Extracting Representative Information on Intra-Organizational Blogging Platforms

Author(s): Xunhua Guo, Qiang Wei, Guoqing Chen, Jin Zhang and Dandan Qiao
Source: MIS Quarterly , December 2017, Vol. 41, No. 4 (December 2017), pp. 1105-1128
Published by: Management Information Systems Research Center, University of Minnesota

Stable URL: https://www.jstor.org/stable/10.2307/26630287

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms

Management Information Systems Research Center, University of Minnesota is collaborating with

JSTOR to digitize, preserve and extend access to MIS Quarterly

This content downloaded from

203.218.254.164 on Wed, 18 Oct 2023 08:47:49 +00:00
All use subject to https://about.jstor.org/terms
 RESEARCH ARTICLE

EXTRACTING REPRESENTATIVE INFORMATION ON

INTRA-ORGANIZATIONAL BLOGGING PLATFORMS1
Xunhua Guo, Qiang Wei, and Guoqing Chen
Research Center for Contemporary Management, School of Economics and Management, Tsinghua University,
Beijing 100084 CHINA {guoxh@sem.tsinghua.edu.cn} {weiq@sem.tsinghua.edu.cn}
{chengq@sem.tsinghua.edu.cn}

Jin Zhang
School of Business, Renmin University of China,
Beijing 100872 CHINA {zhangjin@rbs.ruc.edu.cn}

Dandan Qiao
Research Center for Contemporary Management, School of Economics and Management, Tsinghua University,
Beijing 100084 CHINA {qiaodd.12@sem.tsinghua.edu.cn}

Content generated on intra-organizational blogging platforms may help managers understand the emerging
ideas, issues, and opportunities of their companies, whereas the difficulty is how to go beyond the overload of
information and obtain an overall view. This paper proposes a system framework for extracting representative
information from intra-organizational blogging platforms, as well as the REPSET (REPresentative SET)
method, which serves as the core component of the extraction system. Drawing from a novel clustering
technique, REPSET is designed to identify a small set of items that largely represent the diversified content of
a huge information base. Building on REPSET, an extraction system enables managers to locate representative
articles that may serve as starting points for comprehensively understanding the hot topics, prevailing thoughts,
and emerging opinions among employees. Empirical evaluations are conducted based on the massive database
accumulated on an internal blogging platform at a large telecommunications company. The results from data
experiments and user evaluations demonstrate that REPSET and the extraction system upon which it is based
can provide outstanding performance, in comparison with benchmark methods.

Keywords: Intra-organizational blogging platforms, representative information extraction, enterprise social

media, business intelligence, design science

Introduction1 mation production on the Internet since the beginning of this

As one of the typical social media applications, blogs have
been prevailing and serving as a substantial source of infor-
1
Sumit Sarkar was the accepting senior editor for this paper. Manish
Agrawal served as the associate editor.
The appendices for this paper are located in the “Online Supplements”
section of the MIS Quarterly’s website (http://www.misq.org).
century (Nardi et al. 2004). Recently, blogs have also been
introduced into corporate contexts along with other social
media technologies such as wikis and group-messaging,
creating a new form of collaborative environment (Bente and
Karla 2009; Lu et al. 2015; McAfee 2006). It is widely be-
lieved that enterprise blogging platforms would help build a
flexible intra-organizational networking stage that may effec-
tively facilitate knowledge sharing and support emergent
innovations (Li et al. 2015; McAfee 2006; Suh et al. 2011).

MIS Quarterly Vol. 41 No. 4, pp. 1105-1127/December 2017 1105

This content downloaded from

203.218.254.164 on Wed, 18 Oct 2023 08:47:49 +00:00
All use subject to https://about.jstor.org/terms
Guo et al./Extracting Representative Information on Blogging Platforms
Many companies are investing heavily in internal blogging
platforms and encouraging their employees to maintain blogs
(Chui et al. 2012; Liao et al. 2011; Luo et al. 2015).
When employees have accepted an internal blogging platform
and are regularly posting their thoughts and writings on it,
information accumulated on such a crowd-based platform is
potentially useful for managers to support their managerial
work. Existing research has revealed that blog posts often
reflect real feelings of the authors (Nardi et al. 2004) and can
be considered as a reliable source of information (Aggarwal
et al. 2012). By reading internal blogs, decision makers will
be able to leverage the shared knowledge originally distrib-
uted across the organization (Li et al. 2015; Lu et al. 2015)
and react to emerging business opportunities or changing
organizational conditions in a more timely fashion.
The difficulty, however, lies in the huge amount of data gen-
erated on such a platform. An internal blogging system
typically produces thousands of posts every week. It is
almost impossible for a decision maker to read or even
browse all of the articles. The explosion of content creates a
challenge that, while access to knowledge and information is
facilitated, causes quality signals to become so faint and dif-
fused that cognitive overload occurs (Bawden and Robinson
2009). In other words, one critical obstacle to the gold mine
of valuable blog content is how to look beyond the overload
of information and obtain a clear view of the diversified
content on the platform. Specifically, to exploit the potential
value of internal blogging platforms, decision makers face a
fundamental question: Which articles will they read, given
the large amount of articles generated every day, in order to
better understand the prevailing thoughts and emerging
opinions among employees?
The information needs of managers in organizational contexts
are notably different from those in public blogging contexts.
On public blogs, users are usually concerned about the
content that fits with their own interest, as well as widely
discussed topics (e.g., top rated or most commented). In
contrast, in organizational contexts, to facilitate the mana-
gerial use of blog content, managers would prefer to be more
aware of diverse voices as well as emerging ideas and
opinions that might potentially be meaningful to the organi-
zation, no matter how popular they are or whether they fit
with the managers’ personal interest. In this sense, the quality
of representativeness of the extracted information is critical.
Literally, representativeness can be defined as the degree to
which a small set of items reflect the diverse content of a
large information base. A simple example for the need of
representativeness is that an employee posts an article that
implicitly conveys a novel idea for improving customer ser-
1106 MIS Quarterly Vol. 41 No. 4/December 2017
This content downloaded from
203.218.254.164 on Wed, 18 Oct 2023 08:47:49 +00:00
All use subject to https://about.jstor.org/terms
vice, but few other people read and discuss the idea. In this
case, the article will not be recommended by a popularity-
based system. It will also not be recommended to a manager
by a preference-based system if it does not fit with the man-
ager’s pre-stored preferences. However, the idea might prove
to be innovative if it is extracted and exposed to the manager.
Another example is that an employee posts an article
expressing some negative feelings. Using popularity-based
mechanisms, the article will not be recommended to the
manager unless it is widely discussed, but the manager may
find it useful to notice the existence of such articles at an early
stage. In other words, suppose that there is a box containing
100 balls, among which 99 are red (common articles) and 1 is
green (the special article). Now we want to pick out 10 balls
from the box to represent the population in the box. A
random sampling or a popularity-based extraction is likely to
result in 10 red balls. However, if a method can pick out 9
red balls and 1 green ball, this result can be deemed as more
“representative” because it informs us that there are green
balls existing as minority in the box.
Since existing methods are mostly designed for mining public
blogs, they usually focus on popularity and personalization,
whereas representativeness is rarely considered. For intra-
organizational contexts, it is meaningful to develop a novel
approach that is centered on representativeness. Therefore,
the objective of this study is to propose a system framework
capable of extracting a small number of articles that can
largely represent the diversified content generated on an
organizational blogging platform. To achieve this goal, we
develop REPSET (REPresentative SET), a business intelli-
gence (BI) method serving as the core component of the
extraction system. Featuring a novel clustering technique,
REPSET is designed for effectively identifying a small set of
items, which can represent the diversified content of a large
information base. To validate the effectiveness of REPSET,
empirical evaluations were conducted based on the massive
database accumulated on an internal blogging platform at a
large telecommunications company. An illustrative example
is provided depicting how our proposed method can facilitate
enhanced diversity in coverage, resulting in more representa-
tive results. Results from data experiments and user evalua-
tions demonstrate that REPSET and the extraction system
based on it can provide outstanding performance, in com-
parison with benchmark methods.
From the perspective of design science research (Goes 2014;
Gregor and Hevner 2013; Hevner et al. 2004), our work can
be seen as positioned in the “exaptation” quadrant of the
design science knowledge contribution framework (Gregor
and Hevner 2013), as it focuses on the relatively new problem
of representative information extraction on internal blogs and
develops REPSET as an exaptation of related BI methods in
 Guo et al./Extracting Representative Information on Blogging Platforms
response to the particular challenges of that problem. There-
fore, this study aims to make contributions in the form of
design artifacts including the REPSET method as well as the
formalization and instantiation of the representative informa-
tion extraction framework. Based on this motivation, this
work has been conducted in line with the design science
methodological principles (Hevner et al. 2004).
Literature Review
In the BI literature, there are efforts to develop reliable
methods for monitoring and analyzing online content. For
instance, research on online opinion mining aims to determine
the attitude of the public toward news, products, or services
(Turney and Littman 2003; Wright 2009). Methods have
been developed for collecting and analyzing social network
content such as blogs (Chau and Xu 2012), extracting small
sentiments from a relatively large message body (Turney and
Littman 2003), and discovering relationships between infor-
mation items (Ku et al. 2009). Meanwhile, models and
systems have been built based on various techniques to
achieve diverse managerial goals, such as classifying and
summarizing product reviews (Hu and Liu 2004; Zhang et al.
2016), predicting sales performance (Liu et al. 2007), esti-
mating the effectiveness of communication (Abbasi and Chen
2008), and providing decision support for business mergers
and acquisitions (Lau et al. 2012). Underlying most of the
research on online content mining, the most widely addressed
BI techniques are topic clustering (Pons-Porrata et al. 2007;
Zhao et al. 2005), which is the unsupervised classification of
a given document set into clusters such that documents within
each cluster are more similar between each other than those
in different clusters, and text summarization (Das and Martins
2007; Gupta and Lehal 2010), which aims to condense the
source text into a shorter version while preserving its
information content and overall meaning.
Previous research on topic clustering is usually based on
partitional methods, such as K-Means (Jain et al. 1999) and
graph-partition techniques (Han et al. 1998), density-based
methods (Ester et al. 1996), or hierarchical methods (Berkhin
2006). Among them, the partitional and density-based
methods have a linear time complexity, which is very effi-
cient. However, such methods may produce inferior clusters
in some cases and the number of clusters is required to be pre-
determined, which is usually difficult in practice, as well as in
the context of this study. Locality sensitive hashing (LHS) is
another technique related to partitional methods (Gionis et al.
1999; Indyk and Motwani 1998). LHS maps similar data
objects to the same buckets (Leskovec et al. 2014), so it can
be applied to partition data objects. However, like other
This content downloaded from
203.218.254.164 on Wed, 18 Oct 2023 08:47:49 +00:00
All use subject to https://about.jstor.org/terms
partitional clustering methods, the number of hash tables
needs to be predetermined, which may restrict its application
in the research scenario of this study. Compared to partitional
methods, the complexity of hierarchical clustering methods is
higher. Meanwhile, in hierarchical methods, if a data object
is assigned to a wrong cluster at a stage, it will never be re-
allocated in the following clustering process, missing the
possible opportunity to further refine the clustering results.
Furthermore, topic clustering in online content analysis
usually focuses on finding the most common and dominant
opinions, which is different from our emphasis on represen-
tativeness. For instance, K-Means clustering, the most com-
monly used partitional method, suffers from at least three
well-known issues: (1) it tends to generate spatially equi-
distant clusters; (2) it fails to effectively consider different
data densities; and (3) it fails to consider data sets with non-
globular structures effectively (Han et al. 2011). While
certain methods fare better on some of these issues (for
instance, DBScan is better for the density problem and nearest
neighbor-based methods have alleviated the non-globular
problem), no existing method effectively addresses all three
of these issues. Considering that the issues of different sizes
and densities are even more crucial and common in the
context of representative information extraction, the deficien-
cies of such methods may be amplified with the emergence of
novel ideas and opinions that are still not popular and of small
sizes and densities.
Another type of techniques related to topic clustering is topic
modeling, such as LDA (Blei et al. 2003), hLDA (Blei et al.
2004), and DTM (Blei and Lafferty 2006). These techniques
typically decompose documents into topic terms, so that every
document is associated to a topic with a certain probability.
The probability that a document belongs to a topic is higher
when the document contains more discussion on that topic
(Blei 2012). Therefore, for each topic, the document with the
largest probability may be selected as the representative item.
These methods offer the advantage that the relationship
between a document and the topic that it represents can be
effectively explained. However, topic modeling methods
generally focus on the frequencies of terms. Therefore, they
tend to obtain topics based on frequently used terms while the
“long tail” terms are likely to be ignored.
Additionally, text summarization methods may also be con-
sidered as a foundation for representative information
extraction to some extent. In the existing literature, two types
of text summarization techniques have been well developed,
namely extractive methods and abstractive methods (Aligu-
liyev 2009; Cai and Li 2011). Between them, extractive
methods are more well developed. The disadvantage of such
methods, however, is that the extracted text summaries often
MIS Quarterly Vol. 41 No. 4/December 2017 1107
Guo et al./Extracting Representative Information on Blogging Platforms
lead to problems in overall coherence and are very difficult
for users to understand in practice, as has been repeatedly
pointed out by the existing literature (Gupta and Lehal 2010).
In contrast, abstractive summarization relies on advanced
language generation techniques to produce grammatical sum-
maries, but the quality of the summaries heavily depends on
sophisticated language processing tools that are not always
adaptable in various contexts, making abstractive methods
less usable and less studied than extractive methods (Cai and
Li 2011). Furthermore, with both types of summarization
methods, when the summaries are extracted, the integrity of
every single article cannot be maintained and authorship
information usually cannot be included. Consequently, it is
difficult for readers to locate the sources of particular senti-
ments or to find a typical individual article that contains the
sentiments.
Therefore, although text summarization can be used to obtain
good information coverage in the extraction of representative
content, it is still far from perfect because it suffers from the
above-mentioned disadvantageous attributes. In this sense,
we regard it as a worthwhile attempt to develop a novel
method to achieve representative information extraction
through a different path. In this paper, we focus on the
method for extracting representative items. This new method
is not intended to replace sentiment analysis or text summari-
zation. Instead, we believe that the efforts on text summari-
zation and representative set extraction can develop in
parallel, from different perspectives.
Design Framework
Our vision for the representative information extraction
framework was initially shaped through our on-site observa-
tion on the use of two internal blogging systems in practice.
Subsequently, we conducted a series of field interviews to
refine the framework as well as to derive the specific design
requirements. Prior design science studies have used inter-
views to understand artifact use cases, strengths and weak-
nesses, and potential areas for improvement (Adomavicius et
al. 2008; McLaren et al. 2011). From 2010 to 2012, we
investigated a large Chinese telecommunications company
(X), a multinational software company (M), and a collabora-
tion software vendor (Y). Both X and M had deployed and
were using the internal blogging systems actively. Company
Y was a vendor of internal blogging systems and had imple-
mented their products in more than 10 customer companies.
We visited companies X (four times) and M (one time) and
observed the use of the internal blogging systems on-site.
Our interviewees included five persons from X (CEO, one
department head, two team managers, and one engineer), one
person from M (department head), and two persons from Y
1108 MIS Quarterly Vol. 41 No. 4/December 2017
This content downloaded from
203.218.254.164 on Wed, 18 Oct 2023 08:47:49 +00:00
All use subject to https://about.jstor.org/terms
(CEO and engineer). In the interviews, the envisioned frame-
work was conveyed to the interviewees. All of the inter-
viewees replied with a positive confirmation that such a
representative information extraction system was desirable.
We then asked them to describe what features they might
expect for such extraction systems. Based on their responses,
we developed the design requirements for the extraction
method.
The overall design framework is for a system that can extract
representative information from corporate internal blogging
platforms using a business intelligence approach. As a highly
conceptualized illustration, the framework is shown in
Figure 1. It highlights the relationship between the corporate
internal blogging platform and the representative information
extraction system, of which the central component is the
extraction engine. In the following, we will discuss the blog-
ging platform and the extraction engine respectively, and then
elaborate on the technical requirements for the extraction
method implemented in the engine.
The Enterprise Blogging Platform
The extraction system is designed as an add-on to the cor-
porate internal blogging platform. Generally, blogs are fre-
quently modified web pages, in which dated entries (articles)
are listed in reverse chronological sequences (Jackson et al.
2007). Readers are usually allowed to post comments to
specific articles in a blog. An article may contain both text
and graphical elements, and even multimedia information
such as video and music. In this paper, we focus on the text
content of blog articles.
The blogging platform usually provides some predefined
mechanisms for organizing the blogs and articles in a struc-
tured way. Typically, blogs are categorized according to the
organizational structure. Additionally, a set of article cate-
gories (or tags) are defined. The author of a blog article can
choose to attach the article to particular categories. Such
structures may help readers browse the contents on the plat-
form. Moreover, on some blogging platforms, readers are
allowed to rate the articles that they have read. Based on
reader ratings, top rated articles will be recommended to a
new reader, in case that he/she cannot decide which articles
to read first. Platform administrators may also rate the articles
and mark some of them as “recommended” based on their
own judgment. All of these mechanisms are designed in order
to help readers more easily seize the essential articles that
deserve reading. Although there are auxiliary mechanisms
such as category structures, reader ratings, and administrator
recommendations, it remains difficult for readers, especially
 Guo et al./Extracting Representative Information on Blogging Platforms
Figure 1. Design Framework for the Extraction System
senior managers who do not have much time to spend on
browsing the blogging platform, to quickly and accurately
find out the most information-rich items. In this regard, the
extraction system should be of important help.
The Representativeness Extraction Engine
The core component of an extraction system is the extraction
engine, which fetches data from the blogging platform,
analyzes the content of blog articles together with platform
usage records (including records about user browsing,
reading, and commenting), and outputs representative articles
as results. The representative articles will be presented to
users who are concerned with grasping the overall situation of
the organization. Herein, representative articles are defined
as a small set of blog articles, which, to a large extent, cover
the content of all the blog articles posted on the blogging
platform within a specific time period.
Representative articles extracted by the system are com-
parable to the top-rated articles generated by the user rating
mechanism, considering the fact that they are both small sets
of articles recommended to readers to help them select what
to read. The difference, however, is that representative
articles are selected with the aim of representing the whole
population of blog articles posted within a specific time
period, while top-rated articles are simply the posts most
favored by general readers. From the perspective of a high
level decision maker, apart from knowing what is most
favored in the organization, it is at least equivalently
important to capture an overview on the whole platform.
This content downloaded from
203.218.254.164 on Wed, 18 Oct 2023 08:47:49 +00:00
All use subject to https://about.jstor.org/terms
Requirements for the Extraction Method
As described above, the principal requirements for the extrac-
tion engine were derived by studying two blogging systems in
practice and interviewing typical users. In the interviews, the
interviewees were asked to describe the features that they
would expect for such extraction systems. In light of the
responses, the three most desirable features turned out to be
(1) From the perspective of functionality, the extracted
articles should be indeed representative, which means
that they should cover the content of the targeted article
population to a large extent.
(2) From the perspective of usability, users would like to
freely specify the number of representative articles to
display and the results should be presented in real time.
For example, a user usually prefers to read 10 represen-
tative articles every day, but occasionally he/she may
choose to read 20 or more articles when he/she has more
time. Similarly, user A may prefer to read 10 articles
every time but user B may prefer to read 30. Therefore,
the system should allow the users to select the number of
articles in real time according to their needs, rather than
fixing it to a predetermined value.
(3) From the perspective of information granularity, users
would like to “drill down” on representative articles (i.e.,
after reading a presentative article, they may want to read
some more articles that are representationally related)
and such queries should also be responded to in real time.
MIS Quarterly Vol. 41 No. 4/December 2017 1109
Guo et al./Extracting Representative Information on Blogging Platforms

Documents Document Document Clusters , kth Representative

Dendrogram Level of the Dendrogram Documents
Similarity
Matrix
Level 1
Construction

Level 2 S
E
Clustering L
… E
Level k C
… T
I

…
…
… … O
N
…

Level n-1 …

Level n …

Figure 2. REPSET Procedure

Based on these design requirements, we propose a BI method Similarity Matrix Construction

called REPSET (REPresentative SET), which incorporates a
novel document clustering technique. Evaluated with mul-
tiple measures about representativeness, REPSET is proved
to be an effective method supporting the extraction system.
In the following sections, we will describe the REPSET
method and present the empirical evaluations.
The REPSET Method
According to the definition of representativeness, the target of
representative article extraction is to find a small set of
articles that, to a large extent, reflect the diverse content of the
blogging system. Generally, this target can be achieved with
two major steps: (1) categorize all of articles based on their
similarities of content between each other; and (2) select a
good representative article for each category. If the cate-
gorization of articles and the selection of representative items
are more accurate, the degree of representativeness of the
resulting set will be higher. Accordingly, the overall proce-
dure of the REPSET method is shown in Figure 2. As illus-
trated, in the first step, the degrees of similarity between
documents are calculated. That is, given n documents, an
n × n similarity matrix would be constructed. Next, a clus-
tering algorithm is utilized to cluster the documents into a
dendrogram with different granularities of clusters. Finally,
for all given levels in the document dendrogram, a repre-
sentative document is selected from each cluster according to
the degrees of similarity. From the document dendrogram,
the manager could obtain various levels of clustering results
and representative articles according to the need, instead of
determining the cluster number k in advance.
The following three subsections will introduce the three
components of REPSET in detail, respectively.
The similarity between two documents can be calculated
using the cosine-similarity measure (Salton 1971), which is
widely used in web/text mining. Using Lucene, a commonly
used word-splitting tool, each document can be converted into
a set of words. With the standard treatment in the TF/IDF
model (Salton 1971), a feature word list can be derived, based
on which the corresponding keyword vector space can be
generated. Each document d then can be mapped as a
keyword vector denoted as [ω1, ω2, …, ωp] with the TF/IDF
model. An alternative way is to use “shingles” (contiguous
subsequences of tokens), instead of keywords, to denote a
document (Broder et al. 1997). In our study, keyword vectors
are adopted to reduce computational complexity. Thereafter,
the cosine-similarity between every two documents can be
calculated based on their corresponding keyword vectors.
Specifically, the cosine-similarity between any two docu-
ments d1 and d2 (d1 = [ω11, ω21, …, ωp1] and d2 = [ω12, ω22, …,
ωp2]) in a p-dimensional vector space can be calculated with
the following Sim measure:
Sim(d1 , d 2 ) = cos(θ between vectors d1 and d 2 )
d1 • d 2
=
|| d1 || × || d 2 || (1)
ω11 × ω12 + ω21 × ω22 + ... + ω p1 × ω p 2
=
ω112 + ω212 + ... + ω p21 × ω122 + ω222 + ... + ω p2 2
Given a set D with n documents, an n × n similarity matrix M
can be calculated with Equation (1). In M, the element at row
i and column j represents the similarity between the docu-
ments i and j.

1110 MIS Quarterly Vol. 41 No. 4/December 2017

This content downloaded from

203.218.254.164 on Wed, 18 Oct 2023 08:47:49 +00:00
All use subject to https://about.jstor.org/terms
 Guo et al./Extracting Representative Information on Blogging Platforms

Clustering In Equation (2), nkm and nkn represent the number of docu-
ments in clusters Ckm and Ckn, respectively. At stage k, two
Based on the three desirable features for the extraction engine, clusters with the maximum average similarity, Ckp and Ckq,
we derived three principal requirements for the clustering will be merged into a new cluster, Ckpq, which is demonstrated
algorithm: (1) the quality (accuracy) of clustering results as Equation (3).
should be high, because this is the basis for high quality of
representative extraction; (2) the number of clusters should
not be arbitrarily determined in advance (it should be able to
(
Sim C pk , Cqk = ) {
Cmk ,Cnk ∈
max
C1k ,C2k ,,Cnk−k +1 } (
 Sim Cmk , Cnk 
 ) (3)
be flexibly adjusted when wanted, without having to restart
the whole procedure of clustering), so that the system can
Furthermore, given a threshold λ (0 # λ # 1), documents with
respond immediately once the user has selected a different
average in-cluster similarities less than λ (near boundaries)
number of representative articles; and (3) there should be a
will be marked as boundary objects (to be considered to
multilevel structure in the resulting clusters, so that the
reallocate). Specifically, in the new cluster Ckpq, documents
navigation features such as drill down for representative
meeting the following condition will be marked as dboundary:
articles can be facilitated.

 Sim(d ) < λ , λ ∈[0,1]

1
Based on examining existing algorithms that are commonly boundary , d x (4)
used for clustering analysis, we found that they cannot suffi- n kpq k
d x ∈C pq
ciently satisfy these three requirements simultaneously. In the
partitional methods, the number of clusters needs to be arbi-
where nkpq represents the number of documents in the cluster
trarily set before the clustering algorithm is executed. In other
Ckpq. Notably, the value of the parameter λ may affect the
words, one execution of the partitional algorithm can only
accuracy of the clustering algorithms. Before the extraction
result in one fixed and pre-given number of clusters. Mean-
of representative information, the value of λ can be tuned in
while, partitional methods are not applicable for varying sizes,
light of the following measure:
densities, and structures. In the density-based methods, the
minimum numbers of points and radius need to be configured k
 1 
in advance, which indirectly fixes the number of clusters. In
addition, both partitional clustering and density-based clus-   |C | p di ∈C p
{Sim(d , O )}
i p
tering typically create one-level partitioning of the data p =1
CS = k (5)
{ ( )}
objects at once and cannot provide various levels of repre-  
sentative information. In contrast, the hierarchical clustering   max
q = 1,2,, k , p ≠ q
Sim Op , Oq 

can meet both requirements (2) and (3). However, the p =1
accuracy of hierarchical clustering is relatively low (Stein-
bach et al. 2000). Table 1 summarizes the key features of the In the equation, the numerator represents the aggregate in-
clustering methods. cluster similarity after clustering. Cp denotes a cluster and Op
is the centroid document of Cp. The denominator represents
Therefore, in REPSET, a novel method is proposed based on the aggregate between-clusters similarity after clustering,
hierarchical clustering in order to retain the two desirable where Op and Oq denote the centroid documents of two
features. Additionally, a heuristic strategy, namely the back- clusters, respectively. The clustering result is better when in-
ward strategy for documents near boundaries, is incorporated cluster similarities are high and between-clusters similarities
to improve the quality of clustering. The detailed procedure are low. In the clustering process, given a certain λ value,
of the clustering algorithm is as shown in Figure 3. boundary documents can be determined and reallocated when
necessary. After the reallocation procedure, the CS value is
Given a set D with n documents, in the proposed clustering calculated accordingly. In this sense, the value of λ will
algorithm, each document is initially considered to be a single influence the clustering result and therefore have impacts on
cluster. Therefore, at the first stage, there are n clusters, the final CS value. A higher CS value indicates a more appro-
denoted as C11, C12, …C1n. Let Ckj denote the jth cluster in the priate value of λ. In other words, the proper input parameter
kth stage in the clustering process of REPSET. At each stage, of λ in REPSET can be determined by seeking for the maxi-
the average similarity of each pair of clusters Ckm, Ckn is mized CS value. Since the expression of CS does not expli-
calculated as follows: citly include λ, a commonly used parameter-tuning technique
called stepwise self-tuning (Zelnik-Manor and Perona 2004)
( ) (
Sim Cmk , Cnk = 1 / nmknnk ) 

d x ∈Cmk   d y ∈Cnk
( ) 
Sim d x , d y 

(2) is adopted to achieve this goal. This technique has been
widely applied in parameter tuning for various studies, such

MIS Quarterly Vol. 41 No. 4/December 2017 1111

This content downloaded from

203.218.254.164 on Wed, 18 Oct 2023 08:47:49 +00:00
All use subject to https://about.jstor.org/terms
Guo et al./Extracting Representative Information on Blogging Platforms

Table 1. Features of Clustering Methods

Number of Multilevel Applicable for Varying Sizes,
Methods Accuracy Clusters Clusters Densities, and Structures
Partitional clustering medium-high fixed no no
Density-based clustering medium-high fixed no yes
Hierarchical clustering low-medium flexible yes yes

No Output:
Document
dendrogram
th
Any Record the k
Input:
document to level of the
D, M
merge? dendrogram
Yes Backward
strategy
Identifying the
Merging two Reallocating
boundaries of
closest documents near
the new
clusters boundaries
cluster

Figure 3. Clustering Algorithm

as cluster numbers and other related parameters in clustering
(Gu and Zhou 2009; Mandala et al. 2014; Seref et al. 2013),
topic numbers in LDA (Blei et al. 2003; Griffiths and Stey-
vers 2004), confidence and support degrees in association-rule
classification (Chen et al. 2006; Coenen and Leng 2007),
normalization parameters in information retrieval (He and
Ounis 2003), and regularization parameters in logistic regres-
sion (Simon et al. 2011).
With stepwise self-tuning, values of λ with equal intervals are
applied on a training dataset that is randomly sampled from
the data set D. Typically, with a step length of 0.1, the value
of λ can be set as {0.1, 0.2, 0.3, …, 0.9}, respectively.
Notably, in the clustering process of REPSET, CS values can
be calculated at every level of the dendrogram. However,
because the sample data set for tuning λ is usually much
smaller than the whole data set D, it is not feasible to tune λ
for every level independently. Therefore, a unified value of
λ will be used for all levels of the dendrogram. In the tuning
process, for each given value of λ, the clustering algorithm is
run on the sample data set, during which CS values are
recorded at all levels in the dendrogram. Subsequently, the
average value of CS is calculated as the indicator for choosing
λ. The λ value that results in the highest average CS value is
chosen as the final input parameter for REPSET to perform
the clustering in the dataset D.
gain the highest average similarity, if necessary. This is
called a backward strategy. Specifically, at stage k, there are
(n – k +1) clusters, denoted as Ck1,Ck2, …, Ck(n-k+1). For the
marked documents dboundary, REPSET first calculates the
average similarities between dboundary and the (n – k +1)
clusters, which is shown as Equation (6).
(
Sim dboundary , C kj = ) 1
n kj  Sim(d boundary , d x ) (6)
d x ∈C kj
If the average similarity between dboundary and Ckj is the largest
among all the clusters, then dboundary is assigned to the cluster
Ckj.
In each iteration of updating clustering, two clusters with the
largest similarity between each other is merged into one, so
the number of clusters decreases by 1. Therefore, in stage k,
the documents in D can be clustered into (n – k + 1) sets of
similar documents. The (n – k + 1)th level of the document
dendrogram is constructed by these sets, in which the docu-
ments near boundaries have been appropriately processed.
After n iterations of the proposed algorithm, the clusters will
converge to 1 and the whole dendrogram will have been well
constructed. Every level of this dendrogram will consist of
document clusters with good clustering quality.

The marked boundary documents will be reevaluated, and The backward strategy in REPSET is designed to overcome
reallocated to some new cluster, in which the document will the disadvantage of traditional hierarchical clustering methods

1112 MIS Quarterly Vol. 41 No. 4/December 2017

This content downloaded from

203.218.254.164 on Wed, 18 Oct 2023 08:47:49 +00:00
All use subject to https://about.jstor.org/terms
 Guo et al./Extracting Representative Information on Blogging Platforms
Figure 4. Backward Strategy Example
that the clusters cannot be dynamically adjusted during the
construction process of the dendrogram. This heuristic
strategy aims to improve the clustering quality by assigning
boundary objects to more appropriate clusters, based on the
rationale that the boundary objects in a newly merged cluster
are likely to belong to another cluster. Specifically, in the
traditional hierarchical clustering process, in each iteration,
two items or clusters with the highest similarities will be
merged. Consider the example in Figure 4, where each circle
denotes an article. From the overall perspective, the desired
clustering result should put all white circles into one cluster
and all black ones to another. However, in the step-wise
dendrogram, because the distance between di and dj is very
close (high similarity), these two articles will first be merged
into one cluster and will never be separated in the subsequent
procedures. In contrast, the backward strategy in REPSET
can effectively solve this problem. Suppose that di and dj have
been merged in Clusterij, and this cluster is then merged with
the other black articles forming Clusterblack, while the other
white articles are clustered as Clusterwhite. In this case, for
each document in Clusterblack, REPSET will calculate the
average similarity between it and the other documents, and
then accordingly mark out the boundary articles in this
cluster. In the example, article di will turn out to have a low
average similarity. Therefore, REPSET will recalculate the
average similarities between di and all existing clusters. Con-
sequently, REPSET will find that di has the largest average
similarity to the Clusterwhite, which means di is more likely to
be an article in Clusterwhite. Therefore, di is reallocated into
Clusterwhite. Thus, the clustering accuracy will be improved.
A more detailed illustrative example is provided in Appen-
dix A.
As compared with traditional hierarchical clustering methods
such as agglomerative clustering, the novelty of REPSET
clustering lies in the heuristic design of the backward strategy
that significantly improves the clustering accuracy. There-
fore, it will be capable of overcoming the major weakness of
hierarchical clustering algorithms and obtain higher accuracy.
Meanwhile, although the computation time is a weakness of
This content downloaded from
203.218.254.164 on Wed, 18 Oct 2023 08:47:49 +00:00
All use subject to https://about.jstor.org/terms
hierarchical methods, as well as REPSET, this is not a big
issue in our context, since the clustering procedure only needs
to run one time on a daily basis and it can be scheduled as an
automatic task that runs at night. Particularly, as REPSET
allows real time flexible query and there is no need to re-do
the clustering when a new number of representative articles is
set, the computation time issue is even more marginal for this
method.
Furthermore, the method is able to capture more diversified
content and, consequently, innovative ideas and emerging
opinions are likely to be included in the extracted item set.
Considering an extreme case of an innovative idea that is new
and rarely discussed, it will have very low similarity levels
with other articles. Therefore, in the clustering procedure of
REPSET, this item is not likely to be merged with other items
or clusters, because of its low similarities with others. In
other words, this innovative idea will be maintained indepen-
dent of other items or clusters all through the clustering
process and eventually become an individual cluster. Subse-
quently, when selecting representative items, since this
innovative idea is the only one item in its cluster, it will be
certainly extracted.
Representative Document Selection
Given a set of documents, the document with the highest
average similarity to the other documents will offer the
highest content coverage in the set (Liu and Croft 2008).
Therefore, such a document can be selected as the represen-
tative in a cluster.
Particularly, the clustering process of REPSET can provide
the whole dendrogram after one time running and every level
of this dendrogram contains document clusters with good
clustering quality. At each level of the dendrogram, REPSET
can extract the corresponding representative documents in the
clusters, forming the representative subset of documents at
this level. Therefore, after running one time, REPSET can
MIS Quarterly Vol. 41 No. 4/December 2017 1113
Guo et al./Extracting Representative Information on Blogging Platforms

Input:
Document Set D = {d1, d2, …, dn} // D is the set of n documents
Output:
Dendrogram C = {C1, C2, …, Cn}; // The document dendrogram C
// Ck represents the kth level of C
Set of representative documents // A series of representative documents
RepSet = {R1,..., Rki, …, Rn}; // Rki : the representative documents for Cki
Begin:
1. RepSet = ∅; // Initially, RepSet is set empty
2. Number_of_Clusters = n; // Initially, it could regarded as n clusters
3. Clusteri = {di}, i = 1, 2, …, n; // Each cluster containing one document
4. Cn = {Cluster1,…, Clustern}; // The nth level of the dendrogram
5. Level = n; // record the level of the dendrogram
6. M = Cosine_Similarity(D); // Calculate the similarity matrix M
7. While Number_of_Clusters > 0 Do // Clustering procedure
8. {
9. Level = Level – 1;
10. CLevel = CLevel+1;
11. While(Avg_Sim(Clusterp, Clusterq)
= Maxi, j = 1, 2, …, Number_of_Clusters, i ≠ j(Avg_Sim(Clusteri, Clusterj) Do
12. {
13. Clusterp = Merge(Clusterp, Clusterq); // Update sets
14. Reallocate_Boundaries(); // Backward strategy
15. Number_of_Clusters --;
16. }
17. CLevel = Remove(Clusterq); // Record the kth level of the dendrogram
18. }
19. For Level = 1 to n // Select representative documents
20. {
21. For Clusteri in CLevel
22. RLevel = RLevel + Select(Clusteri); // Update RLevel
23. RepSet = RepSet + RLevel;
24. }
25. Output(RepSet); // Output the representative information

Figure 5. REPSET Procedure Pseudo-Codes

provide representative subsets at all levels of the dendrogram,
guaranteeing real time flexible queries.
Algorithm Summary
Figure 5 presents the structural pseudo-codes for the proposed
method. In the codes, lines 1–6 represent the components of
initialization and similarity matrix construction. Lines 7–18
correspond to the component of the clustering procedure.
Lines 19–25 are the component of representative document
selection.
The overall time complexity of REPSET is low-level poly-
nomial. At the first stage (initialization and similarity matrix
construction), the main computational consumption is the
calculation of the similarity matrix, which is of O(n2) time
complexity. At the second stage (clustering), suppose that the
average scale of boundary documents in newly generated
clusters is navg, under the best-case situation, where there is no
document to be reallocated throughout the process, the
clustering procedure degenerates into a typical hierarchical
clustering method that is of O(n2logn) time complexity (Liu
et al. 2007). Under the worst-case situation, where every
boundary document in the newly merged cluster satisfies the
backward strategy condition in each step of the procedure, the
time complexity of the clustering procedure will be
O(n3navglogn). At the third stage (representative document
selection), the time complexity is O(n2). Therefore, the over-
all time complexity of REPSET is between O(n2logn) (best-

1114 MIS Quarterly Vol. 41 No. 4/December 2017

This content downloaded from

203.218.254.164 on Wed, 18 Oct 2023 08:47:49 +00:00
All use subject to https://about.jstor.org/terms
 Guo et al./Extracting Representative Information on Blogging Platforms
case) and O(n3navglogn) (worst-case). Since navg is less than n,
it can be concluded for simplicity that the average-case time
complexity of REPSET is between O(n2logn) and O(n4logn).
Benchmark Data Evaluation
In order to evaluate the performance of the clustering
algorithm in REPSET, we compare it with 10 widely used
clustering algorithms, including four K-Means type methods
(RBR, DIRECT, Simple K-Means, and X-Means) (Marmanis
and Babenko 2009; Pelleg and Moore 2000; Zhao et al. 2005),
four variances of the agglomerative method with different
intercluster similarity calculation methods (Agglo(group),
Agglo(max), Agglo(min), and Agglo(centroids)) (Zhao et al.
2005), a graph partitioning method (GRAPH) (Zhao et al.
2005), and a density-based method (DBscan) (Ester et al.
1996), with respect to two major measures (i.e., entropy and
purity) (Bezdek and Pal 1998; Liu 2007), on some benchmark
datasets. Moreover, to evaluate the effectiveness of the pro-
posed backward strategy in clustering, REPSET with no
backward strategy is also considered in the comparative
experiments. The 10 algorithms cover the main categories
with respect to the mainstream clustering methods (Liu 2007).
Among them, RBR, DIRECT, Simple K-Means and Agglo
are deemed effective document clustering algorithms (Stein-
bach et al. 2000) and are adopted by the well-known CLUTO
clustering platform (Zhao et al. 2005). Graph is a typical
graph clustering algorithm and DBscan is designed for dis-
covery of clusters with arbitrary shape and good efficiency on
large databases (Ester et al. 1996). On this basis, these 10
algorithms are suitable for the comparative experiment.
The entropy measure is based on information-theoretic con-
siderations, which represents the difference of information
entropy between the clustering by an algorithm and the
actually true clustering, the smaller the better. In particular,
assume that the given testing dataset D is composed of
clusters C+1, C+2, …, C+k+ (true clustering), and a clustering
algorithm is applied to find clusters C1, C2, …, Ck in the
dataset D, then the entropy measure can be defined as
Equation (7).
k  k+ | C p ∩ C p+ + |  |C p ∩ C + + |  
|C p | 
 ( )
p 
1
Entropy = − log
n  log k + Cp  Cp   (7)
p =1  p + =1  
In Equation (7), n represents the scale of the dataset D. The
signal “| |” means the size of the corresponding set. For
example, |Cp| represents the size of Cp.
The other measure for evaluating the performance of clus-
tering algorithms is purity, which is based on counting the
This content downloaded from
203.218.254.164 on Wed, 18 Oct 2023 08:47:49 +00:00
All use subject to https://about.jstor.org/terms
pairs of data points n, with which two clustering results agree/
disagree. The purity measure represents the degree that all
clusters contain true documents, the larger the better.
Specifically, given a testing dataset D, the true clusters C+1,
C+2, …, C+k+, and the clustering results C1, C2, …, Ck, the
purity measure is defined as Equation (8).
1 k
Purity = max p+ =1,2,...,k+ | Cp ∩Cp++ |
n p=1
(8)
The metrics of entropy and purity evaluate the consistency
between the clusters generated by an algorithm and the actual
cluster labels assigned by experts, which are widely used in
the evaluation of clustering (Bezdek and Pal 1998; Liu 2007).
In order to conduct a reliable comparison, 30 data sets are
collected from multiple benchmark databases specifically for
clustering and text mining algorithms (Boley et al. 1999;
Chen et al. 2006; Han et al. 1998; Moore et al. 1997; Wu et
al. 2010), including 15 data sets from the UCI Machine
Learning Repository and 15 from the test-bed of the CLUTO
Data Clustering Software. The data sets are randomly chosen
from the databases, covering various kinds of data types and
attributes. For RBR, DIRECT, AGGLO (all variances), and
GRAPH, the CLUTO software package (Zhao et al. 2005) is
adopted. The programs for DBscan and K-Means are pro-
vided by a commonly used text book (Marmanis and Babenko
2009). The program for X-Means is adapted from the original
codes provided by the authors of the algorithm (Pelleg and
Moore 2000). For agglomerative clustering, four types of
intercluster similarity (group average, min, max, and distance
between centroids) are used.
For each given data set, the target number of clusters is set to
the ground truth cluster number in the data set. Formally, for
a data set Di, there are ki clusters indicated by the label field.
In the experiment, for the three K-Means type methods (RBR,
DIRECT, and Simple K-Means) and the graph partitioning
method (graph), the parameter k is directly set to ki. For X-
Means, a range of (ki – 1, ki + 1) is set to guarantee that it
outputs the results of ki clusters. For DBscan, its parameters
eps and minpts are tuned to make the cluster number equal ki.
For the four variances of the agglomerative methods and
REPSET, only the particular level of the document dendro-
gram containing ki clusters is extracted as the clustering
results.
The generalized results of the benchmark testing are listed in
Table 2. As shown in the table, the REPSET clustering
method produces 12 minimum entropy values and 15 maxi-
mum purity values in the 30 instances, which significantly
outperforms the other 11 algorithms. The average rank values
also support the fact.
MIS Quarterly Vol. 41 No. 4/December 2017 1115
Guo et al./Extracting Representative Information on Blogging Platforms

Table 2. Entropy and Purity Measures

Entropy Purity
Number of Minimum Average Rank Number of Maximum Average
30 Datasets Entropy Value Purity Rank Value
RBR 6 3.30 5 3.97
DIRECT 5 3.62 5 4.52
AGGLO(group) 0 6.30 5 6.27
AGGLO(max) 0 9.03 3 8.45
AGGLO(min) 0 7.77 3 7.72
AGGLO(centroids) 0 8.82 3 8.17
GRAPH 5 4.07 6 4.48
DBscan 0 9.10 3 8.42
K-Means 6 5.03 8 4.98
X-Means 1 6.43 6 6.07
REPSET 12 2.53 15 2.97
REPSET(no Backward) 0 5.80 2 5.77
Note: The range of rank values for the algorithms is 1 to 12, where 1 represents the best (i.e., the minimum entropy or the maximum purity), and
12 represents the worst (i.e., the maximum entropy or the minimum purity).

Furthermore, statistical analysis can also be conducted to test
whether the proposed REPSET method is reliably better.
Considering that the 30 datasets are from two different data
sources and may not satisfy the identical distribution assump-
tion, non-parameter Friedman test is considered (Vapnik
1998), which will test whether the rankings of the algorithms
concerned are significantly different. Table 3 shows the
Friedman test results on both the entropy and purity measures.
These results are consistent with those of Table 2. Briefly, on
the benchmark datasets, REPSET is superior to other tradi-
tional text clustering algorithms in ranks, indicating that
REPSET is technically effective and advantageous over major
existing ones for text clustering. In addition, it is worth
noting that REPSET is significantly better than REPSET with
no backward strategy, revealing that the proposed backward
strategy plays an effective role in the clustering process and
can significantly promote the accuracy level of clustering.
Notably, the backward strategy in REPSET will lead to data
object movements during the clustering process. Because the
number of boundary objects is usually small, the amount of
such movements is also small. The proportions of data ob-
jects that have been reallocated during the clustering process
of REPSET are provided in Appendix B. On average, the
reallocation proportion in the clustering of the 30 benchmark
datasets is 1.29%. The appendix also includes a box plot
showing the average reallocation proportion across the 30
benchmark data sets in every stage of iteration.
Experiments were also conducted to evaluate the time and
space complexity of the REPSET method. REPSET is mainly
composed of two phases. The first phase is to cluster the
original data objects to construct a dendrogram. The second
phase is to extract representative items at every level in the
dendrogram. The first phase can be executed offline as a
batch processing task. After the dendrogram is generated and
stored, the time needed to read the dendrogram and return a
specified level is very short at the millisecond level. There-
fore, the second phase is immediately responsive to users’
online queries. In the following, the time and space com-
plexity experiments only focus on the first phase.
Different sizes of testing data collections were constructed by
duplicating a dataset randomly selected from the 30 bench-
mark datasets described above. The testing platform was a
Windows 7 system on a PC with an Intel Core i3-2100 CPU
(3.1 GHz) and 4G RAM. The programs were implemented in
java. The running time of REPSET on different scales of data
sets is represented in Figure 6(a), in comparison with typical
partitional, density, and hierarchical methods. The curves
show that REPSET is slightly slower than existing methods,
indicating that it might not be efficient for processing real-
time stream data. However, the overall trend reveals a low-
level polynomial complexity, which is acceptable for offline
batch processing tasks such as representative article extraction
in blogging systems. The memory consumption is presented
in Figure 6(b), revealing a low-level polynomial complexity
curve as well, which indicates that REPSET has the scala-
bility to deal with large-scale datasets.

1116 MIS Quarterly Vol. 41 No. 4/December 2017

This content downloaded from

203.218.254.164 on Wed, 18 Oct 2023 08:47:49 +00:00
All use subject to https://about.jstor.org/terms
 Guo et al./Extracting Representative Information on Blogging Platforms

Table 3: Friedman Test Statistical Results on Entropy and Purity Measures

Entropy Chi-Square Sig. Purity Chi-Square Sig.
REPSET < RBR 4.172 * REPSET > RBR 4.167 *
REPSET < DIRECT 8.533 *** REPSET > DIRECT 11.560 ***
REPSET < AGGLO (group) 25.138 *** REPSET > AGGLO (group) 17.640 ***
REPSET < AGGLO (max) 26.133 *** REPSET > AGGLO (max) 22.154 ***
REPSET < AGGLO (min) 22.533 *** REPSET > AGGLO (min) 21.160 ***
REPSET < AGGLO (centroids) 30.000 *** REPSET > AGGLO (centroids) 22.154 ***
REPSET < GRAPH 3.846 * REPSET > GRAPH 3.846 *
REPSET < DBscan 26.133 *** REPSET > DBscan 22.154 ***
REPSET < K-Means 6.533 * REPSET > K-Means 9.846 **
REPSET < X-Means 10.800 *** REPSET > X-Means 6.259 *
REPSET < REPSET(no Backward) 22.533 *** REPSET > REPSET(no Backward) 18.615 ***

Note: *p < 0.05, **p < 0.01, ***p < 0.001

(a) (b)

Figure 6. Running Time of REPSET on Different Scales of Datasets

Empirical Evaluations In 2006, an internal blogging platform was deployed in Com-

Empirical evaluations of REPSET were conducted based on
the massive database accumulated on an internal blogging
platform in practice. Multiple approaches including data
experiments and user evaluations were utilized.
Research Site
Empirical data for our search were provided by Company X,
a large telecommunications service provider in China, whose
parent company is a Fortune 500 telecommunications firm.
As of 2013, Company X had over 40,000 employees and pro-
vided mobile services for more than 90 million customers.
The annual revenue of the company was about USD 13.5
billion, holding 8% of market share of mobile telecommuni-
cations service in China.
pany X, aimed at enhancing the communication among work-
groups and employees, promoting knowledge sharing, and
providing a stage for self-expression. Each blog on the plat-
form is maintained by a specific business workgroup in the
company. Since establishment, effective policies were made
to encourage the use of the blogging platform. As of October
2013, almost all workgroups in the company were main-
taining their blogs on the platform, among which about 3,000
blogs were active (with at least one article posted per week).
Around 5,000 articles plus about 10,000 comments were
produced every day.
We evaluated the function of REPSET to extract a small set
of representative articles from all the blog articles posted
within a specific time period. This was done by comparing
the articles selected by REPSET with articles recommended
by popularity-based methods and those obtained by other

MIS Quarterly Vol. 41 No. 4/December 2017 1117

This content downloaded from

203.218.254.164 on Wed, 18 Oct 2023 08:47:49 +00:00
All use subject to https://about.jstor.org/terms
Guo et al./Extracting Representative Information on Blogging Platforms
clustering methods. In light of the findings from research on
search engines that users usually only pay attention to the first
10 resulting articles of the first page (Granka et al. 2004), we
initially restricted the size of the representative set to 10
articles.
Using a subset of data from the blogging platform, an illus-
trative example is provided in Appendix C, which compares
the results of representative article extraction between
REPSET and X-Means, a typical traditional clustering
method. The comparison intuitively shows that REPSET can
facilitate enhanced diversity in coverage and therefore capture
more representative results. In the following, more consoli-
dated experiments will be presented to evaluate the perfor-
mance of REPSET.
Measures
The REPSET method is composed of two major phases. The
first phase is to cluster the original blog articles and the
second is to select a representative article for each cluster.
Therefore, the evaluation of REPSET also includes two
phases. As presented in the “Benchmark Data Evaluation”
subsection, the measures of entropy and purity are used for
the first phase, namely to evaluate the accuracy of clustering.
These are the most commonly used measures for evaluating
clustering accuracy.
The second phase of evaluation is focused on the quality of
representative information extraction, namely how well the
extracted items may represent the whole population. In this
regard, the measures of coverage and redundancy are used.
These measures have been developed and formulated by a
series of related literature on e-commerce and information
retrieval (Ma and Wei 2012; Pan et al. 2005). Generally,
coverage is defined as the degree to which a small set of items
cover the contents of the whole body of a large information
base. The other measure for representativeness is redun-
dancy, which is defined as the degree to which the contents of
the items in an extracted set interleave with each other (Zhang
et al. 2002). Since the target of the extraction method is to
arrive at a small set of information items with good represen-
tativeness, it is also important that the resulting set should be
concise, so that its readers can capture as much information as
possible in a limited time. Therefore, a high degree of redun-
dancy would indicate a low level of extraction performance.
To comprehensively evaluate the overall performance
including both coverage and redundancy, the balanced F-
measure (F1 measure) may also be considered (Arazy and
Woo 2007; Powers 2011). The F1 measure is commonly used
in testing the accuracy of classification analysis, the higher
the better. In our research, the F1 measure can be calculated
as the harmonic mean of coverage and redundancy.
1118 MIS Quarterly Vol. 41 No. 4/December 2017
This content downloaded from
203.218.254.164 on Wed, 18 Oct 2023 08:47:49 +00:00
All use subject to https://about.jstor.org/terms
In the context of blog article extraction, the coverage,
redundancy, and F1 measures of the resulting articles can be
evaluated from two perspectives. The first is the de facto
perspective. The de facto results of the measures are defined
as the degrees of coverage and redundancy obtained by
calculating the textual similarities between blog articles. The
other perspective is user perception. The perceived results of
the evaluation measures are defined as the degrees of cover-
age and redundancy reported by readers of blog articles. In
the spirit of the measures generally used in research on
reading behavior (Allington 1984), perceived coverage may
be measured by the number of meaningful words that a reader
remembers after reading the articles. For perceived redun-
dancy, self-reported scales are usually utilized.
Benchmark Methods and Evaluation Plan
We compared the 10 representative articles generated by
REPSET with 3 types of reader-based recommended articles.
The first is the top 10 rated articles based on reader ratings;
the second is the top 10 most read articles; and the third is the
top 10 most commented articles. Furthermore, based on the
results of the benchmark data evaluation, we chose four clus-
tering algorithms, namely Direct, Graph, RBR, and X-Means,
to analyze the company blogs and compared them with
REPSET. Three typical topic modeling methods, namely
LDA, hLDA, and DTM, were also considered. Additionally,
randomly generated articles were also taken into account as
a method for comparison. Altogether, 11 benchmark methods
were involved.
Therefore, the evaluation plan was to compare REPSET with
each of the benchmark methods in terms of three de facto
measures (coverage, redundancy, and F1) and three perceived
measures (coverage, redundancy, and F1). Two types of
experiments were involved. First, we conducted data experi-
ments to evaluate the de facto results of the measures.
Second, we conducted user experiments to evaluate the per-
ceived results of the measures. The de facto and perceived
results will verify each other and provide a solid base for the
evaluation conclusions.
Data Experiments
The de facto performances of the methods were evaluated
using data experiments. As defined, de facto degrees of
coverage and redundancy were obtained by calculating the
textual similarities between blog articles. Given two blog
article sets, D = {d1, d2, …, dn} and D = {d1, d2, …, dm}, and
suppose that Sim(d1, d2) indicates the similarity between
articles d and d', the de facto coverage of D' over D is calcu-
lated as follows (Ma and Wei 2012):
 Guo et al./Extracting Representative Information on Blogging Platforms

Table 4. De facto Performance (Overall)

Method De facto Coverage De facto Redundancy De facto F1
REPSET 0.3735 0.2274 0.5036
Random 0.284 0.3731 0.3909
Top Rated 0.1927 0.5015 0.278
Most Read 0.2874 0.4628 0.3745
Most Commented 0.1301 0.4982 0.2066
Graph 0.3258 0.6746 0.3256
X-Means 0.3646 0.4682 0.4326
RBR 0.4022 0.4367 0.4693
Direct 0.3999 0.48 0.4521
LDA 0.2843 0.5103 0.3597
hLDA 0.3115 0.4054 0.4088
DTM 0.2718 0.4228 0.3696

rC ( D′ , D) =   max (Sim(d , d ′))

d ∈D
d ′∈D ′
| D|
Generally speaking, de facto coverage represents the percent-
age of information conveyed in D that is reflected by the
information conveyed in D'. For any blog article set D = {d1,
d2, …, dn}, the de facto redundancy of D is calculated as (Ma
and Wei 2012):
 
rR ( D) =  1 − 1  Sim(d , d ′)
d ∈D d ′∈D
| D|
Following the definition of the F1 measure, the de facto F1 of
article set D' is calculated as
rF1 ( D′ , D) = 2 ⋅
( )
rC ( D′ , D) ⋅ 1 − rR ( D′ )
(
rC ( D′ , D) + 1 − rR ( D′ ) )
For the experiments, we selected the blogging data of a typi-
cal region of Company X, during the time period of July
2010, with a total of 2,239 blog articles in 62 blogs (namely
62 workgroups), in which 590 employees were involved as
authors and/or readers. Using REPSET and the benchmark
methods, respectively, we extracted 10 articles from the 2,239
blog posts. In line with the above formulae, the de facto
coverage, redundancy and F1 measures of the 12 methods
were calculated, as shown in Table 4.
Intuitively, we can see from the table that the de facto cover-
age of REPSET articles is higher than 9 out of the 11 bench-
mark methods. The de facto coverage degrees of RBR and
Direct articles are higher than REPSET, but the differences
(9)
are small. Meanwhile, the de facto redundancy of REPSET
articles is the lowest. As the comprehensive measure, the F1
of REPSET is clearly higher than all the benchmark methods,
indicating that REPSET achieves the best performance. To
statistically test the performance differences, we split the
dataset of 2,239 articles into 11 groups along the timeline,
with each group consisting of no less than 200 articles. Then
we used each method to generate 10 representative articles for
each group and compared them in the corresponding group.
We used the paired-samples t-test method to evaluate the
(10) differences. The testing results are shown in Table 5, which
illustrates that in 25 out of the 33 cases, the results of
REPSET are superior to those of the benchmark methods.
In terms of coverage and redundancy, REPSET significantly
outperforms 7 and 9 (out of 11) benchmark methods, respec-
tively. Overall, and more importantly, in terms of the com-
prehensive F1 measure, REPSET significantly outperforms
(11)
most of the benchmark methods except for RBR and X-
Means. RBR and X-Means achieve a comparable level of
performance as REPSET, largely because of their high
clustering accuracy (Marmanis and Babenko 2009).
However, they are both partitional clustering methods and can
hardly satisfy the requirements for flexible representative
article numbers and multilevel clusters, as discussed above in
the “Clustering” subsection. Meanwhile, topic modeling
methods do not perform well in these results. In topic
modeling, because topic labels are determined by a number of
top-frequency terms (Blei et al. 2003), the resulting topics are
usually dominated by high-frequency words, while low-
frequency words are neglected. In contrast, although
REPSET is also dependent on term frequency for the cosine-
similarity measure, the impacts of low-frequency words are
still preserved all through the process of clustering.

MIS Quarterly Vol. 41 No. 4/December 2017 1119

This content downloaded from

203.218.254.164 on Wed, 18 Oct 2023 08:47:49 +00:00
All use subject to https://about.jstor.org/terms
Guo et al./Extracting Representative Information on Blogging Platforms

Table 5. Significance Testing Results (de facto Measures)

Measure Benchmark Method t-value Significance
REPSET > Random 11.315 ***
REPSET > Top Rated 11.229 ***
REPSET > Most Read 9.424 ***
REPSET > Most Commented 10.082 ***
REPSET > Graph -1.935
de facto Coverage REPSET > RBR -6.047 ***
REPSET > X-Means -1.646
REPSET > Direct -3.518 **
REPSET > LDA 8.574 ***
REPSET > hLDA 6.000 ***
REPSET > DTM 5.803 ***
REPSET < Random -3.470 **
REPSET < Top Rated -8.037 ***
REPSET < Most Read -5.130 ***
REPSET < Most Commented -6.107 ***
REPSET < Graph -4.527 **
de facto Redundancy REPSET < X-Means -3.416 **
REPSET < RBR -2.401 *
REPSET < Direct -5.516 ***
REPSET < LDA -9.845 ***
REPSET < hLDA -1.664
REPSET < DTM -2.051
REPSET > Random 11.830 ***
REPSET > Top Rated 11.438 ***
REPSET > Most Read 8.875 ***
REPSET > Most Commented 11.975 ***
REPSET > Graph 3.065 *
de facto F1-Measure REPSET > X-Means 1.977
REPSET > RBR -1.100
REPSET > Direct 4.754 **
REPSET > LDA 10.733 ***
REPSET > hLDA 6.268 ***
REPSET > DTM 6.024 ***
Note: *p < 0.05; **p < 0.01; ***p < 0.001

In these experiments, the number of representative articles
was set to 10. It might be questioned whether a similar level
of performance can be achieved when different numbers of
items are extracted. As a robustness check, we re-executed
the empirical data experiments 15 times with the number of
extracted articles varying from 10 to 150 (see Appendix D for
details). It turned out that REPSET performed well across all
of the 15 cases of varying numbers of extracted articles.
Results from Friedman tests show that REPSET significantly
outperforms all of the 11 benchmark methods in terms of the
F1 measure.
In addition, considering that a target of representativeness
extraction is to capture the special articles that have low
similarities with others, a comparative experiment was also
conducted to illustrate the capability of the similarity-based
clustering methods in this regard. In this experiment, all of
the articles were ranked according to their average similarities
to other articles in the whole dataset. The articles ranked in
the last 10% were gathered to form a testing set of least-
common articles. For each set of 10 articles extracted by
REPSET, Direct, RBR, Graph, and X-Means, the coverage
values on the least-common article set were calculated using

1120 MIS Quarterly Vol. 41 No. 4/December 2017

This content downloaded from

203.218.254.164 on Wed, 18 Oct 2023 08:47:49 +00:00
All use subject to https://about.jstor.org/terms
 Guo et al./Extracting Representative Information on Blogging Platforms
Equation (9), respectively. The results show that the coverage
value achieved by REPSET was 0.180, while those achieved
by Direct, RBR, Graph, and X-Means were 0.135, 0.138,
0.044, and 0.140, respectively, indicating that REPSET
managed to capture distinguishably more least-common
content than the other methods.
User Experiments
Although the results of blog data analysis have demonstrated
the performance of the proposed method, aiming to secure
higher quality for the evaluations, we decided to go beyond
that and incorporate user experiments so that the results of the
two types of evaluation can be cross validated. Using the
same data set, user experiments were conducted among volun-
tary participants recruited from a prestigious university in
China. The experiment process was designed as follows.
First, participating subjects were randomly divided into
groups, which were assigned to REPSET and the benchmark
methods respectively. Thereafter, each subject was asked to
review the 10 articles of his/her group independently in 20
minutes. The subjects were not aware of the group to which
they belonged, and they did not know whether the articles
they were reviewing came from REPSET extraction or other
methods. After a subject had read the 10 articles assigned,
he/she was asked to fill in a short questionnaire consisting of
four scale questions about perceived redundancy. After that,
the subjects were given a word list consisting of 267 feature
words collected from all the 2,239 blog articles. These blog
articles were first preprocessed with a series of processing
strategies including stopword removing, stemming, and case
folding, according to the common practice of document
indexing (Witten et al. 1999). They were then segmented into
keywords using a widely used Chinese language processing
toolkit called NLPIR (http://ictclas.nlpir.org/). Moreover,
infrequent keywords were removed, because they usually
have little semantic meaning and may introduce ambiguity
(Sedding and Kazakov 2004). Finally, all of the remaining
keywords constitute the word list of 267 feature words. The
subjects were asked to mark out all of the feature words that
were mentioned (explicitly or implicitly) in the 10 articles that
they had just read. They were not allowed to refer back to the
articles when they were marking out the words. In the end,
the subjects were also asked to write down any of their
feelings or thoughts about the 10 articles they had read. The
experiment script and screenshots for the experiment system
UI are provided in Appendix E and Appendix F, respectively.
Perceived coverage was then measured by calculating the
percentage of words in the list of feature words that were
marked as mentioned. This is consistent with the research in
reading behavior in which the degree of content coverage is
This content downloaded from
203.218.254.164 on Wed, 18 Oct 2023 08:47:49 +00:00
All use subject to https://about.jstor.org/terms
measured with the number of new words learned (Allington
1984). Perceived redundancy was reported by the short
questionnaire that the subjects answered. The four question-
naire items for measuring perceived redundancy, in the form
of seven-point Likert-type scales, were developed through
two rounds of discussion and a pilot study to ensure its
validity.
Two rounds of experiments were conducted. The first round
was conducted in December 2012, during which we recruited
212 qualified participants from the students of the university
and compared REPSET with top rated, most read, most com-
mented articles as well as the Graph and X-Means methods.
The second round was conducted in January 2015, during
which 166 qualified students were recruited and REPSET was
compared with the Direct, RBR, LDA, hLDA, DTM, and
random methods. According to self-reporting, volunteers that
were not familiar with blogs were excluded. The experiments
were conducted in the Behavior Lab of a national research
center at the university. When dividing the participants into
the comparison groups, attributes including gender, age, and
grade were controlled, so that there are no significant differ-
ences between the groups. Although the blog articles were
collected from a company, students did not have much diffi-
culty in understanding the contents. The business of the
company is mobile phone services, with which most students
were very familiar. In our experiments, we asked the subjects
whether there was anything they did not understand after they
read the articles. Nobody answered that he/she felt diffi-
culties, and very few people raised questions. In the rare
cases when subjects felt confused, our research assistants
gave explanations. All of the participants finished the entire
process of the experiment smoothly in the time allocated.
Tests using the Cronbach Alpha indicator show that the ques-
tionnaire data are reliable.
Experiment results are shown in Table 6, where the numbers
for perceived coverage are the average percentages of feature
words marked as mentioned in the corresponding group, the
higher the better. The numbers for perceived redundancy are
the average scores of the seven-point Likert-type scales, the
lower the better. The perceived F1 measure is the harmonic
mean of perceived coverage and the reversed value of per-
ceived redundancy (normalized to the range of [0,1]).
Specifically,
Perceived F1 =
( Perceived Coverage) ×  7 − Perceived6 Redundancy  (12)
2×
+ 
7 − Perceived Redundancy 
( Perceived Coverage) 6


MIS Quarterly Vol. 41 No. 4/December 2017 1121
Guo et al./Extracting Representative Information on Blogging Platforms

Table 6. Perceived Performance

Perceived Coverage Perceived Redundancy Perceived F1
REPSET (first round) 51.80% 2.50 0.61
REPSET (second round) 54.73% 3.03 0.56
Random 39.46% 3.95 0.42
Top Rated 36.15% 3.66 0.51
Most Read 41.41% 3.84 0.51
Most Commented 17.17% 4.54 0.45
Graph 36.27% 3.79 0.50
X-Means 42.44% 2.98 0.51
RBR 49.22% 3.13 0.52
Direct 46.93% 3.51 0.47
LDA 30.95% 2.90 0.39
hLDA 48.25% 4.26 0.42
DTM 47.64% 2.51 0.56

The table illustrates that REPSET results in high perceived
coverage, low perceived redundancy, and the highest F1
values in both rounds, which are desirable. We also used the
independent-samples t-test method to statistically test the
differences. The testing results are shown in Table 7, which
illustrates that in 22 out of the 33 cases, the differences are
significant. Particularly, in terms of the comprehensive F1
measure, the performance advantages of REPSET over the
benchmark methods are mostly significant, except for RBR,
Direct, and DTM. These are basically consistent with the
data experiments and demonstrate the outstanding perfor-
mance of REPSET.
In summary, considering the comprehensive results that we
have obtained from the multiple commonly used design
evaluation methods (Gregor and Hevner 2013), it is reason-
able to conclude that REPSET has demonstrated an
outstanding performance.
Concluding Remarks
Satisfying user information needs on internal blog platforms
is a broad topic area that deserves much research efforts in
various directions. Our work on the information extraction
framework and the REPSET method offers a new and mean-
ingful approach to utilizing blogging information in an
organizational contexts, which is different from other mech-
anisms such as the popularity-based or preference-based
methods. Specifically, popularity-based methods such as
most-read and most-commented blogs are focused on pro-
viding the information in which most people are interested.
In contrast, representative information extraction is focused
on extending the range of information covered by the
extracted set. In practice, multiple methods can be used
alongside each other to provide multiple perspectives.
The research presented in this paper draws upon the promi-
nent needs for representative information in organizational
context. Nevertheless, the design artifacts proposed in this
paper may be adapted and applied to other scenarios where
information representativeness is a concern. Apart from
internal blogs, when managers and employees seek to lever-
age blogs on the external web for customer-oriented insights
(Chau and Xu 2012), the representativeness extraction
problem will also be a compelling application need. Mean-
while, situations such as online news, discussion boards,
online magazines, and newsletters can also benefit from
showing a more representative sample of contents. Most
news websites (such as cnn.com) carry a section on most-read
news stories, but readers may find it useful if the top represen-
tative articles are also provided. Furthermore, information
overload scenarios in offline contexts such as corporate
reports, news articles, and legal documents can also poten-
tially benefit.
Our research has been positioned as an exaptation study
(Gregor and Hevner 2013) and executed in line with the prin-
ciples of design science (Hevner et al. 2004). We formulated
the relatively new problem of representative information
extraction on internal blogs, which is meaningful for tackling
information overload and utilizing accumulated data for
managerial effectiveness. Considering the rapid development
of social network applications (Kane et al. 2014), this topic

1122 MIS Quarterly Vol. 41 No. 4/December 2017

This content downloaded from

203.218.254.164 on Wed, 18 Oct 2023 08:47:49 +00:00
All use subject to https://about.jstor.org/terms
 Guo et al./Extracting Representative Information on Blogging Platforms

Table 7. Significance Testing Results (Perceived Measures)

Measure Benchmark Method t value Significance
REPSET > Random 2.698 **
REPSET > Top Rated 4.096 ***
REPSET > Most Read 2.462 *
REPSET > Most Commented 10.337 ***
REPSET > Graph 4.009 ***
Perceived Coverage REPSET > X-Means 2.174 *
REPSET > RBR 0.948
REPSET > Direct 1.271
REPSET > LDA 4.075 ***
REPSET > hLDA 0.965
REPSET > DTM 1.194
REPSET < Random -2.666 *
REPSET < Top Rated -4.123 ***
REPSET < Most Read -5.634 ***
REPSET < Most Commented -10.049 ***
REPSET < Graph -5.417 ***
Perceived Redundancy REPSET < X-Means -2.019 *
REPSET < RBR -0.279
REPSET < Direct -1.427
REPSET < LDA 0.438
REPSET < hLDA -3.740 ***
REPSET < DTM 1.832
REPSET > Random 3.2106 **
REPSET > Top Rated 4.1196 ***
REPSET > Most Read 4.9261 ***
REPSET > Most Commented 3.9703 ***
REPSET > Graph 4.7732 ***
Perceived F1–Measure REPSET > X-Means 2.1558 *
REPSET > RBR 0.9773
REPSET > Direct 1.907
REPSET > LDA 3.7258 ***
REPSET > hLDA 3.6432 ***
REPSET > DTM 0.0224
Note: *p < 0.05; **p < 0.01; ***p < 0.001

may inspire more extensive efforts in related areas. We drew
upon existing BI techniques to design a solution. After ana-
lyzing the requirements for the solution, we found that they
may not be met by directly applying existing methods. There-
fore, the novel technique of REPSET was proposed and
incorporated into the design framework. The postulated arti-
fact was evaluated using multiple approaches involving both
objective data and subjective perceptions, with all experi-
ments carefully designed, organized, and executed, in order
to guarantee the validity and reliability of the results. In this
way, our contributions are two-fold: (1) we contribute to the
design science literature with the formulated research question
of representative content extraction on internal blogs, as well
as the design framework and extended clustering technique to
address this problem; (2) we contribute to the business intelli-
gence literature with the novel clustering method developed
through the course of solution design, which offers greater
understanding of clustering research and may be generalized
to other contexts.
One limitation of our current work is that we did not empiri-
cally test whether and how the extracted representative

MIS Quarterly Vol. 41 No. 4/December 2017 1123

This content downloaded from

203.218.254.164 on Wed, 18 Oct 2023 08:47:49 +00:00
All use subject to https://about.jstor.org/terms
Guo et al./Extracting Representative Information on Blogging Platforms
articles can eventually help with managerial decision making.
Another limitation is that we only considered internal blog-
ging applications, while other forms of enterprise social media
platforms could also have been studied. Meanwhile, although
the novel clustering method proposed in REPSET performed
well in our research context, the validation of its advantages
in general demands future work to test it in other contexts.
Additionally, the time efficiency of the proposed method
needs to be improved, if it is to be applied in real-time stream
data processing scenarios such as extracting representative
information from microblogging platforms like Twitter.
Another direction for future research is to explore other BI
techniques that could also be helpful for further exploiting the
massive data accumulated on enterprise social media plat-
forms, so as to support management and decision making on
various aspects.
Acknowledgments
Xunhua Guo was supported by the National Natural Science Foun-
dation of China (grant numbers 71490721, 71572092). Qiang Wei,
Guoqing Chen, and Dandan Qiao were supported by the National
Natural Science Foundation of China (grant numbers 71110107027,
71490724, 71372044) and the MOE Project of Key Research Insti-
tute of Humanities and Social Sciences at Universities of China
(grant number 12JJD630001). Jin Zhang was supported by the
National Natural Science Foundation of China (grant number
71402186).
References
Abbasi, A., and Chen, H. 2008. “Cybergate: A Design Framework
and System for Text Analysis of Computer-Mediated Communi-
cation,” MIS Quarterly (32:4), pp. 811-837.
Adomavicius, G., Bockstedt, J. C., Gupta, A., and Kauffman, R. J.
2008. “Making Sense of Technology Trends in the Information
Technology Landscape: A Design Science Approach,” MIS
Quarterly (32:4), pp. 779-809.
Aggarwal, R., Gopal, R., Sankaranarayanan, R., and Singh, P. V.
2012. “Blog, Blogger, and the Firm: Can Negative Employee
Posts Lead to Positive Outcomes?,” Information Systems
Research (23:2), pp. 306-322 (doi: 10.1287/isre.1110.0360).
Aliguliyev, R. M. 2009. “A New Sentence Similarity Measure and
Sentence Based Extractive Technique for Automatic Text Sum-
marization,” Expert Systems with Applications (36:4), pp.
7764-7772 (doi: 10.1016/j.eswa.2008.11.022).
Allington, R. L. 1984. “Content Coverage and Contextual Reading
in Reading Groups,” Journal of Literacy Research (16:2), pp.
85-96 (doi: 10.1080/10862968409547506).
Arazy, O., and Woo, C. 2007. “Enhancing Information Retrieval
through Statistical Natural Language Processing: A Study of
Collocation Indexing,” MIS Quarterly (31:3), pp. 525-546.
Bawden, D., and Robinson, L. 2009. “The Dark Side of Infor-
mation: Overload, Anxiety and Other Paradoxes and Path-
1124 MIS Quarterly Vol. 41 No. 4/December 2017
This content downloaded from
203.218.254.164 on Wed, 18 Oct 2023 08:47:49 +00:00
All use subject to https://about.jstor.org/terms
ologies,” Journal of Information Science (35:2), pp. 180-191
(doi: 10.1177/0165551508095781).
Bente, S., and Karla, J. 2009. “Enterprise Social Network Plat-
forms as a Management Tool in Complex Technical Systems,” in
Proceedings of the 15th Americas Conference on Information
Systems, San Francisco, August 6.
Berkhin, P. 2006. “A Survey of Clustering Data Mining Tech-
niques,” in Grouping Multidimensional Data, J. Kogan,
C. Nicholas, and M. Teboulle (eds.), Berlin: Springer pp. 25-71
(doi: 10.1007/3-540-28349-8_2).
Bezdek, J. C., and Pal, N. R. 1998. “Some New Indexes of Cluster
Validity,” IEEE Transactions on Systems, Man, and Cybernetics,
Part B: Cybernetics (28:3), pp. 301-315 (doi: 10.1109/3477.
678624).
Blei, D. M. 2012. “Probabilistic Topic Models,” Communications
of the ACM (55:4), pp. 77-84 (doi: 10.1145/2133806.2133826).
Blei, D. M., and Lafferty, J. D. 2006. “Dynamic Topic Models,” in
Proceedings of the 23rd International Conference on Machine
Learning, New York: ACM, pp. 113-120 (doi: 10.1145/
1143844.1143859).
Blei, D. M., Griffiths, D., Jordan, M. I., and Tenenbaum, J. B. 2004.
“Hierarchical Topic Models and the Nested Chinese Restaurant
Process,” in Advances in Neural Information Processing Systems
16, Cambridge, MA: MIT Press, pp. 17-24.
Blei, D. M., Ng, A. Y., and Jordan, M. I. 2003. “Latent Dirichlet
Allocation,” The Journal of Machine Learning Research (3), pp.
993-1022.
Boley, D., Gini, M., Gross, R., Han, E.-H., Hastings, K.,
Karypis, G., Kumar, V., Mobasher, B., and Moore, J. 1999.
“Partitioning-Based Clustering for Web Document Categori-
zation,” Decision Support Systems (27:3), pp. 329-341 (doi:
10.1016/S0167-9236(99)00055-X).
Broder, A. Z., Glassman, S. C., Manasse, M. S., and Zweig, G.
1997. “Syntactic Clustering of the Web,” Computer Networks
and ISDN Systems (29:8), pp. 1157-1166 (doi: 10.1016/S0169-
7552(97)00031-7).
Cai, X., and Li, W. 2011. “A Spectral Analysis Approach to Docu-
ment Summarization: Clustering and Ranking Sentences Simul-
taneously,” Information Sciences (181:18), pp. 3816-3827 (doi:
10.1016/j.ins.2011.04.052).
Chau, M., and Xu, J. 2012. “Business Intelligence in Blogs:
Understanding Consumer Interactions and Communities,” MIS
Quarterly (36:4), pp. 1189-1216.
Chen, G., Liu, H., Yu, L., Wei, Q., and Zhang, X. 2006. “A New
Approach to Classification Based on Association Rule Mining,”
Decision Support Systems (42:2), pp. 674-689 (doi: 10.1016/
j.dss.2005.03.005).
Chui, M., Manyika, J., Bughin, J., Dobbs, R., Roxburgh, C.,
Sarrazin, H., Sands, G., and Westergren, M. 2012. The Social
Economy: Unlocking Value and Productivity through Social
Technologies, McKinsey Global Institute (http://www.mckinsey.
com/industries/high-tech/our-insights/the-social-economy).
Coenen, F., and Leng, P. 2007. “The Effect of Threshold Values on
Association Rule Based Classification Accuracy,” Data &
Knowledge Engineering (60:2), pp. 345-360 (doi: 10.1016/
j.datak.2006.02.005).
 Guo et al./Extracting Representative Information on Blogging Platforms
Das, D., and Martins, A. F. T. 2007. “A Survey on Automatic Text
Summarization,” Literature Survey for Language and Statistics
II course, Carnegie Mellon University.
Ester, M., Kriegel, H., Sander, J., and Xu, X. 1996. “A Density-
Based Algorithm for Discovering Clusters in Large Spatial
Databases with Noise,” in Proceedings of the 2nd International
Conference on Knowledge Discovery and Data Mining, Portland,
OR, August 2-4, AAAI Press, pp. 226-231.
Gionis, A., Indyk, P., and Motwani, R. 1999. “Similarity Search in
High Dimensions via Hashing,” in Proceedings of the 25th
International Conference on Very Large Data Bases, Edinburgh,
Scotland, pp. 518-529.
Goes, P. B. 2014. “Editor’s Comments: Design Science Research
in Top Information Systems Journals,” MIS Quarterly (38:1), pp.
iii-viii.
Granka, L. A., Joachims, T., and Gay, G. 2004. “Eye-Tracking
Analysis of User Behavior in WWW Search,” in Proceedings of
the 27th Annual International ACM SIGIR Conference on
Research and Development in Information Retrieval, New York:
ACM, pp. 478-479 (doi: 10.1145/1008992.1009079).
Gregor, S., and Hevner, A. R. 2013. “Positioning and Presenting
Design Science Research for Maximum Impact,” MIS Quarterly
(37:2), pp. 337-355.
Griffiths, T. L., and Steyvers, M. 2004. “Finding Scientific
Topics,” Proceedings of the National Academy of Sciences
(101:Supplement 1), pp. 5228-5235 (doi: 10.1073/pnas.
0307752101).
Gu, Q., and Zhou, J. 2009. “Co-clustering on Manifolds,” in
Proceedings of the 15th ACM SIGKDD International Conference
on Knowledge Discovery and Data Mining, New York: ACM,
pp. 359-368 (doi: 10.1145/1557019.1557063).
Gupta, V., and Lehal, G. S. 2010. “A Survey of Text Summariza-
tion Extractive Techniques,” Journal of Emerging Technologies
in Web Intelligence (2:3), pp. 258-268.
Han, E.-H., Karypis, G., Kumar, V., and Mobasher, B. 1998.
“Hypergraph Based Clustering in High-Dimensional Data Sets:
A Summary of Results,” IEEE Bulletin of the Technical Com-
mittee on Data Engineering (21:1), pp. 15-22.
Han, J., Pei, J., and Kamber, M. 2011. Data Mining: Concepts and
Techniques, Amsterdam: Elsevier.
He, B., and Ounis, I. 2003. “A Study of Parameter Tuning for
Term Frequency Normalization,” in Proceedings of the 12th
International Conference on Information and Knowledge
Management, New York: ACM, pp. 10-16 (doi: 10.1145/
956863.956867).
Hevner, A. R., March, S. T., Park, J., and Ram, S. 2004. “Design
Science in Information Systems Research,” MIS Quarterly
(28:1), pp. 75-105.
Hu, M., and Liu, B. 2004. “Mining and Summarizing Customer
Reviews,” in Proceedings of the 10th ACM SIGKDD Inter-
national Conference on Knowledge Discovery and Data Mining,
New York: ACM, pp. 168-177 (doi: 10.1145/1014052.
1014073).
Indyk, P., and Motwani, R. 1998. “Approximate Nearest Neigh-
bors: Towards Removing the Curse of Dimensionality,” in
Proceedings of the 30th Annual ACM Symposium on Theory of
This content downloaded from
203.218.254.164 on Wed, 18 Oct 2023 08:47:49 +00:00
All use subject to https://about.jstor.org/terms
Computing, New York: ACM, pp. 604-613 (doi: 10.1145/
276698.276876).
Jackson, A., Yates, J., and Orlikowski, W. 2007. “Corporate
Blogging: Building Community through Persistent Digital Talk,”
in Proceedings of the 40th Annual Hawaii International
Conference on System Sciences, Waikoloa, HI, pp. 80-80 (doi:
10.1109/HICSS.2007.155).
Jain, A. K., Murty, M. N., and Flynn, P. J. 1999. “Data Clustering:
A Review,” ACM Computing Surveys (31), pp. 264-323 (doi:
10.1145/331499.331504).
Kane, G. C., Alavi, M., Labianca, G., and Borgatti, S. P. 2014.
“What’s Different about Social Media Networks? A Framework
and Research Agenda,” MIS Quarterly (38:1), pp. 275-304.
Ku, L.-W., Ho, H.-W., and Chen, H.-H. 2009. “Opinion Mining
and Relationship Discovery Using CopeOpi Opinion Analysis
System,” Journal of American Society for Information Science
and Technology (60:7), pp. 1486-1503 (doi: http://dx.doi.org/
10.1002/asi.v60:7).
Lau, R. Y. K., Liao, S. S. Y., Wong, K. F., and Chiu, D. K. W.
2012. “Web 2.0 Environmental Scanning and Adaptive Decision
Support for Business Mergers and Acquisitions,” MIS Quarterly
(36:4), pp. 1239-1268.
Leskovec, J., Rajaraman, A., and Ullman, J. D. 2014. Mining of
Massive Datasets, Cambridge, UK: Cambridge University Press.
Li, N., Guo, X., Chen, G., and Luo, N. 2015. “Reading Behavior
on Intra-Organizational Blogging Systems: A Group-Level
Analysis through the Lens of Social Capital Theory,” Information
& Management (52:7), pp. 870-881 (doi: 10.1016/j.im.2015.
03.004).
Liao, Q., Pan, S., Lai, J. C., and Yang, C. 2011. “Enterprise Blog-
ging in a Global Context: Comparing Chinese and American
Practices,” in Proceedings of the ACM 2011 Conference on
Computer Supported Cooperative Work, New York: ACM, pp.
35-44 (doi: 10.1145/1958824.1958831).
Liu, B. 2007. Web Data Mining: Exploring Hyperlinks, Contents
and Usage Data, Heidelberg: Springer.
Liu, X., and Croft, W. B. 2008. “Evaluating Text Representations
for Retrieval of the Best Group of Documents,” in Advances in
Information Retrieval, C. Macdonald, I. Ounis, V. Plachouras,
I. Ruthven, and R. W. White (eds.), Berlin: Springer, pp.
454-462 (doi: 10.1007/978-3-540-78646-7_43).
Liu, Y., Huang, X., An, A., and Yu, X. 2007. “ARSA: A
Sentiment-Aware Model for Predicting Sales Performance Using
Blogs,” in Proceedings of the 30th Annual International ACM
SIGIR Conference on Research and Development in Information
Retrieval, New York: ACM, pp. 607-614 (doi: http://doi.acm.
org/10.1145/1277741.1277845).
Lu, B., Guo, X., Luo, N., and Chen, G. 2015. “Corporate Blogging
and Job Performance: Effects of Work-Related and Nonwork-
Related Participation,” Journal of Management Information
Systems (32:4), pp. 285-314 (doi: 10.1080/07421222.2015.
1138573).
Luo, N., Guo, X., Zhang, J., Chen, G., and Zhang, N. 2015.
“Understanding the Continued Use of Intra-Organizational Blogs:
An Adaptive Habituation Model,” Computers in Human
Behavior (50:9), pp. 57-65 (doi: 10.1016/j.chb.2015.03.070).
MIS Quarterly Vol. 41 No. 4/December 2017 1125
Guo et al./Extracting Representative Information on Blogging Platforms
Ma, B., and Wei, Q. 2012. “Measuring the Coverage and Redun-
dancy of Information Search Services on e-Commerce Plat-
forms,” Electronic Commerce Research and Applications (11:6),
pp. 560-569 (doi: 10.1016/j.elerap.2012.09.001).
Mandala, S., Kumara, S., and Chatterjee, K. 2014. “A Game-
Theoretic Approach to Graph Clustering,” INFORMS Journal on
Computing (26:3), pp. 629-643 (doi: 10.1287/ijoc.2013.0588).
Marmanis, H., and Babenko, D. 2009. Algorithms of the Intelligent
Web (1st ed.), Greenwich, CT: Manning Publications Co.
McAfee, A. P. 2006. “Enterprise 2.0: The Dawn of Emergent
Collaboration,” MIT Sloan Management Review (47:3), pp.
21-28.
McLaren, T., Head, M., Yuan, Y., and Chan, Y. 2011. “A Multi-
level Model for Measuring Fit Between a Firm’s Competitive
Strategies and Information Systems Capabilities,” MIS Quarterly
(35:4), pp. 909-929.
Moore, J., Han, E., Boley, D., Gini, M., Gross, R., Hastings, K.,
Karypis, G., Kumar, V., and Mobasher, B. 1997. “Web Page
Categorization and Feature Selection Using Association Rule and
Principal Component Clustering,” (available at http://europepmc.
org/abstract/CIT/233736/reload=0).
Nardi, B. A., Schiano, D. J., Gumbrecht, M., and Swartz, L. 2004.
“Why We Blog,” Communications of the ACM (47:12), pp.
41-46.
Pan, F., Wang, W., Tung, A. K. H., and Yang, J. 2005. “Finding
Representative Set from Massive Data,” in Fifth IEEE
International Conference on Data Mining, p. 8 (doi: 10.1109/
ICDM.2005.69).
Pelleg, D., and Moore, A. 2000. “X-means: Extending K-means
with Efficient Estimation of the Number of Clusters,” in Pro-
ceedings of the 17th International Conference on Machine
Learning, San Francisco: Morgan Kaufmann, pp. 727-734.
Pons-Porrata, A., Berlanga-Llavori, R., and Ruiz-Shulcloper, J.
2007. “Topic Discovery Based on Text Mining Techniques,”
Information Processing & Management (43:3), pp. 752-768 (doi:
DOI: 10.1016/j.ipm.2006.06.001).
Powers, D. M. 2011. “Evaluation: from Precision, Recall and F-
measure to ROC, Informedness, Markedness and Correlation,”
Journal of Machine Learning Technologies (2:1), pp. 31-63.
Salton, G. 1971. The SMART Retrieval System: Experiments in
Automatic Document Processing, Englewood Cliffs, NJ.:
Prentice-Hall.
Sedding, J., and Kazakov, D. 2004. “WordNet-Based Text
Document Clustering,” in Proceedings of the 3rd Workshop on
Robust Methods in Analysis of Natural Language Data,
Stroudsburg, PA: Association for Computational Linguistics, pp.
104-113 (available at http://dl.acm.org/citation.cfm?id=
1621445.1621458).
Seref, O., Fan, Y.-J., and Chaovalitwongse, W. A. 2013. “Mathe-
matical Programming Formulations and Algorithms for Discrete
k-Median Clustering of Time-Series Data,” INFORMS Journal
on Computing (26:1), pp. 160-172 (doi: 10.1287/ijoc.
2013.0554).
Simon, G. J., Kumar, V., and Li, P. W. 2011. “A Simple Statistical
Model and Association Rule Filtering for Classification,” in
Proceedings of the 17th ACM SIGKDD International Conference
1126 MIS Quarterly Vol. 41 No. 4/December 2017
This content downloaded from
203.218.254.164 on Wed, 18 Oct 2023 08:47:49 +00:00
All use subject to https://about.jstor.org/terms
on Knowledge Discovery and Data Mining, New York: ACM,
pp. 823-831 (doi: 10.1145/2020408.2020550).
Steinbach, M., Karypis, G., and Kumar, V. 2000. “A Comparison
of Document Clustering Techniques,” in Proceedings of the KDD
Workshop on Text Mining, New York: ACM, pp. 525-526.
Suh, A., Shin, K., Ahuja, M., and Kim, M. S. 2011. “The Influence
of Virtuality on Social Networks Within and Across Work
Groups: A Multilevel Approach,” Journal of Management
Information Systems (28:1), pp. 351-386.
Turney, P. D., and Littman, M. L. 2003. “Measuring Praise and
Criticism: Inference of Semantic Orientation from Association,”
ACM Transations on Information Systems (21:4), pp. 315-346
(doi: http://doi.acm.org/10.1145/944012.944013).
Vapnik, V. N. 1998. Statistical Learning Theory, New York:
Wiley.
Witten, I. H., Moffat, A., and Bell, T. C. 1999. Managing Giga-
bytes: Compressing and Indexing Documents and Images, San
Francisco: Morgan Kaufmann.
Wright, A. 2009. “Our Sentiments, Exactly,” Communications of
the ACM (52:4), pp. 14-15 (doi: http://doi.acm.org/10.1145/
1498765.1498772).
Wu, J., Yuan, H., Xiong, H., and Chen, G. 2010. “Validation of
Overlapping Clustering: A Random Clustering Perspective,”
Information Sciences (180:22), pp. 4353-4369 (doi: 10.1016/
j.ins.2010.07.028).
Zelnik-Manor, L., and Perona, P. 2004. “Self-Tuning Spectral
Clustering,” in Advances in Neural Information Processing
Systems, Cambridge, MA: MIT Press, pp. 1601-1608.
Zhang, Y., Callan, J., and Minka, T. 2002. “Novelty and Redun-
dancy Detection in Adaptive Filtering,” in Proceedings of the 25th
Annual International ACM SIGIR Conference on Research and
Development in Information Retrieval, New York: ACM, pp.
81-88 (doi: http://doi.acm.org/10.1145/564376.564393).
Zhang, Z., Chen, G., Zhang, J., Guo, X., and Wei, Q. 2016.
“Providing Consistent Opinions from Online Reviews: A
Heuristic Stepwise Optimization Approach,” INFORMS Journal
on Computing (28:2), pp. 236-250 (doi: 10.1287/ijoc.2015.
0672).
Zhao, Y., Karypis, G., and Fayyad, U. 2005. “Hierarchical Clus-
tering Algorithms for Document Datasets,” Data Mining and
Knowledge Discovery (10:3), pp. 141-168 (doi: doi:10.1007/
s10618-005-0361-3).
About the Authors
Xunhua Guo is an associate professor of Information Systems at the
School of Economics and Management, Tsinghua University. He
received his Ph.D. degree from Tsinghua University in 2005. His
research takes behavioral and design science approaches to topics on
electronic commerce, social networks, and business intelligence. He
has published in journals including Communications of the ACM,
Decision Sciences, Decision Support Systems, Information & Man-
agement, Information Systems Journal, INFORMS Journal on
Computing, Journal of Information Technology, and Journal of MIS.
 Guo et al./Extracting Representative Information on Blogging Platforms

He serves as Executive Board Member and Vice General-Secretary
of China Association for Information Systems (CNAIS).
Qiang Wei is an associate professor of Information Systems at the
School of Economics and Management, Tsinghua University,
Beijing, China. His research interests cover big data analytics, busi-
ness intelligence, online advertising, web mining, online recommen-
dation, etc. His work has been published in journals including ACM
Transactions on Knowledge Discovery from Data, Decision
Sciences, INFORMS Journal on Computing, Decision Support
Systems, and Information Sciences. He serves as an associate editor
for Decision Support Systems and Electronic Commerce Research,
and secretary or committee member of several academic societies.
Qiang was the corresponding author for this paper.
Guoqing Chen is EMC Chair Professor of Information Systems at
the School of Economics and Management, Tsinghua University
(Beijing, China). He received his Ph.D. in managerial informatics
from the Catholic University of Leuven (KUL, Belgium) in 1992.
He was appointed China's national Changjiang Scholar Professor in
2005, and awarded IFSA Fellow in 2009. He served as the founding
president of China Association for Information Systems (CNAIS,
2005–2013). He has numerous publications in journals, books, and
conference proceedings, including a number of Harvard Business
School cases on IT/IS in Chinese companies. His research interests
include business intelligence and analytics, e-business and IT-
enabled innovation, and fuzzy logic applications.
Jin Zhang is an assistant professor at the School of Business,
Renmin University of China. He received his Ph.D. from the School
of Economics and Management, Tsinghua University. His current
research interests include web data mining, competitive intelligence,
and information retrieval. His work has been published in journals
including INFORMS Journal on Computing, IEEE Transactions on
Neural Networks and Learning Systems, ACM Transactions on
Knowledge Discovery from Data, Decision Support Systems, and
Information & Management.
Dandan Qiao is currently pursuing her Ph.D. at the School of
Economics and Management, Tsinghua University, Beijing, China.
Her research focuses on business intelligence, data mining, and
human behavior analytics. Her work has been published in journals
including ACM Transactions on Knowledge Discovery from Data
and Information & Management.

MIS Quarterly Vol. 41 No. 4/December 2017 1127

This content downloaded from

203.218.254.164 on Wed, 18 Oct 2023 08:47:49 +00:00
All use subject to https://about.jstor.org/terms
1128 MIS Quarterly Vol. 41 No. 4/December 2017

This content downloaded from

203.218.254.164 on Wed, 18 Oct 2023 08:47:49 +00:00
All use subject to https://about.jstor.org/terms
 RESEARCH ARTICLE

EXTRACTING REPRESENTATIVE INFORMATION ON

INTRA-ORGANIZATIONAL BLOGGING PLATFORMS
Xunhua Guo, Qiang Wei, and Guoqing Chen
Research Center for Contemporary Management, School of Economics and Management, Tsinghua University,
Beijing 100084 CHINA {guoxh@sem.tsinghua.edu.cn} {weiq@sem.tsinghua.edu.cn}
{chengq@sem.tsinghua.edu.cn}

Jin Zhang
School of Business, Renmin University of China,
Beijing 100872 CHINA {zhangjin@rbs.ruc.edu.cn}

Dandan Qiao
Research Center for Contemporary Management, School of Economics and Management, Tsinghua University,
Beijing 100084 CHINA {qiaodd.12@sem.tsinghua.edu.cn}

Appendix A
An Illustrative Example for the Clustering Process of REPSET

Figure A1 illustrates the clustering process of the REPSET method. There are nine documents represented as five round points and four square
points. The distances between data points correspond to the pairwise similarities between documents. Intuitively, these nine documents can
be divided into two clusters, i.e., the round point cluster and the square point cluster.

At the beginning, each point is treated as an individual cluster. In each step, two clusters with the largest similarity are merged into a new one.
The backward strategy is applied after the generation of a new cluster. In the example, in steps (b)–(f), no boundary document is found in the
generated cluster and reallocation does not occur. In step (g), two documents located in the middle are merged into one cluster, since they have
the largest similarity. If the clustering process continues as the traditional hierarchical clustering method, the nine documents will finally be
grouped into two clusters as shown in step (h), which turns out to be inaccurate, because one square document is assigned to a cluster in which
most documents are round. In contrast, in REPSET, the square document will be marked as a boundary object (outlined in red), since its
average similarity to the whole cluster is less than λ. The backward strategy will then reevaluate the red document and reallocate it if needed.
As shown in (i), because the red document’s similarity to the left cluster is than that to the right cluster, it is reallocated to the left cluster.

MIS Quarterly Vol. 41 No. 4–Appendices/December 2017 A1

This content downloaded from

203.218.254.164 on Wed, 18 Oct 2023 08:47:49 +00:00
All use subject to https://about.jstor.org/terms
Guo et al./Extracting Representative Information on Blogging Platforms

Figure A1. Illustrative Example for REPSET Clustering

A2 MIS Quarterly Vol. 41 No. 4–Appendices/December 2017

This content downloaded from

203.218.254.164 on Wed, 18 Oct 2023 08:47:49 +00:00
All use subject to https://about.jstor.org/terms
 Guo et al./Extracting Representative Information on Blogging Platforms

Appendix B
Proportions of Data Objects Reallocated in REPSET Clustering

Table B1: Proportions of Data Objects Reallocated in REPSET Clustering

Dataset No. Reallocation Proportion Dataset No. Reallocation Proportion
1 0.91% 16 3.23%
2 0.15% 17 1.17%
3 0.00% 18 0.37%
4 1.35% 19 0.25%
5 0.01% 20 5.79%
6 4.42% 21 1.52%
7 2.16% 22 1.16%
8 0.00% 23 0.04%
9 3.47% 24 0.78%
10 0.35% 25 1.08%
11 0.38% 26 1.16%
12 0.51% 27 0.60%
13 0.47% 28 0.18%
14 4.94% 29 0.00%
15 0.30% 30 2.37%
Average 1.29%

Figure B1. Proportions of Data Objects Reallocated in Iteration Stages

MIS Quarterly Vol. 41 No. 4–Appendices/December 2017 A3

This content downloaded from

203.218.254.164 on Wed, 18 Oct 2023 08:47:49 +00:00
All use subject to https://about.jstor.org/terms
Guo et al./Extracting Representative Information on Blogging Platforms

Appendix C
Representative Article Extraction Example

As an example, we compare the results of representative article extraction between REPSET and X-Means, a typical traditional clustering
method. We used the blogging data of a typical region of Company X, during the time period of July 2010, with a total of 2,239 blog articles.
Using REPSET and X-Means, respectively, we extracted 10 articles from the data set. Figure C1 shows the comparison of the sizes of clusters
generated by the two methods. It can be seen that the sizes of clusters generated by X-Means is much more even than those of REPSET. These
results illustrate the fact that X-Means tends to select clusters that are spatially equidistant, which is deficient for representativeness extraction.
In contrast, the sizes of clusters generated by REPSET are more discrepant, indicating the capability of REPSET to capture more diverse
content.

Figure C1. Cluster Size Comparison

Table C1 lists the titles of the representative articles extracted by REPSET and X-Means, respectively. It can be seen that the articles extracted
by X-Means are mostly life-related and emotional essays. In the whole data set, about 70% of articles are life-related and 30% are work-related.
In the results of X-Means, work-related articles are drown in life-related content and fail to be revealed. Meanwhile, most of the extracted life-
related articles are similar to some extent. In contrast, in the results of REPSET, six extracted articles are work-related, covering diverse topics
including work balance, industry analysis, and customer alerts. The four life-related articles are also diverse, covering topics such as love,
dining, and life advices. Such results intuitively show that REPSET can facilitate enhanced diversity in coverage and therefore capture more
representative results.

A4 MIS Quarterly Vol. 41 No. 4–Appendices/December 2017

This content downloaded from

203.218.254.164 on Wed, 18 Oct 2023 08:47:49 +00:00
All use subject to https://about.jstor.org/terms
 Guo et al./Extracting Representative Information on Blogging Platforms

Table C1. Representative Article Extraction Results

Articles Extracted by REPSET
No. Article Title Category
1 In the south, love is not only conveyed with poetry Life
2 Nice places for dining in the Shunde city Life
3 How to balance project work and post work Work
4 Fourteen advices for girls born in the 1980s Life
5 Industry Watch: China Mobile is facing five challenges in the field of information technology Work
6 The most comprehensive methods for recovering from drunk Life
7 Sharing the morning business meeting memo for July 6, 2010 Work
8 Training plan for new employees Work
9 Work plan, 2nd week, July, Ronggui customer services Work
10 VIP customer alert numbers Work
Articles Extracted by X-Means
No. Article Title Category
1 Happiness blind Life
2 There is a love that cannot be waited, there is a love that cannot be hurt. Life
3 Life is like donkeys, dogs, and monkeys. Life
4 Workplace is like a battlefield Work
5 To lose Life-work
6 Sadness is a kind of beauty Life
7 Love is only one more stroke than hate Life
8 Someone to allow me to be unreasonable Life
9 Losing affection Life
10 Spider nets Life

MIS Quarterly Vol. 41 No. 4–Appendices/December 2017 A5

This content downloaded from

203.218.254.164 on Wed, 18 Oct 2023 08:47:49 +00:00
All use subject to https://about.jstor.org/terms
Guo et al./Extracting Representative Information on Blogging Platforms

Appendix D
Empirical Data Experiments Results with Varying Numbers of
Representative Articles

CLUSTER TOP MOST MOST

NUM MEASURE REPSET RANDOM RATED READ COMMENTED GRAPH
10 de facto F1-measure 0.504 0.391 0.278 0.374 0.205 0.326
10 de facto Coverage 0.374 0.284 0.193 0.287 0.130 0.326
10 de facto Redundancy 0.227 0.373 0.502 0.463 0.521 0.675
15 de facto F1-measure 0.442 0.4 0.286 0.348 0.212 0.356
15 de facto Coverage 0.34 0.331 0.219 0.311 0.146 0.382
15 de facto Redundancy 0.371 0.492 0.588 0.606 0.619 0.667
20 de facto F1-measure 0.444 0.396 0.274 0.333 0.211 0.368
20 de facto Coverage 0.379 0.357 0.235 0.341 0.162 0.409
20 de facto Redundancy 0.465 0.554 0.671 0.675 0.697 0.665
25 de facto F1-measure 0.458 0.391 0.258 0.313 0.206 0.382
25 de facto Coverage 0.437 0.373 0.246 0.353 0.171 0.457
25 de facto Redundancy 0.519 0.588 0.73 0.718 0.741 0.672
30 de facto F1-measure 0.438 0.371 0.245 0.295 0.194 0.359
30 de facto Coverage 0.445 0.4 0.271 0.36 0.177 0.477
30 de facto Redundancy 0.569 0.653 0.776 0.75 0.785 0.712
35 de facto F1-measure 0.433 0.349 0.239 0.27 0.183 0.338
35 de facto Coverage 0.458 0.407 0.281 0.369 0.182 0.492
35 de facto Redundancy 0.59 0.694 0.792 0.787 0.817 0.742
40 de facto F1-measure 0.401 0.349 0.261 0.249 0.177 0.337
40 de facto Coverage 0.466 0.427 0.373 0.382 0.188 0.51
40 de facto Redundancy 0.648 0.705 0.799 0.815 0.832 0.748
45 de facto F1-measure 0.387 0.33 0.245 0.238 0.174 0.318
45 de facto Coverage 0.459 0.44 0.379 0.389 0.196 0.52
45 de facto Redundancy 0.665 0.735 0.819 0.828 0.844 0.771
50 de facto F1-measure 0.378 0.312 0.233 0.223 0.164 0.307
50 de facto Coverage 0.465 0.442 0.384 0.394 0.2 0.53
50 de facto Redundancy 0.682 0.76 0.833 0.844 0.861 0.784
55 de facto F1-measure 0.371 0.299 0.216 0.217 0.214 0.296
55 de facto Coverage 0.474 0.456 0.392 0.407 0.336 0.534
55 de facto Redundancy 0.696 0.778 0.851 0.852 0.843 0.795
60 de facto F1-measure 0.357 0.302 0.215 0.209 0.211 0.279
60 de facto Coverage 0.483 0.456 0.397 0.413 0.356 0.536
60 de facto Redundancy 0.717 0.775 0.853 0.86 0.85 0.812
70 de facto F1-measure 0.33 0.268 0.193 0.187 0.198 0.251
70 de facto Coverage 0.521 0.483 0.408 0.426 0.382 0.549
70 de facto Redundancy 0.759 0.815 0.873 0.88 0.866 0.837
80 de facto F1-measure 0.311 0.251 0.199 0.17 0.192 0.227
80 de facto Coverage 0.533 0.49 0.415 0.433 0.403 0.557

A6 MIS Quarterly Vol. 41 No. 4–Appendices/December 2017

This content downloaded from

203.218.254.164 on Wed, 18 Oct 2023 08:47:49 +00:00
All use subject to https://about.jstor.org/terms
 Guo et al./Extracting Representative Information on Blogging Platforms

CLUSTER TOP MOST MOST

NUM MEASURE REPSET RANDOM RATED READ COMMENTED GRAPH
80 de facto Redundancy 0.78 0.831 0.869 0.894 0.874 0.857
90 de facto F1-measure 0.306 0.238 0.197 0.155 0.176 0.212
90 de facto Coverage 0.54 0.499 0.425 0.44 0.416 0.563
90 de facto Redundancy 0.787 0.844 0.872 0.906 0.888 0.87
100 de facto F1-measure 0.285 0.231 0.187 0.143 0.168 0.192
100 de facto Coverage 0.547 0.51 0.433 0.448 0.426 0.565
100 de facto Redundancy 0.808 0.851 0.881 0.915 0.895 0.885
110 de facto F1-measure 0.276 0.215 0.172 0.135 0.174 0.195
110 de facto Coverage 0.555 0.52 0.437 0.454 0.433 0.572
110 de facto Redundancy 0.817 0.864 0.893 0.921 0.891 0.882
120 de facto F1-measure 0.25 0.192 0.162 0.151 0.163 0.178
120 de facto Coverage 0.559 0.528 0.445 0.459 0.44 0.577
120 de facto Redundancy 0.839 0.883 0.901 0.91 0.9 0.895
130 de facto F1-measure 0.243 0.193 0.16 0.144 0.152 0.174
130 de facto Coverage 0.563 0.535 0.453 0.465 0.445 0.584
130 de facto Redundancy 0.845 0.882 0.903 0.914 0.908 0.898
140 de facto F1-measure 0.225 0.187 0.154 0.138 0.143 0.163
140 de facto Coverage 0.57 0.543 0.463 0.473 0.453 0.59
140 de facto Redundancy 0.86 0.887 0.907 0.919 0.915 0.906
150 de facto F1-measure 0.215 0.171 0.146 0.135 0.137 0.153
150 de facto Coverage 0.575 0.551 0.467 0.478 0.459 0.592
150 de facto Redundancy 0.868 0.899 0.913 0.921 0.92 0.912
CLUSTER
NUM MEASURE XMEANS RBR DIRECT LDA HLDA DTM
10 de facto F1-measure 0.433 0.469 0.452 0.36 0.409 0.37
10 de facto Coverage 0.365 0.402 0.4 0.284 0.311 0.272
10 de facto Redundancy 0.468 0.437 0.48 0.51 0.405 0.423
15 de facto F1-measure 0.424 0.474 0.467 0.339 0.395 0.369
15 de facto Coverage 0.391 0.438 0.431 0.317 0.324 0.28
15 de facto Redundancy 0.536 0.483 0.491 0.635 0.497 0.456
20 de facto F1-measure 0.396 0.464 0.467 0.294 0.379 0.394
20 de facto Coverage 0.421 0.463 0.463 0.326 0.34 0.353
20 de facto Redundancy 0.627 0.536 0.529 0.732 0.572 0.553
25 de facto F1-measure 0.364 0.436 0.419 0.286 0.359 0.385
25 de facto Coverage 0.434 0.483 0.486 0.337 0.352 0.363
25 de facto Redundancy 0.686 0.603 0.632 0.752 0.632 0.591
30 de facto F1-measure 0.336 0.41 0.43 0.287 0.348 0.429
30 de facto Coverage 0.446 0.497 0.495 0.356 0.365 0.391
30 de facto Redundancy 0.731 0.652 0.621 0.76 0.668 0.525
35 de facto F1-measure 0.322 0.382 0.389 0.258 0.334 0.42
35 de facto Coverage 0.47 0.511 0.508 0.363 0.374 0.406
35 de facto Redundancy 0.755 0.696 0.685 0.8 0.698 0.565
40 de facto F1-measure 0.305 0.38 0.372 0.252 0.32 0.389
40 de facto Coverage 0.478 0.521 0.518 0.38 0.381 0.383

MIS Quarterly Vol. 41 No. 4–Appendices/December 2017 A7

This content downloaded from

203.218.254.164 on Wed, 18 Oct 2023 08:47:49 +00:00
All use subject to https://about.jstor.org/terms
Guo et al./Extracting Representative Information on Blogging Platforms

CLUSTER
NUM MEASURE XMEANS RBR DIRECT LDA HLDA DTM
40 de facto Redundancy 0.776 0.701 0.71 0.812 0.725 0.604
45 de facto F1-measure 0.287 0.349 0.371 0.249 0.296 0.396
45 de facto Coverage 0.488 0.531 0.529 0.456 0.391 0.424
45 de facto Redundancy 0.797 0.74 0.715 0.829 0.762 0.63
50 de facto F1-measure 0.268 0.359 0.369 0.205 0.277 0.337
50 de facto Coverage 0.494 0.538 0.537 0.386 0.398 0.449
50 de facto Redundancy 0.816 0.73 0.719 0.86 0.788 0.73
55 de facto F1-measure 0.259 0.342 0.346 0.206 0.259 0.357
55 de facto Coverage 0.503 0.545 0.545 0.396 0.399 0.451
55 de facto Redundancy 0.825 0.751 0.747 0.861 0.809 0.705
60 de facto F1-measure 0.254 0.333 0.356 0.179 0.254 0.352
60 de facto Coverage 0.505 0.554 0.552 0.396 0.413 0.471
60 de facto Redundancy 0.83 0.762 0.737 0.884 0.817 0.72
70 de facto F1-measure 0.23 0.304 0.338 0.182 0.267 0.354
70 de facto Coverage 0.532 0.564 0.562 0.488 0.437 0.483
70 de facto Redundancy 0.853 0.792 0.758 0.888 0.808 0.72
80 de facto F1-measure 0.213 0.296 0.307 0.165 0.245 0.292
80 de facto Coverage 0.546 0.574 0.57 0.493 0.444 0.49
80 de facto Redundancy 0.868 0.801 0.79 0.901 0.831 0.792
90 de facto F1-measure 0.194 0.288 0.301 0.146 0.233 0.301
90 de facto Coverage 0.558 0.583 0.582 0.506 0.452 0.51
90 de facto Redundancy 0.883 0.809 0.797 0.914 0.843 0.787
100 de facto F1-measure 0.183 0.263 0.283 0.139 0.219 0.303
100 de facto Coverage 0.565 0.592 0.588 0.513 0.455 0.494
100 de facto Redundancy 0.891 0.831 0.814 0.92 0.856 0.781
110 de facto F1-measure 0.174 0.269 0.27 0.132 0.206 0.247
110 de facto Coverage 0.572 0.6 0.598 0.52 0.46 0.519
110 de facto Redundancy 0.898 0.827 0.826 0.924 0.867 0.838
120 de facto F1-measure 0.162 0.244 0.244 0.121 0.194 0.244
120 de facto Coverage 0.58 0.607 0.605 0.523 0.465 0.535
120 de facto Redundancy 0.906 0.847 0.847 0.932 0.877 0.842
130 de facto F1-measure 0.156 0.225 0.239 0.105 0.185 0.232
130 de facto Coverage 0.581 0.613 0.61 0.452 0.47 0.519
130 de facto Redundancy 0.91 0.863 0.852 0.941 0.885 0.851
140 de facto F1-measure 0.145 0.228 0.233 0.102 0.172 0.215
140 de facto Coverage 0.586 0.621 0.617 0.461 0.473 0.545
140 de facto Redundancy 0.918 0.861 0.856 0.943 0.895 0.866
150 de facto F1-measure 0.133 0.207 0.229 0.098 0.163 0.231
150 de facto Coverage 0.589 0.625 0.623 0.54 0.476 0.54
150 de facto Redundancy 0.925 0.876 0.86 0.946 0.901 0.853

A8 MIS Quarterly Vol. 41 No. 4–Appendices/December 2017

This content downloaded from

203.218.254.164 on Wed, 18 Oct 2023 08:47:49 +00:00
All use subject to https://about.jstor.org/terms
 Guo et al./Extracting Representative Information on Blogging Platforms

Appendix E
Experiment Script (Translated from Chinese)

Experiments on Blog Article Reading

NO:

Dear students,

Welcome to participate in this experiment!

The following are ten articles selected from a company's internal blogging system. Please carefully read these blog articles for 20 minutes.
When the time is up, you will be asked to answer several questions mark a number of words according to what you have read.

Basic information:
1. Name , Gender , Age
2. Major , Grade
3. Your familiarity with blogging system
A. Completely unfamiliar B. Unfamiliar C. Neutral D. Familiar E. Very familiar

(After providing the basic information, the articles are displayed and the subjects are required to read the articles for 20 minutes. When the
time is up, the system closes the article and presents the following questions.)

For each of the following question, please select the number that corresponds with your reading experience. The number “1” represents
“strongly disagree,” while the number “7” represents “strongly agree.”

Please select
Strongly Slightly Slightly Strongly
Questions Disagree Disagree Disagree Neutral Agree Agree Agree
This group of blog articles has much repetition. 1 2 3 4 5 6 7
This group of blog articles has similar topics. 1 2 3 4 5 6 7
This group of blog articles has a wide range of contents. 1 2 3 4 5 6 7
This group of blog articles provides rich information. 1 2 3 4 5 6 7

If you have any feelings, opinions or comments on this group of blog articles please feel free to write them down in the following space:

MIS Quarterly Vol. 41 No. 4–Appendices/December 2017 A9

This content downloaded from

203.218.254.164 on Wed, 18 Oct 2023 08:47:49 +00:00
All use subject to https://about.jstor.org/terms
Guo et al./Extracting Representative Information on Blogging Platforms

Please label the following words based on your reading.

If you have browsed any information related with the word, please tick “%”, otherwise no label is needed.

100 139 2010 BOSS G3 GPRS

http arrangement case method transaction organization
warranty standard performance others department participate
take part in operation inquire product surpass growth
success grade member charge error final
unit cause zone place second third
first phone store adjustment effect dynamic
sms happen send discovery development program
method manner fee branch allocation analysis
share minute Foshan service responsible accessory
change thanks post tell the work company
function communication purchase key management Guangdong
specification rule process kids number cooperation
bill joy environment reply home conference
activity opportunity foundation group plan record
quarter home value reduce inspection simple
set up proposal health reward drop exchange
accept end solve explanation introduction progress
manager experience experience spirit competition account
opening happy check client control happiness
binding difficulty leliu leave understanding utilize
leadership process satisfaction monthly password free
goal internal content ability strive train
coordinate friend brand balance usually platform
business reception situation mood area channel
global pass life staff daily Ronggui social
application close identity body life life
time world market thing receive collect
cell phone familiar data Shunde attitude discussion
package put forward improve provide upgrade remind
experience condition call communicate communications notification
colleague unity complaints team spread recommendation
expand network website future hope habit
love system afternoon show on site relevant
enjoy project consumption sales effect assist
thanks mood mentality information week industry
form happiness demand propaganda choose student
learning pressure business opinion awareness factor
marketing operate affect own forever user
preferential outstanding pre-store staff reason operations
online responsibility increase gift about correct
policy support knowledge execution index guide
formulate quality intelligence china center terminal
emphasis initiative theme status consultation charges
data resources comprehensive

A10 MIS Quarterly Vol. 41 No. 4–Appendices/December 2017

This content downloaded from

203.218.254.164 on Wed, 18 Oct 2023 08:47:49 +00:00
All use subject to https://about.jstor.org/terms
 Guo et al./Extracting Representative Information on Blogging Platforms

Appendix F
Screenshots for the Experiment System UI

MIS Quarterly Vol. 41 No. 4–Appendices/December 2017 A11

This content downloaded from

203.218.254.164 on Wed, 18 Oct 2023 08:47:49 +00:00
All use subject to https://about.jstor.org/terms
Guo et al./Extracting Representative Information on Blogging Platforms

A12 MIS Quarterly Vol. 41 No. 4–Appendices/December 2017

This content downloaded from

203.218.254.164 on Wed, 18 Oct 2023 08:47:49 +00:00
All use subject to https://about.jstor.org/terms