈ODS i∈CFSPp k ∈MKSCp t∈CTSC j o∈ODS

Balance at 2-stage product conversion

 
DFip(1)  ∑ ∑ ∑ x(1) − ∑ ∑ ∑ y (1) − ∑ ∑ ∑ y ( 2) = 0
 h∈MFSCi t∈CTSMh o hiptmo  j∈DCSCi t∈CTSCi o∈ODS ijptmo k∈MKSCi t∈CTSCi o∈ODS ikptmo

∀m, i ∈ CFS, p ∈ PRSCi .

(10.22)

Balance at single-stage production

∑ ∑ ∑ (1)
∑ y ijptmo = ∑ ∑ z(jlpm
2)
, ∀m, p ∉ PRSS. (10.23)
i∈CFSPp j∈DCSCi t∈CTSCi o∈ODS j∈DCSPp l∈DCSD j

258 Big Data Analytics Using Multiple Criteria Decision-Making Models

Balance at distribution

∑ ∑ (1)
∑ y ijptmo + ∑ (1)
zljpm = ∑ (1)
zljpm + ∑ z(jkpm
2)
, ∀m, j , p ∈ PRSDj (10.24)
i∈CFSD j t∈CTSCi o∈ODS l∈DCSA j l∈DCSD j k ∈MKSD j

∑ (1)
zljpm = ∑ z(jlpm
1)
+ ∑ z(jkpm
2)
, ∀m, j ∈ DCSTOMARKETS, p ∈ PRSDj . (10.25)
l∈DCSA j l∈DCSD j k ∈MKSD j

Manufacturing capacity

    
∑  MPThptmo  ∑ x(hiptmo
1)
∑ MPThptmo  ∑ x(hkptmo
 + p∈PRST
2) ( 1)
  ≤ ∑o MChtmo γ hto

p∈PRSTt  i∈CFSMh t  k∈MKSMh 
∀m, o, h ∈ MFS, t ∈ CTSMh . (10.26)

Conversion capacity

  (1)   ( 2) 
∑ CPTiptmo  ∑ y ijptmo
 p∈PRST  + p∈PRST
∑ CPTiptmo  ∑ y ikptmo   ≤ o∈∑ CCitmo γ (ito2)
 t  j∈DCSCi t  k∈MKSCi  ODS

(10.27)
∀m, o, i ∈ CFS, t ∈ CTSCi .

Manufacturing and conversion binary

Equations 10.28 and 10.29 make sure that at least one production line t is
active in order to open a manufacturing facility h or converting facility i.
Note that the binary variables are indexed based on the opening time period
o, indicating that there may be multiple options to open production lines
as well as manufacturing or converting facilities at different time periods.
Consequently, the constraints below apply to each of the production network
design options given by index o.

γ (hto
1)
≤ δ (ho1) , ∀o, h ∈ MFS, t ∈ CTSMh (10.28)

γ (ito2) ≤ δ (io2) , ∀o, i ∈ CFS, t ∈ CTSCi . (10.29)

Distribution binary
Similarly, Equation 10.30 states a shipping capacity constraint in cubic
meters for the flows out of the distribution facility j considering if it is open
or closed. The capacity of a distribution facility may vary at different time
periods m. As well, the formulation provides for the ability to evaluate open-
ing a distribution center at different time periods, considering that its capac-
ity may differ as well depending on the opening period o.

( 3 ) (1) ( 2) ( 2)
∑ ∑ DFjlp z jlpm + ∑ ∑ DFjkp z jkpm − ∑ SC jmoδ (jo3 ) ≤ 0
l∈DCSD j p∈PRSDl k ∈MKSD j p∈PRSK k o∈ODS

∀m, j ∈ DCS. (10.30)

Designing Resilient Global Supply Chain Networks 259

Production extensions binary

For each of the potential time periods when a production line can start
operations, Equations 10.31 and 10.32 enforce that a new production line t′
proposed to be installed at a production facility as a capacity expansion is
only activated when all the existing machines of similar technology t are
operating. If at least one production line of type t is idle at a given plant,
then no capacity expansions can be done. In other words, a given capacity
extension cannot open at any given period of time if the current installed
equipment has idle capacity. Note that these constraints can be relaxed if the
analyst wants to evaluate the replacement of equipment; in this case, new
production lines could be opened even if it implies that existing equipment
stays idle.

γ (ht1)′o ≤ γ (hto
1)
, ∀( h, o, t , t ′ ) ∈ MFSTThtt′ (10.31)

γ (it1′)o ≤ γ (ito1) , ∀(i , o, t , t ′ ) ∈ CFSTTitt′ . (10.32)

Plant/production line constraint binary

∑ γ (ito2) ≥ δ (io2) ∀ i, o (10.33)

t

∑ γ (hto
1)
≥ δ (ho1) ∀ h, o. (10.34)
t

In a similar way, above equations make sure that at least one production
line t is active in order to open a manufacturing facility h or converting
facility i, applying this restriction to every possible opening period o for a
­production line or facility.
Time period binary

∑ δ (ho1) ≤ 1 ∀h (10.35)
o

∑ δ (io2) ≤ 1 ∀i (10.36)
o

∑ δ (jo3 ) ≤ 1 ∀j (10.37)
o

∑ γ (hto
1)
≤1 ∀h (10.38)
o

∑ γ (ito2) ≤ 1 ∀ i. (10.39)
o

The above equations are included to constrain that a given facility or pro-
duction line that has different start-up options opens at only one of the time
periods (i.e., can open only once).
260 Big Data Analytics Using Multiple Criteria Decision-Making Models

All decision variables are continuous and nonnegative, except for δ(1)ho,
γ (hto
1)
, δ(2)io, γ (ito2) , and δ(3)jo that are binary.

10.4.6  Objective Function Considering Currency Exchange Rates

When designing supply chain structures to support commercial activities
within international environments, country-specific monetary factors such
as currency exchange rates may have an impact on the financial performance
of multinational firms. Perhaps a particular business unit of a multinational
firm may collect revenues in one currency and have costs in different or sev-
eral currencies. The impact of currency exchange rates fluctuation on sales
price, local operating costs, local and imported raw material costs, among
others may be decisive to choose where to locate production and distribu-
tion facilities as well as to determine sourcing strategies. This has significant
implications for companies with international supply chains and markets.
Recognizing and considering these implications in strategy making activi-
ties now could mean the difference between success and failure in the future.
Therefore, it is considered imperative to include the effect of exchange rate in
this optimization model.
With this purpose, the parameter Ecm is added to the model to represent
the currency exchange rate for country c at period m. This variable is defined
to include a deterministic best estimate of the future behavior of the cur-
rency exchange rate for a given country. Index c = (1, …, nc) relates to the
set of countries (COS). This new set is broken down in subsets correspond-
ing to manufacturing facilities (COSh), converting plants (COSi), DCs (COSj),
and markets (COSk), establishing the relationship between these nodes and
the country where they are located. This parameter is included in the gross
profit goal constraint to convert financial parameters defined in multiple
currencies from different countries to a single currency of comparison. The
next chapter will deal with the variability of this parameter. For now, the
modified deterministic goal constraint is presented next.

f1 = Max. gross profit

1 
∑ ∑ ∑ ∑ ∑ z(jkpm
2)
Pkpm Ecm
T µ(1)  c∈COSk k∈MKS p∈PRSKk j∈DCSKk m∈PDS
+ ∑ ∑ ∑ ∑ ∑ ∑ ( 2)
∑ y ikptmo Pkpm Ecm (10.40)
c∈COSk k ∈MKS p∈PRSK k i∈CFSK k t∈CTSCi m∈PDS o∈ODS
( 2)
+ ∑ ∑ ∑ ∑ ∑ ∑ ∑ y P E
hkptmo kpm cm
c∈COSk k ∈MKS p∈PRSK k h∈MFSK k t∈CTSMh m∈PDS o∈ODS

− ∑ ∑ ∑ ∑ Ecm FMhmoδ (ho1)

c∈COSh h∈MFS m∈PDS o∈ODS

− ∑ ∑ ∑ ∑ Ecm FCimoδ (io2) (10.41)

c∈COSi i∈CFS m∈PDS o∈ODS
(3)
− ∑ ∑ ∑ ∑ Ecm FD δ jmo jo
c∈COSj j∈DCS m∈PDS o∈ODS

Designing Resilient Global Supply Chain Networks 261

− ∑ ∑ ∑ ∑ ∑ Ecm FMMhtm γ (hto

1)
− ∑ ∑ ∑ ∑ ∑ Ecm FCMitm γ (ito2) (10.42)
o∈COSh h∈MFS t∈CTSMh m∈PDS o∈ODS c∈COSi i∈CFS t∈CTSCi m o

 
− ∑ ∑ ∑ ∑ ∑ Ecm MRChpm  ∑ ∑ x(1) + ∑ ∑ x( 2)
 i∈CFSMh t∈CTSMh hiptmo k∈MKSMh t∈CTSMh hkptmo 
(10.43)
o∈COSh m∈PDS h∈MFS p∈PRSMh o∈ODS

 
− ∑ ∑ ∑ ∑ ∑ ∑ Ecm MMCtmo MPTptmo  ∑ x(hiptmo
 i∈CFSMh
1)
+ ∑ x(hkptmo
2)
 (10.44)
o∈COSh m∈PDS o∈ODS h∈MFS t∈CTSMh p∈PRSTt k ∈MKSMh

 
− ∑ ∑ ∑ ∑ EcmCRCipm ∑
c∈COSj i∈CFS p∈PRSCi m∈PDS t∈CTSCi o∈ODS
∑  ∑ (1)
∑ y ijptmo
 j∈DCSCi m∈PDS
+ ∑ ( 2)
∑ y ikptmo
k ∈MKSCi m∈PDS 
(10.45)

 
− ∑ ∑ ∑ ∑ ∑
c∈COSi m∈PDS o∈ODS i∈CFS t∈CTSCi p∈PRSTt
(1)
∑ EcmCMCtmoCPTptmo  ∑ y ijptmo
 j∈DCSCi
( 2)
+ ∑ y ikptmo
k ∈MKSCi
 (10.46)

 ( 2) 
− ∑ ∑ ∑ EcmVOC j  ∑ ∑ z(jlpm
1) ( 2))
DFjkp + ∑ ∑ z(jkpm
2)
DFjkp (10.47)
c∈COSj m∈PDS j∈DCS  l∈DCSC j p∈PRSDl k ∈MKSD j p∈PRSK k 

( 2)
− ∑ ∑ ∑ ∑ ∑ ∑ ∑ EcmThipm x(hiptmo
1)
c∈COSh m∈PDS h∈MFS i∈CFSMh p∈PRSCi t∈CTSCh o∈ODS
( 4 ) ( 1)
− ∑ ∑ ∑ ∑ ∑ ∑ ∑ EcmTijpm y ijptmo
c∈COSi m∈PDS i∈CFS j∈DCSCi p∈PRSD j t∈CTSCi o∈ODS (10.48)
( 6 ) (1)
− ∑ ∑ ∑ ∑ ∑ E T z
cm jlpm jlpm
c∈COSj m∈PDS j∈DCS l∈DCSD j p∈PRSDl

(3)
− ∑ ∑ ∑ ∑ ∑ ∑ ∑ EcmThkpm x(hkptmo
2)
c∈COSh m∈PDS h∈MFS k ∈MKSMh p∈PRSK k t∈CTSCh o∈ODS
( 5) ( 2)
− ∑ ∑ ∑ ∑ ∑ ∑ ∑ EcmTikpm y ikptmo (10.49)
c∈COSi m∈PDS i∈CFS k ∈MKSCi p∈PRSK k t∈CTSCi o∈ODS
(7 ) ( 2)
− ∑ ∑ ∑ ∑ ∑ E T z
cm jkpm jkpm
c∈COSj m∈PDS j∈DCS k ∈MKSD j p∈PRSK k

( 2)
− ∑ ∑ ∑ ∑ ∑ ∑ ∑ Ecm MUChpm I hipm x(hiptmo
1)
c∈COSh m∈PDS h∈MFS p∈PRSMh i∈CFSMh t∈CTSMh o∈ODS
( 4 ) (1)
− ∑ ∑ ∑ ∑ ∑ ∑ ∑ EcmCUCipm I ijpm y ijptmo (10.50)
c∈COSi m∈PDS i∈CFS p∈PRSCi j∈DCSDi t∈CTSCi o∈ODS

− ∑ ∑ ∑ ∑ ∑ Ecm DUC jpm I (jlpm

6 ) (1)
z jlpm
c∈COSj m∈PDS j∈DCS k ∈PRSD j l∈DCSD j

( 2) ( 2) (1)
− ∑ ∑ ∑ ∑ ∑ ∑ ∑ EcmTI hip m I hipm x hiptmo
c∈COSh m∈PDS h∈MFS i∈CFSMh p∈PRSCi t∈CTSMh o∈ODS
( 4 ) ( 4 ) (1) (10.51)
− ∑ ∑ ∑ ∑ ∑ ∑ ∑ EcmTI ijpm I ijpm y ijptmo
c∈COSi m∈PDS i∈CFS j∈DCSCi p∈PRSD j t∈CTSCi o∈ODS

6 ) ( 6 ) ( 1) 
− ∑ ∑ ∑ ∑ ∑ EcmTI (jlpm I jlpm z jlpm  (10.52)
c∈COSj m∈PDS j∈DCS l∈DCSD j p∈PRSDl 

+ d1− − d1+ = 1. (10.53)

262 Big Data Analytics Using Multiple Criteria Decision-Making Models

10.5  Data Collection

Before conducting the analysis, significant effort was required to collect data
of both types, historical and planned. Because of the large scale of the analy-
ses, it is important to highlight the effort dedicated to build the databases
for the different supply chain network scenarios. Each scenario analysis
required extracting thousands of market-, finance-, and operations-related
data records from the firm’s business systems as well as obtaining informa-
tion from external sources. At least three employees worked full time directly
gathering or requesting information, as well as organizing it appropriately
to run the optimization models. In addition, at least a dozen people were
contacted to provide information. The historical data was obtained from the
company’s ERP system at very low levels of granularity to have the flexibility
of aggregating it as required by the optimization model. Sales volume and
prices were obtained at stock keeping unit (SKU) and customer levels and
then aggregated to the product and customer zone level. Production volume
and unit costs were extracted at the SKU and production line levels and then
aggregated to a product for each production line. Plant cost information was
generated at the financial account level by each cost center associated with
each production line and then classified as plant overheads, production line
fixed and variable costs, and raw material costs. Transportation and cross-
sourcing cost was defined for each arc of the network at the SKU level and
then aggregated to products. Also, mass conversion factors were determined
to handle different volume units of measure used at the different echelons
of the supply chain. Projected information was obtained from the company’s
most recent business plans, including forecasted demand volume, price, and
cost projections. All this data was stored, analyzed, and processed using the
company’s internal web services enabled by a single worldwide data center
in North America. The mathematical formulation was coded in ILOG and
solved using a CPLEX solver. The optimization model consisted of approxi-
mately 7500 variables, from which 300 were binary and around 7000 were
constraints. Three persons were involved in coding the models in ILOG and
then using the CPLEX software to solve the problems. In general, optimal
solutions were obtained very efficiently, taking less than three minutes for
each scenario analysis to run.

10.6  Case Study

This section presents the case study results for a multiperiod scenario con-
sidering exchange rate. The objective is to expand capacity to strengthen the
position of a line of products in the marketplace for a particular country.
Designing Resilient Global Supply Chain Networks 263

In the last three years, the sales growth for this business has been 21% per
year. Recent customer demand forecasts indicate that with the currently
operating capacity plus a recent installation of one machine in a regional
plant outside the country, this business unit will need to import from outside
the region approximately 17% of the total volume over the next 5 years. The
incremental cost of import ranges from $12 to 22/Standard Unit versus pro-
ducing in the regional plant. The purpose is to determine the most efficient
supply chain network configuration and sourcing plans that best support
this expected growth. The analysis consists of comparing the status quo to
three new different scenarios as results from a preliminary analysis suggest
that the status quo scenario is not convenient for the long-term sustainability
of the business because of the supply disruption risk of importing 17% of the
demand over the next 5 years, besides incurring higher product costs.
The first scenario consists of adding one more machine in the regional
plant leveraging centralization efficiencies. The regional plant has four
machines currently running and a fifth machine that has been installed and
will soon begin operating. The proposal is to add a sixth machine. The sec-
ond scenario proposes installing a new machine at a local DC in a neighbor-
ing country. With this, relevant savings could be achieved from reductions
in freights and duties; however, additional fixed operating costs may become
necessary and operational efficiency in the local facility could be lower than
in the regional plant. Results of these scenarios illustrate how installing a
new machine, regardless of its location, significantly reduces the volume
of imports from outside the region, having a positive impact on the cost of
goods and reducing the overall risk of supply chain disruptions. On one
side, installing the new machine in the regional plant instead of a newly
opened plant locally increases the overall production volume due to a better
productivity learning curve in the regional plant, reducing the imports out
of the region. On the other hand, installing it in a new plant locally would
reduce the imports within the region but slightly increase the imports out-
side the region. Note that increasing the production capacity in the regional
plant implies that the local market would need to continue being sourced
with imported goods from the regional plant, which has an increased risk of
­supply chain disruptions than producing locally.
The third scenario suggests installing the new machine in the local facility
and transferring the recently installed machine from the regional plant. In
this case, it is expected that the incremental fixed operating costs would get
diluted in the higher throughput and that better operating efficiency could
be reached. For Scenario 3, it is assumed that the new machines will have
a faster learning curve and better productivity levels in the regional plant,
leveraging the know-how and product mix efficiency because the high num-
ber of machines would allow to reduce the number of changes required to
produce different product types. Installing the two machines in the local dis-
tribution center would still cost more than installing them in the currently
operating regional plant. This is because existing overheads in the regional
264 Big Data Analytics Using Multiple Criteria Decision-Making Models

plant get diluted in much higher production volume even if new machines
are not installed, compared with additional headcount and infrastructure
required in a new production local facility. Moreover, it is important to deter-
mine the best time to install the new machine. This can happen immediately
or the following year, balancing between machine utilization rate and the
associated incremental fixed operating costs. Any of these options would
allow sourcing within the region for all of the market requirements, so the
increased costs from imports outside the region could be reduced and sup-
ply chain risk of disruptions minimized. The objective is to maximize gross
profit and minimize risk. Results from the analysis indicate that install-
ing two machines locally would significantly reduce the imports out of the
region and almost eliminate the imports within the region, giving a very
high autonomy to the local market and therefore strongly minimizing the
risk of supply disruption.
In conclusion, the three new proposals generate savings compared with
the current supply chain network design requiring only between 2% and
4% of imports from outside the region. The scenario of adding one new
machine to the regional plant generates more than twice the savings
obtained from the other alternatives. In addition, it provides at least half
the invested capital payback than installing one or two assets locally. These
results are driven by the fact that when installing assets locally the reduc-
tions in transportation and duties do not set off higher production costs.
The unit cost of producing locally is higher due to lower productivity and
increased operating costs. Therefore, neither opening a new production
facility locally nor continuing importing from outside the region is benefi-
cial considering profits and risk. The best solution to support the expected
business growth locally is to continue with a centralized strategy in the
regional plant moving forward with the newly installed machine plus
­adding a new one.
As a complementary analysis indicated, delaying 1 year on the installa-
tion of the new machine in the regional plant affects its production volume.
Considering the high cost of importing goods out of the region, besides
the operational convenience of this solution, the results indicate that it is
­economically favorable putting in operation and starting amortizing the
new machine the first year.
The analysis above was conducted in dollars. A second round of a­ nalysis
was performed in local currency. Given the expected monetary deprecia-
tion trend in the regional plant host country and a more stable currency
in the local one, the results reinforced the previous recommendation. The
­currency exchange rate of the country, where the regional plant is located,
presents a steep depreciation rate forecast reaching almost 16% accumu-
lated in 4 years, while the exchange rate from the country proposed as
the location for the new plant presents only a 2% depreciation projection
within the same timeframe. The impact of such currency depreciation dif-
ference ­provides a competitive cost advantage for exports considering that a
Designing Resilient Global Supply Chain Networks 265

significant proportion of the product cost is associated with locally incurred

operating expenses as well as some key raw materials sourced from local
suppliers.
Following this exercise, similar analyses were performed using this model
in other regions of Latin America, North America, and China, with the
objective of optimizing manufacturing and distribution networks as a base
of business strategic plans over multiple years and considering the impact of
foreign currency exchange rates from multiple countries and their impact on
business results.

10.7  Conclusions and Future Research

This research provides relevant insights from modeling to implementation
designing resilient complex global supply chain systems, a contribution to
academia and industry research. In this work, diverse techniques classi-
fied as multicriteria mixed-integer programming and discrete multicriteria
selection methods were combined to develop a global supply chain design
model. It provides a resilient solution for supporting supply chain strategic
decisions related to footprint design and tactical plans within a highly com-
plex global environment. Focus was on optimizing conflicting criteria such
as financial, customer service, risk, and strategic measures. These models
were able to handle the complexity of the system (i.e., multiple products sold
in several markets in different countries that required dealing with differ-
ent currencies and commercial practices across multiple periods of time).
Decisions regarding the optimal location, relocation, and allocation of pro-
duction and distribution facilities as well as specific assets were efficiently
addressed.
In addition, decisions regarding production and distribution plans were
defined balancing demand and available capacity along multiple periods of
time. By incorporating multiple period, the models provided the ability to
consider the impact of changes through time in demand, machine produc-
tivity, facility capacity, prices, costs, and other parameters. On the strategic
side, the models allowed the evaluation of different alternatives for the mod-
ification of the supply chain design at different time periods. On the tactical
side, the models allowed a better evaluation of the balance between supply
and demand as well as more accurate cross-sourcing strategy definition.
Although the model presented above embodies a robust solution for sup-
ply chain optimization, extensions can focus on determining the efficient
frontier of conflicting criteria, treating the uncertainty associated with key
elements of the problem using more sophisticated methods such as stochas-
tic programming, and on expanding the scope on multiple supply chain
directions, from including external partners as customers and vendors to
266 Big Data Analytics Using Multiple Criteria Decision-Making Models

optimizing more operational aspects of supply chain as sales and opera-

tions planning, distribution networks optimization, as well as inventory
management.
Future research can be done on big data mining, particularly on distrib-
uted mining as data sources increase complexity and become more diverse
and decision makers need to act faster, finding user-friendly visualizations
as data gets bigger, and implementing hidden big data techniques as large
quantities of data get lost as it is untagged and unstructured, aiming to
take these analysis to have more higher frequency and granularity as data
becomes more diverse, larger, and of faster availability.

References
Arntzen, B. C., G. G. Brown, T. P. Harrison, and L. L. Trafton, 1995, Global supply
chain management at digital equipment corporation, Interfaces, 25, 69–93.
Ashayeri, J. and J. M. J. Rongen, 1987, Central distribution in Europe: A multi-criteria
approach to location selection, The International Journal of Logistics Management,
9(1), 97–106.
Brown, G. G., G. W. Graves, and M. D. Honczarenko, 1987, Design and operation of a
multi-commodity production/distribution system using primal goal decompo-
sition, Management Science, 33(11), 1469–1480.
Camm, J. D., T. E. Chorman, and F. A. Dull, 1997, Blending OR/MS judgment and GIS:
Restructuring P&G’s supply chain, Interfaces, 27, 128–142.
Friedman, T., 2005, The World is Flat: A Brief History of the Twenty-First Century,
Farrar, Straus & Giroux, New York.
Gaur, S. and A. R. Ravindran, 2006, A Bi-Criteria model for the inventory aggregation
problem under risk pooling, Computers & Industrial Engineering, 51, 482–501.
Geoffrion, A. and G. Graves, 1974, Multicommodity distribution system design by
benders decomposition, Management Science, 29, 822–844.
Kambatla, K., G. Kollias, V. Kumar, and A. Grama, 2014, Trends in big data analytics,
Journal of Parallel and Distributed Computing, 74(7), 2561–2573.
Korpela, J., K. Kyläheiko, and A. Lehmusvaara, 2002, An analytic approach to pro-
duction capacity allocation and supply chain design, International Journal of
Production Economics, 78, 187–195.
Martin, C. H., D. C. Dent, and J. C. Eckhart, 1993, Integrated production, distribution,
and inventory planning at Libbey–Owens–Ford, Interfaces, 23, 68–78.
Masud, A. M. and A. R. Ravindran, 2008, Multiple criteria decision making. In
Operations Research and Management Science Handbook, A. Ravi Ravindran
(Editor). CRC Press, Boca Raton, FL, 2008.
Melachrinoudis, E. and H. Min, 1999, The dynamic relocation and phase-out of a
hybrid, two-echelon plant/warehousing facility: A multiple objective approach,
European Journal of OR, 123, 1–15.
Melachrinoudis, E., H. Min, and A. Messac, 2000, The relocation of a manufacturing/
distribution facility from supply chain perspectives: A physical programming
approach, Multi-criteria Applications, 10, 15–39.
Designing Resilient Global Supply Chain Networks 267

Pooley, J., 1994, Integrated production and distribution planning at Ault foods,
Interfaces, 24, 113–121.
Portillo, R. C., 2009, Resilient Global Supply Chain Network Design. PhD dissertation,
Pennsylvania State University, University Park, PA, 2009.
Portillo, R. C., 2016, Designing resilient global supply chain networks. In Multiple
Criteria Decision Making in Supply Chain Management, A. Ravi Ravindran (Editor),
CRC Press, Boca Raton, FL, 2016.
Rao, U., A. Scheller, and S. Tayur, 2000, Development of a rapid-response supply
chain at caterpillar, Operations Research, 48(2), 189–204.
Ravindran, A. R., 2008, Operations Research and Management Science Handbook. CRC
Press, Taylor and Francis Group, Boca Raton, FL.
Ravindran, A. R., K. M. Ragsdell, and G. V. Reklaitis, 2006, Engineering Optimization:
Methods and Applications. Wiley, Hoboken, NJ.
Ravindran, A. R. and D. P. Warsing, Jr., 2013, Supply Chain Engineering: Models and
Applications. CRC Press, Boca Raton, FL.
Sousa, R., N. Shah, and L. G. Papageorgiou, 2008, Supply chain design and multi-
level planning—An industrial case, Computers and Chemical Engineering, 32,
2643–2663.
Tsiakis, P. and L. G. Papageorgiou, 2008, Optimal production allocation and distribu-
tion supply chain networks, International Journal of Production Economics, 111,
468–483.
Van Roy, T. J. and L. A. Wolsey, 1985, Valid inequalities and separation for uncapaci-
tated fixed charge networks, Operations Research, 4, 105–112.
Wadhwa, V. and A. R. Ravindran, 2007, Vendor selection in outsourcing, Computers
& OR, 34, 3725–3737.
11
MCDM-Based Modeling Framework for
Continuous Performance Evaluation of
Employees to Offer Reward and Recognition

S. S. Sreejith and Muthu Mathirajan

CONTENTS
11.1 Introduction................................................................................................. 269
11.2 Identification of Multiple Variables/Multiple Criteria for
CPEE to Offer R&R..................................................................................... 271
11.3 Performance Evaluation of Employees Using MCDM Methods.......... 282
11.4 Development of an MCDM-Based Modeling Framework for
CPEE to Offer R&R..................................................................................... 288
11.5 Validation of the Proposed MCDM Modeling Framework for
CPEE to Offer R&R..................................................................................... 291
11.6 Managerial Implications............................................................................ 295
11.7 Conclusion................................................................................................... 298
References.............................................................................................................. 299

11.1 Introduction
Periodic performance evaluations of employees are required in order to
measure the contribution level of employee toward organizational objectives
and also to calibrate individual performance. Performance evaluations of
employees in organizations are customarily conducted via a formal perfor-
mance appraisal system (PAS) (Gruenfeld and Weissenberg 1966, Rosen and
Abraham 1966). In general, the PAS is administered on an annual basis to
evaluate the performance of employee over a period of past 1 year (Bassett
and Meyer 1968). Such annual PAS has been subject to criticism from the
industry citing long frequency, among other drawbacks (Henderson 1980,
Ilgen et al. 1981). Some industries such as Information Technology (IT) have
tried to shorten this frequency by making it a half-yearly PAS. Nevertheless,
the resentment with the existing PAS persists. Some of the most common
issues with the existing PAS on annual or half-yearly basis are: recency error,

269
270 Big Data Analytics Using Multiple Criteria Decision-Making Models

bias, favoritism, subjectivity, selective memory, etc. (Bowman 1999, Facteau

and Craig 2001, Gray 2002). In order to minimize some of these issues, it has
been repetitively recommended that the performance evaluation of employ-
ees should be conducted in a continuous manner (Garafano and Salas 2005,
Kondrasuk 2011).
Literature endorses the necessity of having a continuous performance
evaluation of employees (CPEE), rather than periodic appraisals (Boice and
Kleiner 1997, Schraeder et  al. 2007, Palaiologos et  al. 2011, Sreejith 2015).
Despite the theoretical compulsion, the idea of CPEE has not known to be
translated into action in organizations. One reason for such inaction could
be due to a lack of model, which details the intricacies of having a system to
successfully carry out CPEE. As CPEE is a frequent phenomenon (assumed
to be weekly or fortnightly in this study), there has to be periodic tangible
outputs so as to increase the acceptability of the CPEE process. One such
output can be to identify the best-performing employee and offer reward
and recognition (R&R).
R&R is regarded as a motivational method to indicate appreciation and to
boost better performance (Cacioppe 1999). It can be considered that reward
is a materialistic part of the appreciation while recognition is the spiritual
part. Examples of reward could be certificates, mementos, souvenirs, etc.,
while recognition could be public appreciation, notifying the contribution of
employees through organization-wide newsletter or a simple pat on the back
of the employee. Although in principle, “reward” and “recognition” indicate
different concepts (Hansen et al. 2002), in this chapter they are considered
in unison and mutually inclusive and is referred to as R&R. In addition, in
this study, the means of offering the R&R is assumed to be varied across
organization.
It is obvious that, given a chance, every employee expects to get R&R for dis-
playing superior performance (Brun and Dugas 2008). Moreover, it has been
observed that performance of employee increases when deserving employees
are provided proper and timely R&R (Bradler et al. 2012). In addition, in order
to be effective, R&R should be offered to employees immediately after notic-
ing the superior performance. However, there is no known formal system/
process/framework whereby deserving employees receive proper and timely
R&R. To offer proper and timely R&R in a continuous manner, employees
need to be evaluated on a continuous basis and ranked according to their
performance. The CPEE process could frequently generate a rank of well-
performing employees, who can be offered R&R. Hence, there needs to be
a CPEE system/framework to identify the eligible employee(s) to offer R&R.
One such framework for CPEE to offer R&R is attempted in this study.
The proposed framework for CPEE to offer R&R is anticipated to work
much better in a dynamic industry with younger workforce, such as IT indus-
try. This is because IT industry is comprised of younger workforce (Arora
et al. 2001), who anticipate immediate feedback regarding their performance
(Smola and Sutton 2002), and due to the alarming rate of voluntary attrition
MCDM-Based Modeling Framework 271

(Bhatnagar 2006, Economic Times 2015). It can be presumed that if the pro-
posed framework for CPEE to offer R&R is properly implemented, it could
address the issue of voluntary attrition to some extend and offer better moti-
vation for employees to perform better. This study focuses on developing a
framework for CPEE system to offer R&R for the IT industry where employ-
ees are represented by software engineers (SEs) and project managers (PMs).
Based on the analysis of the literature and based on the observation in
the industries, the performance evaluation of employees is assessed based
on a combination of multiple numerous variables (criteria). Accordingly, the
CPEE to offer R&R should be based on multiple variables/multiple criteria.
From the literature review on CPEE, particularly for SEs, it appears that there
are no variables/criteria that focus on CPEE to offer R&R. Due to this, mul-
tiple variables/multiple criteria suggested in the literature for the traditional
PAS (among all industries) are collected to understand the types of multiple
criteria being considered in PAS. Based on the understanding from these,
subsequently suitable explorative and descriptive research methods have
been carried out to identify the required set of variables/criteria for CPEE
to offer R&R, particularly for SEs. In order to effectively utilize these identi-
fied multiple variables/multiple criteria for evaluating the performance of
employees, a suitable multicriteria decision-making (MCDM) method is nec-
essary. Accordingly in this study, two MCDM methods, analytical hierarchy
process (AHP) and modified Pugh matrix method (MPMM), are considered
and appropriately implemented to develop the proposed framework for
CPEE to offer R&R.
The rest of the chapter is organized as follows: Identification of a compre-
hensive list of variables/criteria from (a) literature review, (b) exploratory,
and (c) descriptive research methods and determining the main criteria for
CPEE to offer R&R are detailed in Section 11.2. The literature focusing on
the analysis of the performance of employees using MCDM methods are
highlighted in Section 11.3. The development of MCDM-based modeling
framework for CPEE to offer R&R is elaborated in Section 11.4. A suitable
numerical example is developed to demonstrate the workability of the pro-
posed framework for CPEE to offer R&R in Section 11.5. Section 11.6 dis-
cusses the managerial implications of the proposed MCDM-based modeling
framework for CPEE to offer R&R. The study concludes by highlighting the
contributions, limitation, and further research of the study in Section 11.7.

11.2 Identification of Multiple Variables/Multiple

Criteria for CPEE to Offer R&R
In general, performance evaluation of employees using PAS commences
by setting forth a set of defined variables/criteria. As there is no known
272 Big Data Analytics Using Multiple Criteria Decision-Making Models

literature specifying the variables for CPEE to offer R&R, the literature deal-
ing with performance evaluation of employees using the traditional PAS are
reviewed. Based on this review, a list of 44 variables is identified from the
literature and is shown in column 3 of Table 11.1. In addition, a set of six
variables are intuitively proposed by the researcher for the purpose of per-
formance evaluation of employees (Sreejith 2016) and these are also listed in
column 4 of Table 11.1.
The first 51 variables listed (based on columns 3 and 4) in Table 11.1 are not
exclusively identified for IT industry, so it cannot be confidently assumed
that all these 51 variables are relevant for performance evaluation of SEs in
the IT industry. In addition, in general, all the listed 44 variables in Table 11.1
(column 3) pertaining to the traditional PAS cannot be blindly assumed to
hold good specifically for continuous evaluation of employee. Hence, it is
necessary to conduct exploratory and/or descriptive research methods to
identify the required variables/criteria for CPEE to offer R&R, directly from
the employees of IT industry. Accordingly, a Caselet approach is carried out
and seven SEs are interviewed with suitably prepared Caselet schedule to
identify a set of variables/criteria and individual Caselets are developed.
Due to the brevity of the chapter, the Caselets developed are not presented
in this chapter. From the analysis of the seven Caselets developed, 27 unique
variables/criteria are identified for CPEE of SEs.
As the inference and/or finding obtained from Caselet approach cannot
be generalized, another phase of exploratory research based on semistruc-
tured interviews is conducted among 58 SEs by developing an appropriate
interview schedule using the finding from the Caselet approach. At the end
of this phase, 35 variables/criteria are identified based on the opinion from
SEs for CPEE to offer R&R and they are presented in Table 11.1 (5th column).
The list of variables/criteria identified based on 58 SEs for CPEE to offer
R&R is required to be cross-verified from the administrative employee’s (i.e.,
PM’s) perspective for the purpose of offering R&R. In order to cross-check
and confirm the variables/criteria identified from SEs to offer R&R, 31 PMs
are interviewed by developing a suitable interview schedule based on the
35 variables identified from SEs’ perspectives. At the end of this stage, 29
variables are confirmed from the list of variables identified from SEs’ per-
spective by the PMs. In addition, the PMs added four new variables/criteria
from their perspectives to offer R&R. Accordingly, the list of 33 variables (i.e.,
29 variables confirmed by PMs from the list of variables identified based on
SEs’ perspective and four variables exclusively included based on PM’s per-
spective) is considered based on PMs’ perspective and the same is presented
in Table 11.1 (6th column).
By comparing the list of variables presented in column 2 to column 6
of the Table 11.1, in this study, 33 variables, which are exactly accepted by
PMs, are considered to offer R&R. The 33 variables considered for offering
R&R are listed with description in Table 11.2. Accordingly, in this study, the
CPEE to offer R&R is considered as a function of all the 33 variables listed
TABLE 11.1
A Consolidated List of Variables Identified for CPEE to Offer R&R
Variables/Criteria To Be Considered for CPEE to
Offer R&R
Final Set of
Variables Identified from Software Project Variables/Criteria
Name of the Variable/ Considered Researcher’s Engineers’ Managers’ Considered for
No. Criterion from Literature Perspective Perspective Perspective CPEE to Offer R&R
1 Age ✓ – ✓ ✓ ✓
2 Gender ✓ – – – –
3 Marital status ✓ – – – –
4 Education ✓ – ✓ ✓ ✓
5 University/institution ✓ – ✓ ✓ ✓
MCDM-Based Modeling Framework

6 Experience ✓ – ✓ ✓ ✓
7 Tenure ✓ – ✓ ✓ ✓
8 Personality ✓ – – – –
9 Parent’s education ✓ – ✓ – –
10 Child status ✓ – – – –
11 Quantity of work ✓ – – – –
12 Timeline adherence ✓ – ✓ ✓ ✓
13 Customer interaction ✓ – ✓ ✓ ✓
14 Target achievement ✓ – – – –
15 Timely reporting ✓ – ✓ ✓ ✓
16 Documentation ✓ – ✓ ✓ ✓
17 Reviewing ✓ – ✓ ✓ ✓
18 Analytical ability ✓ – – ✓ ✓
19 Work planning ✓ – – – –

(Continued)
273
 274
TABLE 11.1 (Continued)
A Consolidated List of Variables Identified for CPEE to Offer R&R
Variables/Criteria To Be Considered for CPEE to
Offer R&R
Final Set of
Variables Identified from Software Project Variables/Criteria
Name of the Variable/ Considered Researcher’s Engineers’ Managers’ Considered for
No. Criterion from Literature Perspective Perspective Perspective CPEE to Offer R&R
20 Creativity ✓ – ✓ ✓ ✓
21 Communication skills ✓ – ✓ ✓ ✓
22 Knowledge updation ✓ – ✓ ✓ ✓
23 Initiative ✓ – ✓ ✓ ✓
24 Understanding big picture ✓ – ✓ ✓ ✓
25 Additional responsibilities ✓ – ✓ ✓ ✓
26 Presentation skills ✓ – ✓ ✓ ✓
27 Negotiation skills ✓ – – ✓ ✓
28 Ideas/suggestions ✓ – ✓ ✓ ✓
29 Innovation ✓ – – – –
30 Patents/publications ✓ – – – –
31 Self-learning ✓ – ✓ ✓ ✓
32 Leadership ✓ – ✓ ✓ ✓
33 Team cooperation ✓ – ✓ ✓ ✓
34 Punctuality ✓ – ✓ ✓ ✓
35 Mentoring ✓ – – ✓ ✓
36 Perseverance ✓ – – – –
37 Humor sense ✓ – – – –
38 Critical thinking ✓ – – – –
(Continued)
Big Data Analytics Using Multiple Criteria Decision-Making Models
TABLE 11.1 (Continued)
A Consolidated List of Variables Identified for CPEE to Offer R&R
Variables/Criteria To Be Considered for CPEE to
Offer R&R
Final Set of
Variables Identified from Software Project Variables/Criteria
Name of the Variable/ Considered Researcher’s Engineers’ Managers’ Considered for
No. Criterion from Literature Perspective Perspective Perspective CPEE to Offer R&R
39 Passion ✓ – – – –
40 Resilience ✓ – – – –
41 Commitment ✓ – ✓ ✓ ✓
42 Knowledge sharing ✓ – ✓ ✓ ✓
43 Proactiveness ✓ – – – –
MCDM-Based Modeling Framework

44 Code of conduct ✓ – ✓ ✓ ✓
45 Social volunteering – ✓ – – –
46 Agility – ✓ ✓ – –
47 Corporate social responsibility – ✓ – – –
48 Business domain knowledge – ✓ – – –
49 Multitasking – ✓ ✓ – –
50 Parent’s occupation – ✓ ✓ – –
51 Parent’s domicile – ✓ ✓ – –
52 Improving morale – – ✓ ✓ ✓
53 Quality of the job – – ✓ ✓ ✓
54 Process adherence – – ✓ ✓ ✓
55 Cost saving – – ✓ ✓ ✓
56 Cocurricular activities – – ✓ – –
57 Best practice – – – ✓ ✓
Total number of variables/criteria 44 07 35 33 33
275
276 Big Data Analytics Using Multiple Criteria Decision-Making Models

TABLE 11.2
Criteria Considered Based on the Perspectives of Both SEs and PMs for CPEE to
Offer R&R
Variables/Criteria
Considered for CPEE to
No. Offer R&R Description
1 Age Age of the software engineer (SE)
2 Education Completed highest education level of SE
3 University University/institution where the highest education was
obtained
4 Tenure Number of years spent in the current organization
5 Experience Total years of relevant experience in a similar profile
6 Quality of the job The output produced should conform to the
requirements/expectation
7 Timeline adherence The task should be completed on time or ahead of time
8 Process adherence Standard process for the job execution should be adhered
9 Customer interaction Ability to communicate with the client/customer to
convey and elicit required information
10 Documentation Creating, updating, and maintaining all documents
relating to the job
11 Reviewing Willingness and ability to review other documents,
codes, etc.
12 Timely reporting Reporting the progress or defects on time, so that
corrective action can be taken with minimal loss
13 Analytical ability Ability to think in a logical and analytical manner
14 Best practice Display best practice in process and quality
15 Communication skills Ability to convey ideas orally and verbally
16 Ideas and suggestions Recommend valid and implementable improvement
suggestions
17 Knowledge updation Keeping oneself updated with the knowledge in the
fields by certifications and other relevant qualifications
18 Negotiation skills Ability to confer with another person/department in the
team/organization in order to come to terms or reach
an agreement
19 Cost saving Demonstrate measures to save cost for the project and
organization
20 Presentation skills Ability to present the ideas to audience
21 Understanding big picture Ability to comprehend and understand the big picture of
the job assigned
22 Additional responsibilities Willingness to take up additional administrative
responsibilities that fall outside the normal scope of job
(like interviewing, auditing, etc.)
23 Creativity Ability to think and come up with some creative
solutions/process
24 Initiative Proactiveness, innovating processes, which influence the
project
(Continued)
MCDM-Based Modeling Framework 277

TABLE 11.2 (Continued)

Criteria Considered Based on the Perspectives of Both SEs and PMs for CPEE to
Offer R&R
Variables/Criteria
Considered for CPEE to
No. Offer R&R Description
25 Self-learning Willingness to learn required knowledge and skills
required for the job on one’s own, rather than depend
on formal training
26 Knowledge sharing Willingness to share knowledge with other team
members
27 Team cooperation Ability to be flexible, accommodative, and work in a
team
28 Mentoring Genuinely interested in and working toward developing
the skills and abilities of a junior member
29 Commitment The dedication, ownership, and accountability shown in
completing the task
30 Code of conduct Disciplined behavior; comply with the organization’s
code of conduct, and uphold organizational values
31 Punctuality Attending the workplace and meetings on time,
regularly
32 Leadership Ability to step up and direct the team to the right course
of action
33 Improving morale Take effort to improve the team spirit

in Table 11.2. Upon carefully scrutinizing the variables listed in Table 11.2,
it can be observed that the first five variables: age, education, university,
tenure, and experience can be grouped to the demographic characteristic of
the SEs and called DCSE in this study, while the remaining 28 variables
can be grouped to performance of SEs and they are called PSE in this study.
Though the final list of 33 variables identified based on (a) analyzing the
literature review, and (b) both perspectives’ of SEs and PMs using explor-
atory research methods, the importance of these variables for CPEE to offer
R&R is not ascertained based on large-scale opinion of SEs. Accordingly, this
is done from the opinion sought from a sample of 443 SEs from 12 different
IT organizations by carrying out descriptive research.
For conducting the descriptive research, a questionnaire is designed incor-
porating the 33 variables and the respondents are asked to rate the impor-
tance of each of these variables for CPEE to offer R&R. The responses are
sought on a 7-point Likert scale (7 being extremely important). Using the
data obtained from 443 respondents, a bivariate analysis is conducted using
t-test to identify the significance of five demographic variables which may
influence the CPEE to offer R&R. The details on the bivariate analysis with
t-test for the significance of demographic variables on the influence of CPEE
of SEs are shown in Table 11.3. From Table 11.3, it is clear that only three
demographic variables viz. education (E), university (U), and experience (X)
 278

TABLE 11.3
Bivariate Analysis and the Significant Demographic Variables in DCSE
Demographic Bifurcation Proactive Prompt Resourceful Responsible Diagnostic Dynamic
Variable Characteristics N Mean (SD) Mean (SD) Mean (SD) Mean (SD) Mean (SD) Mean (SD)
Age ≤23 years 218 22.9 (3.44) 23.763 (3.054) 24.645 (2.79) 34.503 (4.77) 28.813 (3.45) 18.504 (1.84)
>23 years 225 23.802 (2.8) 23.114 (2.89) 24.870 (2.66) 34.24 (4.68) 28.41 (3.38) 18.395 (1.72)
t = 1.59 t = 2.23* t = 0.87 t = 0.59 t = 1.24 t = 0.64
Education UG 311 22.174 (3.401) 23.647 (2.69) 24.902 (2.97) 34.017 (4.23) 28.11 (3.74) 18.222 (1.9)
PG 132 23.784 (2.78) 22.925 (3.07) 24.261 (2.4) 34.982 (4.95) 28.956 (3.56) 18.871 (1.82)
t = 4.79** t = 2.47* t = 2.19* t = 2.08* t = 2.21* t = 3.33**
University IIX 27 23.781 (2.6) 24.48 (3.15) 25.63 (3.12) 35.288 (5.02) 30.04 (4.18) 19.104 (1.88)
Non-IIX 416 23.047 (3.07) 23.072 (3.06) 24.477 (2.89) 33.47 (4.53) 28.205 (3.48) 18.377 (1.75)
t = 2.76* t = 2.31* t = 1.95 t = 2.00* t = 2.62** t = 2.08*
Tenure ≤3 years 262 23.115 (3.27) 23.440 (3.43) 24.512 (2.64) 34.63 (4.47) 28.422 (3.55) 18.51 (1.83)
>3 years 181 22.94 (3.0) 23.911 (2.88) 24.808 (2.87) 34.895 (4.86) 28.83 (3.72) 18.79 (1.81)
t = 0.57 t = 1.51 t = 1.12 t = 0.59 t = 1.16 t = 1.56
Experience ≤3 years 280 23.101 (3.1) 23.911 (2.84) 24.205 (2.44) 34.2.1 (4.56) 28.3 (3.29) 18.314 (1.86)
>3 years 163 23.718 (2.55) 23.27 (3.1) 24.73 (2.74) 35.589 (4.8) 29.41 (3.86) 18.86 (1.78)
t = 2.15* t = 2.21* t = 2.09* t = 3.03** t = 3.21** t = 3.03**
*significant at p < 0.05, **significant at p < 0.01.
Big Data Analytics Using Multiple Criteria Decision-Making Models
MCDM-Based Modeling Framework 279

have significant influence in all the six main criteria. Accordingly, these three
significant demographic variables (E, U, and X) are considered to constitute
DCSE.
The importance of the 28 variables (i.e., variables numbered from 6 to 33
in Table 11.2), grouped under PSE, is subjected to statistical analysis using
factor analysis and structural equation modeling (SEM). The factor analysis
is performed so as to identify the latent structure of the 28 variables, if any,
so as to group them under the same factor. The factor analysis yielded six
factors and they are named as: proactive, prompt, resourceful, responsible, diag-
nostic, and dynamic. The corresponding variables being manifested by each
of these six factors are shown in Table 11.4. These six factors are considered
as the main criteria against which SEs will be evaluated to ascertain their
performance.
Summarizing the analysis based on t-test and factor analysis, the three
demographic variables (i.e., DCSE) and the six factors/main criteria: proac-
tive, prompt, resourceful, responsible, diagnostic, and dynamic related to PSE for
SEs are considered to be important for CPEE to offer R&R. Based on these, a
framework is proposed for CPEE to offer R&R and is shown in Figure 11.1.
Furthermore, the following hypotheses are also proposed based on the
framework shown in Figure 11.1:

H1: DCSE is positively related to R&R.

H2: DCSE is positively related to each of the six main factors/criteria
in PSE.
H3: Each of the six main factors/criteria in PSE is positively related to
R&R.

In order to test the proposed hypotheses, a partial least square (PLS)-based

SEM is performed using R software (version 3.2.3). The path coefficients and
corresponding t-values (given in parentheses) are shown in Figure 11.2.
Upon analyzing the path diagram and the corresponding path coefficients
linking each of the six factors related to PSE with that of DCSE, it is observed
that DCSE is positively influencing all the main criteria in PSE, while it has
a negative path coefficient with R&R. It indicates that the hypothesis H1 is
rejected and hypothesis H2 is accepted. From this, it is inferred that (i) DCSE
has no direct effect on R&R, and (ii) DCSE has positive and direct influence
on PES. This is taken into consideration while developing the performance
evaluation model. Further, it is also noticed that PSE and each of the main
six factors/criteria in PSE have positive path coefficients relating to R&R. As
the path coefficients confirm the proposed hypothesis H3, it is concluded that
each of the six main factors/criteria in PSE is positively related to R&R.
The result of the H2 indicates that PSE is found to be influenced by DCSE.
Although DCSE is not directly related to R&R, DCSE is considered to have a
moderating effect on the relation between PSE and R&R from the point of view
 280

TABLE 11.4
Main Factor/Criterion-Wise List of Variables Grouped under PSE by Factor Analysis
Factor 5:
Factor 1: Proactive Factor 2: Prompt Factor 3: Resourceful Factor 4: Responsible Diagnostic Factor 6: Dynamic
Knowledge updation Timeline adherence Understanding big picture Additional responsibilities Quality of the job Customer interaction
Initiative Timely reporting Ideas and suggestions Knowledge sharing Documentation Communication
Self-learning Process adherence Creativity Commitment Analytical ability Negotiation
Leadership Punctuality Cost saving Teamwork Reviewing –
– – – Mentoring Presentation –
– – – Improving morale – –
Note: The variables “Best Practice” and “Code of Conduct” listed under PES were not seen in the above table because these two variables are not con-
sidered for factor analysis owing to low KMO value.
Big Data Analytics Using Multiple Criteria Decision-Making Models
MCDM-Based Modeling Framework 281

PSE

Proactive

Prompt

Resourceful
DCSE
(E, U, X) R&R
Responsible

Diagnostic

Dynamic

FIGURE 11.1
A proposed framework for CPEE to offer R&R.

PSE

Proactive
0.7947 0.1737
(3.279**) (4.809**)

Prompt 0.0747 –0.505

0.7248 (3.143**) (2.903**)
(4.22**)
0.0702
0.595 Resourceful (2.022*)
(2.914**)
DCSE 0.3839
0.2626 R&R
(E, U, X) (3.42*)
(3.119**)
Responsible 0.0932
(1.941*)
0.6547
(6.525**)

0.9035 Diagnostic
(4.196**) 0.162
(2.605**)

Dynamic

FIGURE 11.2
Proposed framework for CPEE to offer R&R with the path coefficients.
282 Big Data Analytics Using Multiple Criteria Decision-Making Models

PSE

Proactive

Moderates
Prompt
Influences
DCSE Resourceful R&R

Responsible

Diagnostic

Dynamic

FIGURE 11.3
Final version of the framework for CPEE to offer R&R.

of the performance evaluation of an employee. That is, the DCSE influences

the PSE as well as regulates the direct relation between PSE and R&R. Based
on the result of these hypotheses, the framework shown in Figure 11.1 is mod-
ified and the final framework for CPEE to offer R&R is shown in Figure 11.3.
From Figure 11.3, it is clear that multiple factors/criteria are needed for CPEE
to offer R&R. Particularly these multiple criteria have to be jointly evaluated
so as to determine the overall performance of the SE on these criteria.
As multiple criteria are involved for the CPEE, such evaluation can be
effectively done using MCDM method(s). This necessitates the need of iden-
tifying suitable MCDM method(s), considering the factors/criteria related to
PSE and the moderating factor related to DCSE, for CPEE to offer R&R. To
identify and implement suitable MCDM method(s) for CPEE to offer R&R,
various MCDM methods considered in the literature for performance evalu-
ation of employee in general are reviewed and the same is discussed in the
next section briefly.

11.3 Performance Evaluation of Employees

Using MCDM Methods
Although the MCDM methods have gained prominence since 1970s (Zionts
and Wallenius 1976), it appears that the application of such methods for
performance evaluation of employees gained momentum only during the
last decade. Though our aim in this study is not to give a complete and
exhaustive list of MCDM methods considered in the literature for perfor-
mance evaluation of employees, a sample of MCDM methods considered for
MCDM-Based Modeling Framework 283

performance evaluation of employees is presented to highlight the need of

MCDM method(s) for CPEE of SEs to offer R&R.
A popular MCDM method called analytical hierarchy process (AHP) is
introduced by Thomas L. Saaty during 1970s. The application of AHP for
performance evaluation of employees through a hypothetical example is
demonstrated in Saaty (1990). In the hypothetical example, the researcher
evaluated three employees and created the hierarchy for evaluation and
the relative priorities through pairwise comparison. The final scores are
obtained for the three employees.
A real-time application of AHP is described by Taylor III et  al. (1998) to
evaluate candidates for the position of Dean at Texas A&M University at
Kingsville. They have followed AHP with Saaty’s scale to develop priorities
and have compared a total of 33 employees who are considered for promo-
tion. They faced difficulty in making pairwise comparison and divided the
33 employees considered for evaluation into three groups and the suitable
candidate is chosen among a pool of prospective candidates. They have rec-
ommended that the number of employees being compared (i.e., the number
of alternatives) should be limited to a maximum of eight.
Colson (2000) utilizes the outranking MCDM methods: ELmination Et
Choice Translating Reality (ELECTRE) and Preference Ranking Organization
METHod for Enrichment Evaluations (PROMETHEE) to evaluate the judg-
ments made by decision makers on the selection of a suitable student to be
considered for a scientific award. ARGOS (Aid to the Ranking to be made
by a Group of decision makers using an Outranking Support), a software, is
developed in order to validate the group decision. A model for performance
evaluation of employees using AHP is proposed by Hemaida and Everett
(2003). They developed the model in order to rank the employees for the pur-
pose of salary allocation (i.e., the best-performing employee would receive a
greater increase in pay). They have claimed that the director is able to use the
proposed AHP to evaluate four employees and allocate the salary increase
according to their performance. The authors have asserted that AHP pro-
vides a sense of objectivity with significantly different conclusion from the
previous traditional evaluation methods used in the firm.
The application of AHP in a Malaysian service organization to evaluate the
employees is detailed by Islam and Rasad (2006). The supervisors are pro-
vided training on using AHP and they are instructed to evaluate employees
accordingly. Overall ranking of each employee is obtained, and they are able
to identify the best-performing employees during the appraisal period. They
have highlighted that the human resource department of the organization
testified that the evaluation using AHP has generated a more reliable set of
information. They recommended that performance evaluations should be an
ongoing process. However, they do not provide any recommendation about
how to use the AHP for CPEE.
A combination of two MCDM methods is attempted by Han and Ji (2009)
to address the problem of performance evaluation of employees. They have
284 Big Data Analytics Using Multiple Criteria Decision-Making Models

developed a theoretical synthetic evaluation method based on AHP and fuzzy

mathematics. By effectively combining the two techniques, the researchers
confirm that a comprehensive performance evaluation of employees can be
carried out for making the performance evaluation process more scientific
and systematic.
Data envelopment analysis (DEA) is another MCDM method, which can be
used to evaluate employee performance. Manoharan et al. (2009) detail one
such process of performance evaluation using DEA to evaluate 18 employees
of an Indian automotive organization. The factors considered by them for
evaluation are classified into input and output factors. Based on these factors,
they are able to rank the employees according to their performance. They
underline that well-defined procedures and guidelines incorporated in DEA
provide the performance appraisal and also reduce bias and subjectivity.
A fuzzy model for performance evaluation of employees using Choquet
Integral (CI) and Measuring Attractiveness by a Categorical Based Evaluation
TecHnique (MACBETH) is developed by Gurbuz (2010). This fuzzy model is
validated in a medical equipment manufacturing organization, and a final
ranking of five employees is produced. The author asserts that this model
addressed the drawback of AHP (i.e., the assumption of independence among
criteria/alternatives), by incorporating the interaction effect of criteria.
A decision support system is developed by Taufiq and Sugiharto (2011) using
AHP for employee evaluation in a healthcare organization in Indonesia. The
main objective of the performance evaluation is to select the employee for
promotion purpose. The authors claim that the management is able to iden-
tify the suitable employee for promotion. A popular ­outranking methods for
evaluating and ranking the employees using MCDM technique—ELECTRE—
is used to compare the performance of five employees in a telecom organiza-
tion, based on seven criteria (Afshari et al. 2010). They have used ELECTRE
method along with AHP technique to sort the employees according to their
performance and rank them.
Pugh matrix method (PMM) is another MCDM method, which satisfies
the pairwise comparison conditions of outranking methods put forward by
Bouyssou (1996). Nada (2010) assesses an organizational innovation process
and employee contribution toward organizational innovation. The author
evaluates the employee participation in organizational innovation process
using PMM. Specifically the author uses PMM method to outrank the ideas
suggested by the employees. Although this is not widely used as a com-
prehensive performance evaluation of employees method, PMM method is
found to be simple compared with other outranking methods (Scharge 2010).
A conceptual framework based on MCDM methods, fuzzy AHP (FAHP),
simple additive weighting (SAW), and technique for order of preference by simi-
larity to ideal solution (TOPSIS) is proposed by Ardabili (2011) to evaluate the
performance of bank staff. The author proposes that FAHP could be used for
pairwise comparison of the criteria. Further, SAW and TOPSIS are proposed for
evaluating the employees. The performance of vocational teachers is evaluated
MCDM-Based Modeling Framework 285

by Jati (2011) using MCDM methods: AHP and PROMETHEE II, particularly for
evaluating three school teachers based on 10 criteria. AHP is used to determine
the weights of the criteria using pairwise comparison. Further, the author uses
PROMETHEE II to calculate preference functions for all pairs of alternatives,
across the ten criteria considered in that study. The net outranking flow is cal-
culated and based on which the performance of the three teachers is ranked.
Wu et al. (2012) uses a fuzzy MCDM approach to evaluate the employee
performance of aircraft maintenance staff. They use FAHP and Vise
Kriterijumska Optimizacija I Kompromisno Resenje (VIKOR) to develop a fuzzy
MCDM method to evaluate 51 employees based on four performance dimen-
sions. They observe that the ranking result obtained from the fuzzy MCDM
method is differed from the traditional evaluation model. They also claim
that their model brought in better fairness to the evaluation process.
Using DEA, a study is conducted to evaluate the performance efficien-
cies of banking professionals in Slovenia (Zbranek 2013) considering salary,
working conditions, and benefits as input variables, and work motivation,
job satisfaction, and organizational commitment as output factors. They
observe that 12 employees among 60 are fully efficient and the remaining
48 are recommended for training. Islam et al. (2013) uses FAHP and TOPSIS
to evaluate the performance of banking professionals based on five crite-
ria. They use FAHP to calculate the weights of these criteria and subcriteria.
Further, they evaluate the employees using TOPSIS by assigning triangular
fuzzy numbers to the subcriteria. Based on these scores, the employee with
a score closest to the ideal solution and farthest from the nonideal solution is
identified for efficient one and the ones who need training.
Morte et al. (2013) conducted the performance evaluation of 31 road drivers
in a Portuguese logistic company. As the number of employees (alternatives)
is large, they do not use AHP so as to eliminate the pairwise comparison
process. Instead they use PROMETHEE and Methodologia Multicriterio para
Apoio a Selecao de Sistemas de Informacao (MMASSI) for evaluating two groups
consisting of two and three decision makers, respectively, on 11 criteria and
31 employees. These evaluations are then compared with a set of self-evalua-
tion made by the employee themselves to finally rank the employees accord-
ing to their performance.
Gurbuz and Albayrak (2014) consider both analytical network processing
(ANP) and CI to evaluate the employees of a pharmaceutical organization.
They highlight that the traditional methods for performance evaluation of
employees are highly subjective and do not provide a comprehensive infor-
mation about the employee performance. They propose a framework based
on ANP and CI based on three major criteria: sales-related performance
criteria, customer-related performance criteria, and relations-related perfor-
mance criteria. In addition, they consider organizational climate and demo-
graphic variables as moderating factor. Finally, they recommend that it is
important to understand the interdependencies of the criteria and alterna-
tives which influence the decision-making process.
286 Big Data Analytics Using Multiple Criteria Decision-Making Models

A comparative study of various performance evaluation methods such as

BARS, 360° method, assessment center, etc. as well as MCDM methods such
as AHP, FAHP, fuzzy TOPSIS, etc. is done by Shaout and Yousif (2014). It
appears that they have not provided an in-depth comparative analysis on the
MCDM techniques considered for comparison and they have not suggested
the conditions under which the suitable MCDM techniques can be adopted.
Recently, Sahoo et al. (2016) used DEA method to identify the top researchers
in various focus areas of management, based on six evaluation criteria and
have identified the top 10 subject matter experts in eight academic areas.
A summary on the performance evaluation of employees using MCDM
methods discussed in this section is presented in Table 11.5. It can be observed
from Table 11.5 that AHP and its variants (FAHP) are widely used across
industries for performance evaluation of employees. Further, from Table 11.5, it
can be observed that there are few other methods, commonly called outrank-
ing methods, used for performance evaluation of employees. In general, the
outranking methods build a preference function (called outranking relation)
among the alternatives which are evaluated on several criteria (Bouyssou 1996,
Bouyssou and Vincke 1997). This relation is built through a series of pairwise
comparisons of alternatives. This would enable the decision maker to conclude
that alternative A is better than B, if a majority of the evaluation criteria sup-
ports this conclusion (Bouyssou 2009). Some of the MCDM methods satisfying
this property of building an outranking relation, which are already mentioned
in Table 11.5, are: ELECTRE, PMM, TOPSIS, PROMETHEE, etc.
Inferring from Table 11.5, in general, the MCDM method(s) are mostly
used to evaluate employees for the purpose of selection of a new employee,
offering salary revision, offering promotion, etc. However, there is no known
literature for CPEE for choosing the best-performing employee to provide R&R,
particularly for SEs. In order to address this gap, an MCDM-based model
for CPEE to offer R&R is attempted in this chapter. As the proposed CPEE
system is a repetitive process, the choice of MCDM method(s) to offer R&R
should satisfy the following characteristics:

1. The choice of the MCDM method(s) should complete the perfor-

mance evaluation for offering R&R in the shortest possible time
(time-duration).
2. The proposed MCDM-based modeling framework for CPEE to
offer R&R should be easy to understand and easy to use by the PM
(easiness).
3. Over a period of time, one or a few criteria considered would become
irrelevant or not sufficient and due to this a few criteria have to be
removed or have to be added to from the list of variables/criteria
considered in the proposed MCDM-based modeling framework for
CPEE to offer R&R. Thus, the selected MCDM method(s) should be
flexible enough to add/remove criteria and/or alternatives (flexibility).
TABLE 11.5
Sample Applications of MCDM Methods in Performance Evaluation of Employees
No. MCDM Method Purpose Industry Reference
1 AHP Selection and promotion Service Taylor III et al. (1998)
2 AHP Salary revision Service Hemaida and Everett (2003)
3 AHP Performance appraisal Service Islam and Rasad (2006)
4 AHP Promotion Service Taufiq and Sugiharto (2011)
5 FAHP Performance appraisal Theoretical study Han and Ji (2009)
6 DEA Performance appraisal Manufacturing Manoharan et al. (2009)
MCDM-Based Modeling Framework

7 DEA Performance efficiency Service Zbranek (2013)

8 DEA Subject expert identification Service Sahoo et al. (2016)
9 ELECTRE Performance appraisal Manufacturing Afshari et al. (2010)
10 PMM Evaluation of ideas from employees Manufacturing Nada (2010)
11 PROMETHEE Performance prediction and selection Service Jati (2011)
12 FAHP and VIKOR Performance appraisal Service Wu et al. (2012)
13 FAHP, TOPSIS Performance appraisal Service Islam et al. (2013)
14 ELECTRE, PROMETHEE Student award Service Colson (2000)
15 CI and MACHBETH Performance appraisal Manufacturing Gurbuz (2010)
16 PROMETHEE and MMASSI Performance appraisal Service Morte et al. (2013)
17 ANP and CI Performance appraisal Manufacturing Gurbuz and Albayrak (2014)
18 FAHP, SAW, and TOPSIS Performance appraisal Service Ardabili (2011)
287
288 Big Data Analytics Using Multiple Criteria Decision-Making Models

Based on the expected characteristics of MCDM method(s), as stated

above, for CPEE to offer R&R and frequent and successful implementation of
AHP for performance evaluation of employees, in this study, we propose to
develop a modeling framework based on AHP and an MPMM for CPEE to
offer R&R. The MCDM, AHP, is considered for computing weights based on
the importance of the six main factors/criteria: proactive, prompt, resourceful,
responsible, diagnostic, and dynamic, and the moderating factor: DCSE consid-
ered for CPEE to offer R&R. Once the importance of the main factors/criteria
along with the moderating factors is obtained in the form of weights, each SE
(i.e., each alternative) is compared to select the best-performing SEs to offer
R&R using the MPMM. The complete implementation details of AHP and
MPMM for CPEE to offer R&R are discussed in the next section.

11.4 Development of an MCDM-Based Modeling

Framework for CPEE to Offer R&R
In this section, the implementation details of the proposed MCDM-based
modeling framework for CPEE to offer R&R are presented in the form of
step-by-step procedure as follows:

Step 1: Determine the importance of the variables/criteria of DCSE using AHP

For k variables in DCSE, the relative importance or weights (wj) can be

determined using AHP, using Saaty’s basic scale.

Step 2: Determine the normalized demographic score (NDS) for each SE

A decision matrix can be developed to assign uniform range of ordinal val-

ues (dij) for m SEs to each of the demographic variables, considered for DCSE.
For each of the m SEs, the sum of products of weights (wj) and ordinal value
(dij) for three demographic variables would give the demographic score (DS)
and it can be represented as:
k

DSi = ∑w * d
j
j ij for all i = 1, 2,…., m. (11.1)

The normalized DS (NDS) of the m SEs is calculated as follows:

DSi
NDSi = for all i = 1, 2,…., m, (11.2)
TDS

where TDS (total demographic score) = ∑ im=1 DSi .

MCDM-Based Modeling Framework 289

Step 3: Determine the importance of each of the main factors/criteria of PSE

using AHP

For each of the main factors/criteria of PSE, the relative importance or

weights (vj) can be determined using AHP with Saaty’s basic scale.

Step 4: Evaluate the SEs based on the main criteria of PSE using MPMM

In the PMM, one alternative is selected as a baseline (B), and all other
alternatives are compared against B, with respect to each criterion (Pugh
1991). The comparison is denoted as −1 for worse score, 0 for equal score and
+1 for better score. This evaluation results in a column vector with scores
for all the alternatives other than the baseline (whose score will be zero).
The alternative with the highest positive score would be chosen as the best
alternative.
The PMM has been criticized highlighting three major limitations (Mullur
et al. 2003): (i) the criteria weights are not incorporated, (ii) the rating scale
is too small, and (iii) the selection of random baseline alternative could lead
to bias. These limitations are addressed in this chapter by (i) using AHP to
calculate the criteria weights, (ii) by increasing the evaluation range of the
scale similar to a 5-point Likert scale, and (iii) by selecting all alternatives to
be baseline and all other alternatives could be evaluated against the baseline.
This implies a modification to the original PMM and referred as modified
PMM (MPMM) in this chapter.
The evaluation scale for the MPMM for comparing the performance
of two SEs A and B (where B is the baseline) would be by providing a
qualitative measure of performance evaluation (g ij) for a given criterion,
j such that:

 +2, if A performs much better than B in criterion j


 +1, if A performs slightly better than B in criterion j
gij = 0, if A performs as good as B in criterion j
−1, if B performs slightly better than A in criterion j

−2, if B performs much better than A in criterion j

The gij is calculated for m SEs with respect to each of the main criteria con-
sidered for CPEE to offer R&R. Using the values of gij and vj for n criteria, the
performance score (PS) of SEs can be calculated as:

PSi = ∑v * g
J
j ij for all i = 1, 2,… , m; and j = 1, 2,… , n. (11.3)

290 Big Data Analytics Using Multiple Criteria Decision-Making Models

The normalized PS (NPS) for each of m SEs can be calculated as follows:

PSi
NPSi = for all i − 1, 2, …., m, (11.4)
TPS

where TPS (total performance score) = ∑ im=1 PSi .

In the original PMM, the NPSi for m SEs with baseline SE as B1 results in an
m × 1 column matrix (named as C1) with the first set of NPSi for m SEs. In the
MPMM, all m alternatives serve as baseline and hence it results in an m × m
matrix such that:

 a11 a12 … a1m 

 a21 a22 … a2m 
A= ,
… … … … 
a amm 
 m1 am 2 …

where aij = NPSij which represents the NPS of ith SE when jth SE is the
baseline.

Step 5: Determine the final performance score (FPS) of SEs

After executing Step 4 using the MPMM, it would result in m NPS score for
every SE. The mean of NPS needs to be ascertained, such that

 m
NPSij 
NPSi = 

∑ m 
 for all i = 1, 2, … , m.

j =1 

For each of the m SEs, the final performance score (FPSi) can be obtained by
incorporating the moderating effect of DCSE with the six main factors relat-
ing to PSE. This can be computed in this study as follows:

NPSi
FPSi = for all i = 1, 2,…, m. (11.5)
NDSi

At the end of Step 5, the FPSi of m SEs is considered for ranking the given
set of SEs. Accordingly, the SE who has highest FPS can be offered R&R.
If more than one has equal highest score, then everyone should be offered
R&R. This completes one cycle (evaluation period) of CPEE of SEs.
The FPSi of m SEs should be stored in an R&R database (RRDB), which
could be accessed at a later point of time. Furthermore, at the end of every
cycle (i.e., evaluation period) of CPEE, the FPSi score of m SEs needs to be
MCDM-Based Modeling Framework 291

reset to zero. After initializing the FPSi score of m SEs, a fresh evaluation
needs to be carried out and the process needs to be repeated from Step 4.
The validation of the proposed MCDM-based modeling framework for
CPEE of SEs for R&R presented here is illustrated with a numerical example
in the following section.

11.5 Validation of the Proposed MCDM Modeling

Framework for CPEE to Offer R&R
In order to validate the proposed MCDM-based modeling framework for
CPEE to offer R&R, a hypothetical team of eight SEs with similar job pro-
file is considered. Indicative demographic characteristics with respect to
the variables: education (E), university (U) and experience (X) of the eight
SEs (that is the required data for measuring the moderating factor/criterion:
DCSE) are shown in Table 11.6. With these data, the performance evaluation
process of eight SEs is detailed as follows:

Step 1: Determine the importance of the demographics variables: E, U, and X

of DCSE using AHP

The initial evaluation process commences by determining the importance

for the demographic variables: E, U, and X of DCSE. The weights are obtained
for these demographic variables using AHP with Saaty’s basic scale and the
same is presented in Table 11.7. From Table 11.7, it can be observed that the
demographic variable experience (X) demands more importance compared
to the demographic variables: education (E) and university (U).

TABLE 11.6
Data on Demographic Characteristics of Eight Software Engineers for the
Numerical Example
Demographic Characteristics of Each Software Engineer
Name of the Experience (X)
No. Software Engineer Education (E) University (U) in Years
1 Anita BTech Cochin University 5
2 Benny BE Amrita University 2
3 Casper BTech NIT, Trichy 2
4 Deepak MCA Anna University 1
5 Esther MTech IIT Delhi 2
6 Faizel BE Manipal University 4
7 Gopi ME VTU 6
8 Harsha BE Anna University 1
292 Big Data Analytics Using Multiple Criteria Decision-Making Models

TABLE 11.7
Computed Weights for Demographic Variables of DCSE Using AHP
Relative Importance
Demographic of Demographic
Variables E U X Variables by Weights
Education (E) 1 2 1/3 0.216
University (U) 1/2 1 1/7 0.103
Experience (X) 3 7 1 0.682

In this study, the relative importance (i.e., the normalized weights) obtained
for the demographic variables: E, U, and X would remain unchanged for the
PAS period unless a new demographic variable becomes significant for mea-
suring DCSE or one of the demographic variables considered for measuring
DCSE loses its significance for measuring DCSE. So, using these normalized
weights for the demographic variables: E, U, and X of DCSE, the team of
eight SEs can be evaluated so as to ascertain their demographic score (DS).
This is detailed in the next step.

Step 2: Determine the demographic score (DS) for each SE

After ascertaining the weights for demographic variables, each of the eight
SEs are rated on an ordinal scale, with respect to their demographic charac-
teristics (Table 11.6). The ordinal ratings for each of the eight SEs are obtained
according to the following scale:

 2, if Post Graduate ( Masters)

Education, E = 
1, if Under Graduate (Bachelors)
2, if IIX(IIT/IISc/IIIT/NIT )
University , U = 
1, if Non-IIX
2, if >3 years
Experience , X =  .
1, if <= 3 years

Based on the scale defined for the demographic variables and the data
given in Table 11.6, the ordinal ratings are obtained for each of the eight
SEs and the same is given in Table 11.8. In addition, using Equations 11.1
and 11.2, both DS and NDS for each SE are calculated and presented in
Table 11.8. The SE-wise NDS presented in Table 11.8 could be fixed for
utmost 1 year (or for the PAS period), and hence the Step 2 need not to be
repeated unless there is a change in the team of SEs considered for perfor-
mance evaluation.
MCDM-Based Modeling Framework 293

TABLE 11.8
Software Engineer-Wise Ordinal Ratings and the Computed Score on DS
and NDS
Weights for the Demographic Computed
Name of the Variable: Score
Software
Engineer E: 0.216 U: 0.103 X: 0.682 DS NDS
Anita 1 1 2 1.682 0.154
Benny 1 1 1 1.000 0.092
Casper 1 2 1 1.103 0.101
Deepak 2 1 1 1.216 0.112
Esther 2 2 1 1.318 0.121
Faizel 1 1 2 1.682 0.154
Gopi 2 1 2 1.897 0.174
Harsha 1 1 1 1.000 0.092

Step 3: Determine importance of the six factors/main criteria of PSE using AHP

The six main criteria defined to represent PSE need to be weighted for their
relative importance. This is done using AHP utilizing Saaty’s basic scale.
The Saaty’s basic scale for the six main criteria and the weight obtained from
AHP toward the relative importance (i.e., normalized weights) of each of the
six main criteria are shown in Table 11.9.
The computed weights for the six main criteria considered for CPEE to
offer R&R, presented in Table 11.9, can be fixed for a relatively long duration
(or for the PAS period), unless some criteria become irrelevant or some new
criteria need to be added as part of PSE. Using these computed weights for
the six main criteria, the performance of SEs can be evaluated using MPMM
approach. This is detailed in the next step.

Step 4: Evaluate PSE based on MPMM to determine the PS of each SE

After computing the importance of weights for each of the six main criteria
of PSE, the team of eight SEs is compared against each of the main criteria for
evaluating their performance using MPMM. For proceeding with MPMM,
the qualitative measure of performance evaluation (gij) for a given criterion j,
as defined in Section 11.4, is followed to generate the data for each of the eight
SEs against each of the criteria.
The pairwise comparison and the qualitative measure are presented in
Table 11.10. Using the data given in Table 11.10, each of the SEs is considered
as baseline employee and the MPMM process is applied to obtain the m × m
matrix (i.e., 8 × 8 matrix for the numerical example) with NPS score of m (i.e.,
8) SEs and is shown in Table 11.11 (columns 2–9).
 294

TABLE 11.9
Computed Weights for the Six Factors/Main Criteria, Representing PSE Using AHP
Main Criteria Representing PSE Relative
Main Importance of
Criteria Proactive Prompt Resourceful Responsible Diagnostic Dynamic Main Criteria
Proactive 1 1/7 1/3 1 3 1/3 0.088
Prompt 7 1 5 3 4 3 0.406
Resourceful 3 1/5 1 3 1/2 1 0.151
Responsible 1 1/3 1/3 1 4 2 0.141
Diagnostic 1/3 1/4 2 1/4 1 1/2 0.086
Dynamic 3 1/3 1 1/2 2 1 0.128
Big Data Analytics Using Multiple Criteria Decision-Making Models
MCDM-Based Modeling Framework 295

TABLE 11.10
Performance Evaluation of SE Using MPMM
Name of Main Criteria Representing PSE with Its Weight (i.e., Importance) in Bracket
the
Software Proactive Prompt Resourceful Responsible Diagnostic Dynamic
Engineer (0.088) (0.406) (0.151) (0.141) (0.086) (0.128)
Anita 0 0 0 0 0 0
Benny −1 0 1 2 0 −1
Casper 1 1 0 −1 0 1
Deepak 2 0 1 2 −1 −2
Esther −2 2 0 1 1 −1
Faizel 1 0 0 −2 2 −1
Gopi −1 −1 1 0 2 0
Harsha −1 1 1 −1 0 1

Finally the SE-wise mean of NPS (i.e., NPS) is computed and presented in
Table 11.11 (column 10). From Table 11.11, particularly the column titled NPS, it
can be observed that Esther has the highest NPS and ranks first. Accordingly,
Esther can be offered R&R, if the demographic profile of SEs is not considered. In
order to be judicious and equitable in evaluation, it is important to consider
the moderating effect of NDS on the NPS. This is explained in the next step.

Step 5: Calculate the FPS by incorporating the moderating effect of NDS on NPS

Each SE’s NPS is moderated by respective NDS presented in Table 11.8 so as

to obtain an unbiased FPS for each SE using Equation 11.5. Accordingly, the FPS
for all the eight SEs are calculated and are given in Table 11.11 (last Column).
From the last column of Table 11.11, it can be observed that after incorporat-
ing the moderating effect of DCSE, Harsha emerges as the top ranker. After
Harsha is identified as the SE with the top FPS for offering R&R, one cycle/
evaluation period (could be weekly or fortnightly) of CPEE will be completed.
The FPSi of m (i.e., 8) SEs should be stored in RRDB, which could be accessed at
a later point of time. Finally, as each cycle of CPEE to offer R&R is independent,
the score on FPS at the commencement of each cycle needs to be reset to zero.

11.6  Managerial Implications

The emphasis in the CPEE to offer R&R is in the word “continuous.” The
proposed MCDM-based modeling framework in this study is very simple so
any PM or the concerned administrator of an organization can conduct the
performance evaluation using the proposed MCDM models and decide an
 296

TABLE 11.11
The NPS and FPS for Each Software Engineer
NPS Obtained Using MPMM Considering Baseline Employee as
Software
Engineer Anita Benny Casper Deepak Esther Faizel Gopi Harsha NPS FPS
Anita 0.000 0.309 0.116 0.028 −0.075 0.406 0.278 −0.174 0.111 0.720
Benny 0.118 0.000 −0.129 0.340 −0.084 −0.024 0.000 0.205 0.053 0.582
Casper 0.263 0.317 0.000 −0.020 0.105 0.326 0.391 0.046 0.179 1.765
Deepak 0.146 0.328 0.014 0.000 0.411 −0.128 0.455 0.417 0.205 1.841
Esther 0.344 0.928 −0.046 0.283 0.000 0.234 0.014 0.000 0.220 1.816
Faizel −0.081 −0.509 0.372 −0.021 0.340 0.000 −0.279 0.000 −0.022 −0.144
Gopi −0.094 −0.165 0.055 0.188 −0.001 0.000 0.000 0.505 0.061 0.350
Harsha 0.249 −0.209 0.641 0.201 0.183 0.149 0.151 0.000 0.171 1.859
Big Data Analytics Using Multiple Criteria Decision-Making Models
MCDM-Based Modeling Framework 297

employee to offer R&R. The proposed R&R system loses its main objective if
it is not offered on a timely and continuous basis. The frequency (i.e., evalua-
tion frequency) of the CPEE may be recommended as weekly or fortnightly.
One can also think of having the FPS as transparent to every employee of
the team, involved in the evaluation system. When there is a continuous
and transparent output from the system (i.e., timely R&R based on proper
and transparent evaluation), the employee (i.e., SE) might feel energized and
motivated to exhibit better performance. In addition, the PM has to decide an
appropriate R&R to the deserving employee.
The numerical example illustrated in the previous section enables the PM
to identify the best-performing SE(s) based on the relative maximum FPS
after incorporating their respective demographic and performance-related
factors. Instead of having relative FPS, the PM may specify a threshold FPS,
beyond which all SEs can be offered R&R. Moreover, the proposed frame-
work can also be used to identify the relatively worst-performing SEs and
based on this the PM may recommend the specific SE(s) with the least FPS
for special training to get performance improvement.
The FPS obtained during every performance evaluation cycle serves
two purposes, as shown in Figure 11.4. The obvious purpose is to identify
the best-performing SE(s) to offer immediate R&R. The second purpose is
to cumulatively store the individual FPS of all SEs at every performance
evaluation cycle in RRDB. This cumulative score in RRDB can be linked to
organization’s existing PAS and can be used appropriately during the orga-
nization’s regular PAS. In addition to the cumulative FPS, data-indicate the
best- and worst-performing SE(s) over a longer period of time.
Given that the performance evaluation cycle is related to a shorter period
of time, the FPS of SEs during one performance evaluation cycle does not
really convey comprehensive information about the actual performance
of SEs. Hence, the CPEE system has to be a repetitive process with equal
frequency. It is also important that the PM needs to conduct the CPEE in

Select the best-

performing SE based
Generation of on relatively maximum
FPS based on FPS to offer R&R
CPEE for
reward and
recognition
(R&R) Store the cumulative
FPS to an RRDB Organization’s
which gets updated existing PAS
after every cycle of
CPEE

FIGURE 11.4
Purpose of the proposed MCDM-based modeling framework for CPEE to offer R&R.
298 Big Data Analytics Using Multiple Criteria Decision-Making Models

an independent manner, that is, the FPS obtained during one cycle (perfor-
mance evaluation cycle) should not influence the CPEE during the next cycle.
To address this, after the top performing SE is offered R&R, at the end of one
evaluation period, the FPS of all SEs in the team shall be set to zero and a
fresh evaluation needs to be conducted for the next cycle. Such independent
evaluation process when repeated in a continuous manner would produce a
performance trend, which can easily be captured and interpreted using time
series analysis.
The success of the proposed MCDM framework requires a commitment
from the PM. The PM should tenaciously execute the CPEE process and iden-
tify the best-performing SE to offer R&R. Conducting the CPEE on a weekly
or fortnightly frequency generates a large amount of data. The cumulative
FPS data could serve as an objective input during the periodic appraisal pro-
cess. Further, if a new PM assumes charge during the middle of an appraisal
cycle, the data in the RRDB makes it easier for the new PM to understand the
performance distribution in the team.

11.7 Conclusion
A new research problem on CPEE to offer R&R for an organization is
attempted in this chapter, particularly focusing on IT organizations. A mod-
eling framework for CPEE of SEs is proposed in order to offer R&R. In order
to propose the modeling framework for CPEE to offer R&R, appropriate
exploratory and descriptive research processes are carried out to identify
suitable and adequate variables/criteria. Accordingly, 33 variables/criteria
are identified. These variables/criteria are grouped into demographic vari-
ables (five numbers), which explain the DCSE and performance-related vari-
ables (28 numbers), which explain the direct PSE.
Based on the statistical test, out of five variables/criteria considered under
DCSE, only three variables education (E), university (U), and experience
(X) are statistically significant in representing the DCSE. Furthermore, the
performance-related 28 variables/criteria are grouped into six main criteria/
factors using factor analysis and in this study they are named as proactive,
prompt, resourceful, responsible, diagnostic, and dynamic of PSE.
The AHP is used in this study to compute SE-wise demographic score,
called normalized demographic score (NDS) and PSs, called normalized
performance score (NPS). In addition, the proposed MPMM is appropriately
implemented with the weighted scores: vj obtained from AHP and finally
obtained the mean NPS for each of the SEs. For introducing the moderating
effect of DCSE and ranking the SEs, the data on NPS and NDS are used to
obtain the score on FPS for each of the SEs. Finally, the SE who has the high-
est FPS will become the top performer for getting R&R.
MCDM-Based Modeling Framework 299

This study makes the following contributions to the existing literature:

• A unique set of variables/criteria, related to demographic character-

istics and performance evaluation characteristics of SEs, are identi-
fied for CPEE to offer R&R.
• A conceptual framework for CPEE to offer R&R is proposed, based
on descriptive analysis and SEM analysis on the demographic and
performance evaluation characteristics of SEs.
• After analysis on the existing literature on MCDM methods for
performance evaluation of employees in general, MCDM methods,
AHP and MPMM, have been implemented for demonstrating the
proposed conceptual framework for CPEE to offer R&R.
• The proposed MCDM-based modeling framework for CPEE to offer
R&R is demonstrated by developing a hypothetical data set.
• The score on cumulative FPS obtained from the MCDM-based mod-
eling framework for CPEE to offer R&R is preserved for the possibil-
ity appropriately utilizing in the existing PAS.

Though the MCDM methods, AHP and MPMM, are successfully imple-
mented for demonstrating the workability of the framework for CPEE of
SEs to offer R&R, identifying different MCDM method(s) for its applicabil-
ity toward CPEE to offer R&R and following by a systematic process for
comparing various possible MCDM methods are the immediate research
directions in this area. The development of interactive Excel-based soft-
ware could be another possible extension for the research considered in
this study.

References
Afshari, A. R., M. Mojahed, R. M. Yusuff, T. S. Hong, and M. Y. Ismail. 2010. Personnel
selection using ELECTRE. Journal of Applied Sciences. 10(23): 3068–3075.
Ardabili, F. S. 2011. New framework for modeling performance evaluation for bank
staff departments. Australian Journal of Basic and Applied Sciences. 10: 1037–1043.
Arora, A., V. S. Arunachala, J. Asundi, and R. Fernandes. 2001. The Indian software
services industry. Research Policy. 30: 1267–1287.
Bassett, G. A. and H. H. Meyer. 1968. Performance appraisal based on self-review.
Personnel Psychology. 21(4): 421–430.
Bhatnagar, S. 2006. Indian software industry. V. Chandra (Ed.) 95–124. Technology
Adaptation and Exports: How Some Developing Countries Got It Right. World Bank
Publications: Washington, DC.
Boice, D. H. and B. H. Kleiner. 1997. Designing effective performance appraisal sys-
tems. Work Study. 46(6): 197–201.
300 Big Data Analytics Using Multiple Criteria Decision-Making Models

Bouyssou, D. 1996. Outranking relations: Do they have special properties? Journal of

Multi-Criteria Decision Analysis. 5: 99–111.
Bouyssou, D. 2009. Outranking methods. C. A. Floudas and P. M. Pardalos (Eds.)
2887–2983. Encyclopedia of Optimization. Springer: New York.
Bouyssou, D. and P. Vincke. 1997. Ranking alternatives on the basis of preference
relations: A progress report with special emphasis on outranking relations.
Journal of Multi-Criteria Decision Analysis. 6: 77–85.
Bowman, J. S. 1999. Performance appraisal: Verisimilitude trumps veracity. Public
Personnel Management. 28(4): 557–576.
Bradler, C., R. Dur, S. Neckermann, and A. Non. 2012. Employee Recognition and
Performance: A Field Experiment. CESIFO Working Paper No. 4164, Munich,
Germany. https://www.cesifo-group.de/DocDL/cesifo1_wp4164.pdf (accessed
February 22, 2014).
Brun, J. and N. Dugas. 2008. An analysis of employee recognition: Perspectives
on human resource practices. The International Journal of Human Resource
Management. 17(4): 716–730.
Burge, S. 2009. The Systems Engineering Tool Box. http://www.burgehugheswalsh.
co.uk/uploaded/1/documents/pugh-matrix-v1.1.pdf (accessed March 12, 2015).
Cacioppe, R. 1999. Using team–individual reward and recognition strategies to drive
organizational success. Leadership & Organization Development Journal. 20(6):
322–331.
Colson, G. 2000. The OR’s prize winner and the software ARGOS: How a Multijudge
and Multicriteria Ranking GDSS Helps a Jury to Attribute a Scientific Award.
Computers & Operations Research. 27: 741–755.
Economic Times. 2015. Indian IT in FY15: What went wrong with TCS, Wipro & others
as they struggled to grow revenues? http://articles.economictimes.indiatimes.
com/2015-04-27/news/61578067_1_indian-it-industry-leader-tcs-anant-gupta/2
(accessed April 29, 2015).
Facteau, J. D. and S. B. Craig. 2001. Are performance appraisal ratings from different
sources comparable? Journal of Applied Psychology. 86(2): 215–227.
Garafano, C. M. and E. Salas. 2005. What influences continuous employee develop-
ment decisions. Human Resource Management Review. 15: 281–304.
Gray, G. 2002. Performance appraisals don’t work. Industrial Management. 44(2): 15–17.
Gruenfeld, L. W. and P. Weissenberg. 1966. Supervisory characteristics and attitudes
towards performance appraisals. Personnel Psychology. 19(2): 143–151.
Gurbuz, T. 2010. Multiple criteria human performance evaluation using Choquet
integral. International Journal of Computational Intelligence Systems. 3(3): 290–300.
Gurbuz, T. and E. Y. Albayrak. 2014. An engineering approach to human resources
performance evaluation: Hybrid MCDM application with interactions. Applied
Soft Computing. 21: 365–375.
Han, N. and X. Ji. 2009. The study on performance evaluation based on AHP-
fuzzy. Proceedings of the Eighth International Conference on Machine Learning
and Cybernetics, Baoding, China: 2750–2753. http://ieeexplore.ieee.org/stamp/
stamp.jsp?tp=&arnumber=5212097 (accessed January 12, 2015).
Hansen, F., M. Michelle, and R. B. Hansen. 2002. Rewards and recognition in
employee motivation. Compensation and Benefits Review. 34(4): 64–72.
Hemaida, R. and S. Everett. 2003. Performance evaluation of employees using the
Analytic Hierarchy Process. Journal of Management Information and Decision
Sciences (Academy of Information and Management Sciences Journal). 6(1–2): 67–76.
MCDM-Based Modeling Framework 301

Henderson, R. I. 1980. Performance Appraisal. Prentice-Hall: Englewood Cliffs, NJ.

Ilgen, D. R., R. B. Peterson, B. A. Martin, and D. A. Boeschin. 1981. Supervisor
and subordinate reactions to performance appraisal sessions. Organizational
Behavior and Human Performance. 28: 311–330.
Islam, R. and S. M. Rasad. 2006. Performance evaluation of employees by the AHP: A
case study. Asia Pacific Management Review. 11(3): 163–176.
Islam, S., G. Kabir, and T. Yesmin. 2013. Integrating analytic hierarchy process with
TOPSIS method for performance appraisal of private banks under fuzzy envi-
ronment. Studies in System Science. 1(4): 57–70.
Jati, H. 2011. Study on performance appraisal method of vocational education teachers
using PROMETHEE II. Proceedings of International Conference on Vocational
Education and Teaching (ICVET 2011), Yogyakarta, Indonesia: 139–145. http://
eprints.uny.ac.id/5926/1/ICVET_paper.pdf (accessed November 3, 2014).
Kondrasuk, J. 2011. So what should an ideal performance appraisal look like? Journal
of Applied Business and Economics. 12(1): 57–71.
Manoharan, T. R., C. Muralidharan, and S. G. Deshmukh. 2009. Employee perfor-
mance appraisal using data envelopment analysis: A case study. Research and
Practice in Human Resource Management. 17(1): 92–111.
Morte, R., T. Pereira, and D. B. M. M. Fontes. 2013. MCDA applied to performance
analysis and evaluation of road drivers: A case study in the road trans-
port company. Proceedings of the Third International Conference on Business
Sustainability, Póvoa de Varzim, Portugal. https://repositorio-aberto.up.pt/bit-
stream/10216/70496/2/13674.pdf (accessed May 15, 2015).
Mullur, A., C. Mattson, and A. Messac. 2003. New decision matrix based approach
for concept selection using linear physical programming. Proceedings of
44th AIAA/ASME/ASCE/AHS Structures, Structural Dynamics, and Materials
Conference, Paper No. AIAA 2003-1446, Norfolk. www.ee.uidaho.edu/ee/
classes/ECE482-F03/DM/LP-DM.pdf (accessed August 19, 2014).
Nada, N. 2010. A framework for systematic application and measurement of the inno-
vation management process. The Journal of Knowledge Economy and Knowledge
Management. 5: 57–69.
Palaiologos, A., P. Papazekos, and L. Panayotopoulou. 2011. Organizational jus-
tice and employee satisfaction in performance appraisal. Journal of European
Industrial Training. 35(8): 826–840.
Pugh, S. 1991. Total Design: Integrated Methods for Successful Product Engineering.
Addison Wesley: New York.
Rosen, A. and G. E. Abraham. 1966. Attitudes of nurses towards a performance
appraisal system. Nursing Research. 15(4): 317–322.
Saaty, T. L. 1990. How to make a decision: The analytic hierarchy process. European
Journal of Operations Research. 48: 9–26.
Sahoo, B. K., R. Singh, B. Mishra, and K. Sankaran. 2016. Research productivity in man-
agement schools of India during 1968–2015: A directional benefit-of-doubt model
analysis. Omega. 66(A): 118–139, http://dx.doi.org/10.1016/j.omega.2016.02.004i
Scharge, D. P. 2010. Product Lifecycle Engineering (PLE): An Application. Encyclopedia of
Aerospace Engineering. John Wiley & Sons Ltd: Atlanta.
Schraeder, M., B. J. Becton, and R. Portis. 2007. A critical examination of performance
appraisals. The Journal for Quality and Participation. 30(1): 20–25.
Shaout, A. and M. K. Yousif. 2014. Performance evaluation: Methods and techniques
survey. International Journal of Computer and Information Technology. 3(5): 966–979.
302 Big Data Analytics Using Multiple Criteria Decision-Making Models

Smola, K. W. and C. D. Sutton. 2002. Generational differences: Revisiting genera-

tional work values for the new millennium. Journal of Organizational Behaviour.
23: 363–382.
Sreejith, S. S. 2015. Performance evaluation in IT needs to move from the manufactur-
ing model. Human Resource Management International Digest. 23(1): 32–34.
Sreejith, S. S. 2016. Development of a Multi-Criteria Decision Making Model for
Continuous Evaluation of Employee to Offer Reward and Recognition.
Unpublished Ph.D. Thesis, Department of Management Studies, Indian
Institute of Science, Bangalore, India.
Taufiq, R. and A. Sugiharto. 2011. The decision support system design of employee
performance appraisal using analytical hierarchy process (AHP) method.
Proceedings of the 1st International Conference on Information Systems for Business
Competitiveness (ICISBC), Semerang, Indonesia: 283–288.
Taylor, III, F. A., A. F. Ketcham, and D. Hoffman. 1998. Personnel evaluation with
AHP. Management Decision. 36(10): 679–685.
Wu, H., J. Chen, and I. Chen. 2012. Performance evaluation of aircraft mainte-
nance staff using a fuzzy MCDM approach. International Journal of Innovative
Computing, Information and Control. 8(6): 3919–3937.
Zbranek, P. 2013. Data envelopment analysis as a tool for performance evaluation of
employees. Acta Oeconomica et Informatica. 26(1): 12–21.
Zionts, S. and J. Wallenius. 1976. An interactive programming method for solving
multiple criteria problem. Management Science. 22(6): 652–663.
12
Use of DEA for Studying the Link
between Environmental and
Manufacturing Performance

Ramakrishnan Ramanathan

CONTENTS
12.1 Introduction................................................................................................. 303
12.2 Literature Review.......................................................................................304
12.2.1 Data Envelopment Analysis..........................................................304
12.2.2 DEA as an MCDM Tool..................................................................306
12.2.3 Relationship between Environmental Performance
and Financial Performance............................................................ 307
12.3 Data and Analysis.......................................................................................308
12.3.1 Sample and Data Collection..........................................................308
12.3.1.1 Manufacturing Performance (Manufacturing
Efficiency Scores)..............................................................309
12.3.1.2 Environmental Expenditure........................................... 309
12.3.1.3 Control Variables.............................................................. 310
12.3.2 Regression Model............................................................................ 310
12.4 Summary and Conclusions....................................................................... 311
References.............................................................................................................. 312

12.1 Introduction
We presently live in an era where data are being generated continuously
and in several forms. These data have been called as the next big innovation
(Gobble, 2013), and data analysts strive to make business sense of such data
by analyzing using appropriate tools (Bose, 2009). It is important that appro-
priate tools that have the ability to use such large data and generate useful
business insights are explored and made available to data scientists. In this
regard, this book and this chapter focus on the use of multicriteria decision-
making (MCDM) methods to help data scientists make sense of data. In this
chapter, we illustrate specifically how data envelopment analysis (DEA),
an MCDM tool, can be advantageously employed to help in economic and

303
304 Big Data Analytics Using Multiple Criteria Decision-Making Models

policy analysis. The specific problem we focus is a well-researched topic

in the field of environmental policy. This problem focuses on the relation-
ship between environmental expenditure and manufacturing performance.
Since we are interested in using publicly available data, we do not focus on
firms but instead focus on manufacturing sectors at aggregated level.

12.2  Literature Review

The literature review first describes DEA and provides arguments on why it
should be considered as an MCDM tool. Then the literature on environmen-
tal expenditure and its relationship with performance are reviewed.

12.2.1  Data Envelopment Analysis

DEA is briefly discussed in this section. More elaborate discussions are avail-
able elsewhere (Charnes et al., 1994; Ramanathan, 2003). DEA has been suc-
cessfully employed for assessing the relative performance of a set of firms,
usually called as the decision-making units (DMUs), which use a variety of
identical inputs to produce a variety of identical outputs. It has been recog-
nized as a benchmarking tool (Charnes et al., 1994).
Assume that there are N DMUs, and that the DMUs under consideration
convert I inputs to J outputs. In particular, let the mth DMU produce outputs
ymj using xmi inputs. The objective of the DEA exercise is to identify the DMUs
that produce the largest amount of outputs by consuming the least amounts
of inputs, subject to the limits imposed by the performance of other similar
DMUs. A DMU is deemed to be efficient if the ratio of weighted sum of out-
puts to the weighted sum of inputs is the highest. Hence, the DEA program
maximizes the ratio of weighted outputs to weighted inputs for the DMU
under consideration subject to the condition that the similar ratios for all
DMUs be less than or equal to one. Thus, a model for calculating the effi-
ciency of mth DMU (called the base DMU) is the following:

∑
J
vmj y mj
j =1
max
∑
I
umi xmi
i =1
such that
(12.1)
∑
J
vmj y nj
j =1
0≤ ≤ 1; n = 1, 2, … , N
∑
I
umi xni
i =1
vmj , umi ≥ 0; i = 1, 2, … , I ; j = 1, 2, … , J ,

DEA for Environmental–Manufacturing Link 305

where the subscript i stands for inputs, j stands for outputs, and n stands
for  the DMUs. The variables vmj and u mi are the weights (also called
­multipliers) to be determined by the above mathematical program, and
the subscript m indicates the base DMU. Soon after formulating Model
(12.1), its authors suggested that the nonnegativity restrictions should be
replaced by strict positivity constraints to ensure that all of the known
inputs and outputs have positive weight values (Charnes et  al., 1979).
The  optimal value of the objective function is the DEA efficiency score
assigned to the mth DMU. If the efficiency score is 1 (or 100%), the mth
DMU satisfies the necessary condition to be DEA efficient and is said to
be located on efficiency frontier; otherwise, it is DEA inefficient. Note
that the efficiency is relative to the performance of other DMUs under
consideration.
It is difficult to solve the above program because of its fractional objec-
tive function. However, if either the denominator or numerator of the ratio
is forced to be unity, then the objective function will become linear, and a
linear programming problem can be obtained. For example, by setting the
denominator of the ratio equal to unity, one can obtain the following output
maximization linear programming problem. Note that by setting the numera-
tor equal to unity, it is equally possible to produce input minimization linear
programming problem.

max ∑v
j =1
y
mj mj

such that
I

∑u
i =1
x
mi mi = 1; (12.2)

J I

∑v
j =1
y −
mj nj ∑u
i =1
x ≤ 0; n = 1, 2,… , N
mi ni

vmj , umi ≥ 0; i = 1, 2,… , I ; j = 1, 2,… , J .

Model (12.2) is called the output maximizing multiplier version in the DEA
literature. A complete DEA model involves solving N such programs
(Model 12.2), each for a base DMU (m = 1, 2, …, N), to get the efficiency
scores of all the DMUs. In each program, the objective function and
the  first constraint are changed while the remaining constraints are the
same.
Computation of efficiency score is usually done with the dual of Model
(12.2). The dual constructs a piecewise linear approximation to the true fron-
tier by minimizing the quantities of the different inputs to meet the stated
levels of the different outputs. The dual is given below:
306 Big Data Analytics Using Multiple Criteria Decision-Making Models

min θ m
such that
N

∑y λ
n =1
nj n ≥ y mj ; j = 1, 2, … , J
(12.3)
N

∑x λ
n =1
ni n ≤ θ m xmi ; i = 1, 2, … , I

λ n ≥ 0; n = 1, 2, … , N ; θ m free.

Model (12.3) is usually called input-oriented envelopment version in the DEA

literature. Analogously, an output-oriented envelopment version could be devel-
oped as the dual of input minimization linear programming problem men-
tioned earlier. Model (12.3) rates a particular DMU (mth DMU here). This
DMU is relatively efficient if and only if the optimal values of its efficiency
ratio, θm, equal unity.
Two different assumptions could be made while computing efficiency
scores using DEA—constant returns to scale (CRS) and variable returns to
scale (VRS). The assumption of CRS is said to prevail when an increase of
all inputs by 1% leads to an increase of all outputs by 1% (Golany and Thore,
1997). Model (12.3) assumes CRS. However, VRS can be incorporated in it
by appending the convexity constraint: ∑ nN=1 λ n = 1 (Banker et al., 1984). The
assumption of VRS is said to prevail when the CRS assumption is not sat-
isfied. It has been proved that DEA efficiency scores computed with CRS
assumption (usually called the CRS efficiency scores) are less than or equal
to the corresponding VRS efficiency scores (Charnes et  al., 1994) owing to
the difference in scale size of DMUs. VRS efficiency of a DMU measures its
pure technical efficiency, while CRS efficiency accounts for both technical
efficiency and efficiency loss when the DMU does not operate in its most
productive scale size (Charnes et al., 1994). The ratio of CRS to VRS efficiency
scores is called the scale efficiency. Thus, scale efficiency of a DMU operating
in its most productive scale size is 1.
DEA has the ability to give a single index of performance, usually called
the efficiency score, synthesizing diverse characteristics of different DMUs.
Due to this ability, DEA has found a number of applications as a bench-
marking tool and for measuring comparative performance of organizations,
industries, schools, banks, as well as nations (Emrouznejad et al., 2008).

12.2.2  DEA as an MCDM Tool

As highlighted in Chapter 2, DEA has now been accepted as an MCDM tool.
It has been proved that DEA scores provide a ranking of MCDM alternatives
when maximization criteria are considered as outputs in a DEA model, and
minimization criteria are considered as inputs (Joro et al., 1998; Belton and
DEA for Environmental–Manufacturing Link 307

Stewart, 1999). In this study, we use this MCDM perspective of DEA to rank
manufacturing sectors in terms of their ability to produce maximum outputs
by consuming minimum inputs.

12.2.3 Relationship between Environmental Performance

and Financial Performance
Firms spend money and efforts in order to ensure that they meet the
r­equirements  of environmental regulations formulated by governments
all over the world. Traditionally, this is considered a burden on firms,
particularly on manufacturing, since pollution abatement and restric-
tions on the use of certain materials raise the cost of operations, thereby
reducing profitability and productivity (Christiansen and Haveman,
1981). However, in the early 1990s, following a new wave of environmen-
tal concern in the public and political sphere, Porter (1991), among others,
suggested that environmental regulations might in fact be beneficial for
businesses: if properly designed and received with a “dynamic mindset,”
regulations could prompt a move toward leaner manufacturing practices,
more efficient energy and resource use, etc. This is based on the notion,
drawn from industrial ecology, that pollution and discarded waste is a
sign of inefficiency in production processes (Frosch 1982). More efficient
use of energy, leaner and cleaner production processes, and the recycling
and reuse of expired products will lower costs and reduce both the input
of new resources (a cost) and the output of undesirable resources such as
pollution (whose abatement represents a cost), thereby meeting the regu-
lations and yielding a benefit for the firm. This is the so-called win–win
argument or Porter hypothesis.
Porter’s hypothesis has been tested extensively on firms with varying
results. For example, a positive relationship between environmental per-
formance and financial performance has been reported by Hart and Ahuja
(1996), Waddock and Graves (1997), Russo and Fouts (1997), Margolis et al.
(2007), Callan and Thomas (2009), and Peloza (2009). Support for a positive
link has also been highlighted in case studies (Porter and van der Linde, 1995;
Rugman and Verbeke, 2000) as well. A negative relationship was reported
by Konar and Cohen (2001), Moore (2001), Sarkis and Cordeiro (2001), and
Brammer et al. (2006).
A notable feature of previous studies is that most of these studies are based
on primary data collected either from questionnaire surveys or interviews.
There has been no attempt to use publicly available data sources. In line with
the concept of big data, we attempt to collect publicly available data sources
to test Porter’s hypothesis in this study. In line with the focus on MCDM, we
measure manufacturing performance using multiple criteria with the help
of DEA.
Several previous studies have used environmental protection expen-
diture by firms as a proxy for environmental performance of firms
308 Big Data Analytics Using Multiple Criteria Decision-Making Models

(Majumdar and Marcus, 2001; Ramanathan et  al., 2010). Thus, based on
Porter’s hypothesis and previous literature, we propose the following
hypothesis:
Hypothesis 1. Environmental protection expenditure to meet environ-
mental regulations is significantly positively related to performance.

12.3  Data and Analysis

12.3.1  Sample and Data Collection
A governmental regulation is applicable to all the firms in a particular sector.
However, the influence of regulations will differ from sector to sector—more
polluting sectors will face higher level of regulation. The question of how
environmental regulations affected performance of firms in highly polluting
and lowly polluting sectors is of much policy interest. Hence, we focus on
sector level in this study.
We have considered 12 sectors in this study (Table 12.1). We have used
data on these 12 sectors for a period of 5 years (2002–2006). Thus our sample
size is 60. Though we have used data from multiple years, we consider them
as cross-sectional data. This approach is similar to the ones used in the
economics literature, for example, by Besley and Burgess (2004) mentioned
above.

TABLE 12.1
Sectors Analyzed
SIC Code Description SIC Code Description
10–14 Mining and quarrying 26 Manufacture of other
nonmetallic mineral products
15–16 Manufacture of food, beverages, 27–28 Manufacture of basic metals
and tobacco products and fabricated metal products
21–22 Manufacture of pulp, paper and 29 Manufacture of machinery and
paper products publishing, equipment not elsewhere
and printing classified
23 Manufacture of coke, refined 30–33 Manufacture of electrical and
petroleum products, and optical equipment
nuclear fuel
24 Manufacture of chemicals, 34–35 Manufacture of transport
chemical products, and equipment
man-made fibers
25 Manufacture of rubber and 40–41 Electricity, gas, and water
plastic products supply
DEA for Environmental–Manufacturing Link 309

12.3.1.1  Manufacturing Performance (Manufacturing Efficiency Scores)

As highlighted earlier, we have used DEA to measure manufacturing perfor-
mance of various sectors. This approach is similar to that of Majumdar and
Marcus (2001). The inputs and outputs used in the calculation of these manu-
facturing efficiency scores are shown in Table 12.2. Similar to the approach
by Majumdar and Marcus (2001), we have used constant-returns-to-scale
input-minimization DEA for calculating the efficiency scores.

12.3.1.2  Environmental Expenditure

We obtained sector-level data from the Department for Environment, Food,
and Rural Affairs (DEFRA) for both the operating and capital expenditure
associated with pollution abatement in several different media. These data
are based on the U.K. Environmental Protection Expenditure by Industry
Survey 2006, which is a survey of a stratified random sample of 7850 com-
panies belonging to various industrial sectors with 20.4% response rate
(DEFRA, 2008). The survey found that gross spending on environmental
­protection in 2006 by the U.K. industry amounted to an estimated £4.2 ­billion,
and that operating expenditure accounted for 71% of the total environmen-
tal p
­ rotection expenditure. The primary spending sectors as per the survey
were ­electricity and gas (37% of total spend), food, beverages, and tobacco
products (12% of total spend), and basic metals and metal products (8% of
total spend). The survey also found that the use of environmental manage-
ment systems was more widespread in the larger companies.
In our study, we use two measures of pollution control expenditure.
Operating expenditure (OPEX) covers in-house expenditure associated with
the operation of pollution control abatement equipment and payments to
external organizations for environmental services, including, labor costs,
leasing payments, maintenance costs for equipment, and the treatment
and disposal of waste. Capital expenditure (CAPEX) covers expenditure on

TABLE 12.2
DEA Inputs and Outputs for Measuring Manufacturing Efficiency and Descriptive
Statistics
Inputs or Outputs Mean Std. Dev. Min. Max.
Inputs
Compensation of employees (£ millions) 8704 4332 2048 14,870
Net capital stock (£ billions) 24.96 21.78 6.1 88.6
Intermediate consumption (£ millions) 25,035 12,110 6429 45,790
Outputs
Gross value added (£ millions) 14,596 6597 2377 32,202
Gross fixed capital formation (£ millions) 4816 7221 0 20,917
310 Big Data Analytics Using Multiple Criteria Decision-Making Models

end-of-pipe pollution control equipment and on integrated processes—new

or modified production facilities that have been designed so that environ-
mental protection is an integrated part of the process.
We have used data on pollution abatement expenditure for waste, water
pollution, air pollution, and all other pollution. We have aggregated the oper-
ating and capital expenditure for a particular pollution (waste, water, air, and
other) to get a single measure of pollution abatement for use in our analysis.

12.3.1.3  Control Variables

To control for the potential relationship between sector size and efficiency,
we include the number of employees in each sector. We also incorporate
R&D expenditure and energy consumption as control variables (Majumdar
and Marcus, 2001). Table 12.3 shows the summary statistics for the variables
of interest and the correlations between.

12.3.2  Regression Model

We verify our hypothesis using statistical regression. The dependent vari-
able (manufacturing efficiency scores) is regressed on the various measures
of environmental expenditure and the three control variables outlined above.
Results are shown in Table 12.4. Note that the regression has a high R2 value,

TABLE 12.3
Summary Statistics and Correlation Coefficients
Variable Correlation Coefficients
1 2 3 4 5 6 7
1.  Manufacturing efficiency 1
2. Number of employees (hundred 0.38*** 1
thousands)
3. Energy consumption (million tonnes −0.55*** −0.09 1
of oil equivalent)
4. Other pollution abatement −0.49*** 0.02 0.00 1
expenditure (£ hundred millions)
5. Waste pollution abatement −0.13 0.40*** −0.03 0.18 1
expenditure (£ hundred millions)
6. Air pollution abatement expenditure −0.48*** 0.02 0.26** 0.43*** 0.05 1
(£ hundred millions)
7. Water pollution abatement −0.20 0.48*** 0.18 0.41*** 0.23* 0.37*** 1
expenditure (£ hundred millions)
Mean 77.73 6.51 2.85 0.45 0.77 0.38 0.83
Std. Dev. 19.62 3.58 2.42 0.39 0.53 0.31 1.01
Min 32 1.22 0.38 0.06 0.03 0.04 0.13
Max 100 15.75 9.15 1.8 2.97 1.27 5.09
***p < 0.01; **p < 0.05; *p < 0.1.
DEA for Environmental–Manufacturing Link 311

TABLE 12.4
Results of the Simple Regression Model (Dependent
Variable: Manufacturing Efficiency)
Variables Regression Coefficient
Controls
Energy consumption −3.63***
Other pollution expenditure −17.08***
Number of employees 2.69***
Direct Effects
Waste expenditure −8.99***
Air expenditure −11.55**
Water expenditure −1.71
R2 0.73
F 23.45***
***p < 0.01; **p < 0.05.

indicating that 73% of the variation in the dependent variable is explained

by the independent variables. The regression is statistically significant as
shown by the F-test. Expenditure incurred on waste and air is significant in
explaining performance but has negative values. The expenditure incurred
on water is not influencing performance significantly.
Our results thus fall short of validating Hypothesis 1 and further analysis
is needed to identify the reasons. We believe that the key lies in understand-
ing the different regulation regimes followed in air/water/waste regulations
(Majumdar and Marcus, 2001). However, such more detailed analysis is not
attempted in this chapter since the focus of this chapter is to demonstrate
how DEA could be combined with statistics for environment policy analysis.
Obviously, this forms scope for further work.

12.4  Summary and Conclusions

We have illustrated in this study how DEA, an MCDM tool, can be usefully
combined with big data sources to perform further analysis for understand-
ing an important issue of environmental policy. We measured manufacturing
efficiency of various sectors using DEA and hypothesized that environmen-
tal protection expenditure will be significantly related to efficiency. However,
our illustration has shown that the results do not support the hypothesis but
we have pointed out that more detailed analysis on the flexibility of environ-
mental regulations in various media (air, water, and waste) may be required
to make more sense of available data.
312 Big Data Analytics Using Multiple Criteria Decision-Making Models

Despite the fact that our hypothesis is not validated, we believe that this
chapter shows how MCDM models can be used to make business sense of
publicly available big data for policy analysis.

References
Banker, R.D., Charnes, A., and Cooper, W.W. 1984. Some models for estimating tech-
nical and scale efficiencies in data envelopment analysis. Management Science,
30(9), 1078–1092.
Belton, V. and Stewart, T.J. 1999. DEA and MCDA: Competing or complementary
approaches? In: Meskens, N. and Roubens, M. (Eds.). Advances in Decision
Analysis, Springer Science & Business Media, Dordrecht, Netherlands, 87–104.
Besley, T. and Burgess, R. 2004. Can labor regulation hinder economic performance?
Evidence from India. Quarterly Journal of Economics, 119, 91–134.
Bose, R. 2009. Advanced analytics: Opportunities and challenges. Industrial
Management & Data Systems, 109(2), 155–172.
Brammer, S., Brooks, C., and Pavelin, S. 2006. Corporate social performance and stock
returns: UK evidence from disaggregate measures. Financial Management, 35(3),
97–116
Callan, S.J. and Thomas, J.M. 2009. Corporate financial performance and corpo-
rate social performance: An update and reinvestigation. Corporate Social
Responsibility and Environmental Management, 16, 61–78
Charnes, A., Cooper, W.W., Lewin, A.Y., and Seiford, L.M. 1994. Data Envelopment
Analysis: Theory, Methodology and Applications, Kluwer, Boston.
Charnes, A., Cooper, W.W., and Rhodes, E. 1979. Short communication: Measuring
the efficiency of decision making units. European Journal of Operational Research,
3, 339–339.
Christainsen, G.B. and Haveman, R.H. 1981. The contribution of environmental
regulations to the slowdown in productivity growth. Journal of Environmental
Economics and Management, 8(4), 381–390.
DEFRA. 2008. UK environmental protection by industry survey. Available at: https://
www.gov.uk/government/statistics/environmental-protection-expenditure-
survey, Last accessed: September 29, 2015.
Emrouznejad, A., Parker, B.R., and Tavares, G. 2008. Evaluation of research in effi-
ciency and productivity: A survey and analysis of the first 30 years of scholarly
literature in DEA. Socio-Economic Planning Sciences, 42(3), 151–157.
Frosch, R.A. 1982. Industrial ecology: A philosophical introduction. Proceedings of the
National Academy of Sciences, 89, 800–803.
Gobble, M.M. 2013. Big data: The next big thing in innovation. Research Technology
Management, 56(1), 64–66.
Golany, B. and Thore, S. 1997. The economic and social performance of nations:
Efficiency and returns to scale. Socio-Economic Planning Sciences, 31(3), 191–204.
Hart, S.L. and Ahuja, G. 1996. Does it pay to be green? An empirical examination of
the relationship between emission reduction and firm performance. Business
Strategy and the Environment, 5, 30–37.
DEA for Environmental–Manufacturing Link 313

Joro, T., Korhonen, P., and Wallenius, J. 1998. Structural comparison of data envelop-
ment analysis and multiple objective linear programming. Management Science,
40, 962–970.
Konar, S. and Cohen, M.A. 2001. Does the market value environmental performance?
The Review of Economics and Statistics, 83(2), 281–289.
Majumdar, S.K. and Marcus, A.A. 2001. Rules versus discretion: The productivity
consequences of flexible regulation. Academy of Management Journal, 44, 170–179.
Margolis, J., Elfenbein, H.A., and Walsh, J. 2007. Does It Pay To Be Good? A Meta-
Analysis and Redirection of Research on the Relationship between Corporate Social
and Financial Performance. Mimeo, Harvard Business School, Boston, MA.
Moore, G. 2001. Corporate social and financial performance: An investigation of the
UK supermarket industry. Journal of Business Ethics, 34, 299–315.
Peloza, J. 2009. The challenge of measuring financial impacts from investments in
corporate social performance. Journal of Management, 35(6), 1518–1541.
Porter, M.E. 1991. America’s green strategy. Scientific American, 264, 168.
Porter, M.E. and van der Linde, C. 1995. Toward a new conception of the environ-
ment-competitiveness relationship. Journal of Economic Perspectives, 9, 97–118.
Ramanathan, R. 2003. An Introduction to Data Envelopment Analysis. Sage, New Delhi.
Ramanathan, R., Black, A., Nath, P., and Muyldermans, L. 2010. Impact of environ-
mental regulations on innovation and performance in the UK industrial sector,
Special Topic Forum on Using Archival and Secondary Data Sources in Supply
Chain Management Research. Management Decision (Special issue on Daring to
Care: A Basis for Responsible Management), 48(10), 1493–1513.
Rugman, A.M. and Verbeke, A. 2000. Six cases of Corporate Strategic responses to
environmental regulation. European Management Journal, 18(4), 377–385.
Russo, M.V. and Fouts, P.A. 1997. A resource-based perspective on corporate environ-
mental performance and profitability. Academy of Management Journal, 40, 534–559.
Sarkis, J. and Cordeiro, J.J. 2001. An empirical evaluation of environmental efficien-
cies and firm performance: Pollution prevention technologies versus end-of-
pipe practice. European Journal of Operational Research, 135(1), 102–113.
Waddock, S.A. and Graves, S.B. 1997. The corporate social performance-financial per-
formance link. Strategic Management Journal, 18(4), 303–319.
13
An Integrated Multicriteria Decision-
Making Model for New Product
Portfolio Management

Pulipaka Kiranmayi and Muthu Mathirajan

CONTENTS
13.1 Introduction................................................................................................. 316
13.1.1 Characteristics of New Product Process (NPP).......................... 316
13.1.2 New Product Management........................................................... 317
13.2 New Product Portfolio Management....................................................... 318
13.2.1 Evaluative Dimensions for NPPM................................................ 319
13.3 Methodologies/Models for NPPM........................................................... 322
13.3.1 Quantitative Models....................................................................... 323
13.3.1.1 Quantitative Models with Single Criterion.................. 324
13.3.1.2 Quantitative Models with Multiple Criteria................ 325
13.3.2 Qualitative Models......................................................................... 326
13.3.2.1 Strategic Approaches, Scoring Models,
and Checklists................................................................... 326
13.3.2.2 Balanced Scorecard.......................................................... 326
13.3.2.3 Analytical Hierarchy Approaches/Analytical
Network Process.............................................................. 327
13.4 Development of Integrated DEA–BSC Model for NPPM...................... 328
13.4.1 BSC for Achieving Strategic and Balanced NPP........................ 328
13.4.2 DEA Approach for Ranking and Prioritization
of Portfolios................................................................................329
13.4.2.1 Base CCR Model............................................................... 330
13.4.3 Proposed Integrated MCDM Model for NPPM......................... 331
13.4.3.1 Phase 1: Development of BSC Evaluation Index
System................................................................................ 332
13.4.3.2 Phase 2: Determination of Balance Constraints
for DEA–BSC Model........................................................ 336
13.4.3.3 Phase 3: Development of the Proposed Integrated
DEA–BSC Model.............................................................. 337
13.5 Validation of the Proposed Integrated DEA–BSC Model for NPPM.....338

315
316 Big Data Analytics Using Multiple Criteria Decision-Making Models

13.6 Summary......................................................................................................342
Appendix A: An Analytical Hierarchical Process for BSC.............................343
Appendix B........................................................................................................... 347
References.............................................................................................................. 349

13.1 Introduction

A successful new product does more good for an organization than any-
thing else that can happen.
Crawford

“Product is a multi-dimensional concept, so change in one or more dimen-

sions or as a whole is considered to be a new product, depending upon the
extent of innovation into it” (Crawford and Di Benedetto 2008). In present
global competition, increasing customer requirements and rapidly changing
market requirements has made development of new product as an essential
task of any organization. The continuous development and market introduc-
tion of new products are important determinants of sustained organization
performance (Blundell et al. 1999). The ability to target the right customers in
the right way is a challenge. In addition to this, the complex characteristics of
the new product process (NPP) is one of the reasons for organizations to lose
touch with the reality of what consumers want and misspend huge amount
of investments on products that fail.

13.1.1  Characteristics of New Product Process (NPP)

New Product Process (NPP) generally is divided into five phases of develop-
ment and they are (1) opportunity identification and selection, (2) concept
generation, (3) concept/project evaluation and selection (PES), (4) develop-
ment, and (5) launch. Of all the five phases of NPP, the third phase: PES
plays a major key role in success of new product. The main purposes of the
PES phase are to (1) help an organization in deciding whether organization
should go forward with that particular project/product or not, (2) manage
the process in an optimized way as well as indicate better portfolio man-
agement by sorting the concepts and identifying the best ones, and (3) even
encourage cross-functional communication for new product success.
From implementation perspective, it appears that practitioners make two
types of errors in case of PES decisions: First, to consider a nonsuccessful proj-
ect to be successful and risk the huge investments; second, to consider a suc-
cessful project to be nonsuccessful and miss the opportunity of earning huge
profits. In case of traditional products, first error is considered to be more
serious compared with second. However, in the case of new product both the
errors would do equal damage to an organization, particularly the second
error can lead to the chances of wiping out the organization from the market
An Integrated Multicriteria Decision-Making Model 317

if competitor comes out with that particular project idea. In this scenario,
effective and accurate decision making and managing of NPP is essential.
In  literature, the managing processes, decision-making perspective of new
product development are referred to as new product management (NPM).

13.1.2  New Product Management

Ulrich and Eppinger (2004) defined NPM as “the set of activities begin-
ning with the perception of a market opportunity and ending in the pro-
duction, sale, and delivery of a product.” Over the years this definition has
been evolved and according to Loch and Kavadias (2002) “New Product
Management (NPM) consists of the activities of the firm that lead to a stream
of new or changed product market offerings overtime. This includes the
­generation of opportunities, their selection and transformation into artifacts
and activities offered to customers and the institutionalization of improve-
ments in new product development activities themselves.”
NPM requires a static plan to sustain and win in a war for market territory,
waging battles alongside alliance partners against competitors, and conquer-
ing market segments with products, services, and solutions. Efficient NPM
outlines what market segment to target with what products, and what posi-
tion to defend against competitors. NPM includes decisions about configura-
tion and development of internal resources to build and defend the desired
new products from the phase of idea selection to launch. NPM not only assists
in making accurate decisions but also allows an organization to achieve stra-
tegic objectives and vision through development of new products. However,
studies focusing on decision-making aspects of NPM are not as widespread
(Yahaya and Abu-Bakar 2007). For all these specific purposes, an efficient man-
agement system for formulation of new product portfolio is essential for any
organization to succeed. In the literature, the management system and deci-
sion making for formulation of portfolio is termed as new product portfolio
management (NPPM). Often managers rate NPPM as the weakest NPM area
and explicit evaluative dimensions are lacking (Cooper et al. 2001; McNally
et al. 2009) in most of the reported portfolio framework. With these, the main
objective of this study is to identify and study significance of different evalua-
tive dimensions and subsequently build an appropriate and efficient decision-
making model of PES for NPPM. Particularly, we intend to study different
methodologies involved in literature, and identify limitation, and make an
attempt to address the research gap. In addition, it is observed from literature
that the integration of qualitative and quantitative methods is significant for
future NPPM methods that seek practical implications. Accordingly, in this
study, we propose to integrate balanced scorecard (BSC) and data envelop-
ment analysis (DEA) methodologies for better NPPM performance.
The structural flow of this study is as follows: NPPM and the required
evaluative dimensions for NPPM are discussed in Section 13.2. Literature on
different methodologies implemented in NPPM is summarized in Section
318 Big Data Analytics Using Multiple Criteria Decision-Making Models

13.3. In Section 13.4, a base BSC system along with proposed BSC index sys-
tem for NPPM is first discussed and subsequently the proposed integrated
DEA–BSC model is presented. The workability of the integrated decision-
making model: DEA–BSC along with numerical example is presented in
Section 13.5.

13.2  New Product Portfolio Management

In general, “Product Portfolio Management (PPM) is the centralized manage-
ment of the processes, methods, and technologies used by project managers
and project management offices (PMOs) to analyze and collectively man-
age current or proposed projects based on numerous key characteristics.”
PPM ascertains program and project managers with the capabilities needed
to manage the time, resources, skills, and budgets necessary to accomplish
all interrelated tasks.
NPPM in general is defined as “The dynamic decision process wherein the
lists of active new products or R&D projects are constantly revised. In this pro-
cess, new products or projects are evaluated, selected, and prioritized. Existing
products or projects may be accelerated, killed, or deprioritized and resources
are allocated to the active projects.” In a competitive market along with goal
of developing a new product, one has to keep in mind multiple goals such as
reaching the market first, competitors’ position, market sentiment, and so on. In
this scenario, best mix of products or projects that ensures strategic alignment,
balance of portfolio, and potential gain is compulsion. This can be achieved
most efficiently through an NPPM that provides both high-level and detailed
information on all products or projects for educated decision making—keep-
ing the entire portfolio aligned with overall corporate strategies and objectives
and securing the highest return on investments (ROIs). Based on the literature
review and from practitioners’ point of view, NPPM decisions are considered
complex and significant. The reasons for the same are identified as follows:

• Decision maker may be interested in working toward multiple

objectives but progress in one direction impedes progress in
­
others.  So in this case, different perspectives lead to different
­
conclusions.
• New product development is characterized by a tremendous degree
of complexity and uncertainty and involves choosing between
­different products competing for the same funding.
• Multiple nature of conflicting objectives (i.e., maximize return, max-
imize R&D productivity, minimize uncertainty, etc.) leads to consid-
erations of multiple variables for taking a single decision.
An Integrated Multicriteria Decision-Making Model 319

• Decisions taken for the formulation of NPP determine the competi-

tive position of business, product sales, and market share.
• Decisions implemented in NPPM play a major role in forging the
link between product or project selection and organization strategy.
As the portfolio is the expression of organizations’ strategy, thus it is
significantly important for NPP to align with organization strategy.
• Efficient NPPM achieves focus of not doing too many projects for
the limited resources available but to allocate these resources to the
efficient and profitable product or projects.
• Efficient NPPM achieves the right balance between long- and short-
term product or projects, and high-risk and low-risk ones, consistent
with the business goals.

Based on the reasons stated here on the complexity and the significance of
NPPM, one needs to consider multiple evaluative dimensions for obtaining
efficient NPPM. The reality and the literature revealed that product/project
managers use different evaluation dimensions/criteria for NPPM. In addition,
the ability of decision models for evaluating accurately the best set of prod-
ucts or projects varies depending on dimensions used and weights applied to
these dimensions. So the identification of evaluative dimensions is inevitable.
In this study, we made an attempt to identify and study the significance of
different evaluative dimensions and its impact on efficiency of NPPM from
the analysis of the literature. Accordingly, the evaluative dimensions identi-
fied through the related research studies on NPPM, R&D portfolio manage-
ment, and PES for NPPM are discussed in detail in the following section.

13.2.1  Evaluative Dimensions for NPPM

There is a rich literature available for identification of evaluative factors
or dimensions which affect NPPM performance. Carbonell-Foulquié et  al.
(2004) did a study on the evaluative dimensions and weights used in the for-
mulation of NPP. The evaluative dimensions identified were named: techni-
cal feasibility, strategic fit (SF), customer acceptance, financial performance,
and market opportunity. Cooper et al. (2001) report that managers use three
broad dimensions: value maximization, portfolio and innovation balance
(PIB), and SF to evaluate the organizations portfolio of new product projects.
Organizations that fail to manage their NPPM activities strategically are not
only running their business from a position of disadvantage but are risking
their future (Fitzsimmons et al. 1991). Thus, critical role of NPPM in the sur-
vival and success of organization and the need for managing it strategically
is being increasingly recognized in both academic (Brown and Eisenhardt
1995; Griffin and Hauser 1996; Krishnan and Ulrich 2001) and practitioner
(Gates 1999; Chesbrough and Teece 2002) point of views. Thus, the alignment
between organization’s missions/objectives and execution of NPP is a tactical
320 Big Data Analytics Using Multiple Criteria Decision-Making Models

challenge faced by every organization. Managers usually perceive that their

new product projects are generally aligned with their organization’s strate-
gic objectives and goals (Cooper et al. 2004), which is important because this
dimension correlates strongly with new product’s performance.
Osawa and Murakami (2002) proposed a methodology to evaluate R&D
projects in terms of SF and financial credibility. The implications of this
study supported the inference: Adding SF as an important evaluative dimen-
sion improves NPPM performance, given in the study by Ronkainen (1985).
Balance is a critical NPPM dimension, as it is second most strongly corre-
lated practice after value maximization with superior new product develop-
ment performance (Cooper et al. 2004). Many researches (Cooper et al. 1997;
Graves et al. 2000) have focused on the dimension of project portfolio balance,
in terms of achieving balance among the set of projects that are selected in
the portfolio. Oh et al. (2012) considered portfolio balance, but the limitation
was that they considered balance between the set of projects which were yet
to be selected for the portfolio and did not consider existing projects and the
amount of innovation or type of innovation involved. However, in real-life
scenario, there is a requirement to achieve balance among multiple direc-
tions (i.e., product innovation, process innovation, common product plat-
forms, etc.). Though there is no strong empirical study which concludes that
PIB increases the efficiency of NPPM, we strongly believe that PIB should be
considered as one of the evaluative dimension for NPPM.
NPPM’s critical task is allocation of resources between different innovation
projects ranging from radical innovation to basic incremental innovation and
each of these projects poses conflicting directions in terms of organizational
strategy. Organizations prefer to develop more than one product in order to
sustain the competition; as a result, interdependency between the projects
increases. Simultaneously, resource allocation becomes complex when the
number of projects and interdependency between them increases. Many
researchers have developed different frameworks to increase effectiveness of
NPPM with major concentration on resource allocation. For example, Loch
and Kavadias (2002) focus on the optimal resource allocation across New
Product Development (NPD) programs. But they do not consider how the
types of the NPD investment or the investment horizon impact the allocation
decision. Chao and Kavadias (2008) use concept of strategic buckets in order to
allocate limited resources throughout project portfolio. Optimizing the avail-
able resources for the set of projects with the aim of increasing the efficiency
of NPPM is very crucial for organization’s success.
It is observed from the literature that cost–revenue or development cost
evaluation or benefit–cost analysis is one of the evaluative dimensions, which
is focused from past two to three decades and received major ­attention from
academicians as well as practitioners. For example, Chiu and Park (1994)
used the expected net present value (NPV) to evaluate the conditions for an
R&D Project’s success or failure. Mahmoodzadeh et al. (2007) reviewed four
common methods: NPV, rate of return, benefit–cost analysis, and payback
An Integrated Multicriteria Decision-Making Model 321

period (PBP) by using analytical hierarchy approaches (AHP). Lockett and

Stratford (1987) and Ghasemzadeh and Archer (2000) focused on investments
and other evaluative dimension: cost–revenue. Cooper et  al. (2004) state,
cost–revenue estimation (CRE) is a crucial dimension which distinguishes
high-performing organizations in the area of NPD from low-performing
organizations. The limitations of these research studies are that they did
not consider multiple projects and overemphasized only on one particular
evaluative dimension of cost–revenue/cost–benefit analysis.
Uncertainty management is an integral part of NPP and different
approaches exist in literature to define and analyze uncertainty dimension.
Recently, researchers have concentrated on minimizing risk and uncertain-
ties involved in NPP (Kahraman et  al. 2007; Mahmoodzadeh et  al. 2007;
Chiang and Che 2010). Kahraman et al. (2007) considered two types of risks,
namely systematic risks (financial and technical) and unsystematic risks
(managerial and personnel) in their hierarchical structure of decision criteria
and employed heuristic multiattribute utility function.
Chiang and Che (2010) concentrated on risk and uncertainty-evaluative
dimension, that is, expected revenue risk, manufacturability risk, and time-to-
market risk, which were calculated using Bayesian belief network. Feyzioğlu
and Büyüközkan (2006) highlighted that NPPM decision, especially necessary
at early stages of the NPD, as it contains considerable amount of uncertainty-
causing elements, which confuses decision maker to reach the target perfor-
mance. At this point, decision maker has to make a lot of decisions based on
inadequate information about the project, vagueness in challenging issues,
no dependency on previous data leads to uncertainty and increase the risk
involved in development of product. However, these studies did not consider
other evaluative dimensions along with risks and uncertainty dimension for
PES decision in NPPM. In this study, the wide spectrum of risks and uncer-
tainties are categorized and considered in accordance with other dimensions.
Cooper (1994) stated that an organization which emphasizes only on a par-
ticular evaluative dimension for NPPM is linked with poorer performance.
Accordingly, there are few studies concentrating on multiple dimensions. For
example, Cooper et al. (1997) report that managers use three broad dimensions
to evaluate the organizations portfolio of new product projects: value maxi-
mization, portfolio optimization, and SF. Ozer (2005) presented an integrated
framework for understanding various factors which affect decision making
in NPP. The study carried out by Mohanty et al. (2005) considered four evalu-
ative dimensions: SF, cost–revenue, resource allocation, and risk and uncer-
tainty for R&D project selection. Bhattacharyya et al. (2011) developed a fuzzy
R&D portfolio selection model by considering three objectives: minimization
of risk, minimization of project cost, and maximization of project outcome.
Based on the analysis of the literature review presented here, it is observed
that there are five evaluative dimensions: SF, PIB, optimized resource alloca-
tion (ORA), CRE, and risk–uncertainty estimation (RUE) are considered in
different sets and combination (Table 13.1).
322 Big Data Analytics Using Multiple Criteria Decision-Making Models

TABLE 13.1
Summary of Closely Related Literature on Evaluative Dimensions of NPPM
Evaluative Dimensions of Study
Strategic Portfolio Risk– Cost– Resource
Related Studies Fit Balance Uncertainty Revenue Allocation
R&D Project Evaluation and Selection
Osawa and Murakami (2002) + +
Mohanty et al. (2005) + + + +
Eilat et al. (2008) + + + +

R&D Project Portfolio Formulation

Eilat et al. (2006) + + +
Wang and Hwang (2007) + +
Bhattacharyya et al. (2011) + + +
Abbassi et al. (2014) + + +

New Product Project Evaluation and Selection

Thieme et al. (2000) + +
Feyzioğlu and Büyüközkan (2006) + + +
Mahmoodzadeh et al. (2007) + +
Chiang and Che (2010) + + +

New Product Portfolio Formulation

Oh et al. (2012) + + + + +

Table 13.1 clearly indicates that no one has considered the entire five
e­ valuative dimensions together for PES decision in NPPM. In addition,
it appears that there is only one study: Oh et  al. (2012) considering PIB
­evaluative dimension, but not in the perspective of this study and par-
ticularly they considered only portfolio balance. Furthermore, there is
no significant study which is carried out to analyze the interrelationship
between these ­evaluative dimensions for the significance of NPPM. Finally,
in a­ ddition to these research gaps, there exists a gap where these identified
evaluative dimensions are not modeled for implementation in industry.
In this study, we attempt to develop an explicit decision-making model
which f­ ulfills these research gaps. In the next section, methodologies used
in different studies for development of decision models for NPPM are
presented.

13.3  Methodologies/Models for NPPM

There are many theoretical and practical attempts to develop models for
decision making in case of product or project selection for portfolio. Earlier
An Integrated Multicriteria Decision-Making Model 323

Single criteria
linear programming, ILP,
dynamic programming
Quantitative
models
Multiple criteria
ANN, DEA, goal
programming, MOEA
NPPM decision-
making models

Single criteria
strategic buckets, product
matrices, score techniques
Qualitative
models
Multiple criteria
AHP, ANP, BSC,
promethee

FIGURE 13.1
Classification of new product portfolio management models/methodology.

studies of portfolio management focused on constrained optimization

models based on theoretical operation research. Brown and Eisenhardt
(1995) classify NPD research depending on its methodological approach.
Liao (2005) conducted a literature survey on NPPM methodologies and
applications. Based on the analysis of the literature, models developed
for NPPM can be classified based upon type of study as qualitative and
quantitative, and further can be classified based on a number of criteria
being considered into single criterion and multicriteria. Figure 13.1 depicts
this classification and only recent literature which discusses these classi-
fied models is summarized in Table 13.2. In addition, a brief overview of
the methods/models based on the classification given in Figure 13.1 is pre-
sented as follows.

13.3.1  Quantitative Models

Quantitative models consist of numeric data as input and adopt proce-
dures such as mathematical algorithms, mathematical programming, or
economic/quantitative indices for evaluating NPPM and produce a numeri-
cal output. These quantitative models again can be categorized based on a
number of criteria considered for evaluation. Accordingly, these quantita-
tive models are classified into single-criterion models and multiple-criteria
models.
324 Big Data Analytics Using Multiple Criteria Decision-Making Models

TABLE 13.2
Summary of Recent Related Literature on Models/Methodologies for NPPM
Number of
Approach Criteria Methodology/Model Reference
Quantitative Single Financial indices and Patah and de Carvalho
methods (2007), Ibbs et al. (2004)
Linear programminga Chien (2002)
Integer programming Mavrotas et al. (2006),
Melachrinoudis and
Kozanidis (2002)
Dynamic programminga Kyparisis et al. (1996)
Multiple Goal programming Wey and Wu (2007), Lee and
Kim (2000)
Data envelopment analysis Chang et al. (2014), Kumar
et al. (2007)
Multiobjective Metaxiotis et al. (2012),
evolutionary algorithms Gutjahr et al. (2010)
Qualitative Single Strategic buckets, score Chan and Ip (2010)
techniques
Multiple Multiattribute utility Duarte and Reis (2006)
theorya
Promethee multicriteriaa Halouani et al. (2009)
Balanced scorecard (BSC) Asosheh et al. (2010), Eilat
et al. (2008)
Analytic hierarchy Ayağ and özdemr (2007),
process/analytic network Wang and Hwang (2007),
process Yurdakul (2003)
a Details of these methods/models are not attempted to provide in this study.

13.3.1.1  Quantitative Models with Single Criterion

A major part of literature that concentrates on PES includes single-criterion
quantitative models such as financial models and financial indices, linear
programming, integer programming (IP), and dynamic programming.

13.3.1.1.1  Financial Models and Financial Indices

Financial models and financial indices are the most prominent and are used
as evaluation models for decades by organizations. The existing literature on
Financial Models and Financial Indices for evaluation can be categorized based
on methodology used into two approaches. The first approach consists of
cost–benefit analyses and the second approach consists of ROI-based anal-
yses. The studies related to cost–benefit analyze the relationship between
NPPM investments and the savings and the studies related to ROI analyze
the relationship between NPPM maturity and project performance. Models
that are frequently employed by practioners include: NPV, Internal Rate
of Return (IRR), PBP, and ROI. Remera et al. (1993) conducted a survey on
An Integrated Multicriteria Decision-Making Model 325

financial project evaluation techniques used in industries and mainly con-

cluded that there is a shift from the use of IRR to NPV and a decrease in the
use of the PBP.

13.3.1.1.2  Integer Programming

Most of the earlier reported studies highlighted that organizations develop
and/or use IP models with different optimizing criteria in case of project
selection. In literature, particularly theoretically, it is considered as the best
method to achieve optimized solution where constraints (or) requirements
can be represented in the form of linear relationships. However, in real-life
scenario, there always exist complex constraints which cannot be modeled
as linear relationships. Because of this, most of IP models developed did not
come into use in practical scenarios.

13.3.1.2  Quantitative Models with Multiple Criteria

Multicriteria optimization is the problem of optimizing two or more
­objectives which may be conflicting in nature, subject to certain constraints
simultaneously. It includes several approaches such as Goal Programming
(GP), DEA, and multiobjective evolutionary algorithm (MOEA).

13.3.1.2.1  Goal Programming

GP can be considered as an extension or generalization of linear program-
ming to handle multiple, normally conflicting objectives. Each of these
­objectives is considered as a goal and assigned a target value to be achieved.
So while modeling, the unwanted deviations from this set of target values
are then minimized in prime objective function. This objective function can
be a vector or a weighted sum dependent on the GP variant used. A major
strength of GP is its simplicity and ease of use. GP can handle relatively large
number of variables, constraints, and objectives.

13.3.1.2.2  Data Envelopment Analysis

The basic concept of DEA is to measure the efficiency of a particular decision-
making unit (DMU) against a projected point on an “efficiency frontier.” It is
a mathematical programming technique that calculates the relative e­ fficiency
of multiple DMUs on the basis of observed inputs and outputs, which may
be expressed with different types of metrics. A DMU is considered efficient
when no other DMU can produce more outputs using an equal or less amount
of inputs. Additionally, DEA generalizes the usual efficiency measurement
from a single-input single-output ratio to a multiple-input multiple-output
ratio by using a ratio of weighted sum of outputs to weighted sum of inputs.
Recent literature indicated that DEA is one of the efficient tools for evaluating
projects (Eilat et al. 2008). In addition to evaluation of projects, the DEA also
provides selection and ranking of projects for portfolio.
326 Big Data Analytics Using Multiple Criteria Decision-Making Models

13.3.1.2.3  Multiobjective Evolutionary Algorithms

MOEA can be useful in complex problems, particularly NP-hard problem,
for which no efficient deterministic algorithm exists. The first implementa-
tion of MOEA for project selection and evaluation dates back to the mid-
1980s (Schaffer 1985). Over the years, researchers have developed several
approaches for obtaining solution of multiobjective optimization problems
with the use of Evolutionary Algorithms. One can refer to Metaxiotis and
Liagkouras (2012) for detailed literature survey on different Evolutionary
Algorithms existing and studies that implemented MOEAs.

13.3.2  Qualitative Models

Qualitative models/approaches such as scoring models, checklist, strategic
buckets, and predictive analysis of expert opinions can be modeled for single
or multiple criteria.

13.3.2.1  Strategic Approaches, Scoring Models, and Checklists

An organization in general tends to develop new product based on future
needs. In this scenario, the organization needs to study unique needs and
finally, evaluate them and try to come up with new products that can ful-
fill the most important ones. In such cases, scenario analysis is being used
for identifying future needs and generating new product concepts before
competition. In addition, there are other methodologies such as different
checklists varying with organization goals, strategic buckets, score models,
product complexity matrix, and so on developed by an organization as deci-
sion aid models. The dynamic model utilizes an extensive amount of input
from the multimedia exercises (e.g., online search by consumers, dealer
visits, word-of-mouth communication, magazine reviews), historical data
of similar products, industry sales, managerial judgments, and production
constraints. All these inputs are used to forecast and simulate the market
environment. For detailed description of other approaches for qualitative
models, one can refer to Ozer (1999).

13.3.2.2  Balanced Scorecard

Among other techniques, BSC technique is promoted as an efficient tool for
strategic alignment of projects along with satisfying other objectives and
measures for specific organizational unit (Eilat et  al. 2008). The core idea
of BSC is to display the organizational strategic trajectory through mutual-
driven causal relationship between four perspectives and they are: fi
­ nancial,
customer, internal processes, and learning and growth. Indeed, many
­
organizations have adopted the BSC approach to (a) accomplish critical
An Integrated Multicriteria Decision-Making Model 327

management processes, (b) clarify and translate their vision and strategy,
(c)  communicate and link strategic objectives and measures, (d) plan and
align strategic ­initiatives, (e) enhance strategic feedback and learning, etc.
The significance and implementation of BSC is trending exponentially in
present competitive ­scenario (Eilat et al. 2006).

13.3.2.3  Analytical Hierarchy Approaches/Analytical Network Process

AHP/analytical network process (ANP) methods are special ones as both
quantitative and qualitative and/or judgment criteria can be considered.
In these methods, first the weights of different objectives are determined,
then alternatives are compared on the basis of their contributions to these
objectives, and finally a set of project benefit measures is computed. Once
the alternatives have been arranged on a comparative scale, the decision
maker(s) can proceed from the top of the list, selecting a subset of alterna-
tives until the feasibility constraint is maintained. AHP has an advantage of
allowing a set of complex issues that have an impact on an overall objective
to be compared with the importance of each issue relative to its impact on the
solution of the problem. However, the conventional AHP method assumes
a unidirectional hierarchical relationship among decision levels and attri-
butes; due to this, researches are using ANP, which is considered as a second
generation of AHP. The ANP method has been designed to overcome the
limitation and provide a solution for more complex decision problems with
multidirectional relationships (Feyzioğlu and Büyüközkan 2006; Ayağ and
özdemr 2007).
From the analysis of literature review and to the best of our knowledge,
there are some specific shortcomings of the current methods, particularly for
PES decision in NPPM and they are as follows:

• Current methods of PES in case of NPPM do not consider probabili-

ties and risk dimensions (Cooper et al. 2001).
• Existing methods mostly depend on extensive financial and other
quantitative data (Cooper et al. 2001).
• Major focuses of the current methods are given to financial and
related criteria.
• No study considered both SF and PIB dimensions/criteria along
with other evaluative dimensions.
• Inability to consider both qualitative and quantitative decisions
while developing a model, particularly for PES problem.

Additionally, it is also implied that the integration of qualitative and quan-

titative methods is significant for future NPPM methods that seek practical
implications.
328 Big Data Analytics Using Multiple Criteria Decision-Making Models

13.4  Development of Integrated DEA–BSC Model for NPPM

This study considers both qualitative and quantitative data for PES deci-
sion in NPPM. The proposed methodology is based on relative evaluation
of products or projects or portfolios. Particularly, this study proposes an
evaluation model by appropriately integrating DEA and BSC models (this
integrated model is called DEA–BSC model) for PES decision in NPPM. This
integrated approach serves as an alternative to the conventional multidimen-
sional knapsack approach which obtains an optimal portfolio with respect to
a well-formulated objective function and multiple resource constraints (Eilat
et  al. 2006). The integrated approach proposed in this study has an addi-
tional feature of evaluating alternative new product projects in the presence
of multiple objectives and possible interactions among the projects. Before
detailing the proposed integrated DEA-BSC model, we discuss the base BSC
and base DEA models in the following sections.

13.4.1  BSC for Achieving Strategic and Balanced NPP

Evaluating new product projects involves certain criteria for which man-
agers cannot provide hard data, where decisions are to be made based on
experience, intuition, and opinions. In order to inculcate this feature into
decision-making tool, we propose to develop a BSC model for NPPM.
The BSC model is a management tool composed of a collection of ­evaluation
indicators (such as customer trust, priority level, and supplier’s satisfaction,
etc.) arranged in groups (i.e., evaluation perspective) and denoted as cards.
The BSC model is collection of nonfinancial and financial measures. The
measures are related to four managerial perspectives, that is, marketing,
strategic, operational, and financial, and are aimed to ­provide a comprehen-
sive view to the top management of their NPP. A  ­specific BSC model for
projects was first proposed by Kaplan and Norton (1992).
The BSC model represents a translation of business strategy into a linked set
of evaluation indicators that define both long-term and short-term objectives.
It acts as a mechanism for achieving and obtaining feedback of the objectives
(Kaplan and Norton 2001). Decision makers who rely on BSC model need not
hang to just short-term financial measures as the sole indicators of project
performance are for formulation of portfolio. The major advantage of using
BSC model is that, it minimizes number of measures used, which in turn
minimizes the information overloaded. This helps to summate seemingly
disparate elements of the evaluation and finally provides with suboptimiza-
tion of all the important measures.
The strategies and the lines of action that would enable the organization
to achieve its strategic vision should be translated into each of the four per-
spectives. The organization’s strategies that are formulated in alignment to
the perspectives of learning and growth and in internal processes determine
An Integrated Multicriteria Decision-Making Model 329

• Vision clarity  • Goal fixation

• Gaining consensus • Knowledge
accumulation
• Knowledge transfer

Translating Communicating
the vision and linking

Feedback
Business
and
planning
learning

• Articulating the
vision • Setting targets
• Strategic review • Strategic alignment
and feedback • Resource allocation

FIGURE 13.2
Managing strategy through balance scorecard.

the success of organization. These organization’s strategies help one to meet

organization’s vision and objectives related to the satisfaction of its cus-
tomers and shareholders. For linking strategy and transforming organiza-
tion’s visions into BSC model’s perspectives, an efficient ­process is required.
Accordingly, BSC model involves four processes which c­ ontribute to linking
long-term strategic objectives to short-term actions. The four processes for
managing strategy are: translating the vision, communicating and linking,
business planning, and feedback and learning. Figure 13.2 depicts how these
four processes help in formulation of BSCs.

13.4.2  DEA Approach for Ranking and Prioritization of Portfolios

All the quantitative models for NPPM capture only part of the scenario and
they are often deemed as incomplete. For this reason, an increasing trend
in the implementation and requirement of new performance measurement
­systems other than financial metrics is implied through many research stud-
ies (Andrews 1996; Banker et al. 2000, 2004).
DEA approach is a mathematical programming technique that calculates
relative efficiency of multiple DMUs on the basis of observed inputs and
­outputs. Further, DEA approach is considered to be an efficient methodol-
ogy for multiobjective decision making where qualitative and quantitative
criteria are involved (Chames et al. 1978). The usefulness of DEA approach
330 Big Data Analytics Using Multiple Criteria Decision-Making Models

in evaluating multicriteria systems and providing targets for such system is

reported in a large number of applications (Seiford 1996).
The basic DEA model estimates the relative efficiency of a DMU as the
ratio of weighted outputs to weighted inputs, where the model takes weights
for each DMU. As a result, model identifies its relative efficiency with respect
to an “efficiency frontier,” which is defined by assessing all the DMUs.
However, in case of practical implementation perspective, the virtually
unconstrained weights are not acceptable (Roll and Golany 1993). For this
scenario, restricted DEA approaches were developed to allow control over
the weights in the model. A general approach for controlling factor weights
was developed by Chames et al. (1989). This weight-restricted method gener-
alizes the original DEA model, by acquiring the values for input and output
weights within the given closed cones. For the detailed weight restriction
approaches, one can refer to Cook and Seiford (1978) and Roll and Golany
(1993). The DEA model, proposed in this study, is based on Chames, Cooper,
and Rhodes (CCR) model, which was first proposed by Chames et al. (1978).

13.4.2.1  Base CCR Model

For the development of base CCR model, we assume that there are k projects.
Each project assumes to have varying amounts of l different inputs which
produce m different outputs. The nomenclature involved in the development
of the base CCR model is detailed as follows:

For Base CCR Model:

k—No. of projects Xj = {xij}  i = {1, …, l} input matrix

l—No. of different inputs consumed by a project Yj = {ynj}  n = {1, …, m} output matrix
m—No. of different outputs produced by a project µ = {ui}  i = {1, …, l} input weight vector
ν = {vn}  n = {1, …, m} output weight vector
Let Pj represent project j

Consider, project Pj(j = 1, …, k) which consumes Xj = {xij} of inputs (i = 1,

…, l) and produces Yj = {ynj} of outputs (n = 1, …, m). Further, when one con-
siders k projects, we have l × k matrix of inputs is denoted by X and n × k
matrix of output is represented by Y. Accordingly, the input and output
weights are denoted by the vectors µ = {ui} and ν = {vn}, respectively. The base
CCR model defines the relative efficiency of a specific project P0 as the ratio
of sum of weighted outputs ( Σ n υn y n 0 ) by sum of weighted inputs ( Σ iui xi 0 ).
The objective function of this base CCR models is defined by Equation 13.1.

max Z0 =
∑ vy n
n n0
, (13.1)

µ,ν
∑ ux i
i i0

where Z0 represents the efficiency score of project P0.

An Integrated Multicriteria Decision-Making Model 331

In order to bind the optimization problem presented in Equation 13.1, nor-

malization constraint needs to be forced to the ratio of objective function as
shown below:

∑ vyn
n nj
≤ 1, ∀j (where j = 1, … , k ). (13.2)

∑ ux i
i ij

The constrained optimization problem defined by Equations 13.1 and 13.2,

including positivity constraints of weights, is included into base CCR model.
Furthermore, the equivalent linear programming formulation for the base
CCR model is presented as follows:
Base CCR model [A]

max Z0 =
µ,ν ∑v y n n0

such that,

∑ u x = 1,
i
i i1

∑v y − ∑u x
n
n nj
i
i ij ≤ 0 ∀j

vn ≥ ε , ui ≥ ε.

The constant ε is a small positive number that defines as a lower bound

for the multipliers. By solving the Base CCR model (A) k times, we obtain k
efficiency scores for all DMUs and these can be grouped into two categories.
One group consists of the efficient ones that lay on efficient frontier and the
other group consists of inefficient ones that fall below the frontier.

13.4.3  Proposed Integrated MCDM Model for NPPM

The proposed integrated DEA–BSC model helps to achieve the objectives
of: (a) strategically aligned with organization goals and vision and (b) bal-
ance between innovation level, risk level, cost, and resources. The proposed
integrated DEA–BSC model is based on DEA–BSC model developed by Eilat
et  al. (2006), who were the first to integrate DEA and BSC methodologies,
for PES.
The objective of the proposed integrated DEA–BSC model is to relate the
identified evaluative dimensions to the perspectives of BSC model. Once this
is achieved, one needs to relate the input methodology of the BSC values
332 Big Data Analytics Using Multiple Criteria Decision-Making Models

to DEA model. Additionally, resource constraints, priorities of perspectives,

and limits of each perspective also have to be included into DEA model.
In order to simplify the procedure and guide the development of the pro-
posed integrated DEA–BSC model, we propose to develop a framework.
Accordingly, the development of the proposed integrated DEA–BSC model
is divided into the following three phases and more details are discussed in
the following subsections:

Phase 1: Development of BSC index system: In this phase, first we relate the
identified evaluative dimensions to the perspectives of BSC. In this
study, we proposed seven perspectives and these have to be mea-
sured. In addition, the required evaluation indicators to measure
these perspectives also need to be determined. The details of these
seven perspectives are described in detail in the next section.
Phase 2: Determination of balance constraints for DEA–BSC model:
In  this phase, different perspective balance limits are determined.
Subsequently, upper and lower bounds of each perspective are fixed
and accordingly constraints are drawn which are implemented in
base DEA model.
Phase 3: Development of integrated DEA–BSC model: In this phase, pri-
orities of each perspective to indicator level are determined. These
priorities are considered as weights in DEA model. The balance con-
strains proposed are introduced and respective balance matrixes
are developed and integrated into these balance constraints.
Additionally, resource constraints and other feasibility constraints
are introduced into base DEA model. With these three phases, the
final proposed integrated DEA–BSC model is developed.

The proposed three phases involved in developing the integrated DEA–

BSC model are semantically represented in the form of a framework and are
presented in Figure 13.3. The proposed framework provides a comprehen-
sive way to develop a holistic understanding of every perspective of BSC
impact on NPPM decision by examining the interactions among the afore-
mentioned dimensions. Furthermore, the framework assists in development
of proposed integrated DEA–BSC model.

13.4.3.1  Phase 1: Development of BSC Evaluation Index System

In this study, the BSC evaluation index system is established using the prin-
ciple of “SMART.” According to this principle, the perspectives of BSC are
determined and the respective evaluation indicators. Most of these evalu-
ation i­ndicators are determined based on subjective judgments. Table 13.3
presents a set of evaluation indicators that are considered in this study
along with card labels and objectives for evaluation. All the evaluation
indicators considered in the proposed BSC are identified from literature
An Integrated Multicriteria Decision-Making Model 333

Phase 1
Evaluation perspectives of
BSC
Evaluative Dimensions of
NPPM Customer perspective
Innovation perspective
Internal business perspective Estimation and
Strategic fit
Learning and growth validation of
Portfolio-innovation balance
perspective evaluation indicators
Cost-revenue estimation
Sustainability perspective for every perspective
Optimized resource allocation
Risk-uncertainty estimation Value contribution perspective
Risk perspective

Priority values— Phase 2

Upper and lower
limit values perspective and
indicators

Formulation of Determination of Formulation of

balance constraints balance constraints balance matrices
for DEA model

Phase 3

Integrated Formulation of
Formulation of
DEA-BSC resource constraints
base DEA model
model

FIGURE 13.3
Proposed framework for integrated DEA–BSC model.

(Kiranmayi and Mathirajan 2013). There are in total seven cards considered

in this study. For  every card accounted, there exist certain measures that
provide comprehensive evaluation of the NPPM performance.
The proposed BSC for NPPM considers seven perspectives. Out of seven
perspectives, four are original perspectives: customer, internal business
processes, learning and growth, financial, and three perspectives: innova-
tion perspective, sustainability perspective, and uncertainty perspective
are proposed explicitly for this study. The three proposed perspectives are
expected to emphasize on the new product project features. In addition, in
this study, we introduced an improved value contribution perspective rather
than financial perspective where project cycle and cost are also taken into
account. The details of these seven perspectives are presented as follows.

13.4.3.1.1  Customer Perspective (O1)

Customer perspective focuses on the responses of customers received through
market surveys. As the required evaluation indicators for ­customer perspec-
tive are representative of new products/projects, the data/values of projects
which are closely related or comparable to the project ideas are considered for
334 Big Data Analytics Using Multiple Criteria Decision-Making Models

TABLE 13.3
Project Evaluation Index System of the Proposed BSC for NPP
Card Evaluation
Label Perspective Evaluation Indicator Evaluation Objective
O1 Customer Customer satisfaction Ensure that project meets and
perspective Customer trust satisfy customer needs
Degree of customer need met
O2 Internal business Employee satisfaction Ensure that the implementation
perspective Supplier’s satisfaction of project plan, control, and
Project quality planning and other aspects, etc. optimize the
tracking organizational internal
processes
Internal communication
Congruence
Priority level
O3 Learning and Propriety position Ensure that the implementation
growth Platform for growth of the projects cultivates the
perspective Technical and market organizational core
durability technologies and
competitiveness
Team incentive
Knowledge accumulation
Project management
maturity
O4 Innovation Technology newness Ensure that project acquires the
perspective Process newness cutting edge to succeed
Market newness
O5 Sustainability Ecological Ensure that project/product
perspective Social sustainability is achieved
Economic
O6 Value contribution Project profitability Ensure that the projects are
perspective Project speed-up cycle completed in accordance with
Product sales desired objectives and provide
business value
O7 Uncertainty Operational–technical Ensure that projects are
perspective Organizational completed in accordance with
Financial time and specifications
without any uncertainties and
Marketing
delay in development
I1 Resources Investments –
Human resources
Machinery and equipment

evaluation. Accordingly, customer satisfaction, customer trust, and degree of

customer needs met are considered as the evaluation indicators of this per-
spective. Customer satisfaction addresses responsiveness, service satisfac-
tion, product quality, market stability, and so on. Customer trust guarantees
a part of demand volume, brand value, and customer relationship. Degree of
An Integrated Multicriteria Decision-Making Model 335

customer needs met represents whether there’s a necessity to change the

product features, or add additional features, limitations of product, and so on.

13.4.3.1.2  Internal Business Perspective (O2)

Internal business perspective measures the degree of share that the new
product/project aligns with the core competencies of the organization (i.e.,
missions and objectives). Accordingly, the evaluation indicators: employee
satisfaction, supplier satisfaction, project quality planning and tracking,
internal communication, congruence, and priority level are considered for
internal business perspective. Employee and supplier satisfaction measures
the commitment level and consistency for the development of new project.
Project quality planning and tracking ensures the required actions and
responsibilities of suppliers and employees are accurately and efficiently
carried out and this accounts for smooth operation. Internal communica-
tion and congruence measure the understanding and level of commitment
between top management and operational team members. Priority level of
project decides the level of commitment required for completion of project.

13.4.3.1.3  Learning and Growth Perspective (O3)

Learning and growth perspective ensures that the development process
does not suffer from knowledge asymmetry and transfer of knowledge.
In  ­present increasing global competition, an organization has to keep up
with the advancement in technology. For this case, the learning ability of
team and incentive of team are necessary to be accounted. For higher chances
of organization success, a project should ensure higher chances of market
and technology growth. To summarize, learning and growth perspective
ensures that objectives of all the other perspectives are satisfied.

13.4.3.1.4  Innovation Perspective (O4)

In case of a new project there exist different levels and angles of newness
brought by a project. An organization has to ensure the level of innovation
that is incorporated by a project to a portfolio in order to achieve a balanced
portfolio. This includes the measure of evaluation indicators such as tech-
nology newness, process newness, and market newness. The accounting for
time and market newness ensures that the portfolios do not contain highly
innovative projects and at the same time it should not be overwhelmed with
just incremental innovation projects. Thus, innovation perspective brings
the balance dimension to portfolio.

13.4.3.1.5  Sustainability Perspective (O5)

To the best of our knowledge, sustainability perspective is considered in
BSC index system for the first time by this study. Increasing global com-
petition and rapidly changing customer needs compete for the develop-
ment of new product. In order to succeed in competition and keep up with
trend development, sustainable products or technology are necessary.
336 Big Data Analytics Using Multiple Criteria Decision-Making Models

Sustainability perspective includes estimation of evaluation indicators

such as ecological, social, and economic. Though economical sustainability
is ensured by many organizations, achieving the balance between all the
evaluation indicators of sustainability is very important as this improves
the efficiency of NPPM and increases the probability of success of new
product.

13.4.3.1.6  Value Contribution Perspective (O6)

Financial perspective is a standard card in BSC index system. In this study,
along with evaluation indicators of financial perspective, the performance
indicator project cycle is introduced. In case of new product, first to ­market
or time to market plays a crucial role in success of new product. Thus,
­estimation of project cycle as a whole ensures the time to market. Thus, we
consider this perspective as value contribution perspective. In this perspec-
tive, along with project cycle, project cost and profitability are also measured.
This financial objective serves as main focus for the objectives and measures
in all the other scorecard perspectives.

13.4.3.1.7  Uncertainty Perspective (O7)

Uncertainty perspective ensures development process of new products with
minimized uncertainty. It includes operational–technical, organizational,
financial, and marketing uncertainties. The probabilities of these uncer-
tainty evaluation indicators are estimated using Bayesian belief networks.
Each and every evaluation indicator of uncertainty includes other measures
which determine the total probability. This uncertainty perspective, though
overlooked by many research studies, turned out to be the most prominent
and essential perspective to be considered for PES decision of NPPM (Oh
et al. 2012).
One of the key components of a BSC is establishment of baseline or
benchmark against which the performance is measured. But the baseline
or b
­ enchmark is hard to determine and can be misleading. Since DEA is
based on relative analysis, the projects are evaluated against each other.
In this study, BSC with DEA is integrated to overcome this constraint. Thus,
the evaluation indicators are considered accordingly in formulation of DEA
model.

13.4.3.2  Phase 2: Determination of Balance Constraints for DEA–BSC Model

To determine balance constraints for the proposed DEA–BSC model, we
first brief how BSC evaluation index system is integrated into DEA model.
Each  card (i.e., the evaluation perspective) represents a major dimension
of ­interest for the multiproject organization and the respective evaluation
indicators are accordingly considered as inputs (xi) and outputs (yn) for
DEA model. Accordingly, in the proposed BSC evaluation index system,
An Integrated Multicriteria Decision-Making Model 337

there exist seven output cards (evaluation perspectives) and one input card.
The base nomenclature of BSC index system is as follows:
For BSC index system:

S—No. of input cards

R—No. of output cards
I1…IS —Input of S cards
O1…OR—Output of R cards
[LIs , UIs —Lower and upper bounds for inputs
[LOr , UOr —Lower and upper bounds for output
To reflect the desired balance, we need to set limits regarded as upper
and lower bounds on each cards. To formulate the proposed DEA–BSC
model, we determine the balancing constraints and they are divided into
two groups. They are lower and upper bound constraints for every out-
put (OR) and input (IS) cards. The values are denoted as: [ LIs , U Is —Lower
and upper bounds for inputs; and [ LOr , UOr —Lower and upper bounds for
outputs.
With these, for a project Pj, the constraints (13.3) and (13.4) are included
to ensure the balance among the input and output cards, respectively.
Furthermore, these balance constraints depicts the importance and variabil-
ity of each card of the BSC index system.

LIs ≤
∑ ux
i ∈I s
i ij
≤ U Is ∀j , … (13.3)

∑ uxi
i ij

LOr ≤
∑ vy
nε Or
n nj
≤ UOr ∀j , … (13.4)

∑ vy
n
n nj

13.4.3.3  Phase 3: Development of the Proposed Integrated DEA–BSC Model

For the proposed integrated DEA–BSC model, the set of balancing constraints
(13.3) and (13.4) are added to the base CCR model (A). These constraints ensure
that balance objective is achieved among the output and input cards. Lower
bound constraints of output cards resemble the relative importance of that
particular output card, while upper bound constraints ensure the variability
balance. The formulation by default assumes nonnegativity constraints. With
these, for example, for a given project P1, the linear programming formulation of
the proposed single-level integrated DEA–BSC model (B) is represented below:
338 Big Data Analytics Using Multiple Criteria Decision-Making Models

The proposed single-level integrated DEA–BSC model (B)

max Z1 =
u, ν ∑v y n n1

n
such that,

∑ u x = 1,
i
i i1

∑ v y − ∑ u x ≤ 0,
n
n nj
i
i ij

L ∑ v y − ∑ v y ≤ 0,
Or n n1 n n1
n n ∈Or

∑ v y − U ∑ v y ≤ 0,
n ∈Or
n n1 Or
n
n n1

L ∑ u x − ∑ u x ≤ 0,
Is i i1 i i1
i i ∈I s

∑ u x − U ∑ u x ≤ 0.
i ∈I s
i i1 Is
i
i i1

13.5 Validation of the Proposed Integrated DEA–BSC

Model for NPPM
In order to validate (that is illustrating the workability) the proposed inte-
grated DEA–BSC model, a numerical example with 10 NPD projects is devel-
oped. Before getting into details, we first detail the preparatory steps for the
implementation of proposed integrated DEA–BSC model:
Data:

1. The data required for the evaluation indicators of each of the BSC
perspectives is suitably generated for 10 NPD projects. For gener-
ating values for each of the evaluation indicators (i.e., input and
output values), we assume a range for every evaluation indicators
and the same are presented in Table 13.4. Using the range defined
in Table 13.4, randomly generated values for each of the evaluation
indicators for the 10 NPD projects are given in Table 13.5.
2. In order to obtain weights for the evaluation perspectives and
respective evaluation indicators, we used AHP. For this, an appro-
priate questionnaire was prepared and project/product managers
were approached for their responses (in this study, we collected
An Integrated Multicriteria Decision-Making Model 339

TABLE 13.4
Range Considered for Generating Value for Each of the Evaluation Indicators
Range
Considered
Card Evaluation Evaluation for Indicator’s
Label Perspective Indicator Value
O1 Customer Customer satisfaction (O11) 5–10
perspective Customers’ trust (O12) 6–10
Degree of customer need met (O13) 6–10
O2 Internal business Employee satisfaction (O21) 6–10
perspective Supplier’s satisfaction (O22) 6–10
Project quality planning and tracking (O23) 7–10
Internal communication (O24) 5–10
Congruence (O25) 4–10
Priority level (O26) 5–10
O3 Learning and Propriety position (O31) 2–10
growth Platform for growth (O32) 2–10
perspective Technical and market durability (O33) 4–10
Team incentive (O34) 4–10
Knowledge accumulation (O35) 6–10
Project management maturity (O36) 4–10
O4 Innovation Technology newness (O41) 4–7
perspective Process newness (O42) 4–7
Market newness (O43) 4–7
O5 Sustainability Ecological (O51) 4–10
perspective Social (O52) 4–10
Economic (O53) 7–10
O6 Value contribution Project profitability(millions) (O61) 7–12
perspective Project speed-up cycle (years) (O62) 0.2–1
Product sales (thousand units) (O63) 20–50
O7 Uncertainty Operational–technical (O71) 0.6–0.9
perspective Organizational (O72) 0.5–0.9
(probability of Financial (O73) 0.7–0.9
success)
Marketing (O74) 0.7–0.9
I1 Resources Investments (millions) (I11) 70–100
Human resources (I12) –
Machinery and equipment (I13) –

responses from 104 managers). The responses obtained are given as

input to AHP and obtained the required weights. The AHP process
on the given primary data is presented in Appendix A. The weights
obtained for the evaluation perspectives and evaluation indicators
are presented in Table 13.6.
340 Big Data Analytics Using Multiple Criteria Decision-Making Models

TABLE 13.5
Project Wise, the Value of Evaluation Indicators
Project No.
Card Label 1 2 3 4 5 6 7 8 9 10
I11 73 82 96 87 75 78 96 89 83 91
O1 O11 6 5 7 7 8 6 8 7 9 8
O12 8 6 7 7 6 6 8 7 8 6
O13 8 6 7 8 9 8 7 8 9 7
O2 O21 7 7 6 6 8 7 8 6 8 7
O22 7 8 9 8 7 7 6 6 8 7
O23 7 8 9 9 8 8 7 7 7 9
O24 9 6 7 7 7 5 8 5 8 8
O25 4 6 8 10 5 6 7 4 4 8
O26 5 6 5 8 8 8 7 6 6 8
O3 O31 3 4 5 8 4 8 8 2 6 5
O32 3 5 5 8 9 9 10 5 2 6
O33 9 6 7 7 7 5 8 5 8 8
O34 4 6 8 10 5 6 7 4 4 8
O35 9 8 9 5 8 8 7 8 7 9
O36 7 4 5 8 4 9 8 4 6 5
O4 O41 4 6 7 6 5 6 5 5 5 6
O42 5 6 7 7 5 7 5 7 7 6
O43 5 6 7 5 5 6 7 6 5 7
O5 O51 7 6 5 9 5 8 8 7 8 8
O52 7 7 5 5 8 4 9 8 4 8
O53 7 7 8 7 8 8 8 7 7 8
O6 O61 8 9 11 10 12 11 10 12 8 9
O62 0.9 0.9 0.5 0.6 0.5 0.4 1 0.8 0.6 0.5
O63 25 24 29 38 45 35 40 42 35 44
O7 O71 0.7 0.6 0.9 0.8 0.8 0.6 0.7 0.7 0.9 0.9
O72 0.7 0.8 0.5 0.6 0.5 0.8 0.9 0.7 0.8 0.8
O73 0.7 0.8 0.8 0.7 0.9 0.7 0.8 0.9 0.7 0.8
O74 0.8 0.9 0.9 0.8 0.9 0.8 0.8 0.7 0.9 0.8

3. In order to implement the proposed integrated DEA–BSC model,

we need to determine the bounds for every evaluation perspective.
However, the bounds allocated vary from organization to organiza-
tion and depend on perspectives of the top management and proj-
ect manager. One can consider maximal tolerance level (i.e., 100%)
for limits of bounds, but in order to emphasize the importance of
bounds, we assume the tolerance to be 60%. Accordingly, Table 13.7
presents the lower and upper bounds of every evaluation perspec-
tives considered for the proposed integrated DEA–BSC model.
An Integrated Multicriteria Decision-Making Model 341

TABLE 13.6
Weights of the Evaluation Perspective and Evaluation Indictors
Card Evaluation First-Level Second-Level
Label Perspective Weights Evaluation Indicator Weights
O1 Customer 0.121 Customer satisfaction 0.0452
perspective Customer trust 0.0356
Degree of customers’ need met 0.0402
O2 Internal business 0.101 Employee satisfaction 0.00887
perspective Supplier’s satisfaction 0.00956
Project quality planning and tracking 0.04756
Internal communication 0.01548
Congruence 0.00157
Priority level 0.00896
O3 Learning and 0.096 Propriety position 0.00945
growth Platform for growth 0.00543
perspective Technical and market durability 0.04255
Team incentive 0.00923
Knowledge accumulation 0.02145
Project management maturity 0.00789
O4 Innovation 0.142 Technology newness 0.0756
perspective Process newness 0.0521
Market newness 0.0143
O5 Sustainability 0.068 Ecological 0.0149
perspective Social 0.0156
Economic 0.0375
O6 Value contribution 0.281 Project profitability 0.1457
perspective Project speed-up cycle 0.0478
Product sales 0.0875
O7 Uncertainty 0.191 Operational–technical 0.0145
perspective Organizational 0.0108
Financial 0.0895
Marketing 0.0762
I1 Resources 1 Investments 0.645
Human resources 0.143
Machinery and equipment 0.212

For generating the proposed integrated DEA–BSC model for any given
data, a LINGO set code has been developed and presented in Appendix B.
The data presented in Tables 13.5 through 13.7 can be given as input to the
LINGO set code and the proposed integrated DEA–BSC model can be gener-
ated for every project and solved using LINGO. Finally, the efficiency scores
of each of the projects along with relative ratings (rankings) are obtained and
the same are presented in Table 13.8. From Table 13.8, one can select the set of
­projects for NPP based on the efficiency score.
342 Big Data Analytics Using Multiple Criteria Decision-Making Models

TABLE 13.7
Lower and Upper Balance Bounds of Evaluation
Perspectives of DEA–BSC Model
Lower Upper
Card Label Evaluation Perspective Bound Bound
O1 Customer perspective 0.1 0.7
O2 Internal business perspective 0.2 0.8
O3 Learning and growth perspective 0.2 0.8
O4 Innovation perspective 0.1 0.7
O5 Sustainability perspective 0.1 0.7
O6 Value contribution perspective 0.3 0.9
O7 Uncertainty perspective 0.12 0.72

TABLE 13.8
Project wise the Efficiency Score and Relative
Rating Yielded by the Proposed DEA–BSC Model
Project Efficiency Score Rating
1 0.7432 7
2 0.9221 4
3 0.6043 8
4 1.0000 1
5 0.9910 2
6 1.0000 1
7 0.7496 6
8 0.9452 3
9 1.0000 1
10 0.8457 5

From the numerical example presented in this study, projects 4, 6, and

9 have almost same efficiency score; hence, they are having rating of “1”
(Table 13.8). The third column (i.e., rating) in Table 13.8 represents the rank-
ing of project. The projects with rating “1” have the maximum efficiency, so
a ­decision maker tends to pick up these projects.

13.6 Summary
As significance for the development of new products is rapidly increasing, it
becomes essential for an organization to have an effective and accurate deci-
sion-making process. In order to develop a successful new product, decision
maker/project manager needs to identify right set of new projects/products
and accordingly formulate a New Product Portfolio (NPP). Thus, the decisions
An Integrated Multicriteria Decision-Making Model 343

taken at the phase of PES play a significant role in NPPM performance. There
exist limited studies concentrating on (a) identifying factors/dimensions that
influence decision making, (b) development of decision-making model for
PES, and (c) improving the performance of NPPM. However, it is observed
through literature that NPPM is the weakest research area and identifica-
tion of explicit evaluative factors/dimensions is lacking (Cooper et al. 2001;
McNally et al. 2009). In this study, we address these gaps.
Accordingly, in this study, we have identified five different evaluative
dimensions (i.e., SF, PIB, Resource Allocation, CRE, and RUE) that are essen-
tial in the case of PES for NPPM. It is evident from literature that there exist
very limited number of studies that concentrate on development of MCDM
for PES in case of NPP. We make an attempt to employ all the evaluative
dimensions in the development of the MCDM model for NPPM. To the best
of our knowledge, this is the first study that considers all the five evaluative
dimensions simultaneously in the development of MCDM in this area.
We further briefed different methodologies employed for MCDM model
formulation in case of PES studies. From the discussion, it is identified that
there exist certain limitations in the existing methodologies. In present NPPM
scenario observed, it is essential to accommodate subjective data of a decision
maker into the developed methodology. This probed us to the development
of a methodology for NPPM in which qualitative and quantitative data can
be considered. Thus, we proposed to develop an integrated DEA–BSC model.
The methodology is based on relative evaluation of entities (projects or
portfolios), which is inspired by an integrated DEA–BSC model that was
first presented by Eilat et al. (2008). For this, the identified evaluative dimen-
sions are respectively reorganized into seven evaluation perspectives of BSC
index system. Accordingly, each evaluation indicators are measured using
different scales and metrics. The output of BSC is considered as input for
DEA, along with certain other inputs such as weights or priorities of perspec-
tives, obtained from AHP. Then, the proposed integrated DEA–BSC model
­estimates the efficiency score of each and every project and ranks them
accordingly. Thus, the model proposed in this study provides clarity and
accuracy regarding the subjective data associated to PES decisions.
In future work, we intend to extend the model for hierarchical level of BSC.
We also intend to introduce an accumulation function that takes care of inter-
actions between resources, benefit functions, and output functions. Finally, we
propose to extend this model by introducing dynamic nature into the problem.

Appendix A: An Analytical Hierarchical Process for BSC

The AHP is a popular MCDM methodology which has been applied vastly in
various fields. It was developed by Thomas Saaty (1960) for including qualitative
344 Big Data Analytics Using Multiple Criteria Decision-Making Models

variables for multicriteria decision-making model. AHP allows decision mak-

ers to model a complex problem in a hierarchical structure. In this method, a
simple hierarchical model consists of a goal, criteria, and alternatives. In order
to carry out an AHP analysis, the following iterative steps are involved:

1. The first and foremost activity in AHP is to analyze the problem,

identify criteria/indicators/alternatives. In our case, evaluation per-
spectives and evaluation indicators are considered as criteria and
subcriteria according to AHP terminology. Once these are identified,
evaluation perspectives and indicators are represented in hierarchi-
cal level network. The network diagram of AHP of this study is pre-
sented in Exhibit 13.1.

EXHIBIT 13.1
Hierarchical Network Diagram of BSC
O11: Customers’ satisfaction

O1: Customer perspective O12: Customers’ trust

O13: Degree of customers’ need met

O21: Employee satisfaction

O22: Supplier satisfaction
O23: Project quality planning and tracking
O2: Internal business perspective
O24: Internal communication
O25: Congruence
O26: Priority level

O31: Propriety position

O32: Platform for growth
O33: Technical and market durability
O3: Innovation perspective
O34: Team incentive
O35: Knowledge accumulation
NPPM
performance O36: Project management maturity

O41: Technology newness

O4: Learning and growth perspective O42: Process newness

O43: Market newness

O51: Ecological

O5: Sustainability perspective O52: Social

O53: Economic

O61: Project profitability

O6: Value contribution perspective O62: Project speedup cycle

O63: Product sales

O71: Operational–technical
O72: Organizational
O7: Uncertainty perspective
O73: Financial
O74: Marketing
An Integrated Multicriteria Decision-Making Model 345

2. Once the evaluation indictors are identified, one needs to obtain

pairwise comparison values for these evaluation indicators. A
questionnaire is prepared in order to obtain relative weights for all
the evaluation perspectives as well as evaluation indicators. These
values lead to dominance matrices which are called pairwise com-
parison matrices. Ratio scales are derived in the form of principle
eigenvectors from these matrices. The pairwise comparison matri-
ces of evaluation perspective with respect to goal and evaluation
indicators with respect to evaluation perspectives are presented in
Exhibits 13.2 and 13.3, respectively.
3. The expert opinion results are either aggregated or considered inde-
pendently to obtain scores of pairwise comparison. Through pair-
wise comparison matrices, one can obtain the priorities of evaluation
perspectives and indicators by normalizing the matrices. The prior-
ity vectors along with pairwise normalized matrices are presented
in Exhibits 13.4 and 13.5.
4. The overall priority vector for evaluation indicators is obtained by
vector multiplication of the previously calculated perspective and
indicator priority vectors. The resultant overall priority vector for
indicator is presented in Exhibit 13.6.

EXHIBIT 13.2
Pairwise Comparison Matrix of Evaluation Perspectives
with Respect to NPPM Performance
O1 O2 O3 O4 O5 O6 O7
O1 1 1/4 6 4 1 1/3 1/5
O2 4 1 1/4 1/3 2 1/3 3
O3 1/6 4 1 2 5 1/3 1
O4 1/4 3 1/2 1 1 1/5 1/3
O5 1 1/2 1/5 1 1 1/5 1/3
O6 3 3 3 5 5 1 1/3
O7 5 1/3 1 3 3 3 1

EXHIBIT 13.3
Pairwise Comparison Matrix of Evaluation
Indicators with Respect to Uncertainty
Perspective
O61 O62 O63 O64
O61 1 2 1 1
O62 1/2 1 1/2 1/4
O63 1 2 1 2
O64 1 4 1/2 1
346 Big Data Analytics Using Multiple Criteria Decision-Making Models

EXHIBIT 13.4
Pairwise Normalized Matrix of Evaluation Perspectives with Respect to
NPPM Performance
O1 O2 O3 O4 O5 O6 O7 Priorities
O1 0.069 0.021 0.502 0.245 0.056 0.062 0.032 0.141
O2 0.277 0.083 0.021 0.020 0.111 0.062 0.484 0.151
O3 0.012 0.331 0.084 0.122 0.278 0.062 0.161 0.150
O4 0.017 0.248 0.042 0.061 0.056 0.037 0.054 0.074
O5 0.069 0.041 0.017 0.061 0.056 0.037 0.054 0.048
O6 0.208 0.248 0.251 0.306 0.278 0.185 0.054 0.219
O7 0.347 0.028 0.084 0.184 0.167 0.556 0.161 0.218

EXHIBIT 13.5
Pairwise Normalized Matrix of Evaluation
Indicators with Respect to Uncertainty Perspective
O61 O62 O63 O64 Priorities
O61 0.2 0.13 0.17 0.31 0.2
O62 0.2 0.13 0.17 0.08 0.14
O63 0.4 0.25 0.33 0.31 0.32
O64 0.2 0.5 0.33 0.31 0.34

EXHIBIT 13.6
Overall Priorities of Uncertainty Evaluation Indicators
with Respect to Uncertainty Perspective
Perspective Overall
Priority Priorities Priorities
O61 0.219 0.2 0.0436
O62 0.14 0.0305
O63 0.32 0.0698
O64 0.34 0.0741

5. The matrices are validated by measuring consistency index and con-

sistency ratio. This validation is considered to remove the response
bias associated with pairwise comparisons. For this, one needs to
calculate the maximum eigenvalue, λmax, and the eigenvector, w, for
the matrix. The consistency is performed by calculating the consis-
tency index (CI) and consistency ratio (CR):

λ max − n
CI =
n−1

CI
CR = ,
RI
An Integrated Multicriteria Decision-Making Model 347

where n is the number of items being compared in the matrix, and

RI is random index. If CR is less than 0.1, the experts’ judgments are
consistent. If the consistency test is not passed, the part of the ques-
tionnaire must be done again.

Appendix B
A LINGO Set Code developed for generating the proposed integrated
­DEA–­BSC model:
MODEL:
! DEA-BSC MODEL FOR NPPM ;
! PLEASE NOTE: THIS CODE ISFOR ILLUSTRATION PURPOSE;
! Part of the sample data is presented in this code
SETS:
   DMU/P1 P2 P3 P4 P5 P6 P7 P8 P9 P10/: ! 10 NEW PRODUCT PROJECTS;
     SCORE; ! Each decision making unit in this case new
product project has a score to be computed;
   FACTOR/I1 O11 O12 O13 O21 O22 O23 O24 O25 O26 O31 O32 O33 O34
O35 O36 O41 O42 O43 O51 O52 O53 O61 O62 O63 O71 O72 O73 O74/;
! ALL THE EVALUATION INDICATORS ARE CONSIDERED AS OUTPUT AND
INVESTMENTS AS INPUT;
DXF( DMU, FACTOR):   F; ! F( I, J) = Jth factor of DMU I;
ENDSETS
DATA:
  NINPUTS = 1;   ! The first NINPUTS factors are inputs;
!     The inputs,  the outputs;
F  = 73 6 8 8 -------- 0.7 0.8
82 5 6 6 -------- 0.8 0.9
96 7 7 7 -------- 0.8 0.9
87 7 7 8 -------- 0.7 0.8
75 8 6 9 -------- 0.9 0.9
78 6 6 8 -------- 0.7 0.8
96 8 8 7 -------- 0.8 0.8
89 7 7 8 -------- 0.9 0.7
83 9 8 9 -------- 0.7 0.9
91 8 6 7 -------- 0.8 0.8;
ENDDATA
!----------------------------------------------------------;
SETS:
! Weights used to compute DMU I’s score;
DXFXD(DMU,FACTOR) : W;
ENDSETS
DATA
W = 0.645
348 Big Data Analytics Using Multiple Criteria Decision-Making Models

0.0356
0.0402
0.00887
|
|
|
0.0108
0.0895
0.0762;
ENDDATA
!------------------------------------------;
SETS:
! Lower Bounds used to compute DMU I’s score;
DXFXA(DMU,FACTOR) : LB;
ENDSETS
DATA
LB= 0
0.1
0.1
0.1
0.2
|
|
|
0.12
0.12
0.12
0.12;
ENDDATA
!-------------------------------------------;
SETS:
! Upper Bounds used to compute DMU I’s score;
DXFXB(DMU,FACTOR) : UB;
ENDSETS
DATA
UB= 56
0.7
0.7
0.7
0.8
|
|
|
0.72
0.72
0.72
0.72;
ENDDATA
! The Model;
! Try to make everyone’s score as high as possible;
An Integrated Multicriteria Decision-Making Model 349

MAX = @SUM( DMU: SCORE);

! The LP for each DMU to get its score;
@FOR( DMU( I):
SCORE( I) = @SUM( FACTOR(J)|J #GT# NINPUTS:
F(I, J)* W(I, J));
! Sum of inputs(denominator) = 1;
@SUM( FACTOR( J)| J #LE# NINPUTS:
F( I, J)* W( I, J)) = 1;
! Using DMU I’s weights, no DMU can score better than 1;
@FOR( DMU( K):
@SUM( FACTOR( J)| J #GT# NINPUTS: F( K, J) * W( I, J))
<= @SUM( FACTOR( J)| J #LE# NINPUTS:
F( K, J) * W( I, J)));
! Balance Constraints for outputs;
(LB(I,J)* @SUM( FACTOR( J)| J #GT# NINPUTS: F( I, J) * W( I, J)))
<= @SUM( FACTOR( J)| J #GT# NINPUTS: F( I, J) * W( I, J));
@SUM( FACTOR( J)| J #GT# NINPUTS: F( I, J) * W( I, J))
<= (UB(I,J)* @SUM( FACTOR( J)| J #GT# NINPUTS: F( I, J) *
W( I, J)));
! Similarly Balance Constraint for input can be introduced;
  );
END

References
Abbassi, M., Ashrafi, M., and Tashnizi, E.S. 2014. Selecting balanced portfolios of
R&D projects with interdependencies: A cross-entropy based methodology.
Technovation 34 (1): 54–63.
Andrews, K.Z. 1996. Two kinds of performance measures. Harvard Business Review
74 (1): 8–9.
Asosheh, A., Nalchigar, S., and Jamporazmey, M. 2010. Information technology proj-
ect evaluation: An integrated data envelopment analysis and balanced score-
card approach. Expert Systems with Applications 37 (8): 5931–5938.
Ayağ, Z. and Özdemr, R.G. 2007. An analytic network process-based approach to
concept evaluation in a new product development environment. Journal of
Engineering Design 18 (3): 209–226.
Banker, R.D., Chang, H., and Pizzini, M.J. 2004. The balanced scorecard: Judgmental
effects of performance measures linked to strategy. Accounting Review 79 (1): 1–23.
Banker, R.D., Potter, G., and Srinivasan, D. 2000. An empirical investigation of an
incentive plan that includes non-financial performance measures. Accounting
Review 75 (1): 65–92.
Bhattacharyya, R., Kumar, P., and Kar, S. 2011. Fuzzy R&D portfolio selection of inter-
dependent projects. Computers & Mathematics with Applications 62 (10): 3857–3870.
Blundell, R., Griffith, R., and Van Reenen, J. 1999. Market share, market value and
innovation in a panel of British manufacturing firms. The Review of Economic
Studies 66 (3): 529–554.
350 Big Data Analytics Using Multiple Criteria Decision-Making Models

Brown, S.L. and Eisenhardt, K.M. 1995. Product development: Past research, pres-
ent findings, and future directions. Academy of Management Review 20 (2):
343–378.
Carbonell-Foulquié, P., Munuera-Alemán, J.L., and Rodrıguez-Escudero, A.I. 2004.
Criteria employed for Go/no-Go decisions when developing successful highly
innovative products. Industrial Marketing Management 33 (4): 307–316.
Chames, A., Cooper, W.W., and Rhodes, E. 1978. Measuring the efficiency of decision
making units. European Journal of Operational Research 2: 429–444.
Chames, A., Cooper, W.W., and Rhodes, E. 1989. Cone ratio data envelopment analy-
sis and multi-objective programming. International Journal of Systems Science
20 (7): 1099–1118.
Chan, S. and Ip, W. 2010. A scorecard-Markov model for new product screening deci-
sions. Industrial Management & Data Systems 110 (7): 971–992.
Chang, Y.T., Park, H.S., Jeong, J.B., and Lee, J.W. 2014. Evaluating economic and envi-
ronmental efficiency of global airlines: A SBM-DEA approach. Transportation
Research Part D: Transport and Environment 27: 46–50.
Chao, R.O. and Kavadias, S. 2008. A theoretical framework for managing the new
product development portfolio: When and how to use strategic buckets.
Management Science 54 (5): 907–921.
Chesbrough, H.W. and Teece, D.J. 2002. Organizing for Innovation: When Is Virtual
Virtuous? Harvard Business School Publishings, MA.
Chiang, T.-A. and Che, Z.H. 2010. A fuzzy robust evaluation model for selecting and
ranking NPD projects using Bayesian belief network and weight-restricted
DEA. Expert Systems with Applications 37 (11): 7408–7418.
Chien, C.-F. 2002. A portfolio-evaluation framework for selecting R&D projects. R&D
Management 32: 359–368.
Chiu, C.-Y. and Park, C.S. 1994. Fuzzy cash flow analysis using present worth crite-
rion. The Engineering Economist 39 (2): 113–138.
Cook, W.D. and Seiford, L.M. 1978. R&D project selection in a multi-dimensional
environment: A practical approach. Journal of the Operational Research Society
21: 29–37.
Cooper, R.G. 1994. Perspective third-generation new product processes. Journal of
Product Innovation Management 11 (1): 3–14.
Cooper, R.G., Edgett, S.J., and Kleinschmidt, E.J. 1997. Portfolio management in
new product development: Lessons from the Leaders-II. Research-Technology
Management 40 (6): 43.
Cooper, R.G., Edgett, S., and Kleinschmidt, E. 2001. Portfolio management for new
product development: Results of an industry practices study. R&D Management
31 (4): 361–380.
Cooper, R.G., Edgett, S.J., and Kleinschmidt, E.J. 2004. Benchmarking best NPD
­practices-1. Research-Technology Management 47 (1): 31–43.
Crawford, C.M. and Di Benedetto, C.A. 2008. New Products Management. Tata
McGraw-Hill Education, New York, NY.
Duarte, B.P. and Reis, A. 2006. Developing a projects evaluation system based
on multiple attribute value theory. Computers & Operations Research 33 (5):
1488–1504.
Eilat, H., Golany, B., and Shtub, A. 2006. Constructing and evaluating balanced port-
folios of R&D projects with interactions: A DEA based methodology. European
Journal of Operational Research 172 (3): 1018–1039.
An Integrated Multicriteria Decision-Making Model 351

Eilat, H., Golany, B., and Shtub, A. 2008. R&D project evaluation: An integrated DEA
and balanced scorecard approach. Omega 36 (5): 895–912.
Feyzioğlu, O. and Büyüközkan, G. 2006. Evaluation of new product development
projects using artificial intelligence and fuzzy logic. International Conference on
Knowledge Mining and Computer Science 11: 183–189.
Fitzsimmons, J.A., Kouvelis, P., and Mallick, D.N. 1991. Design strategy and its inter-
face with manufacturing and marketing: A conceptual framework. Journal of
Operations Management 10 (3): 398–415.
Gates, W. 1999. Business @ the Speed of Sound. Warner Books, New York, NY.
Ghasemzadeh, F. and Archer, N.P. 2000. Project portfolio selection through decision
support. Decision Support Systems 29 (1): 73–88.
Graves, S.B., Ringuest, J.L., and Case, R.H. 2000. Formulating optimal R&D
­portfolios. Research-Technology Management 43 (3): 47–51.
Griffin, A. and Hauser, J.R. 1996. Integrating R&D and marketing: A review and anal-
ysis of the literature. Journal of Product Innovation Management 13 (3): 191–215.
Gutjahr, W.J., Katzensteiner, S., Reiter, P., Stummer, C., and Denk, M. 2010. Multi-
objective decision analysis for competence-oriented project portfolio selection.
European Journal of Operational Research 205 (3): 670–679.
Halouani, N., Chabchoub, H., and Martel, J.-M. 2009. Promethee-md-2t method for
project selection. European Journal of Operational Research 195 (3): 841–849.
Ibbs, C.W., Reginato, J., and Kwak, Y.H. 2004. Developing project management capa-
bility: Benchmarking, maturity, modeling, gap analyses and ROI studies.
In: The Wiley Guide to Managing Projects, John Wiley & Sons, Inc., Hoboken, NJ,
pp. 1214–1233.
Kahraman, C., Büyüközkan, G., and Ateş, N.Y. 2007. A two phase multi-attribute
decision-making approach for new product introduction. Information Sciences
177 (7): 1567–1582.
Kaplan, R.S. and Norton, D.P. 1992. The balance scorecard—Measures that drive
­performance. Harvard Business Review 70 (1): 71–79.
Kaplan, R.S. and Norton, D.P. 2001. Transforming the balanced scorecard from per-
formance measurement to strategic management: Part I. Accounting Horizons 15
(1): 87–104.
Kiranmayi, P. and Mathirajan, M. 2013. A theoretical framework for project evalu-
ation and selection in new product management. Proceedings of International
Conference on Sustainable Innovation and Successful Product Development for a
Turbulent Global Market, Indian Institute of Technology-Madras, Chennai, India.
Krishnan, V. and Ulrich, K.T. 2001. Product development decisions: A review of the
literature. Management Science 47 (1): 1–21.
Kumar, D.U., Saranga, H., Ramírez-Márquez, J.E., and Nowicki, D. 2007. Six sigma
project selection using data envelopment analysis. The TQM Magazine 19 (5):
419–441.
Kyparisis, G.J., Gupta, S. K., and Ip, C.-M. 1996. Project selection with discounted
returns and multiple constraints. European Journal of Operational Research
94 (1): 87–96.
Lee J.W. and Kim H.S. 2000. Using analytic network process and goal program-
ming for interdependent information system project selection. Computers &
Operations Research 27: 367–382.
Liao, S.-H. 2005. Expert system methodologies and applications—A decade review
from 1995 to 2004. Expert Systems with Applications 28 (1): 93–103.
352 Big Data Analytics Using Multiple Criteria Decision-Making Models

Loch, C.H. and Kavadias, S. 2002. Dynamic portfolio selection of NPD programs
using marginal returns. Management Science 48 (10): 1227–1241.
Lockett, G. and Stratford, M. 1987. Ranking of research projects: Experiments with
two methods. Omega 15 (5): 395–400.
Mahmoodzadeh, S., Shahrabi, J., Pariazar, M., and Zaeri, M.S. 2007. Project selec-
tion by using fuzzy AHP and TOPSIS technique. World Academy of Science,
Engineering and Technology 30: 333–338.
Mavrotas, G., Diakoulaki, D., and Caloghirou, Y. 2006. Project prioritization under
policy restrictions: A combination of MCDA with 0–1 programming. European
Journal of Operational Research 171 (1): 296–308.
McNally, R.C., Durmusoglu, S.S., Calantone, R.J., and Harmancioglu, N. 2009.
Exploring new product portfolio management decisions: The role of managers’
dispositional traits. Industrial Marketing Management 38 (1): 127–143.
Melachrinoudis, E. and Kozanidis, G. 2002. A mixed integer knapsack model for
allocating funds to highway safety improvements. Transportation Research Part
A: Policy and Practice 36 (9): 789–803.
Metaxiotis, K. and Liagkouras, K. 2012. Multi objective evolutionary algorithms for
portfolio management: A comprehensive literature review. Expert Systems with
Applications 39 (14): 11685–11698.
Mohanty, R.P., Agarwal, R., Choudhury, A.K., and Tiwari, M.K. 2005. A fuzzy ANP-
based approach to R&D project selection: A case study. International Journal of
Production Research 43 (24): 5199–5216.
Oh, J., Yang, J., and Lee, S. 2012. Managing uncertainty to improve decision-making
in NPD portfolio management with a fuzzy expert system. Expert Systems with
Applications 39 (10): 9868–9885.
Osawa, Y. and Murakami, M. 2002. Development and application of a new methodol-
ogy of evaluating industrial R&D projects. R&D Management 32 (1): 79–85.
Ozer, M. 1999. A survey of new product evaluation models. Journal of Product
Innovation Management 16 (1): 77–94.
Ozer, M. 2005. Factors which influence decision making in new product evaluation.
European Journal of Operational Research 163 (3): 784–801.
Patah, L.A. and De Carvalho, M.M. 2007. Measuring the value of project manage-
ment. PICMET 2007 Proceedings, Portland, OR, pp. 2038–2042.
Remera, D.S., Stokdykb, S.B., and Driel, M.V. 1993. Survey of project evaluation
techniques currently used in industry. International Journal of Production
­
Economics 32: 103–l15.
Roll, Y. and Golany, B. 1993. Alternate methods of treating factor weights in DEA.
Omega 21 (1): 99–109.
Ronkainen, I.A. 1985. Criteria changes across product development stages. Industrial
Marketing Management 14 (3): 171–178.
Saaty, T.L. 1977. A scaling method for priorities in hierarchical structures. Journal of
Mathematical Psychology 15(3): 234–281.
Schaffer, J.D. 1985. Some Experiments in Machine Learning Using Vector Evaluated Genetic
Algorithms. Vanderbilt Univ., Nashville, TN.
Seiford, L.M. 1996. Data envelopment analysis: The evolution of the state of the art
(1978–1995). Journal of Productivity Analysis 7: 99–137.
Thieme, R.J., Song, M., and Calantone, R.J. 2000. Artificial neural network decision
support systems for new product development project selection. Journal of
Marketing Research 37 (4): 499–507.
An Integrated Multicriteria Decision-Making Model 353

Ulrich, K.T. and Eppinger, S.D. 2004. Product Design and Development, 3rd edn.
McGraw Hill, NY.
Wang, J. and Hwang, W.-L. 2007. A fuzzy set approach for R&D portfolio selection
using a real options valuation model. Omega 35 (3): 247–257.
Wey, W.-M. and Wu, K.-Y. 2007. Using ANP priorities with goal programming in
resource allocation in transportation.  Mathematical and Computer Modelling 46
(7): 985–1000.
Yahaya, S.-Y. and Abu-Bakar, N. 2007. New product development management issues
and decision-making approaches. Management Decision 45 (7): 1123–1142.
Yurdakul, M. 2003. Measuring long-term performance of a manufacturing firm using
analytic network process (ANP) approach. International Journal of Production
Research 41 (11): 2501–2529.
Index

A Open Source license, 81

Spark, 88–89, 136
ACF, see Autocorrelation factor
Storm, 88–89
ACMM, see Analytics capability
Arcs, 212
maturity model
Area under curve (AUC), 123, 190
AHP, see Analytic hierarchy process
ARIMA method, see Autoregressive
AIC, see Akaike information criteria
integrated moving average
AIM, see Aspiration-level interactive
method
method
Artificial neural networks (ANN), 88,
Air Force, 4
105–111, 121, 180, 184–186
Akaike information criteria (AIC), 140
Aspiration-level interactive method
Akshaya Patra Midday Meal Routing
(AIM), 25
and Transportation Algorithm
Aspiration levels, 51
(AMRUTA), 74
Attributes, 24
Altruistic disposition, 2
AUC, see Area under curve
Amazon, 69
Autocorrelation, 137
American approaches, 39
Autocorrelation factor (ACF), 137
AMRUTA, see Akshaya Patra Midday
Autoregressive integrated moving
Meal Routing and Transportation
average method (ARIMA
Algorithm
method), 136, 137–138, 138–139
Analytical network processing (ANP),
285, 327
B
Analytic hierarchy process (AHP), 25,
37–39, 247, 271, 283, 321, 327 BA, see Business analytics
for BSC, 343–347 Backward pass process, 185
Analytics, 68, 69 Bad loans, 176–177
big data, 82–84 Balance
data capture, store, prepare, analyze, critical NPPM dimension, 320
and share, 76–78 determination of balance constraints
data store, 79–80 for DEA–BSC model, 332,
descriptive, 71–72 336–337
life cycle, 76, 77 Balanced scorecard (BSC), 317, 326–327
NoSql systems, 78–79 for achieving strategic and balanced
predictive, 72–73 NPP, 328–329
prescriptive, 73–74 AHP for, 343–347
types solution and value-add, 68 customer perspective, 333–335
Analytics capability maturity model index system development,
(ACMM), 72 332–336, 337
ANN, see Artificial neural networks innovation perspective, 335
ANP, see Analytical network processing internal business perspective, 335
Apache learning and growth perspective, 335
Flume, 78 sustainability perspective, 335–336
Hadoop, 80, 88–89, 136 uncertainty perspective, 336
Kafka, 78 value contribution perspective, 336

355
356 Index

Banks impact on society, 176 CART, see Classification and regression

Base CCR model, 330–331 trees
Baseline data base, 5 Caselet approach, 272
Batch processing technologies, 88–89 CCE, see Coca-Cola Enterprises
Bayesian information criteria (BIC), 140 CCR model, see Chames, Cooper, and
BCLP, see Bicriteria linear program Rhodes model
Bellman’s optimality principle, 212 CDC, see Centre for Disease Control
Best compromise solution, 44 Centre for Disease Control (CDC), 69
BI, see Business intelligence Chames, Cooper, and Rhodes model
BIC, see Bayesian information criteria (CCR model), 330
Bicriteria linear program (BCLP), 30; base CCR Model, 330–331
see also Linear programming Checklists, 326
(LP) Chennai city network data, 226
Bicriteria optimization of microgrids, Chennai network, multiobjective model
169–171 to, 234–235
Big data, 67, 74–75, 88, 213 nondominated solutions, 241
analytics, 82–84, 88, 136 optimal solution, 239–240
Hadoop ecosystem, 80–82 Pareto-optimal solutions, 241
limitations of traditional Choice translating algorithm, 25; see also
technologies for, 75 Machine-learning algorithms
MapReduce paradigm, 80, 81 Choquet Integral (CI), 102, 103, 284
MCDM in, 57–58 CI, see Choquet Integral; Consistency
preparation and analyzing in, 80 index
Binary classification problem, 119 Classification and regression trees
Binary pairwise comparison, 51 (CART), 180
Blade MRC, 6 CMI, see Customer minutes interrupted
Boosting, 120 Coca-Cola Enterprises (CCE), 74
Borda count, 123–124 Collective exhaustivity, 33
criteria weights using rating and, 129 Column family store, 79
method, 128, 129 Commercial banks, 176
Borrowers, 181 Comparative trade-off ratio, 51
achieving age diversification Compromise
among, 195 programming, 24, 44, 49–50
hard constraints, 195 solution, 50
late payment frequency of, 194 Computer generated data, 9
number of dependents associated Confusion matrix, 122
with, 194 Consistency index (CI), 55, 346
open credits associated with, 194 Consistency ratio (CR), 346
BSC, see Balanced scorecard Constant returns to scale (CRS), 306
Business analytics (BA), 58, 67, 71 Constraints, model
descriptive analytics, 71–72 nonnegativity restriction, 131–132
predictive analytics, 72–73 restriction on workforce capacity, 131
prescriptive analytics, 73–74 worker type requirement, 131
Business intelligence (BI), 68, 82–84 Constraints, set of, 257–260
Continuous performance evaluation of
employees (CPEE), 270
C
bivariate analysis and the significant
Capital expenditure (CAPEX), 309, 310 demographic variables, 278
Caret package, 198 consolidated list of variables, 273–275
Index 357

criteria based on perspectives of both prepare, 76–78

SEs and PMs, 276–277 quality assessment methods, 246
factor/criterion-wise list of with respect to distance and travel
variables, 280 times, 226
framework for CPEE, 281–282 scrubbing, 183
MCDM-based modeling framework set, 148
development for, 288–291 share, 76–78
MCDM modeling framework store, 76–80
validation for, 291–295 waiting time, 232
multiple variables/multiple criteria Data and analysis, 308
identification for, 271–282 control variables, 310
PAS commences, 271–272 environmental expenditure,
t-test, 277–279 309–310
Conversion binary, 258 manufacturing performance, 309
Conveyor design regression model, 310–311
general model for, 9 sample and data collection, 308
pre-and postfire, 10 Data description, 124, 182–183
COS, see Set of countries power outage data, 125, 127
Cost–benefit analyses, 324–325 weather data, 124–125, 126
Cost–revenue estimation (CRE), 321 Data envelopment analysis (DEA), 24,
Country-specific risk factors, 249 52–53, 284, 285, 303, 304–306
CPEE, see Continuous performance approach for ranking and
evaluation of employees prioritization of portfolios, 329
CR, see Consistency ratio base CCR Model, 330–331
CRE, see Cost–revenue estimation data and analysis, 308–311
Crisp boundary, LP-based classifier literature review, 304
with, 93–96 as MCDM Tool, 306–307
Criteria, 23, 24 methodologies, 317, 325
space, 28 proposed integrated DEA–BSC
Criterion, 23, 24 model for NPPM, 328–338
CRS, see Constant returns to scale relationship between environmental
Currency exchange rates, objective and financial performance,
function considering, 260–261 307–308
Customer acceptance, 319 Data-mining
Customer minutes interrupted (CMI), 125 techniques, 73, 156–157
tools, 156
datanodes, 79, 80
D
DCs, see Distribution centers
Data, 338–340 DCSE, see Demographic characteristic
analyze, 76–78 of SE
capture, 76–78 DEA, see Data envelopment analysis
Chennai road network topology, 227 Debt-to-income ratio, 193
cleansing process, 183–184 Decision maker (DM), 22
collection, 262, 308–310 Decision-making units (DMUs), 304, 325
data-driven decision-making flow Decision(s), 249; see also Multi-criteria
diagram, 70 decision making (MCDM)
data-driven decision process, 118 analysis, 22
distance between nodes, 228–229 decision-making problems, 22
FFT, 230–231 space, 28, 30, 44
358 Index

Decision(s) (Continued) Eight-member relocation team, 5

trees algorithms, 120 ELECTRE, see Elimination Et Coix
variables, 28 Traduisant la Realite or
DEFRA, see Department for Environment, elimination
Food, and Rural Affairs Electrical distribution system, 117
Demand Electric power
constraints, 161 outage, 114
fulfillment, 249 supply industry, 116
uncertainty in loan application Electric supply distribution network,
requests, 196 115–119
Demographic characteristic of SE Electric utility companies, 114
(DCSE), 277 Electric utility service management
Demographic score (DS), 288, 292–293 criteria values for prediction
Department for Environment, Food, and models, 128
Rural Affairs (DEFRA), 309 data description, 124–125, 126, 127
Descriptive analytics, 68, 71–72, 118 experimental results, 125
Deviational variables, 46 layout of electric supply distribution
Digital Equipment Corporation, 247 network, 115–119
Dijkstra’s label-setting algorithm, 214 past research, 115
Dijkstra’s static shortest path prediction models using MCDM
algorithm, 212 techniques, 122–124
Discrete multicriteria selection predictive analysis, 119–121
problems (Discrete MCSP), 245 results from L2 metric, 129–130
Discriminant analysis, 180 scaled results, 129
Distributed file systems, 79–80 smart staffing, 130–132
Distribution, 118 Elimination Et Coix Traduisant
binary, 258 la Realite or elimination
line networks, 118 (ELECTRE), 25, 39, 283
Distribution centers (DCs), 245 ELECTRE I, 40
DM, see Decision maker methods, 27, 39–41
DMUs, see Decision-making units TRI method, 40–41
Document database, 79 Elitist multiobjective evolutionary
Dominated model, 144 algorithm, 214
Dominated option, 29 Energy
Dominated solution, 29 design optimization, 157
Downtime, 12, 114 information administration, 116
DS, see Demographic score operation optimization, 160
Duration, 114 storages, 158–159
of outage, 124 Energy models for microgrids, 158
of power interruption, 118 bicriteria optimization of microgrids,
Dynamic decision process, 318 169–171
Dynamic programming, 324 energy operations model, 159–164
microgrids design optimization,
164–166
E
multiperiod energy model, 166–168
Efficiency frontier, 325, 330 Energy operation model, 157, 159
Efficient set, 29 energy operation optimization, 160
Efficient solution, 29–31 microgrid operations optimization
determination, 31–32 example, 161–164
Index 359

objective function, 161 Forward pass risk type, 185

solution for systems operations FP, see False positives
example, 163, 164 FPR, see False-positive rate
Engine overhaul functions, 5 FPS, see Final performance score
Enterprise resource planning (ERP), Free-flow travel time (FFT), 226–227
67–68, 246 French approach, 39
Entropy, 186–187 Frequency, 114
Environmental expenditure, 309–310 F-test, 311
Environmental performance, relationship Fuzzy AHP (FAHP), 284
between financial and, 307–308 Fuzzy MCDM approach, 285
ε-approach, see ε-constraint method Fuzzy measure, proposed LP-based
ε-constraint method, 195–196, 223 classifiers, 103–105
ERP, see Enterprise resource planning Fuzzy R&D portfolio selection
ETL, see Extraction, transformation, and model, 321
loading Fuzzy set theory, 43
European approach, 39
Evolutionary algorithms, 326; see also
G
Machine-learning algorithms
Extraction, transformation, and loading GDM, see Group decision-making
(ETL), 70 Generalized reduced gradient method
(GRG method), 48–49
Generators
F
generator-purchasing costs, 164
Facility-specific risk factors, 249 microgrids, 158–159
FAHP, see Fuzzy AHP GFS, see Google file system
False negatives (FN), 123 Global positioning system (GPS), 156
error, 191 Global supply chain model (GSCM), 247
False-positive rate (FPR), 190 Global supply chain networks
False positives (FP), 122 case study, 262–265
error, 190 data collection, 262
Fast moving consumer goods (FMCG), 136 digital equipment corporation, 247
Feed-forwarding risk type, 185 global supply chain management,
FFT, see Free-flow travel time 243–244
Final performance score (FPS), 290 MCDM problems, 244–245
Financial model scaling, 254–255
indices, 324–325 multicriteria mixed-integer linear
institutions, 176 programming models, 245–246
models, 324–325 multicriteria model, 248–249
performance, 319 multiperiod model, 250–253
perspective, 336 objective function, 254, 260–261
relationship between environmental problem description, 247–248
and financial performance, set of constraints, 257–260
307–308 set of goal constraints, 255–257
Floyd’s algorithm, 9 supply chain network, 249–250
FMCG, see Fast moving consumer goods systematic optimization, 246–247
FN, see False negatives variables, 253–254
FORTRAN GlusterFS, 79
code, 9 Goal constraints, 45–46
programs, 14 set of, 255–257
360 Index

Goal of problem, 23 Hype around analytics, 68

Goal programming (GP), 24, 45, 247, 325; Hypothetical options, 29
see also Linear programming
(LP)
I
component, 254
formulation, 46–48 Ideal solution, 30, 49, 124, 285
multiple objectives, 90 IIASA, see International Institute of
partitioning algorithm for Applied Systems Analysis
preemptive GP, 48–49 IIT, see Indian Institute of Technology
solution techniques, 132 Incomplete information, 53, 54, 58, 70
Google file system (GFS), 79 Index specification, 51
Google’s flu trends, 69 Indian Institute of Technology (IIT),
GP, see Goal programming 226–227, 238
GPS, see Global positioning system Infinitely scalable approach, 75
GPSYS software, 57 Information gain, 186–187
Graph database, 79 Information Technology (IT), 269
GRG method, see Generalized reduced Input minimization linear programming
gradient method problem, 305, 306
Group decision-making (GDM), 22 Integer GP, 48
Group decisions, 55–56 Integer programming (IP), 179, 324, 325
Group technology cell (GTC), 6 Integrated DEA–BSC model for NPPM,
GSCM, see Global supply chain model 328, 332
BSC for achieving strategic and
balanced NPP, 328–329
H
DEA approach for ranking and
Hadoop, 70 prioritization of portfolios,
Apache, 80, 136 329–331
ecosystem, 80–82 determination of balance constraints,
Hadoop-distributed file system (HDFS), 332, 336–337
79, 80 development of BSC index system,
HDFS, see Hadoop-distributed file 332–336
system managing strategy through balance
Heating, ventilation, and air scorecard, 329
conditioning (HVAC), 156 proposed integrated DEA–BSC
Heuristic algorithm, 214 Model, 337–338
Hewlett Packard (HP), 74 proposed integrated MCDM Model
Hierarchical model, 33, 344 for NPPM, 331-
High-risk borrowers, 181, 190, 201, 206 Intensity, 114
Higher-order polynomial degrees, 89, high-intensity failure, 131
95, 98, 99, 111 of outage, 114, 132
Highest paid person’s opinion using predictors, 130
algorithm (HiPPO Interaction styles, 51
algorithm), 70 Interactive articulation of preferences,
Holt–Winters exponential smoothing see Progressive articulation of
model, 137 preferences
Horizontal scalability, 75 Interactive methods, 44, 50–52
HP, see Hewlett Packard Internal Rate of Return (IRR), 324–325
HVAC, see Heating, ventilation, and air International Institute of Applied
conditioning Systems Analysis (IIASA), 53
Index 361

Internet of things (IOTs), 68, 88 Loan approval process, 177, 178

Interval trade-off ratio, 51 predictive analytics approach for, 180
Inventory management, 74, 136, 244 Logistic regression model, 120, 184, 185
IOTs, see Internet of things LP, see Linear programming
IP, see Integer programming
IRR, see Internal Rate of Return
M
IT, see Information Technology
MACBETH, see Measuring
attractiveness by categorical
J
based evaluation technique
JavaScript Object Notation (JSON), 79 Machine-learning
Jet engines, 3–4 classifier, 119
techniques, 181
tools, 156
K
Machine-learning algorithms; see also
Key performance indicators (KPIs), 70 Learning algorithm
Key-value store, 78 ANN, 184–186
input parameters, 197
logistic regression method, 184, 185
L
RF method, 186–187
L2 metric method, 124, 128, 130 SGBDT algorithm, 188–189
Label-correcting algorithm, 212 stacking method, 189
Label-setting algorithm, 212 training and testing using, 184
Labor utilization, 12 MADM, see Multi-attribute decision
Learning algorithm; see also Machine- making
learning algorithms Mahout, 81
evaluation, 189–191 MAHP, see Multiplicative AHP
and growth perspective, 335 Manufacturing performance, 309
Linear GP, 48 MAPE, see Mean absolute percentage
Linear programming (LP), 57, 88, 324; error
see also Goal programming MapReduce paradigm, 80, 81
(GP); Proposed LP model Material handling codes (MH
LP-based two-phase classifier, 89, 90 codes), 7, 8
numerical for constraints in proposed Mathematical model for time-
LP-based classifier, 96–98 dependent shortest path
optimization approach, 159 problem, 217
proposed LP-based classifiers, 92–96, mathematical model, 219–223
103–111 terminology, 217–219
proposed LP-based two-phase Mathematical programming methods,
classifier, 98–102 180s
proposed MILP model, 91–92 MAUT, see Multi-attribute utility theory
research problem description and MAVT, see Multi-attribute value theory
assumptions, 89 MaxAPE, see Maximum absolute
LINGO set code, 341, 347–349 percentage error
Loan Maximizing objective, 24
diversify of, 193 Maximum absolute percentage error
portfolio optimization, 179 (MaxAPE), 136, 139, 142–144
selection problem, 179 MCAP, see Multi-criteria aggregation
types, 176, 181 procedures
362 Index

MCDM, see Multi-criteria decision MODM, see Multi-objective decision

making making
MCMP model, see Multicriteria Modular repair center (MRC), 5, 6
mathematical programming blade, 6
model problems with large, 13
MDPEA, see Multiobjective MOEA, see Multiobjective evolutionary
deterministic pseudo- algorithm
evolutionary algorithm MOLP, see Multi-objective linear
Mean absolute percentage error programming
(MAPE), 136 MOO, see Multi-objective optimization
Measuring attractiveness by categorical MOP, see Multi-objective programming
based evaluation technique MPMM, see Modified Pugh matrix
(MACBETH), 53, 284 method
Membership function, 43 MRC, see Modular repair center
Metadata, 79 Multi-attribute decision making
Method of global criterion, 49–50 (MADM), 24, 34, 56
Methodologia Multicriterio para Apoio AHP, 37–39
a Selecao de Sistemas de ELECTRE methods, 39–41
Informacao (MMASSI), 285 fuzzy set theory, 43
MH codes, see Material handling codes MAUT, 34–35
Microgrid(s), 155, 158 MAVT, 35–36
data-mining methods, 156–157 PROMETHEE methods, 41–43
data for operations, 162 SMART, 36–37
design optimization, 164 Multi-attribute utility theory (MAUT),
dynamic nature of data in, 157 24, 34–35, 37
energy big data, 156 Multi-attribute value theory (MAVT),
energy models for, 158–171 27, 35–36
energy systems design, 165 Multi-criteria aggregation procedures
example for microgrids design, (MCAP), 33–34
165–166 Multi-criteria decision making
information for microgrids design, 166 (MCDM), 22, 33, 82–84, 88, 181,
notations and symbols for, 160 213, 271, 303; see also Multi-
solution for microgrids design, 166 objective decision making
system of, 158 (MODM)
MILP, see Mixed integer linear in Big Data, 57–58
programming characteristics, 32–33
Minimum cost path problem, 212 classification, 24–28
Minimum-size conveyor system, 9 DEA as tool, 306–307
Mixed integer linear programming demographics variables, 291–292
(MILP), 88, 90 development for CPEE, 288–291
MMASSI, see Methodologia DS, 292–293
Multicriterio para Apoio efficient solution, 29–32
a Selecao de Sistemas de employees performance evaluation
Informacao using, 282–288
MMP, see Multi-objective mathematical evaluating PSE based on MPMM,
programming 293–295, 296
Modified Pugh matrix method field, 22
(MPMM), 289 GDM, 22
evaluating PSE, 293–295, 296 hypothetical options, 29
Index 363

identification of multiple variables/ Multicriteria mixed-integer linear

multiple criteria for CPEE, programming models, 244,
271–282 245–246
IT industry, 270–271 Multicriteria model, 248–249
MADM methods, 34–43 Multicriteria ranking techniques,
managerial implications, 295–298 123–124
MCDM-based modeling framework, Multi-objective decision making
269, 297 (MODM), 24, 44; see also
methodologies, 54–57 Multi-criteria decision making
methods, 53–54 (MCDM)
non-dominated solution, 29–31 compromise programming, 50
outranking methods, 27 DEA, 52–53
Pareto optimal solution, 29–31 GP, 45–49
PAS, 269–270 interactive methods, 50–52
problems, 244–245 MCDM methods, 53–54
SODM, 28 method of global criterion and
surveys, 33 compromise programming,
techniques, 122–124 49–50
terminologies, 23–24 MOLP, 45
test for efficiency, 32 problem, 28–29
validation for CPEE to Offer R&R, 291 Multiobjective deterministic pseudo-
Multicriteria leadership and decisions, 1 evolutionary algorithm
reprint of Tinker project article, 3–17 (MDPEA), 139
Tinker projects, 2–3 computational evaluation of, 148
Multicriteria mathematical data set, 148
programming model (MCMP MAPE being primary objective,
model), 118, 179, 245 140–142
achieve age diversification among MaxAPE being primary objective,
borrowers, 195 142–144
debt-to-income ratio, 193 multiobjective netfront for retail
demand uncertainty in loan segment sales data, 148–152s
application requests, 196 netfront with respect to test data set,
diversify of loan types, 193 145–147
hard constraints, 195 netfront with respect to training
impact of portfolio risk, 201–204, 205 period, 144–145
input parameters of, 200–201 step-by-step procedure of, 139
late payment frequency of stop, 147
borrowers, 194 Multiobjective evolutionary algorithm
MCMP model analysis, 201 (MOEA), 325, 326
model, 191 Multiobjective forecasting, 135
model constraints, 193 ARIMA method, 136, 137–138,
model objectives, 192–193 138–139
multiobjective optimization using computational evaluation of MDPEA,
ε-constraint method, 195–196 148–152
notations, 191–192 MDPEA, 139–147
number of dependents associated seasonality, 137
with borrowers, 194 stationarity and nonstationarity, 137
open credits associated with Multi-objective linear programming
borrowers, 194 (MOLP), 24, 45
364 Index

Multiobjective loan portfolio optimization experimental settings, results, and

analysis, 198–200 discussion, 226
banks impact on society, 176 mathematical model for time-
data cleansing process, 183–184 dependent shortest path
data description, 182–183 problem, 217–223
evaluating learning algorithm, MILP model, 215
189–191 multiobjective model to Chennai
experimental results, 197 network, 234–235, 239–241
input parameters of machine- multiobjective optimization
learning algorithms, 197 algorithm, 223–226
input parameters of MCMP model, multiobjective optimization
200–201 problems, 213
literature review, 179 multiobjective shortest path problem,
loan approval process, 177, 178 215–217
MCMP model, 191–196 real-world shortest path problems,
MCMP model analysis, 201–204 213–214
methodology, 181 sample network topology, 232–234,
motivation, 181 237, 238
predicting risk associated with shortest path problem, 212
applicant, 182 Multiobjective shortest path problem,
predictive analytics approach for 215–217
loan approvals, 180 Multiperiod energy model, 157, 166
problem statement and objective, information for day-period energy
177–179 operation, 168
relationship between residential information for night-period energy
prices and bank lending, operation, 168
179–180 multiperiod energy planning
risks associated with lending, 176–177 optimization example, 167–168
training and testing using machine- solution for multiperiod energy
learning algorithms, 184–189 problem, 168
Multi-objective mathematical Multiperiod model, 250
programming (MMP), 25 index sets, 250–251
Multiobjective mixed-integer linear notation, 250
programming, 215–217 parameters, 251–253
Multiobjective model to Chennai Multiple-criteria models, 323
network, 234–235, 239–241 quantitative models with, 325–326
Multiobjective netfront for retail Multiplicative AHP (MAHP), 38
segment sales data, 148–152 Mutual preferential independence, 35, 36
Multi-objective optimization (MOO), 25
algorithm, 223
N
problems, 213
step-by-step procedure, 223–226 namenode, 79
terminology, 223 NDS, see Normalized demographic score
usingε-constraint method, 195–196 Net flow, 159
Multi-objective programming (MOP), 25 Netfront
Multiobjective routing in multiobjective netfront for retail
metropolitan city segment sales data, 148–152
data with respect to distance and to test data set from models, 145–147
travel times, 226–231 to training period, 144–145, 150, 151
Index 365

Net present value (NPV), 320 terminology, 223

New product development programs Nonnegativity
(NPD programs), 320 constraints, 161
New product management (NPM), restriction, 131–132
317–318 Nonpreemptive function, 254
New product portfolio management Nonpreemptive GP, 47–48
(NPPM), 317, 318 Nonseasonal ARIMA models, 136
analytical hierarchical process for Nonstationarity, 137
BSC, 343–347 Normalized demographic score
BSC for achieving strategic and (NDS), 288
balanced NPP, 328–329 Normalize function, 96
DEA approach for ranking and Not only SQL systems (NoSql systems),
prioritization of portfolios, 78–79
329–331 NPD programs, see New product
determination of balance constraints development programs
for DEA–BSC model, 332, NPM, see New product management
336–337 NPP, see New product process
development of BSC index system, NPPM, see New product portfolio
332–336 management
development of integrated DEA–BSC NP time, see Nondeterministic
model, 328 polynomial time
evaluative dimensions for, 319–322 NPV, see Net present value
LINGO set code, 347–349
methodologies/models for, 322
O
proposed integrated DEA–BSC
Model, 337–338 Objective(s), 24
proposed integrated MCDM model function, 130–131, 219–223, 224, 254,
for, 331 260–261, 330–331
qualitative models, 326–327 space, 28, 31
quantitative models, 323–326 Object to Relational mapping (OR
validation of proposed integrated mapping), 78
DEA–BSC model, 338–342 OLTP, see Online transactional
New product process (NPP), 316 processing
BSC for achieving strategic and One-shift operations, 13
balanced, 328–329 OneFS, 79
characteristics of, 316–317 Online transactional processing
Nodes, 212 (OLTP), 77
Non-linear GP, 48–49 Operating expenditure (OPEX), 309
Nondeterministic polynomial time Operational decisions, 244
(NP time), 213 Operations research models (OR
Nondominated model, 144 models), 73
Nondominated set, 213 OPEX, see Operating expenditure
solutions with respect to test data OPF, see Overall pallet factor
set, 109 OPower, 76
solutions with respect to validation Optimal supply chain, 244
data set, 109 Optimized resource allocation
Nondominated solutions, 29–31, 223 (ORA), 321
step-by-step procedure to generate ORA, see Optimized resource allocation
set of, 223–226 Organizational innovation process, 284
366 Index

OR mapping, see Object to Relational PLS, see Partial least square

mapping PMM, see Pugh matrix method
OR models, see Operations research PMOs, see Project management offices
models PMs, see Project managers
Output maximization linear Portfolio and innovation balance (PIB),
programming problem, 305 319, 321
Outranking approaches, 27, 39, 40, 286 Posteriori articulation of preferences, 27
Overall pallet factor (OPF), 17–18 Power
delivery system, 114
generation, 116
P
outage data, 125, 127
PACF, see Partial ACF restoration time, 131
Pairwise comparison, 51 systems, 114
Pareto charts, 118 PPM, see Product Portfolio Management
Pareto optimal solution, 29–31 Precise local trade-off ratio, 51
Pareto set, 213 Prediction models
Partial ACF (PACF), 137 criteria for model evaluation, 123
Partial least square (PLS), 279 multicriteria ranking techniques,
Partitioning gradient-based algorithm 123–124
(PGB algorithm), 48–49 performance of classification model,
PAS, see Performance appraisal system 122–123
Payback period (PBP), 320–321 using MCDM techniques, 122
PB, see Peta byte Predictive analysis, 326
PBC, see Performance-based contracts Predictive analytics, 72–73, 118, 119
PBP, see Payback period ANNs, 121
Performance-based contracts (PBC), 83 approach for loan approvals, 180
Performance appraisal system (PAS), boosting, 120
269–270 decision trees, 120
Performance evaluation logistic regression, 120
applications of MCDM methods, random forest, 120–121
286–287 SVMs, 121
characteristics of MCDM Preemptive function, 254
methods, 288 Preemptive GP, 47–48
employees performance evaluation Preference assessment styles, 51
using, 282 Preference function, 44
of employees using MCDM Preference Ranking Organization
methods, 282 METHod for Enrichment
methods, 286 Evaluations (PROMETHEE), 27,
problem of, 284–285 39, 41–43, 283
TOPSIS, 284–285 Prescriptive analytics, 68, 73–74, 118
Performance of SEs (PSE), 277 PRIME method, 53
PES, see Project evaluation and selection Priori articulation of preferences, 27
Peta byte (PB), 76 Product, 316
PGB algorithm, see Partitioning Production constraints, 161
gradient-based algorithm Production extensions binary, 259
PIB, see Portfolio and innovation balance Product Portfolio Management
Plant cost information, 262 (PPM), 318
Plant/production line constraint Product supply chain, 115
binary, 259 Progressive articulation of preferences, 27
Index 367

Project evaluation and selection R

(PES), 316
R&R, see Reward and recognition
Project management offices (PMOs), 318
R&R database (RRDB), 290
Project managers (PMs), 271
Random forest method (RF method),
PROMETHEE, see Preference Ranking
120–121, 180, 186–187
Organization METHod for
Rank reversal, 38
Enrichment Evaluations
Rate of return, 320–321
Proposed integrated DEA–BSC model,
Rating method, 123
337–338
RDBMS, see Relational database
efficiency score and relative
management systems
rating, 342
Real-world shortest path problems,
evaluation indicators, 339, 340
213–214
evaluation perspective and
Real constraints, 45–46
evaluation indictors, 341
Receiver operating characteristics
validation for NPPM, 338–342
(ROC), 89, 123, 190
Proposed integrated MCDM model, 331
Regression analysis, 180
determination of balance constraints
Regression model, 310–311
for DEA–BSC model, 336–337
Relational database management
development of BSC evaluation
systems (RDBMS), 78
index system, 332–336
Renewable energy resources, 155–156
development of proposed integrated
Return on investments (ROIs), 318,
DEA–BSC model, 337–338
324–325
Proposed LP model; see also Linear
Reward and recognition (R&R), 270
programming (LP)
MCDM-based modeling framework
algorithm for, 100–101
development for CPEE, 288–291
classifier, 92–93
MCDM modeling framework
classifiers with fuzzy measure,
validation for CPEE, 291–295
103–105
multiple variables/multiple criteria
classifier with crisp boundary, 93–96
identification for CPEE, 271–282
numerical for constraints in
RF method, see Random forest method
proposed LP-based two-phase
Ribbon, Edelman Laureate, 3
classifier, 102
Rich Site Summary (RSS), 77
two-phase classifier, 98
“Risk” variable, 182–183
for two-phase classifier, 98–99
Risk–uncertainty estimation (RUE), 321
PSE, see Performance of SEs
ROC, see Receiver operating
Pugh matrix method (PMM), 284
characteristics
Pλ problem, see Weighted objective
ROIs, see Return on investments
problem
RRDB, see R&R database
R software, 198
Q RSS, see Rich Site Summary
RUE, see Risk–uncertainty estimation
Qualitative models, 326
AHP/ANP, 327
S
balanced scorecard, 326–327
strategic approaches, scoring models, Sample-based algorithm, 214
and checklists, 326 Sample network topology, 232
Quantitative models, 323 one-way traffic regulation, 237
with multiple criteria, 325–326 optimal solution, 233
with single criterion, 324–325 travel starts from node 1, 234, 238
368 Index

SARIMA, see Seasonal ARIMA SME, see Small and Medium

SAW, see Simple additive weighting Enterprises
Scalarization method, 223 Society, banks impact on, 176
Scale efficiency, 306 SODM, see Single-objective
Scale-out approach, 75 decision-making
Scale-up approach, 75 Software engineers (SEs), 271
Scaling, model, 254–255 Software implementation, 56–57
Scenario analysis, 326 Space requirements, 17
Scoring models, 326 SQL, see Structured Query Language
Seasonal ARIMA (SARIMA), 138 SRM, see Structural risk minimization
Seasonality, 137 Stacking, 180, 184, 189, 200
SEM, see Structural equation modeling Staffing cost minimization, 130
Service supply chain, 115, 116 Stationarity, 137
Service time distribution, 12 Statistical regression, 311
SEs, see Software engineers Stochastic gradient-boosted decision
Set of countries (COS), 260 trees algorithm (SGBDT
SF, see Strategic fit algorithm), 188–189, 200
SGBDT algorithm, see Stochastic Stock keeping unit (SKU), 262
gradient-boosted decision trees Storage-purchasing costs, 164
algorithm Strategic buckets, 326
Shortest path problem, 212 Strategic decisions, 244
mathematical model for time- Strategic fit (SF), 319, 321
dependent, 217 Strategic optimization, 248
multiobjective, 215–217 Structural equation modeling
real-world, 213–214 (SEM), 279
Short Life and low Inventory in Structural risk minimization
Manufacturing (SLIM), 74 (SRM), 121
Simple additive weighting (SAW), 284 Structured Query Language (SQL), 76
Simple multi-attribute rating technique Supply and demand behavior of
(SMART), 24–25, 35, 36–37 storages, 159
Simulation model, 9, 14 Supply chain network, 249–250
features of model, 12 Support vector machines (SVMs),
MRC, 10 121, 138
sample five-machine MRC Sustainability perspective, 335–336
configuration, 11 SVMs, see Support vector machines
TIPS, 9 Systematic risks, 321
WCD, 11
Single-criterion models, 323, 324–325
T
Single-objective decision-making
(SODM), 28 Tactical optimization, 248
SKU, see Stock keeping unit TAFB, see Tinker Air Force Base
SLIM, see Short Life and low Inventory TBs, see Tera bytes
in Manufacturing Tchebycheff metric, 50
Small and Medium Enterprises (SME), 72 TCP, see Transmission control protocol
SMART, see Simple multi-attribute TDS, see Total demographic score
rating technique Technical feasibility, 319
Smart staffing, 130 Technique for order of preference by
model constraints, 131–132 similarity to ideal solution
objective functions, 130–131 (TOPSIS), 284–285
Index 369

Temperature, 24 TOPSIS, see Technique for order of

Tera bytes (TBs), 76 preference by similarity to
Test data set, 89–90, 93 ideal solution
non-dominated set of solutions Total demographic score (TDS), 288
with, 109 Total performance score (TPS), 290
Theory of bounded rationality, 70, 71 TP, see True positives
Three shift operations, 13 TPR, see True-positive rate
Time dependency TPS, see Total performance score
mathematical model for time- Training data set, 89–90, 93
dependent shortest path Transmission, 118
problem, 217 Transmission control protocol
multiobjective shortest path problem (TCP), 77
with, 215–217 Transmitters, 116
Time period binary, 259 Transportation, 262
Time-varying models, 180 cost, 262
TIMS Interfaces, 2 network, 212, 213
Tinker Air Force Base (TAFB), 3, 6 Traverse time, 215
Tinker Integrated Planning Simulation True-positive rate (TPR), 190
(TIPS), 9 True negatives (TN), 122
Tinker projects, 2–3 True positives (TP), 122
abstract, 3–4 t-test, 277–279
calculating number of machines Type I error, see False positives (FP)
needed, 12–13 Type II error, see False negatives (FN)
central role of TIPS in shop
management, 15
U
computer generated data, 9
data analysis, 6–8 UCF, see University of Central Florida
design constraints, 15 Ultrashort-term forecasting, 157
general model for conveyor design, 9 Uncertainty, 54
input data verification, 15–16 management, 321
material handling characteristics, 8 perspective, 336
MH codes, 7 Units per assemble (UPA), 5, 11
original 1989 paper University of Central Florida (UCF), 1
acknowledgments, 17 University of Oklahoma design team, 17
output, 14–15 Unstructured data, 75
parts per pallet material handling Unsystematic risks, 321
codes, 8 UPA, see Units per assemble
problems with large MRCs, 13 User, microgrids, 158–159
project scope, 4–6 Utility functions, 34
project summary, 16–17 Utility independence, 35
reconstruction period, 6
reprint of, 3–17
V
simulation model, 9–12
validation, 14 Validation
verification, 13–14 data set, 89–90
WCD, 5 data set non-dominated set of
TIPS, see Tinker Integrated Planning solutions with respect to, 109
Simulation of proposed integrated DEA–BSC
TN, see True negatives model, 338–342
370 Index

Value W
contribution perspective, 336
Warmth, 24
function, 35
Weather data, 124–125
theory, 35
sample data table for historical
trade-off, 51
weather, 126
Variable returns to scale (VRS), 306
Weighted objective problem, 31
Variables, 253–254
Work control document (WCD), 5, 6
Variance, 137
Worker type requirement, 131
Variety in big data, 75
Workforce capacity, restriction on, 131
Vector comparison, 51
Worst-case model error, 139–140
Velocity in big data, 75
Veracity in big data, 75
Vertical scalability, 75 X
Very short-term forecasting, 157 XtremeFS, 79
Vise Kriterijumska Optimizacija
I Kompromisno Resenje
Y
(VIKOR), 285
Volume in big data, 74–75 Yet Another Resource Negotiator
VRS, see Variable returns to scale (YARN), 81