See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/262328799

A Graph Analytical Approach for Topic Detection

Article in ACM Transactions on Internet Technology · December 2013

DOI: 10.1145/2542214.2542215

CITATIONS READS

20 301

2 authors, including:

Louiqa Raschid
University of Maryland, College Park
226 PUBLICATIONS 2,979 CITATIONS

SEE PROFILE

All content following this page was uploaded by Louiqa Raschid on 02 May 2016.

The user has requested enhancement of the downloaded file. All in-text references underlined in blue are added to the original document
and are linked to publications on ResearchGate, letting you access and read them immediately.
 A
A Graph Analytical Approach for Fast Topic Detection
Hassan Sayyadi, University of Maryland
Louiqa Rschid, University of Maryland

Scalability and accuracy challenges when processing large noisy data collections such as social media re-
quire new solutions for Topic Detection. We present KeyGraph, an efficient method that improves on the
current topic detection methods by considering keyword co-occurrence. We show that KeyGraph has similar
accuracy when compared to state-of-the-art approaches on small well annotated collections. Further, Key-
Graph can successfully filter noise and identify events in social media collections. An extensive evaluation
using mechanical Turk demonstrated both the accuracy and recall of KeyGraph. The runtime performance
of KeyGraph is significantly better when compare to LDA topic models on large collections.
Categories and Subject Descriptors: H.3.3 [Clustering]: Information Search and Retrieval; E.1 [Graphs
and Networks]: Data Structure; H.2.8 [Data Mining]: Database Applications
General Terms: Algorithms, Performance
Additional Key Words and Phrases: Topic Detection, Network Analysis, Community Detection, KeyGraph
based Topic Detection

1. INTRODUCTION
A topic (event) is defined in the Topic Detection and Tracking (TDT) community as a
specific event or occurrence at a specific time and in a specific place [3]. New Event
Detection (NED) and Retrospective Event Detection (RED) are two groups of event
detection algorithms [30].
There have been many successful TDT approaches reported in the literature [5; 13;
30]. Initial methods typically relied on clustering documents. Consider a core group of
keywords representing a specific topic. Documents that contain these keywords will be
more similar to each other based on a metric such as TF/IDF[23]. Thus, variations of
TF/IDF were used to compute keyword based feature values, and cosine similarly is
used as a similarity (or distance) measure to cluster documents.
The next generation of TDT approaches moved the analysis from document clus-
tering to keyword based features. Recent research has addressed the relationships
between keywords, their occurrence in documents, as well as the dynamics of such
relationships [4; 10; 19; 24; 29]. A variety of machine learning techniques have been
utilized. Topic modeling approaches based on latent Dirichlet allocation (LDA) [7; 26]
exemplify these advances in TDT.
The gold standard algorithms that are either based on document clustering or on
keyword based machine learning have worked well for many collections. However, the
rapid growth of social media, and the corresponding need to handle large document
collections in a noisy environment, has created a need for scalable and accurate TDT
approaches. Social media collections differ from collections of articles along several
metrics including the size distribution of documents and the distribution of unique
keywords in the collections. The challenge is to rapidly filter noise or irrelevant doc-
uments while at the same time being accurate when clustering a large collection of
documents.
In this paper, we present KeyGraph, a novel and efficient approach for topic detec-
tion that is inspired by the keyword co-occurrence graph and efficient network anal-
ysis methods. We explain the intuition for KeyGraph and keyword co-occurrence by
citing the research of Newman et al. [16]. They performed an evaluation of topic co-
herence across many methods and metrics and empirically demonstrated that keyword
co-occurrence, computed via simple pointwise mutual information (PMI), was the most

ACM Transactions on Embedded Computing Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
 A:2

consistent performer, achieving the best or near-best results over multiple datasets. In
some cases the results approached or surpassed human inter-annotator agreement.
Current topic modeling methods focus on word distributions in the document and
they do not explicitly consider word co-occurrence. Extending topic modeling to include
co-occurrence is a computationally daunting challenge. KeyGraph is a first approxima-
tion; it represents a collection of documents as a keyword co-occurrence graph. It then
applies an off-the shelf community detection algorithm to group highly co-occurring
keywords into communities. Each community forms a constellation of keywords and
corresponds to a set of topic features that represents a topic. Our intuition, which we
prove empirically, is that communities in the KeyGraph are a good proxy for topics.
We compare KeyGraph with other solutions on the well known TDT4 dataset. We
further compare KeyGraph with LDA topic modeling using Gibbs Sampling (LDA-GS)
[2] on a large social media dataset. We perform a human evaluation of the accuracy of
topic detection. Our research makes the following contributions:
— We demonstrate that the accuracy of the KeyGraph method for topic detection is
similar to the gold standard algorithms, including kNN [5], GAC [30], two variants
of LDA topic modeling [1; 2] and a probabilistic model [13]), on the well known TDT4
dataset.
— Using the Spinn3r data collection, a large noisy collection of blog posts, we demon-
strate that KeyGraph can filter noisy irrelevant posts and create smaller clusters of
relevant documents for each topic. In contrast, a topic modeling solution is not able
to filter as many irrelevant documents.
A human evaluation using Amazon’s Mechanical Turk marketplace is used for val-
idation of the detected events. The best topics or events chosen by KeyGraph and
LDA topic modeling with Gibbs Sampling (LDA-GS) were evaluated by humans. On
average, only 40% of the documents selected by LDA-GS received votes of relevant
or highly relevant. In contrast, over 77% of the documents selected by KeyGraph re-
ceived votes of relevant or highly relevant. Both methods were shown to have high
recall using a pseudo-recall metric. There was high inter-annotator agreement on the
ratings.
— We demonstrate that KeyGraph (whose running time is linear in the size of the col-
lection) outperforms LDA-GS [2] by an order of magnitude. We use a subset of up to
113,000 documents from the Spinn3r data collection and show that the performance
improvement is increasingly significant as the number of documents exceeds 50,000
documents.
This paper is organized as follows: In the next section, we review related work. Sec-
tion 3 presents the KeyGraph topic detection algorithm. Next, in Section 4, we compare
the accuracy of KeyGraph to other gold standard solutions on the TDT4 dataset. We
also compare KeyGraph and LDA topic modeling on the Spinn3r data collection.
2. RELATED WORK
Event or topic detection problems usually fall in two categories, namely New Event
Detection (NED) and Retrospective Event Detection (RED). While NED algorithms
deal with documents processed online, RED is defined as the offline discovery of events
in a historical corpus [30]. The solution approaches can be classified as document based
clustering and keyword based.
NED models are proposed as a single pass incremental clustering algorithm. For a
newly arrived document, the similarity between the document and known events is
computed, and the maximum similarity will be determined. If the similarity is more
than some a priori threshold, then the document will be assigned to the correspond-
ing event. Otherwise it will be considered a new event. Allen et al. [5] used kNN

ACM Transactions on Embedded Computing Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
 A:3

(k-Nearest Neighbors) with a variation of TF/IDF. They also penalized the threshold
based on the elapsed time interval between the document and the event; the thresh-
old increases with this elapsed time interval. Since future document features are not
known a priori, such online clustering algorithms need to estimate IDF. While Allen
et al. [5] use an auxiliary dataset to estimate IDF, Yang et al. [30] proposed an incre-
mental IDF factor which will be updated for each new incoming document. Brants [8]
introduced incremental TF/IDF and similarity score normalization for NED. Kumaran
et al. [12] improved their results through the use of text classification techniques as
well as by using named entities.
Retrospective Event Detection (RED) was originally defined by Yang et al. [30].
Many text clustering algorithms have been developed for RED. One of the most ac-
curate is by Yang et al. who proposed an agglomerative clustering algorithm, GAC
(augmented Group Average Clustering), to cluster documents in the corpus. They also
applied an iterative bucketing and re-clustering model that was proposed by Cutting
et al. [9] to control the tradeoff between cluster quality and computational efficiency.
Similar to NED, RED models also use clustering algorithms, but they offer different
variations for document representation, and for distance or similarity metrics. For ex-
ample, [25; 28] have shown that a set of significant words, including named entities,
can be used for efficient search result clustering. Mory et al. [14; 15] also propose a
new document representation model in which they extracted the maximally connected
components from the word co-occurrence graph [18]. These components were called
foundations or basic concepts, and are used as document features. Documents repre-
sented with basic concepts were clustered using a single-link clustering algorithm.
Next, we discuss another approach to event detection that will focus on
words/keywords and their correlation. He et al. [10] study feature trends that may
vary over time and they created and analyzed time-series word signals. Next, they
chose correlated word signals as features for topic detection. Similarly, Prabowo et al.
[19] built word co-occurrence graphs and looked at the graph community structure for
monitoring and tracking debates in online communities.
Based on [7; 26], the Latent Dirichlet Allocation (LDA) topic model is a three-level
hierarchical Bayesian model. Each item of a collection is modeled as a finite mixture
over an underlying set of topics. Each topic is, in turn, modeled as an infinite mixture
over an underlying set of topic probabilities. LDA has been widely used and has been
shown to be a successful algorithm for topic modeling. The Mallet toolkit provides
an implementation of LDA which uses Gibbs sampling as the inference engine(LDA-
GS)[2]. Blei also provides another variant of LDA which uses variational expectation
maximization (VEM) for latent Dirichlet allocation (LDA-VEM)[1].
Li et al. [13] proposed a novel solution that included a probabilistic algorithm which
directly exploits time for topic modeling. Documents and events are represented by
features including persons, locations, keywords, and time. The probability of each doc-
ument is defined as:

p(article) = p(time) ∗ p(persons) ∗ p(locations)

∗p(keywords)
p(xi |ej ) = p(timei |ej ) ∗ p(personsi |ej ) ∗ p(locationsi |ej )
∗p(keywordsi |ej )

where xi and ej are ith article and jth event. They computed an estimation of the event
count using the time distribution of the article count. Next, an Expectation Maximiza-
tion (EM) algorithm is used to maximize the log-likelihood of the distributions of these
features and to learn the model parameters. Such an algorithm requires the number

ACM Transactions on Embedded Computing Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
A:4

of events to be given a priori which makes this approach somewhat impractical and
difficult to tune.

3. KEYGRAPH BASED TOPIC DETECTION

Topic modeling represents topics as an infinite mixture over the probability distribu-
tions of sets of words. Our intuition is to consider word co-occurrence as well. Ideally,
we would extend the framework so that a topic is an infinite mixture over the proba-
bility distributions of sets of words such that the probability distribution of word co-
occurrences among these sets of words is maximized. We note that this extension can no
longer be addressed by the current 3 level hierarchical modeling approach of topic mod-
eling. Further, the computational complexity of such an extension which may have to
simultaneously consider the co-occurrences of subsets of keywords is not known. Thus,
as a first approximation to this extension, KeyGraph will focus on the challenge of
maximizing word co-occurrence alone. The problem then becomes almost identical to
community detection over the keyword co-occurrence KeyGraph. KeyGraph, presented
in Algorithm 1, comprises the following phases:

(1) Building the KeyGraph.

(2) Extracting topic features.
(3) Assigning topics to documents.

In the first phase, we construct a keyword co-occurrence graph labeled a KeyGraph.

A KeyGraph has one node for each keyword in the corpus. Edges in a KeyGraph rep-
resent the co-occurrence of the corresponding keywords and are weighted by the count
of the co-occurrences.
The second phase of the algorithm includes community detection and feature extrac-
tion, i.e., a constellation of keywords to represent a topic. From Figure 1, constellations
or sets of keywords which capture some meaningful topic are more densely linked,
while there are fewer links between keywords that participate in different commu-
nities. Community detection approaches apply relational clustering algorithms from
network analysis. Several algorithms for community detection on static graphs are
presented in [17]. We use an off-the shelf algorithm based on the betweenness central-
ity metric that has been shown to perform well.
The third phase of the algorithm is to use the features to assign topics to documents.
Finally, for each pair of topics, if there are multiple documents that are assigned to
both topics, then we assume that these are sub-topics of the same parent topic, and we
merge the sub-topics.
Recall that Li et al. [13] incorporated time into topic modeling. The KeyGraph is dy-
namic since the documents are time stamped. Community detection for dynamic net-
works is an emerging problem. While some solutions are presented in [21; 22], their
accuracy has not been well studied. The complexity and cost of dynamic community de-
tection is also much greater compared to solutions for static networks. Therefore, in our
current solution, we aggregate keyword co-occurrences over time. This allows us to use
an off-the-shelf static community detection algorithm. We further note that our first
priority is to extend topic modeling to consider both word and keyword co-occurrence
distributions. We feel this extension has greater potential compared to temporal fea-
tures, for noisy social media.
Preliminary results of the KeyGraph approach to topic detection were presented
in [24]. Our prior work has been extended to include a comparison against the gold
standard approaches based on accuracy. We also extended the prior research with an
extensive human evaluation and scalability experiments.

ACM Transactions on Embedded Computing Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
 A:5

Fig. 1. A small sample of the KeyGraph

Building the KeyGraph: We first extract keywords from each document and create
one node per unique keyword in the dataset. Keywords can include the terms, noun
phrases, and named entities in the document. An edge ei,j between nodes ni and nj is
added if keywords wi and wj co-occur in at least one document.
After building the initial co-occurrence graph, we calculate the document frequency
(DFi ) for each node ni . Nodes with a document frequency below a pre-specified
threshold are removed. Similarly, we remove all edges ei,j where the keywords wi and
wj co-occur in at most one document. We then compute the conditional probabilities
p(ki |kj ) and p(kj |ki ) for each edge ei,j . The conditional probability is computed as
DF DFe
follows: P (ki |kj ) = DFi∩j
j
= DFi,j
j
. The edges with both conditional probabilities below
a pre-specified threshold are also removed.

Extracting Topic Features: Our intuition is that keywords co-occur when there
is a meaningful topical relationship between them. Based on this analogy, we model
the KeyGraph as a social network of relationships between keywords. The occurrence
of an event will change the community structure of the KeyGraph. The number of
co-occurrences of the topic keyword features will increase due to the increase in the

ACM Transactions on Embedded Computing Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
A:6

Algorithm 1 Graph Analytical approach for Topic Detection

1: for all Document di do
2: Extract words, noun phrases and named entities as document keywords/features
3: end for

%Building the KeyGraph

4: Build the keyword co-occurrence graph
%Extracting Topic Features
5: Extract strongly connected communities of keywords in the KeyGraph as topic fea-
tures
%Assigning Topics to Documents
6: for all Extracted topic feature ft do
7: Find the likelihood of topic t for document d as follow:
cosine(d, ft )
p(t|d) = P
t∈T cosine(d, ft )

8: end for
9: for all Topic ti and tj do
10: if the overlap of ti and tj (between their document set) is greater than µ then
11: Merge ti and tj
12: end if
13: end for

Fig. 2. Document frequency of topics Litigating gay marriage in California and Cyclone in Myanmar

number of documents related to the topic, sharply at the beginning or onset of the
event. Then, depending on the nature of the event, the co-occurrences will decrease
slowly or more rapidly. The corresponding number of relevant documents that will be
grouped with this event will reflect a similar pattern. Figure 2 shows the document
frequency for the two biggest events we found in the Spinn3r dataset (see Table I
for the details of Spinn3r dataset). Both diagrams clearly show the general document
frequency pattern of the topics.
For example, there are several communities of keywords in Figure 1 which are sug-
gestive of specific topics. Within each community, the nodes are highly connected while
connections to nodes outside of the community are rare. For example, the group of key-
words including ”cyclone”, ”Myanmar”, ”foreign aid workers”, etc. creates a community
to represent the topic cyclone in Myanmar. This broader topic includes the more spe-

ACM Transactions on Embedded Computing Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
 A:7

cific topic that occurred a few days after the onset of the cyclone, when the Myanmar
government denied access to foreign aid workers and refused to accept international
aid. We note that the more general and more specific topic correspond to the two peaks
in Figure 2. Another example which can be seen, describes the amendment to the con-
stitution to ban same sex unions in California. A community of keywords including
”Same-sex marriage”, ”California supreme court”, etc. is visible in the graph.”The Sen-
ate approval of $165 billion to fund the wars in Iraq and Afghanistan” and ”Ellen
DeGeneres’s opposition to the gay marriage” are other topics which can be detected by
the communities of keywords in this graph.
The betweenness centrality score is a good measure to find the edges between com-
munities in a network [17]. Betweenness centrality for an edge is defined as the count
of the shortest paths, for all pairs of nodes in the network, that pass through that
edge. Inter-community edges will always obtain a high score, since the shortest paths
between nodes from different communities will always pass through these edges. For
instance, in Figure 1 the edges (”Ellen DeGeneres”, ”gay marriage”),(”Ellen”, ”nomi-
nees”), and (”ban”, ”same sex marriage”) are among the edges with the highest be-
tweenness centrality score.
We use the following procedure to remove inter-community edges. The set of short-
est paths between all pairs of nodes is calculated using a breadth first search (BFS).
Then, the edge with the highest betweenness centrality score is selected to be removed
from the graph. If two edges have the same betweenness centrality score, the one with
the lower conditional probability value will be selected. This process will be repeated
until no edge with a score greater than a pre-specified threshold can be found, then
we can say we removed all inter-community edges from the network. Next, if the Key-
Graph is not fully connected anymore, we will continue this operation for each smaller
connected component recursively until all edges in the KeyGraph have a small be-
tweenness centrality score. Stop conditions will be discussed in detail in section 4.5.
After all inter-community edges have been removed, each connected component of
the KeyGraph represents a hypothesis about an topic. The highly co-occurring key-
words in each component form the features for the topic.
As a natural consequence of aggregating the co-occurrence graph over time, the
number of nodes will not increase but the density of the graph will increase [19].It is
obvious that the new keywords that can be added to the graph is limited but keywords
that never appear together in the past can appear in the same documents for new
topics which makes the graph denser. We note that while the community structure of
the KeyGraph changes over time, our approach aggregates keyword co-occurrences.
Although such aggregation has the potential to miss the true community structure,
our experimental results show that the quality of topic detection with our aggregated
KeyGraph is as good as the best topic detection models in the literature.

Assigning Topics to Documents: After finding the communities of significantly co-

occurring keywords as topic features, the next step is assigning the topic likelihoods
to each document. In previous step, we extracted the keyword community from the
KeyGraph. Each community of keywords represents a synthetic document, called a
feature document, for the corresponding topic. The likelihood of topic t for a document
d is determined by the cosine similarity of d and the corresponding feature document
ft .

cosine(d, ft )
p(t|d) = P
t∈T cosine(d, ft )

ACM Transactions on Embedded Computing Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
A:8

Table I. Details of Dataset 1 (part of the TDT4 Dataset)

Time Oct. 2000 - Jan. 2001
Number of Topics(Events) 71
Number of topic articles 1923
Number of unique articles 1884
Average articles per event 27

We use the following approach to determine the TF value of a keyword in a feature

document. There is an agreement that the most common keywords for a topic should
have the biggest weight. We expect that the more closely the keyword is representa-
tive of a topic, the more it occurs with other keywords on that topic. Hence, we use the
average co-occurrence of each keyword from the feature document with other keywords
of the feature document as TF value. We experimented with other options including us-
ing the equal values (TF=1), using the node degree as TF, etc, but they did not improve
performance.
Since documents are assigned a likelihood or probability distribution over topics, our
model may assign an article to multiple topics unless we explicitly place a restriction
that a document can only be assigned to a single topic. This provides flexibility to our
method to consider multi-topic documents. It also helps to overcome network commu-
nity detection errors that may have occurred in an earlier step.
A common problem with any clustering or community detection algorithm is that a
true cluster can be split. Our soft clustering capability which allows a document to be
assigned to multiple topics can address this problem. Suppose that a KeyGraph com-
munity related to a topic is mistakenly split into two or more sub-communities. This
results in having several sub-feature documents for the topic. After document assign-
ment, there would be a significant overlap in the corresponding document sets. This
occurs because the documents related to the topic contain words from all sub-feature
documents. We allow our algorithm to partition the KeyGraph into a fairly large num-
ber of communities to make sure that we exceed the expected number of topics. We
then merge those topics with a significant overlap in their relevant document sets. We
demonstrate the robustness of our approach as we change multiple parameters in our
experimental evaluation.

4. EVALUATION
We evaluate KeyGraph as well as several gold standard algorithms that have been
reported in the literature. We show that KeyGraph has similar accuracy to the gold
standard algorithms on the well known TDT4 dataset; this is a small and well an-
notated dataset. We demonstrate further that KeyGraph can successfully filter noise
and identify events in a social media collection. KeyGraph has good accuracy and re-
call. The running time of KeyGraph significantly outperforms LDA topic models on
large collections.

4.1. Datasets
TDT4: The TDT4 collection is a well known dataset that has been widely used to com-
pare methods for topic detection [27; 13; 30]. The original data collection includes more
than 15,000 documents. However, only a subset of it has been annotated by humans.
Similar to all other methods, we use the annotated subset of collection for the purposes
of evaluation.
The TDT4 dataset includes 71 human identified topics together with 1923 docu-
ments. Due to document overlap of some topics, there are 1884 unique documents. We
exclude topics that have less than 4 documents. Dataset statistics are in Table I

ACM Transactions on Embedded Computing Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
 A:9

Table II. Details of Dataset 2 (subset of the Spinn3r dataset for

human evaluation)
Time May 1, 2008 - May 7, 2008
Number of topics [unknown]
Number of sources 1905
Number of unique articles 32041
Average articles per source 1.6

Spinn3r: The Spinn3r collection includes approximately 44 million blog posts col-
lected over 3 months in 2008. This includes millions of posts which are on a wide range
of topics. The document size ranges from one sentence to very large documents. The
posts originate from many thousands of sources, and no text extraction or data clean-
ing was performed. Thus, this represents a large and noisy collection. A random subset
of the Spinn3r collection labeled Dataset 2 was used for human evaluation; details of
Dataset 2 are in Table II. A larger subset of up to 113,000 documents was used to
compare the running time of the methods.
Feature Extraction: For both datasets, we tokenize the text and extract all words
from the documents. After removing stop words, we use a Porter Stemmer to stem the
tokens. Next, we extract noun phrases and named entities since both are found to be
informative and important for TDT tasks. We use the Stanford Parser for noun phrase
extraction and IdentiFinder for named entity extraction. To extract named entities, we
use IdentiFinder. Like other named entity recognition tools, IdentiFinder is sensitive
to sentence structure and it is case sensitive. For example, the named entity ”obama”
which starts with a small letter in the document, will very likely be missed. Some
named entities also will be missed because of the sentence structure. To reduce such
false negatives, we use a two pass approach. In the first pass, we run IdentiFinder
on the text collection and collect all named entities recognized by IdentiFinder. We
use this collection as a reference database. In the second pass, for each word in a
document, we consult the reference database to determine if the word is a named
entity. For example, if ”Obama” appears in at least one document in the corpus and is
detected as a named entity by Identifinder, it will be added to the reference database.
Next, in the second pass, we do a case insensitive search to find all occurrences of
”Obama” in the corpus. Hence, all occurrences of ”obama” or ”Obama” in corpus will be
marked as named entities, independent of the letter case and the sentence structure.
This way, we can retrieve some of the named entities which were initially not detected
by IdentiFinder. However, words like ”House” will be a true positive named entity in
some documents but they are false positives in other documents. Our two pass method
may generate some false positives. Our experience is that the number of false positives
generated in the second pass are negligible compared to the many thousands of false
negatives that we avoid using the two pass approach. When processing each document,
we boost the term frequency of keywords that are identified as named entities so that
they are emphasized by our topic detection algorithm. We obtained the best results by
doubling the term frequency of named entities.
Noise: The size (word count) of social media documents has high variance and there
are many short (one sentence) documents. Simultaneously, there is a skew in the dis-
tribution of unique words. Figure 3 illustrates one metric, i.e., the increase in the
number of unique words as we increase the number of documents (and words) in the
2 collections. The number of unique words increases steeply for the Spinn3r collection
whereas the number appears to increase more gradually for TDT4 and to eventually
taper off. The skew is more noticeable in Spinn3r since the documents are smaller and
the word counts are correspondingly lower. The skewed word distribution for Spinn3r
makes it difficult for topic modeling approaches that rely on word distribution to cre-

ACM Transactions on Embedded Computing Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
A:10

Fig. 3. The number of unique words in the TDT4 and Spinn3r datasets

ate accurate clusters. This same skew benefits KeyGraph since an increase in the
number of unique keywords correspondingly reduces word co-occurrence. This allows
KeyGraph to filter irrelevant documents.

4.2. Comparing the Accuracy of KeyGraph to Other Methods on the TDT4 Dataset
We evaluate two variants as follows:
(1) KeyGraph: This a variant of KeyGraph which only uses terms in each document
as features.
(2) KeyGraph+: This a variant of KeyGraph which uses terms in each document as
well as noun phrases and named entities as features.
We compare them with the following best performing topic detection algorithms:
(1) kNN: The k-Nearest Neighbor clustering algorithm is a popular clustering solution
for TDT tasks[5]. For kNN, we report the results reported by Li et al. (Figure 6 from
[13]) under the best threshold.
(2) GAC: This is the Augmented Group Average Clustering model proposed by Yang
et al. [30]. Similarly for GAC, we report the results reported by Li et al. (Figure 6
from [13]) under the best threshold.
(3) LDA-VEM[1]: This LDA variant uses Variational Expectation Maximization (with
default parameters: α and β will be estimated by the algorithm itself for smoothing
over document topic distribution and word distribution, i = 100 for the number of
expectation maximization iterations, and k = 71 for the number of topics which
is known for TDT4 dataset.) LDA-VEM also performs hyper-parameter optimiza-
tion. Similar to LDA-GS, we assign each document to only one topic which has the
highest probability.
(4) LDA-GS[2]: This is the LDA algorithm from the Mallet toolkit which uses Gibbs
sampling (with default parameters: α = 0.5 for smoothing over document topic
distribution, β = 0.01 for smoothing over document word distribution, and k = 71
for the number of topics which is known for TDT4 dataset.) We note that LDA-GS
does hyper-parameter optimization which allows the model to better fit the data.
Asuncion et al. [6] showed that for several datasets that contain between 1,000
and 20,000 documents, topic models converge in less than 500 iteration of Gibbs
sampling. In our experiments, we selected Mallet’s default value i = 1000 for the
number of Gibbs sampling iterations.
(5) LDA-GS(p)[2]: This is a version of LDA-GS in which we pre-process the data
before feeding it to the Mallet software. Note that Mallet also does some pre-

ACM Transactions on Embedded Computing Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
 A:11

Fig. 4. The performance of the five approaches (kNN, LDA-VEM, LDA-GS, GAC, and KeyGraph) on Dataset
1 (TDT4).

possessing including removing stopwords. In LDA-GS(p), stopwords will be re-

moved first. Next, a Porter stemmer will be applied on the remaining words. Each
document will then be represented as a bag of words which will be stored in a
text file (the order of the words from the original document may not be preserved).
LDA-GS(p) uses the same set of features that KeyGraph uses.
(6) LDA-GS(p)+[2]: This version is a version of LDA-GS(p) that uses noun phrases
and named entities as features as well. LDA-GS(p)+ uses the same set of features
that KeyGraph+ uses.
We note that for all variants of LDA and KeyGraph, documents are assigned a prob-
ability distribution over topics. However, our gold standard contains only the topic
assignments for each document, not the topic distributions. In addition all documents
have a topic assigned to them. Hence, we assign each document to at most one topic
that has the highest probability. We may filter the documents that their highest prob-
ability to any topic is less than 0.1. This is the default threshold in Mallet toolkit.
4.2.1. Evaluation Metrics. Each algorithm may generate any number of clusters, but it
is only evaluated on the reference topics. In other words, the cluster that best matches
each of the reference topics is used for evaluation as discussed in [13; 30]. We report on
the average per-topic precision score, the average per-topic recall score, and the macro
average F1 or F1 (mac-avg) score; The F1 score is the harmonic mean of precision and
recall as follows:
F1 = (2 ∗ precision ∗ recall)/(precision + recall)
and the F1 (mac-avg) is obtained by computing per-topic precision and per-topic recall
scores first, and then averaging the corresponding F1 scores [30].
4.2.2. Results. Figure 4 shows the performance of the two variants of KeyGraph and
the 6 other algorithms on Dataset 1 (TDT4). The results for all variants of LDA-GS,
LDA-VEM, and KeyGraph are obtained by running the algorithms on the TDT4
dataset. The results for kNN and GAC are from Figure 6 in [13] which was executed
on the identical dataset (TDT4).
The results in Figure 4 show that GAC and LDA-GS outperform kNN and LDA-
VEM, for all three metrics. KeyGraph+ slightly outperforms GAC and LDA-GS for
precision, scoring 0.82 compared to 0.79 and 0.80 respectively. LDA-GS slightly out-
performs GAC and KeyGraph+ for recall, scoring 0.61 compared to 0.59 and 0.59.

ACM Transactions on Embedded Computing Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
A:12

Based on the F1 (mac-avg) score, GAC with a score of 0.68, LDA-GS with 0.70, and
KeyGraph+ with a score of 0.69 have similar performance overall. While the per-
topic average precision and recall scores for GAC, LDA-GS, and KeyGraph+ appear
to outperform KeyGraph, LDA-GS(p), and LDA-GS(p)+, there was no significant
difference as measured by the Wilcoxon Signed Rank test.
Figure 4 shows LDA-GS outperforms LDA-GS(p) and LDA-GS(p). This suggests
that our approach for incorporating named entities and noun phrases into Mallet has
a side effect that decrease the performance. However, using noun phrases and named
entities as features can slightly improve the performance of both KeyGraph and LDA-
GS algorithms. In summary, the performance of the KeyGraph algorithm on this well
known dataset is similar to the gold standard solutions GAC and LDA-GS.
We note that our comparison did not include the probabilistic model proposed by
Li et al. [13] that used the publication date as a feature. Their solution obtained an
F1 (mac-avg) score of 0.77 on the TDT4 dataset. Thus, the probabilistic model solution
did have a performance improvement compared to GAC and kNN. We further note
that it significantly dominated GAC and kNN on another dataset that was collected
over a much longer time interval [13]; exploiting the publication date of the articles
may explain its superior performance. As we discussed in section 3, we can incorpo-
rate publication time by using a dynamic community detection algorithm for topic fea-
ture extraction from KeyGraph. We plan to study dynamic network analysis as future
research.
We note that the probabilistic model does have the additional drawback that it must
estimate the number of events a priori. KeyGraph does not have this limitation. In
addition, KeyGraph has many features that make it both scalable and accurate for
noisy social media as well be demonstrated next.

4.3. User Evaluation of Social Media Events

We compare the KeyGraph algorithm KeyGraph+ with a topic modeling solution
LDA-GS on a social media dataset. We note that LDA-GS does not use noun phrases
and named entities. However, the experiments in the previous section shows that the
improvement in the performance is not statistically significant. In addition, our ap-
proach for incorporating named entities and noun phrases into Mallet has a side effect
that decrease the performance. Hence, we choose LDA-GS for human evaluation since
it was the best performing LDA algorithm in the previous section. In this experiments
each document will be assigned to at most 1 topic with the highest probability. We fil-
ter the documents that their highest probability to any topic is less than 0.1. This is
also the default threshold in Mallet toolkit.
4.3.1. Dataset Preparation. We chose the Spinn3r Dataset 2 (Table II); with greater
than 32,000 articles. We chose the default parameters for LDA-GS; α = 0.5 for smooth-
ing over topic distribution, and i = 1000 for the number of Gibbs sampling iterations.
The number of topics k is unknown a priori for the Spinn3r collection. We chose k = 200
for LDA-GS. Based on feedback from an expert on topic modeling, a larger number of
topics allows LDA to better handle noisy datasets.
KeyGraph identified 94 topics. The average number of relevant documents for the 94
topics identified by KeyGraph was 178, with a standard deviation of 538. In contrast,
the average number of relevant documents for the 200 topics identified by LDA-GS
was 164, with a standard deviation of 112. These statistics show that the LDA-GS
algorithm, even with hyper-parameter optimization, cannot cope with the many irrel-
evant documents in the dataset. The topics found by LDA-GS are much more uniformly
distributed than KeyGraph. KeyGraph could filter out unrelated documents, and the
topics are less uniformly distributed.

ACM Transactions on Embedded Computing Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
 A:13

Since KeyGraph+ and LDA-GS did not produce the same set of topics or the same
number of topics, one may argue that comparing their results is not reasonable. To
mitigate this criticism, we handpicked an identical number of the best topics, i.e., the
topics whose keywords and documents were more focused and most specific in their
content, for evaluation by humans. These topics were determined by a group of stu-
dents who did not participate in the evaluation. An alternative was to choose topics
of varying sizes. We did not do so since this would appear to benefit KeyGraph since
most of its topics were small. This has the potential to eliminate good topics of LDA-
GS that may be large(r).
4.3.2. Evaluation Metrics. Since we do not have the ground truth, we could not directly
determine precision and recall [13; 30]. However, we report on pseudo-precision and
pseudo-recall.
To evaluate pseudo-precision, we asked users to vote for the documents assigned to
each topic. Both KeyGraph and LDA provide a list of keywords for each topic as part
of their models. We used the list of top keywords to describe each topic. We then asked
users to vote for each document assigned to each topic; we obtained 3 independent
votes for each document. The vote was as follows:
(1) Not Relevant,
(2) Somewhat,
(3) Relevant,
(4) Highly Relevant.
To evaluate pseudo-recall, we asked users to vote for documents that were not as-
signed to the topics. For each of the topics, the set of on-topic documents are those
that are selected by some method; the set of off-topic documents are the ones that are
not selected. We use cosine similarity to select the most similar off-topic document to
match each on-topic document. We use this subset of matching off-topic documents
for pseudo-recall. As users find more of these off-topic documents to be not relevant,
then pseudo-recall increases. We used the list of top keywords identified by KeyGraph
and LDA to describe each topic. We asked users to vote for each matching off-topic
document and obtained 3 independent votes.
4.3.3. Evaluation with Amazon’s Mechanical Turk. We used Amazon’s Mechanical Turk for
the human user evaluation of the performance of the best 14 topics detected by the two
algorithms. For each of the 14 topics for each solution, we selected a random subset
between 50-150 documents. A total of 972 documents were evaluated for the precision
of Top 14 events of KeyGraph and 1132 documents for the precision of top 14 topics of
LDA-GS.
Tables III and IV present the statistics of the topics/documents and the results of
the user evaluation of precision for LDA-GS and KeyGraph+. As noted in the pre-
vious section, KeyGraph can filter irrelevant documents, but LDA-GS cannot. This
is reflected in the best 14 topics as well. The 14 topics included 972 documents for
KeyGraph+ and 2307 documents for LDA-GS. 12 of the Top 14 topics identified by
KeyGraph+ have less than 100 documents each and 8 have less than 65 documents
each. In contrast, the Top 14 topics of LDA-GS had a much larger number of docu-
ments per topic. Only 4 of the 14 topics has less than 100 documents each and 4 topics
had over 200 documents each.
This factor carries through to our second observation on the accuracy of the detected
topics. When averaged over all topics, only 40% of the documents for LDA-GS received
votes of relevant or highly relevant. The inter annotator agreement for LDA-GS votes
was 74%. In contrast, over 77% of the documents for KeyGraph+ received votes of

ACM Transactions on Embedded Computing Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
A:14

Table III. LDA-GS’s pseudo-precision using Mechanical Turk for Spinn3r

( 32,000 documents)

%Highly Relevant
#Evaluated Docs

%Not Relevant

%Somewhat

%Relevant
Event ID

#Docs
LDAGS1 341 120 33.61 25.56 23.33 17.50
LDAGS2 125 74 45.50 22.52 19.82 12.16
LDAGS3 358 100 27.00 33.00 25.33 14.67
LDAGS4 164 80 39.17 28.33 18.75 13.75
LDAGS5 87 87 44.44 28.74 17.62 9.20
LDAGS6 227 120 30.56 31.39 20.56 17.50
LDAGS7 91 91 55.31 23.81 12.09 8.79
LDAGS8 108 52 46.79 21.79 16.03 15.38
LDAGS9 93 66 36.87 26.77 18.18 18.18
LDAGS10 158 58 31.03 36.78 19.54 12.64
LDAGS11 139 79 44.73 25.32 19.41 10.55
LDAGS12 197 100 37.33 24.67 24.00 14.00
LDAGS13 99 60 48.89 19.44 13.33 18.33
LDAGS14 120 45 43.70 28.15 17.04 11.11
Total 2307 1132 49.91 33.24 23.41 17.12
Inter Annotator Agreement = 74%

Table IV. KeyGraph+’s pseudo-precision using Mechanical Turk for

Spinn3r ( 32,000 documents)

%Highly Relevant
#Evaluated Docs

%Not Relevant

%Somewhat

%Relevant
Event ID

#Docs

KEYGRAPH1 83 83 2.81 24.10 53.01 20.08

KEYGRAPH2 87 87 2.30 17.62 62.84 17.24
KEYGRAPH3 141 141 0.71 19.15 57.45 22.70
KEYGRAPH4 46 46 4.35 23.91 52.90 18.84
KEYGRAPH5 63 63 1.06 18.52 57.67 22.75
KEYGRAPH6 39 39 1.71 14.53 64.10 19.66
KEYGRAPH7 81 81 3.70 18.11 55.97 22.22
KEYGRAPH8 41 41 4.88 18.70 55.28 21.14
KEYGRAPH9 51 51 3.92 22.88 50.98 22.22
KEYGRAPH10 62 62 4.30 19.35 52.15 24.19
KEYGRAPH11 31 31 3.23 17.20 59.14 20.43
KEYGRAPH12 39 39 1.71 18.80 54.70 24.79
KEYGRAPH13 77 77 3.46 20.78 54.98 20.78
KEYGRAPH14 131 131 3.82 20.10 51.91 24.17
Total 972 972 3.00 19.55 55.93 21.52
Inter Annotator Agreement = 81%

relevant or highly relevant. The inter annotator agreement for KeyGraph+ votes was
81%.
For pseudo-recall, users voted for the Top 60 matching off-topic documents for each
topic. A total of 840 documents were evaluated for each method. Tables V and VI

ACM Transactions on Embedded Computing Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
 A:15

Table V. LDA-GS’s pseudo-recall using Mechanical Turk for Spinn3r

( 32,000 documents)

%Highly Relevant
#Evaluated Docs

%Not Relevant

%Somewhat

%Relevant
Event ID
LDAGS1 60 70.00 16.11 10.56 3.33
LDAGS2 60 66.67 15.00 11.11 7.22
LDAGS3 60 73.89 10.00 9.44 6.67
LDAGS4 60 90.00 3.89 3.89 2.22
LDAGS5 60 63.89 23.33 6.67 6.11
LDAGS6 60 90.00 3.89 3.89 2.22
LDAGS7 60 90.56 1.67 3.89 3.89
LDAGS8 60 79.44 8.89 8.89 2.78
LDAGS9 60 74.44 18.33 5.56 1.67
LDAGS10 60 61.67 22.22 10.56 5.56
LDAGS11 60 56.67 23.89 14.44 5.00
LDAGS12 60 92.22 5.00 1.11 1.67
LDAGS13 60 33.89 22.22 32.78 11.11
LDAGS14 60 85.56 8.33 5.56 0.56
Total 840 73.49 13.6 9.17 4.29
Inter Annotator Agreement = 86%

present the statistics of the topics/documents and the results. Note that as the number
of off-topic documents judged to be relevant decreases, pseudo-recall increases; this is
because these were off-topic documents that were not chosen by some method. When
averaged over all topics, about 13% of documents for LDA-GS received votes of rele-
vant or highly relevant. The inter annotator agreement for LDA-GS votes was 86%.
Similarly, about 11% of documents for KeyGraph+ received votes of relevant or highly
relevant. The inter annotator agreement for KeyGraph+ was 89%.
Thus, since we do not have ground truth for this dataset, we compared pseudo-
precision and pseudo-recall. These results show that KeyGraph+ can filter noise and
that the precision of KeyGraph+ appears to significantly outperform LDA-GS for
this noisy data collection. The improved precision of KeyGraph+ does not come at the
expense of lower recall; the pseudo-recall for KeyGraph+ is slightly better than for
LDA-GS.

4.4. Running Time of Methods

We discuss the complexity of KeyGraph and the LDA topic modeling with Gibbs sam-
pling (LDA-GS). We then compare the execution time of KeyGraph and LDA-GS on a
large data collection.
Let N be the number of documents in the collection, M the number of unique words
in the collection and M << N . K represents the number of topics. If D shows the max-
imum number of unique keywords in a document, the complexity of the three steps of
the KeyGraph algorithm is as follows: Building the KeyGraph is done in O(N ∗D +M 2 )
operations, which is the running time of a single pass through all documents in the col-
lection and building the KeyGraph (The number of nodes in a KeyGraph is limited to
M and the number of edges to M 2 ). The running time of the second step is equal to
the running time of the community detection algorithm. We used betweenness central-
ity approach which is running in O(M 3 ) operations[17]. The final step is another pass

ACM Transactions on Embedded Computing Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
A:16

Table VI. KeyGraph+’s pseudo-recall using Mechanical Turk for

Spinn3r ( 32,000 documents)

%Highly Relevant
#Evaluated Docs

%Not Relevant

%Somewhat

%Relevant
Event ID
KEYGRAPH1 60 77.78 11.67 6.67 3.89
KEYGRAPH2 60 85.56 9.44 2.78 2.22
KEYGRAPH3 60 40.56 28.89 23.89 6.67
KEYGRAPH4 60 85.00 7.22 5.00 2.78
KEYGRAPH5 60 73.89 15.00 8.89 2.22
KEYGRAPH6 60 48.89 19.44 21.11 10.56
KEYGRAPH7 60 92.78 1.67 1.67 3.89
KEYGRAPH8 60 84.44 8.33 5.56 1.67
KEYGRAPH9 60 86.11 9.44 3.33 1.11
KEYGRAPH10 60 71.11 17.22 7.78 3.89
KEYGRAPH11 60 88.89 7.78 3.33 0.00
KEYGRAPH12 60 85.00 8.89 4.44 1.67
KEYGRAPH13 60 82.78 8.89 5.00 3.33
KEYGRAPH14 60 87.78 7.78 2.78 1.67
Total 840 77.90 11.55 7.30 3.25
Inter Annotator Agreement = 89%

through the data collection which will be done in O(N ∗ D ∗ K) operations. Hence, the
running time can be summarized as the following:

— Building the KeyGraph: O(N ∗ D + M 2 )

— Topic feature extraction: KeyGraph community structure analysis using between-
ness centrality metric [17]: O(M 3 )
— Topic document assignment: O(N ∗ D ∗ K)
— Total Running Time: O(N ∗ D ∗ K + M 3 )

The total running time is linear in terms of the number of documents. Notice that sec-
ond term in the complexity of KeyGraph, M 3 , is the running time of the community
detection on the KeyGraph which is done by betweenness centrality approach. There
are faster community detection algorithm such as the algorithm of Radicchi et al. [20]
and the Kernighan-Lin algorithm [11] which run in O(M 2 ) operations [17]. Using the
faster community detection algorithm can reduce the complexity of the algorithm spe-
cially for small datasets which M is not much smaller than N .
Next we discuss the complexity of the topic modeling approaches. Let I be the
number of iterations for LDA with Variational Expectation Maximization. For each
document, in each iteration, the complexity of the computation is O(D ∗ K). Usu-
ally the inference iterations stops when ELBO (variational likelihood lower bound)
is met. Hence, the complexity of LDA with Variational Expectation Maximization is
O(N ∗ D ∗ K ∗ I). Similarly for LDA with Gibbs sampling, let I be the number of it-
erations, and L be the maximum number of keywords in a document or the length of
a document. For each document, in each iteration, the complexity of the computation
is O(L ∗ K). Gibbs sampling typically needs to be done for an absurdly large number
of iterations (about hundreds or thousands iteration). Hence, the complexity of LDA
with Gibbs sampling is O(N ∗ L ∗ K ∗ I). In summery, the complexity of a LDA Topic
modeling algorithm is as follows:

ACM Transactions on Embedded Computing Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
 A:17

Fig. 5. The running time of KeyGraph Algorithm versus LDA-GS implemented in Mallet toolkit

— LDA with Variational EM: O(N ∗ D ∗ K ∗ I)

— LDA with Gibbs Sampling: O(N ∗ L ∗ K ∗ I)
where I is the number of iterations.
In practice the number of iterations depends on many factors such as the number
of documents and topics as well as the length of the documents. Because this process
is all stochastic, it is hard to give rigorous bounds for the number of iterations before
convergence. Hence, we compare the running time of the KeyGraph and LDA-GS in
our experiments. Figure 5 reports on the execution time of KeyGraph+ and LDA-GS
as we increase the number of documents; we use Spinn3r Dataset. We create subsets
that are an increasing time slice of this dataset where the first subset is 1 day of data,
the second is 2 days of data, etc., up to 25 days of data. The number of documents
varied from 6,559 to 113,000 and is shown on the X axis of Figure 5. The experiment
was performed on a machine with an Intel Core 2 Due 2.93GHz processer and a 4GB
memory running Linux.
Figure 5 shows the total running time of LDA-GS and KeyGraph+. We stopped
running LDA-GS after 72,000 documents since it was very slow. We ran KeyGraph+
on up to 113,000 documents. The execution time of KeyGraph+ appears to increase
in a linear manner as predicted by the complexity analysis. In comparison, the execu-
tion time of LDA-GS increases in a non-linear manner. As the number of documents
exceeds 72,000 the running time of LDA-GS is almost an order of magnitude greater
than KeyGraph+.

4.5. Model Robustness

We studied the sensitivity of KeyGraph algorithm with respect to the parameters used
in the algorithm. If the result of an algorithm is very sensitive to the value of a param-
eter, then it will not be easily applicable to other domains and dataset. For algorithms
that have multi steps, if a sensitive step is not done perfectly it may affect the result of
the following steps significantly. Hence, we are interested to analyze the robustness of
our proposed model to its parameters. Recall KeyGraph community structure analysis
and extracting topic features from KeyGraph is the core of our model and robustness
of our algorithm to these steps is very important. We report the robustness of the al-
gorithm on three parameters: The first two parameters are α and β which control the
termination of the community detection algorithm. As our community detection algo-
rithm removes edges with the high betweenness centrality, α and β identify where to
stop splitting the graph. These are the termination parameters which are required for
any network community detection algorithms. For the betweenness community detec-

ACM Transactions on Embedded Computing Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
A:18

Fig. 6. Sensitivity analysis for parameter α and β that identify when to stop splitting the communities
for community detection on keyGraph, as well as µ which indicates the minimum overlap of cluster to be
merged in the last step of our algorithm.

tion algorithm that we used in our experiments, we stop removing edges if the highest
betweenness centrality score is less than a threshold θ (which uses α) or the size of
the community is less than β. θ is defined as the logarithmic function of the size of a
KeyGraph G: θ = α log |G| + 1.
The third parameter, µ, controls the process of merging topics with an overlap in
their document set in the last step of KeyGraph algorithm. If the document set overlap
of two topics is greater than µ, the topics will be merged.
Intuitively, KeyGraph model should be very robust to the value of α and β since if a
community is mistakenly broken into two or communities, it will be resolved later in
the last step which overlapping topics will be merged.
Here we did an experiment on the TDT4 dataset to evaluate the robustness of Key-
Graph. We ran the KeyGraph algorithm on a very large range of the values for all three
parameters and report F1 (mac-avg) for each run. Figure 6 shows the robustness of our
model on the TDT4 dataset with respect to parameter α and similarly β. We can see
that for a large range of α and β values the F1 (mac-avg) on the TDT4 dataset decrease
only a little bit. These results show that our KeyGraph algorithm is very robust to the
result of the community detection. Figure 6 also shows that our model is completely
robust to the value of µ which indicated the minimum overlap of cluster to be merged
in the last step of our algorithm.

5. CONCLUSIONS AND FUTURE WORK

In this paper, we present a graph analytical approach for fast topic detection, Key-
Graph. We use communities in a word co-occurrence graph as a proxy for topics. Using
an extensive evaluation we demonstrate the accuracy and scalable performance of the
KeyGraph algorithm, in particular for noisy social media datasets. Given the success

ACM Transactions on Embedded Computing Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
 A:19

of this heuristic solution, KeyGraph, we plan to fully explore an extension to topic

modeling to consider both word and keyword co-occurrence distributions.
KeyGraph is naturally time dependent. Our current approach is based on static
community detection and we use an aggregated KeyGraph. The timestamp is there-
fore very important information for event detection which is not exploited. We note
that even without the important time dependent information, the accuracy of Key-
Graph is similar to the best performing topic detection models on clean data, where no
model uses the timestamp. Further, KeyGraph outperforms the gold standard models
for noisy data.
Li et al. [13] showed that their method can exploit time dependent information and
can dominate on a data collection that is obtained over a longer interval. This result
motivates us to retain timestamps in the KeyGraph, and to explore methods from dy-
namic network analysis. Thus, in future work, we will explore the benefits of dynamic
community detection algorithms for event detection in a KeyGraph. Since these meth-
ods are computationally expensive, we will study the tradeoff of added accuracy and
determine incremental approaches to update the KeyGraph.

REFERENCES
C implementation of variational expectation maximization for latent dirichlet allocation (lda).
Http://www.cs.princeton.edu/ blei/lda-c/index.html.
Machine learning for language toolkit. Http://mallet.cs.umass.edu/topics.php.
Topic detection and tracking(tdt) project. Http://www.nist.gov/speech/tests/tdt/.
The future of social networking: Understanding market stratigic and technology developments. In Data-
monitor, 2007.
J. Allan, R. Papka, and V. Lavrenko. On-line new event detection and tracking. In SIGIR98: Proceedings of
ACM SIGIR, 1998.
A. Asuncion, M. Welling, P. Smyth, and Y.-W. Teh. On smoothing and inference for topic models. In UAI,
2009.
D. M. Blei, A. Y. Ng, and M. I. Jordan. Latent dirichlet allocation. J. Mach. Learn. Res., 3:993–1022, 2003.
T. Brants, F. Chen, and A. Farahat. A system for new event detection. In SIGIR ’03: Proceedings of the
26th annual international ACM SIGIR conference on Research and development in informaion retrieval,
pages 330–337, Toronto, Canada, 2003.
D. R. Cutting, D. R. Karger, J. O. Pedersen, and J. W. Tukey. Scatter/gather: A cluster-based approach to
browsing large document collections. pages 318–329, 1992.
Q. He, K. Chang, and E.-P. Lim. Analyzing feature trajectories for event detection. In SIGIR, pages 207–214,
2007.
B. W. Kernighan and S. Lin. An efficient heuristic procedure for partitioning graphs. The Bell system techni-
cal journal, 49(1):291–307, 1970.
G. Kumaran and J. Allan. Text classification and named entities for new event detection. In SIGIR ’04:
Proceedings of the 27th annual international ACM SIGIR conference on Research and development in
information retrieval, pages 297–304, United Kingdom, 2004.
Z. Li, B. Wang, M. Li, and W.-Y. Ma. A probabilistic model for retrospective news event detection. In SIGIR05:
Proceedings of ACM SIGIR, Brazil, 2005.
M. Mori, T. Miura, and I. Shioya. Extracting events from web pages. In AISTA’04: Proceedings of the Inter-
national Conference on Advances in Intelligent Systems - Theory and Applications (AISTA), 2004.
M. Mori, T. Miura, and I. Shioya. Topic detection and tracking for news web pages. In WI ’06: Proceedings of
the 2006 IEEE/WIC/ACM International Conference on Web Intelligence, pages 338–342, 2006.
D. Newman, J. H. Lau, K. Grieser, and T. Baldwin. Automatic evaluation of topic coherence. In Human
Language Technologies: The 2010 Annual Conference of the North American Chapter of the Association
for Computational Linguistics, HLT ’10, pages 100–108, Stroudsburg, PA, USA, 2010. Association for
Computational Linguistics.
M. E. J. Newman. Detecting community structure in networks. The European Physical Journal B - Con-
densed Matter and Complex Systems, 38(2):321–330, March 2004.

ACM Transactions on Embedded Computing Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
 A:20

Y. Ohsawa, N. E. Benson, and M. Yachida. Keygraph: Automatic indexing by co-occurrence graph based on
building construction metaphor. In ADL ’98: Proceedings of the Advances in Digital Libraries Conference,
page 12, Washington, DC, USA, 1998. IEEE Computer Society.
R. Prabowo, M. Thelwall, I. Hellsten, and A. Scharnhorst. Evolving debates in online communication: a
graph analytical approach. Internet Research: Electronic Networking Applications and Policy, 18(5):520–
540.
F. Radicchi, C. Castellano, F. Cecconi, V. Loreto, and D. Parisi. Defining and identifying communities in
networks. Proceedings of the National Academy of Sciences of the United States of America, 101(9):2658–
2663, March 2004.
N. Ruan, R. Jin, V. Lee, and K. Huang. Dynamic module discovery in temporal complex networks. In 2nd
Intl. Workshop on Analysis of Dynamic Networks, in conjunction with SIAM Intl. Conf. on Data Mining,
2009.
N. Ruan, R. Jin, V. Lee, and K. Huang. A sparsification approach for temporal graphical model decomposi-
tion. In ICDM’09: International Conference on Data Mining, 2009.
G. Salton and C. Buckley. Term-weighting approaches in automatic text retrieval. In Information Processing
and Management, volume 24, pages 513–523, 1988.
H. Sayyadi, M. Hurst, and A. Maykov. Event detection and tracking in social streams. In Proceedings of
International Conference on Weblogs and Social Media (ICWSM), 2009.
H. Sayyadi, S. Salehi, and H. Abolhassani. Nesrec: News meta-search result clustering”. In n Proceeding
of CIS2E 06 The International Joint Conferences on Computer, Information, and Systems Sciences, and
Engineering, pages 173–178, USA, 2006.
M. Steyvers and T. Griffiths. Probabilistic Topic Models. Lawrence Erlbaum Associates, 2007.
J. Tantrum, A. Murua, and W. Stuetzle. Hierarchical model-based clustering of large datasets through frac-
tionation and refractionation. Information Systems, 29(4):315 – 326, 2004. Knowledge Discovery and
Data Mining (KDD 2002).
H. Toda and R. Kataoka. A search result clustering method using informatively named entities. In WIDM
’05: Proceedings of the 7th annual ACM international workshop on Web information and data manage-
ment, pages 81–86, Bremen, Germany, 2005.
Y. Yang, N. Bansal, W. Dakka, P. G. Ipeirotis, N. Koudas, and D. Papadias. Query by document. In WSDM,
pages 34–43, 2009.
Y. Yang, T. Pierce, and J. G. Carbonell. A study on retrospective and on-line event detection. In SIGIR98:
Proceedings of ACM SIGIR, 1998.

ACM Transactions on Embedded Computing Systems, Vol. V, No. N, Article A, Publication date: January YYYY.

View publication stats