Community Detection in Social Media: Performance and Application Considerations

Data Min Knowl Disc (2012) 24:515–554
DOI 10.1007/s10618-011-0224-z
Community detection in Social Media

Performance and application considerations
Symeon Papadopoulos · Yiannis Kompatsiaris ·

Athena Vakali · Ploutarchos Spyridonos
Received: 15 January 2010 / Accepted: 20 May 2011 / Published online: 14 June 2011
© The Author(s) 2011
Abstract The proposed survey discusses the topic of community detection in the
context of Social Media. Community detection constitutes a significant tool for
the analysis of complex networks by enabling the study of mesoscopic structures that
are often associated with organizational and functional characteristics of the underly-
ing networks. Community detection has proven to be valuable in a series of domains,
e.g. biology, social sciences, bibliometrics. However, despite the unprecedented scale,
complexity and the dynamic nature of the networks derived from Social Media data,
there has only been limited discussion of community detection in this context. More
specifically, there is hardly any discussion on the performance characteristics of com-
munity detection methods as well as the exploitation of their results in the context of
real-world web mining and information retrieval scenarios. To this end, this survey
first frames the concept of community and the problem of community detection in the
context of Social Media, and provides a compact classification of existing algorithms
based on their methodological principles. The survey places special emphasis on the
performance of existing methods in terms of computational complexity and memory
Responsible editor: Myra Spiliopoulou, Bamshad Mobasher, Olfa Nasraoui, Osmar Zaiane.
S. Papadopoulos (B) · Y. Kompatsiaris

Informatics and Telematics Institute, CERTH, Thessaloniki, Greece
e-mail: [email protected]
Y. Kompatsiaris
S. Papadopoulos · A. Vakali · P. Spyridonos

Department of Informatics, Aristotle University of Thessaloniki, Thessaloniki, Greece
P. Spyridonos
123
516 S. Papadopoulos et al.
requirements. It presents both a theoretical and an experimental comparative discus-

sion of several popular methods. In addition, it discusses the possibility for incremental
application of the methods and proposes five strategies for scaling community detec-
tion to real-world networks of huge scales. Finally, the survey deals with the inter-
pretation and exploitation of community detection results in the context of intelligent
web applications and services.
Keywords Community detection · Large-scale networks · Social Media
1 Introduction
Networks are omnipresent on the Web. The most profound Web network is the Web
itself comprising billions of pages as vertices and their hyperlinks to each other as
edges (Kumar et al. 2000). Moreover, collecting and processing the input of Web
users (e.g. queries, clicks) results in other forms of networks, such as the query graph
(Baeza-Yates 2007). Finally, the widespread use of Social Media applications, such
as Delicious,1 Digg,2 Flickr3 and YouTube,4 is responsible for the creation of even
more networks, ranging from folksonomy networks (Mika 2005) to rich media social
networks (Lin et al. 2009). Since networks originating from Social Media data are of
particular interest to this study, we shall collectively refer to them as Social Media
networks.
Despite the differences of Social Media networks with respect to the entities and
the type of relations they model, they present a significant source of intelligence since
they encode the online activities and inputs of masses of Social Media participants.
Not only is it possible by analyzing such networks to gain insights into the social
phenomena and processes that take place in our world, but one can also extract action-
able knowledge that can be beneficial in several information management and retrieval
tasks, such as online content navigation and recommendation. However, the analysis of
such networks poses serious challenges to data mining methods, since these networks
are almost invariably characterized by huge scales and a highly dynamic nature.
A valuable tool in the analysis of large complex networks is community detec-
tion. The problem that community detection attempts to solve is the identification of
groups of vertices that are more densely connected to each other than to the rest of
the network. Detecting and analyzing the community structure of networks has led
to important findings in a wide range of domains, ranging from biology to social sci-
ences (Girvan and Newman 2002) and the Web (Kumar et al. 1999; Flake et al. 2000).
Such studies have shown that communities constitute meaningful units of organization
and that they provide new insights in the structure and function of the whole network
under study. Recently, there has been increasing interest in applying community detec-
tion on Social Media networks not only as a means of understanding the underlying
1 http://delicious.com.
2 http://digg.com.
3 http://flickr.com.
4 http://www.youtube.com.
123
Community detection in Social Media 517
phenomena taking place in such systems, but also to exploit its results in a wide range
of intelligent services and applications, e.g. automatic event detection in Social Media
content.
Despite the increasing significance of Social Media and the proliferation of methods
for detecting communities, there has been no prior effort to collect and systematically
discuss research efforts with reference to the emerging topic of community detection
in Social Media. In particular, there are two important aspects of the problem that are
not adequately addressed in related survey articles: (a) performance aspects of com-
munity detection methods, namely computational complexity, memory requirements
and possibility for incremental updates of already identified community structure,
(b) interpretation and exploitation of community detection results by Social Media
applications.
Previous related works have studied individual aspects of this problem. For instance,
the survey articles by Fortunato and Castellano (2007) and Fortunato (2009, 2010)
contain an extensive discussion of numerous community detection methods; however,
they are mostly concerned with the methodological foundations of community detec-
tion. Furthermore, they describe community detection in a generic context, mostly
under the statistical physics perspective, thus lacking any association to web mining
and Social Media research. The study by Danon et al. (2005) presents a comparative
discussion on the computational complexity of several community detection methods.
However, it disregards their memory requirements, and other scalability consider-
ations, such as the possibility for incremental computations, and it further lacks a
Social Media context. In addition, Schaeffer (2007) presents a generic overview of
graph clustering, which can be considered as equivalent to community detection.
Finally, Tang and Liu (2010) discuss several social network analysis problems with
emphasis on community detection relating them with the Social Media domain; nev-
ertheless, it is limited to the methodological principles of methods, discussing neither
the performance attributes of methods nor the interpretation and exploitation of their
results.
The lack of prior work systematically dealing with the aforementioned aspects of
community detection in relation to Social Media applications has motivated the present
study, which makes the following contributions:
– Frame the concept of community and the problem of community detection in the
context of Social Media (Sect. 3).
– Present a systematic study and a compact methodological classification of several
community detection methods (Sect. 4.1).
– Study existing methods in view of the scalability challenges posed by the mag-
nitude and dynamic nature of Social Media networks. Compare complexity
and memory requirements of existing methods (Sect. 4.2) as well as incremen-
tal/dynamic computation characteristics (Sect. 4.3).
– Conduct a comparative experimental study (Sect. 4.2.2) benchmarking eight pop-
ular community detection algorithms in terms of execution time, peak memory
usage and attained community structure precision.
– Introduce five scalability strategies that can be employed in order to scale com-
munity detection methods to the magnitude of Social Media data (Sect. 5).
123
– Discuss the interpretation and exploitation of community detection results by

Social Media applications and services (Sect. 6).
Due to the breadth of the topic under study, we do not attempt to be exhaustive with
respect to the number of presented methods, nor do we target at delving into deep
foundational issues of the methods. Instead, we aim at discussing the methodologi-
cal, performance and application aspects of existing techniques under a paradigmatic
and comparative perspective, which we believe will be valuable for applying them in
Social Media mining problems.
2 Background
The term “community” has been extensively used in the literature in different contexts
and with different connotations. Social studies are probably the earliest context, where
the notion of community was used to denote groups of people with shared interests
or activities (e.g. communities of practice). Once networks became widely adopted
as a means to study social interactions and processes (Wasserman and Faust 1994;
Scott 2000), the concept of community was associated to networks of human actors
exhibiting certain characteristic structural properties.
As networks became established as a model for many real-world complex systems,
the concept of community expanded in scope to refer to group structures in a variety of
networks, not necessarily consisting of human actors. For instance, community struc-
ture was studied in the context of protein interaction networks, bibliographical citation
networks, networks of football teams, and others. With the emergence of the Web and
the Web Graph as a prevalent model of web pages and their hyperlinks, communities
were also considered on the Web (Kumar et al. 1999; Flake et al. 2000).
The advent and wide success of Social Media has created a new context for the
notion of community. Currently, there is a variety of online entities residing in the
virtual setting of Social Web applications and there are numerous kinds of interac-
tions and relations among such entities. For that reason, before proceeding with the
definition of communities in Social Media in Sect. 3, we will first describe the concept
of Social Media networks and the elements that they comprise (Sect. 2.1), as well as
the process of their creation (Sect. 2.2).
2.1 Elements of a Social Media network
The ecosystem of Social Media applications comprises a wide range of objects that are
associated to each other through numerous types of interactions and relations. Social
Media networks provide an elegant representation of Social Media data, containing
online objects as their vertices and the relations/interactions among them as edges.5
The vertices of Social Media networks can represent different types of actors, such
as users, content items (e.g. blog posts, photos, videos), and even metadata items
5 There are alternative data models, such as multidimensional data cubes (Chi et al. 2009), for representing
Social Media data. However, those models are not considered in this survey.
123
(e.g. topic categories, tags). In addition, the edges of Social Media networks can be
of different types, such as simple, weighted, directed and multiway (i.e. connecting
more than two entities) depending on the network creation process (discussed below).
In terms of notation, Social Media networks employ the typical graph notation
G = (V, E), where G stands for the whole network, V stands for the set of all verti-
ces and E for the set of all edges. Due to the different types of vertices and edges in
such networks, it is common to consider sets of vertices and edges within V and E
that contain vertices and edges of the same type. For instance, in the case of a photo
sharing and tagging network, one can consider the set of vertices V to comprise the
users, photos and tags of the system, i.e. V = {U, P, T }. Similarly, the set of edges
in such an application would comprise the set of user-photo, photo-tag and user-tag
associations, E = {U P, P T, U T }.
2.2 Social Media network creation
In practice, the creation of Social Media networks starts from a set of transactions
that are performed and recorded in Social Media applications. Every such transaction
typically involves different entities; for instance a tag assignment in Flickr involves a
user, a photo and a tag, while a comment on a blog article involves the commenter, the
blog article and the comment text. In that way, an association (edge) is formed between
the items of the same transaction on an underlying network, so that the resulting Social
Media network constitutes a direct representation of a subset of online transactions.
Since “raw” Social Media networks comprise multiple types of vertices and edges
(some of which can be multi-way, i.e. link more than two vertices), they are mathe-
matically represented by hypergraphs. Alternatively, if each hyper-edge of the graph
is reduced to the pairwise connections among the k different types of nodes, the
hypergraph is reduced to a k-partite graph. The majority of network analysis methods,
and community detection in particular, are not applicable to hypergraphs or k-partite
graphs. For that reason, it is common practice to extract simplified network forms that
depict partial aspects of the complex interactions of the original network. Such net-
works are typically one- or two-mode (i.e. contain only one or two vertex types) and
contain simple edges (i.e. connecting two vertices), which makes possible the applica-
tion of numerous network analysis techniques. In summary, the typical lifecycle of a
Social Media network (Fig. 1) involves its creation from a set of recorded transactions
and its transformation into some suitable form for the analysis that follows.
Folksonomies constitute an extensively studied example of Social Media networks.
A folksonomy comprises three types of entities, namely users, resources and tags (Mika
2005). Starting from the tag assignments of users, i.e. transactions involving a user,
a resource and a tag, a “raw” folksonomy network is formed, comprising three types
Fig. 1 Typical lifecycle of a Social Media network: A set of transactions involving users, content and meta-
data lead to the formation of a “raw” Social Media network. Typically, this network is simplified before
any sophisticated analysis (e.g. community detection) takes place
123
(a) (b)
Fig. 2 Two paradigmatic Social Media networks: (a) folksonomy network (Mika 2005; Schmitz et al.
2006), (b) digg metagraph (Lin et al. 2009)
of vertices and three-way relations among them. Subsequently, simpler (two-mode

or one-mode) network representations are derived by use of projection operations
(Mika 2005; Schmitz et al. 2006). For instance, Fig. 2a presents a toy folksonomy
network created from six tag assignment transactions. The “raw” tri-partite folkson-
omy network is transformed to a simple tag association network by considering an
edge between two tags when they are used to tag the same resource. Other variants of
deriving tag association networks from folksonomies are described in the works by
Cattuto et al. (2008a), and Au Yeung et al. (2009).
A more sophisticated paradigm for Social Media networks, termed metagraph, is
presented by Lin et al. (2009). The vertices of a metagraph are organized in facets
and interactions among different facets, which can be multi-way, constitute the edges
of the metagraph. For instance, a metagraph representation of the Digg Social Media
application considers users, stories, comments, topics and keywords as different fac-
ets of the metagraph representation and further maps user activities, such as votes
and commenting, to edges on the metagraph as illustrated in Fig. 2b. Another net-
work model is proposed by Agichtein et al. (2008) for representing users, questions
and answers in a community question-answering application. In conclusion, different
forms of Social Media networks are possible depending on the transactions of the
Social Media application under study, the modeling requirements of the problem, as
well as the capabilities of the network analysis method at hand.
3 Social Media communities
Due to the abundance of related works and the variety of adopted perspectives, there is
no unique and widely accepted definition of community. Community definitions are
formulated with reference to the network structure of the system under study and are
123
commonly bound to some property either of some set of vertices (local definitions)
or of the whole network (global definitions). However, at a different level, one should
also define a community with respect to the domain under study, which in this survey
comprises the realm of Social Media systems. For that reason, we will first provide a
qualitative definition of Social Media communities and subsequently we will link this
definition to established network-based definitions of quantitative nature.
At the most abstract level, given a Social Media network G = (V, E), a Social
Media community can be defined as a subgraph of the network comprising a set
VC ⊆ V of Social Media entities that are associated with a common element of inter-
est. This element can be as varied as a topic, a real-world person, a place, an event, an
activity or a cause. For instance, in a blogging network, the set of all bloggers, articles,
tags and comments related to the topic of “renewable energy” constitutes the respec-
tive community. Similarly, in a photo sharing application, the set of users, photos and
tags that are associated with the island of Crete form a distinct community.
Social Media communities can further be described as explicit or implicit. Explicit
communities are created as a result of human decision and acquire members based
on human consent. Examples of explicit Social Media communities are Facebook and
Flickr Groups. Implicit communities, on the other hand, are assumed to exist in the
system and “wait” to be discovered. Implicit communities are particularly important
for two reasons: (a) they do not require human effort and attention for their creation
and (b) they enable the study of emerging phenomena within Social Media systems.
This survey focuses on the definition and discovery of implicit communities.
3.1 Classes of implicitly defined communities
Implicit communities are defined with reference to the network structure. The most
established notion of community-ness within a network is based on the principle that
some sets of vertices are more densely connected to each other than to the rest of
the network. Depending on whether this property of vertices is considered locally (on
a connected subset of vertices) or globally (on the whole network), we distinguish
between local and global community definitions. Communities are also defined on
the basis of the result of some principled network-based process (process-based defi-
nitions). Several of these definitions are tabulated and described in the supplementary
material included in the work by Kovács et al. (2010).
3.1.1 Local definitions
Early notions of community emerged from social studies focusing on the concepts
of subgroup cohesiveness and mutuality. Examples of such community definitions
are cliques, n-cliques, n-clubs, n-clans, k-plexes, and k-cores (Wasserman and Faust
1994; Scott 2000). More recently, the structures of LS and Lambda sets were defined
(Borgatti et al. 1990) as community constructs. However, most of the above definitions
are too restrictive and computationally very expensive. Thus, their use is very limited
in a Social Media context.
123
Alternatively, the internal and external vertex and subgraph degrees have been used
to define community-ness. The internal degree of a vertex is the number of edges
that connect it to vertices of the same community. The external degree is defined in a
similar way. The definitions of communities in the strong and weak sense (Radicchi
et al. 2004) are based on the internal and external degrees of vertices belonging to a
community. Moreover, it is possible to define some local measure of community qual-
ity and then quantify the degree of community-ness for a given subgraph. Examples
of such local measures are local and relative density (Šíma and Schaeffer 2006), local
modularity (Clauset 2005), and subgraph modularity (Luo et al. 2006).
3.1.2 Global definitions
Global community definitions consider community structure as a property of the whole

network. There are several important classes of global community definitions. Perhaps,
the most intuitive community definition relies on the number of edges falling between
communities (cut size) as a measure of quality of a given network partition into commu-
nities. Since the absolute number of inter-community edges is problematic, normalized
measures such as Normalized Cut (Shi and Malik 2000) and conductance (Kannan et
al. 2004) have been introduced for quantifying the profoundness of separation between
the communities of a network.
Another class of global community structure is based on the widely used concept
of modularity that quantifies the extent to which a given partition of a network into
communities deviates from the hypothetical state (null model) that the network would
be randomly rewired under the constraint of same-degree for each vertex (Newman
and Grivan 2004). A generalization of modularity was presented by Reichardt and
Bornholdt (2006) and numerous variations of the concept have been introduced for
computing modularity in weighted (Newman 2004b), directed (Arenas et al. 2007)
and bipartite (Barber 2007) networks.
Finally, a wide class of global community definitions relies on some similarity
measure between network vertices. Once pairwise similarities between vertices are
computed, communities are defined as clusters of vertices that are close to each other.
Vertex similarities can be derived by use of numerous methods, such as embedding
graph vertices in n-dimensional Euclidean space (and then use some traditional dis-
tance measure such as Manhattan or Euclidean distance) or using the adjacency matrix
of the graph, e.g. compute Pearson correlation between rows of the adjacency matrix,
or random-walk based similarities (Pons and Latapy 2005). Another vertex similarity
measure is the structural equivalence (Lorrain and White 1971), which expresses the
overlap between the neighborhoods of two vertices (even if they are not connected).
3.1.3 Process-based definitions
An alternative means of defining communities is by considering some community

formation process taking place on the network under study. For instance, the Clique
Percolation Method (Palla et al. 2005) considers a k-clique template that “rolls” on the
network and results in a community consisting of the union of all k-cliques that are
adjacent to each other (i.e. share (k −1) nodes). Other community definitions rely on a
123
dynamic process that is iteratively applied on the network in order to reveal groups of
vertices that form well-separated communities. Van Dongen (2000) describes a flow
diffusion process, namely the Markov Cluster Algorithm, which is iteratively applied
on the network in order to render the underlying community structure conspicuous.
According to it, communities are defined as sets of vertices, in which a random walker
is very likely to get trapped. Another dynamic process used for the definition of com-
munities is the synchronization of a set of phase oscillators on a network (Arenas
et al. 2006): groups of vertices whose oscillators synchronize first are considered to
form communities. Finally, a label propagation scheme based on neighbor majority
voting is devised by Raghavan et al. (2007) to define communities as groups of vertices
forming stable consensus with respect to their label.
3.2 Community attributes
Across all community definitions presented above, a set-based view of communities

was presented: community was seen as a set of vertices and the membership of each
vertex in a network was implicitly assumed to be the result of a boolean decision.
In reality, and especially in the context of Social Media, the concept of community
and community membership may be more complicated. For instance, in several of the
above community definitions, e.g., most of the local ones (Clauset 2005; Luo et al.
2006), the one based on Clique Percolation (Palla et al. 2005), and others (Gregory
2009; Chen et al. 2009b), it is possible for communities to overlap (Fig. 3a). Com-
munity overlap is important for Social Media networks since it is common for Social
Media entities to participate in multiple communities; for instance, a user may be
affiliated to his/her family, friends and professional community.
(a) (b) (c)
(d)
Fig. 3 Several attributes that may characterize community structure: (a) overlap, (b) weighted membership,
(c) vertex roles within/across communities, (d) hierarchical organization
123
In addition, there are other attributes that vertices of a network may have in relation
to communities. For instance, different vertices may participate with varying degrees
in a community depending on their centrality6 within it (Fig. 3b). Moreover, vertices
may have discrete roles: for example, Xu et al. (2007) define two roles (hubs and
outliers) for vertices that are not assigned to any community. Hubs are connected to
multiple communities and act as liaisons, thus enabling interactions among commu-
nities. Outliers are connected to a single community through a single link, therefore
they are usually considered as noise. Community-based vertex roles are also discussed
by Scripps et al. (2007). Specifically, the roles of “loners”, “big fish”, “bridges” and
“ambassadors” are defined (Fig. 3c).
Finally, it is possible to impose hierarchical (Fig. 3d) or multi-scale structures on
communities. Community organization may be considered at different scales in a
variety of systems. For instance, a set of users of a Social Media application may
be organized in a community focused on a very specific topic (e.g. fans of a par-
ticular indie-rock band) and at the same time they may be considered as members
of a broader community (rock music). For Social Media systems, the consideration
of multiple levels of community organization typically does not involve any kind of
hierarchical organization since the constraints imposed by the hierarchical model are
too restrictive for modeling the uncontrolled and emerging nature of Social Media
phenomena.
4 Community detection methods
From the discussion of Sect. 3 it became clear that there is a variety of community def-
initions based on network measures and structures. The variety of methods that have
appeared in literature for detecting communities is even larger, since for each com-
munity definition there are more than one methods claiming to detect the respective
communities. Here, we will summarize the most important classes of such methods,
associate them with the definitions presented in the previous section and compara-
tively discuss their performance requirements in terms of computational complexity
and memory consumption, as well as their dynamic computation characteristics, which
are particularly pertinent for the analysis of Social Media networks. Before proceeding
with the discussion of the methods, we will first delineate the relation of the problem
of community detection with that of graph partitioning and graph clustering.
Community detection and graph partitioning. Graph partitioning is a well-specified

problem: divide the vertices of a graph into n groups of given sizes such that the num-
ber of edges lying between the groups (cut size) is minimum. Community detection is
different from graph partitioning in two fundamental aspects. First, community detec-
tion requires neither the number of groups nor their sizes as input in order to extract
them. Furthermore, the result of community detection may not be a partition, i.e. a set
of vertex sets, whose union is the set of all graph vertices and whose pairwise inter-
sections result always in the empty set. As became apparent in Sect. 3, communities
6 Centrality quantifies how frequently vertices fall in the paths between other vertices.
123
in a network may present overlap and there may be vertices in a network that are not
assigned to any community.
Community detection and graph clustering. Community detection is almost inter-

changeably used with graph clustering (Fortunato 2010; Schaeffer 2007). In both
problems, the aim is to identify groups of vertices on a graph that are better connected
to each other than to the rest of the network. However, a differentiation between the
two problems regards the requirement for knowing the number of communities/clus-
ters that a method is expected to identify. Community detection methods typically do
not require the number of communities to be provided as input, but instead the number
of communities is one of the method outputs.7 In contrast, there are numerous graph
clustering techniques that require the number of clusters to be provided as input. Due
to the immense scale and evolving nature of Social Media, it is almost impossible
to know or even to estimate the number of communities in a Social Media network.
Therefore, graph clustering methods that require the number of clusters to be provided
as input are of limited use for the study of such data.8
In this section, we first present a classification of existing community detection
and graph clustering methods based on their methodological principles (Sect. 4.1).
Furthermore, we provide a discussion and experimental study of their computational
complexity and memory requirements (Sect. 4.2), which are particularly pertinent for
their application on a Social Media scale. Finally, we discuss the potential of applying
existing community detection methods on dynamic networks (Sect. 4.3).
4.1 Methodological classification
Depending on the underlying methodological principle as well as the adopted def-

inition of community, we consider five broad classes of community detection and
graph clustering methods: (a) cohesive subgraph discovery, (b) vertex clustering, (c)
community quality optimization, (d) divisive, and (e) model-based (Fig. 4). For the
sake of self-containment, we are going to briefly refer to each of the aforementioned
method classes here. For a more thorough discussion of their principles, we refer to
the survey articles by Danon et al. (2005), Fortunato and Castellano (2007), Fortunato
(2009, 2010), Porter et al. (2009), and Schaeffer (2007). Also, a useful listing of a
large number of community detection methods appears in the supplementary material
of Kovács et al. (2010).
4.1.1 Cohesive subgraph discovery
The methods of this class presume a specification of the structural properties that
a subgraph of the network should satisfy in order to be considered a community.
7 That does not imply that community detection methods are necessarily parameter-free. They may well
rely on some other parameter.
8 However, in case it is possible to identify a relatively small set of values for the number of clusters to be
detected, it is customary to employ such algorithms within some iterative parameter selection scheme with
the goal of identifying the optimal number of clusters.
123
Fig. 4 A classification of community detection and graph clustering methods according to the adopted
community definition and underlying methodological principle
Once such a subgraph structure is specified, methods involve the enumeration of such
structures in the network under study. The local community definitions presented in
Sect. 3.1, e.g., cliques, n-cliques, k-cores, LS sets and lambda sets, are examples
of such cohesive structures and therefore algorithmic schemes for enumerating such
structures, such as the Bron–Kerbosch algorithm (Bron and Kerbosch 1973) and the
efficient k-core decomposition algorithm of Batagelj and Zaversnik (2003), belong to
this class of community detection methods. In addition, methods such as the Clique
Percolation Method (Palla et al. 2005) and the SCAN algorithm (Xu et al. 2007), which
lead to the discovery of subgraph structures with well-specified properties, fall under
the same class of methods.
4.1.2 Vertex clustering
Such techniques originate from the traditional data clustering research. A typical means
of casting a graph vertex clustering problem to one that can be solved by conventional
data clustering methods (such as k-means and hierarchical agglomerative clustering)
is by embedding graph vertices in a vector space, where pairwise distances between
123
vertices can be calculated. Another popular method is to use the spectrum of the graph
for mapping graph vertices to points in a low-dimensional space, where the cluster
structure is more profound (Donetti and Munoz 2004; Von Luxburg 2006). Other ver-
tex similarity measures such as the structural equivalence (Breiger et al. 1975) and the
neighborhood overlap have been used to compute similarities between graph vertices
(Wasserman and Faust 1994). Finally, a noteworthy method, called Walktrap (Pons
and Latapy 2005), makes use of a random-walk based similarity between vertices and
between communities and uses modularity in a hierarchical agglomerative clustering
scheme to derive an optimal vertex clustering structure.
4.1.3 Community quality optimization
There is a very large number of methods that are founded on the basis of optimizing
some graph-based measure of community quality. Subgraph density and cut-based
measures, such as normalized cut (Shi and Malik 2000) and conductance (Kannan et
al. 2004), were among the first to be used for quantifying the quality of some network
division into clusters. A whole new wave of research was stimulated by the measure
of modularity. Approximate modularity maximization schemes abound in the liter-
ature. Apart from the seminal greedy optimization technique of Newman (2004a),
and speeded up versions of it, such as max-heap based agglomeration (Clauset et al.
2004) and iterative heuristic schemes (Blondel et al. 2008), more sophisticated opti-
mization methods have been devised, for instance, extremal optimization (Duch and
Arenas 2005), simulated annealing (Massen and Doye 2005) and spectral optimization
(Newman 2006). Methods aiming at the optimization of local measures of commu-
nity quality, such as local and subgraph modularity (Clauset 2005; Luo et al. 2006),
also belong to this category. Finally, this category includes methods that exploit the
“hills” and “valleys” in the distribution of network-based node or edge functions, e.g.
the ModuLand framework proposed by Kovács et al. (2010) and the “reachability”
measure by Chen et al. (2009b).
4.1.4 Divisive
These methods rely on the identification of network elements (edges and vertices) that
are positioned between communities. For instance, the seminal algorithm by Girvan
and Newman (2002) progressively removes the edges of a network based on some
edge betweeness measure until communities emerge as disconnected components of
the graph. Several measures of edge betweeness have been devised, for instance,
edge, random-walk, and current-flow betweeness (Newman and Grivan 2004), as well
as information centrality (Fortunato et al. 2004) and the edge clustering coefficient
Radicchi et al. (2004). A similar principle is adopted by vertex removal methods
(Vragović and Louis 2006); such methods remove vertices in order to reveal commu-
nity structure. Finally, min-cut/max-flow methods (Flake et al. 2000; Ino et al. 2005)
adopt a different divisive perspective: they try to identify graph cuts (i.e. sets of edges
that separate the graph in pieces) that have a minimum size.
123
4.1.5 Model-based
This is a broad category of methods that either consider a dynamic process taking
place on the network, which reveals its communities, or they consider an underlying
model of statistical nature that can generate the division of the network into commu-
nities. Examples of dynamic processes are label propagation (Raghavan et al. 2007;
Leung et al. 2009; Gregory 2009), synchronization of Kuramoto oscillators (Arenas
et al. 2006), diffusion flow, better known as Markov Cluster Algorithm (Van Dongen
2000), and the popular spin model by Reichardt and Bornholdt (2006). In addition,
community detection can be cast as a statistical inference problem (Hastings 2006),
assuming some underlying probabilistic model, such as the planted partition model,
that generates the community structure and estimating the parameters of this model.
Other model-based approaches rely on the principle that a good clustering is deter-
mined by a low encoding cost, thus they perform community detection by finding the
cluster structure that results in the lowest possible cluster encoding cost (Chakrabarti
2004; Rosvall and Bergstrom 2008).
4.2 Performance comparison
In assessing the performance of community detection methods, there are two funda-
mental aspects that one needs to consider: (a) the computational complexity and (b)
the requirements of the method in terms of main memory. As will be discussed in the
following (Sect. 4.3), the incremental computation properties of community detection
methods constitute an additional performance consideration, which is increasingly
important in the context of Social Media systems. In this section, we provide a theo-
retical discussion of the first two performance aspects, computational complexity and
memory requirements, for a variety of methods, and report the results of an experi-
mental study that compares the performance of eight popular community detection
methods on a wide variety of synthetic networks.
4.2.1 Theoretical discussion
Computational complexity. It is customary to quantify the computational complexity

of community detection and graph clustering algorithms in terms of graph parameters.
Since many of the algorithms rely on vertex and/or edge iteration schemes, their com-
plexity is commonly expressed in relation to the number of graph vertices n and/or
edges m. In addition, the complexity of algorithms is sometimes dependent on the
average or maximum degree of the graph vertices, denoted by k and kmax respectively.
Sometimes, the complexity also depends on the number C of communities to be dis-
covered. For hierarchical methods that produce a community dendrogram, the depth
d of the dendrogram can also affect the complexity of the method.
In the literature, there are many cases where, instead of the “true” complexity of an
algorithm, one reports its approximate complexity based on the assumption that the
underlying graphs are sparse, i.e. n ≈ m. Since this distinction is frequently not evi-
dent, we are going to present, for each method, the complexity both in the case where
123
Table 1 Comparison of community detection complexity
Method Complex-A Complex-B Scale
Cohesive substructure detection

Bron and Kerbosch (1973) O(3n/3 ) O(3n/3 ) S
k-core detection (Batagelj and Zaversnik 2003) O(n 2 ) O(n) L
SCAN (Xu et al. 2007) O(n 2 ) O(n) L
Vertex clustering
Embedding in space + k-means O(C n 2 ) O(C n 2 ) M
Walktrap (Pons and Latapy 2005) O(n 4 ) O(n 2 log n) M
Donetti and Munoz (2004) O(C n 2 ) O(C n 2 ) M
Community quality optimization
Clauset et al. (2004) O(n 2 d log n) O(n log2 n) M
Extremal optimization (Arenas et al. 2007) O(n 2 log n) O(n 2 log n) M
Spectral optimization (Newman 2006) O(n 3 d) O(n 2 log n) M
Community folding (Blondel et al. 2008) O(n 2 ) O(n) L
Divisive
Girvan and Newman (2002) O(n 5 ) O(n 3 ) S
Information centrality (Fortunato et al. 2004) O(n 7 ) O(n 4 ) S
Edge clustering coefficient (Radicchi et al. 2004) O(n 6 ) O(n 2 ) M
Max flow + Gomory-Hu tree (Ino et al. 2005) O(n 4 log n) O(n 3 log n) S
Model-based
MCL (Van Dongen 2000) O(n 3 ) O(n 3 ) M
Minimum encoding cost (Chakrabarti 2004) O(n 2 ) O(n) L
Label propagation (Raghavan et al. 2007; Leung et al. 2009) O(n 2 ) O(n) L
Infomap (Rosvall and Bergstrom 2008) O(n 2 log n) O(n log n) L
Two bounds are provided, one for general graphs irrespective of density (Complexity-A) and one computed
under the assumption that the graph is sparse (Complexity-B). Furthermore, the scale of graphs for which
each method is appropriate is provided: S stands for small scale (<104 nodes), M stands for medium scale
(<106 nodes), and L for large scale (106 −109 nodes)
no assumption is made about the underlying graph and in the case where the assump-
tion of a sparse graph is made. Table 1 presents a complexity comparison among
several selected graph clustering and community detection methods. From the table,
it appears that vertex clustering and divisive approaches present complexities higher
than quadratic to the number of network vertices, which renders them inapplicable
for large- and mega-scale networks that commonly appear in Social Media systems.
On the other hand, there are several recent methods (Blondel et al. 2008; Leung et al.
2009; Raghavan et al. 2007; Xu et al. 2007) from the other classes that present linear
complexity to the number of network edges. Since the number of edges contained in
real-world Social Media networks can easily exceed the billion scale, it is obvious that
such methods are preferable in terms of complexity for tackling community detection
on such networks.
For several popular methods, it is not possible to derive a closed-form expres-
sion of their complexity, as this depends on the structure of the graph under study,
as well as on particular method parameters. For instance, in the case of the Clique
Percolation Method (Palla et al. 2005) there is no closed-form expression for bound-
ing the algorithm running time, although it is experimentally demonstrated that the
algorithm completes in reasonable time for fairly large networks. In a similar fashion,
123
the complexity of the modularity maximization method based on simulated anneal-

ing (Massen and Doye 2005) cannot be expressed in terms of graph characteristics
(although it is known to be a slow method and applicable only to relatively small
networks).
A whole class of methods whose complexity cannot be expressed in closed form are
the methods that rely on spectral graph properties (Von Luxburg 2006). Such methods
are based on the computation of some or all the eigenvectors of the graph Laplacian in
order to cluster the vertices of the graph. In the worst case, where all eigenvectors are
computed, such methods have a O(n 3 ) complexity. However, most methods compute
only a limited number of eigenvectors (corresponding to the smallest eigenvalues of
the Laplacian) by use of some efficient method (e.g. Lanczos) and are thus much more
efficient. Still, their complexity cannot be expressed in relation to the graph size.
Finally, the complexities of methods that operate at a local level, such as the ones
by Clauset (2005), Luo et al. (2006), Bagrow (2008), Papadopoulos et al. (2009a),
were not included in Table 1, since such methods are used to discover a single com-
munity around a given vertex or set of vertices. Although it is possible in principle
to derive global community detection schemes by bootstrapping such methods, there
is no meaning in comparing the complexity of such schemes with inherently global
approaches. For that reason, and due to the fact that local methods can be used as a
means of scaling community detection to larger networks, we present an additional
discussion of local methods in Sect. 5.2.
Memory requirements. The magnitude of networks formed from Social Media data,
and the development of community detection methods presenting linear complexity
with respect to the network size, has revealed an additional bottleneck to community
detection methods: memory consumption. Most existing techniques consider direct
and instant access to graph elements, since the whole graph is assumed to reside in
main memory. Moreover, additional main memory structures are frequently used by
methods to speed up certain graph analysis operations. For that reason, it is necessary
to assess community detection methods on the basis of their memory requirements.
A large number of community detection methods rely on the edges of the graph
under study to reside in memory. In addition, the assignments of nodes to communities
and their degrees need to be readily accessible by the algorithm (e.g. in the case of
modularity computation). Thus, the bare minimum memory consumption for many
community detection methods scales linearly with the size of the graph9 (2 · m + 2 · n).
This estimate assumes that the graph is unweighted and stored in the form of an edge list
(thus requiring only two elements per edge, the endpoint vertices). In practice, many
community detection implementations rely on a matrix representation of the graph
(adjacency matrix) to be available in memory, thus calling for at least 2 · m + 3 · n in
case of unweighted undirected graphs in case the Yale sparse matrix format is used.
More often than not, methods require additional data structures to be held in mem-
ory in order to speed up certain operations. For instance, the efficient greedy scheme
of Clauset et al. (2004) makes use of a Q matrix instead of a simple adjacency
9 When discussing memory consumption, it is meaningless to use the Big O notation, since we want to
distinguish between a memory consumption of α · m, α ∈ N+ and one of m.
123
matrix that requires both n binary trees and n max-heaps (which are also represented
in the form of trees), thus raising the memory needs of the method. The eigenvec-
tor-based approach for modularity maximization (Newman 2006) poses even higher
requirements for memory, since it relies on the so-called modularity matrix B that is
not sparse and thus takes up n 2 space. Other spectral approaches (Von Luxburg 2006)
store a set C of eigenvectors in main memory (in addition to some graph-derived
sparse matrix, such as the Laplacian) resulting in the need for additional C · n memory
space. Some recent methods tackle the problem of minimizing the Normalized Cut of
a clustering (spectral partitioning), without the need to compute the eigenvectors of
graph-derived matrices (Dhillon et al. 2007).
Another class of methods that inherently suffer from large memory requirements are
those based on vertex clustering. Such methods involve the embedding of graph verti-
ces in some n-dimensional space, thus resulting in a dense n × n matrix, which takes
up n 2 memory space. Moreover, community detection based on statistical inference
(Hastings 2006) also suffers from high memory usage. Approximating the solution to
the belief propagation problem that the method tries to solve requires an additional
C · n memory in order to track the estimated probabilities with which network vertices
belong to each community.
On the other hand, there are numerous community detection methods that are suit-
able even under strict memory constraints. For instance, the label propagation method
of Raghavan et al. (2007) proceeds by inspecting only the neighborhood of a vertex
each time a relabeling operation is performed. Therefore, if needed the method can
work even with as little as n + kmax memory space (community assignment vector
plus vector of largest vertex neighborhood) assuming that the graph is streamed in
memory in the form of vertex adjacency lists. Obviously, local community detection
approaches are also memory efficient, since they process only a portion of the graph at
a time. However, when the local community boundary grows large (e.g. in the case of
gigantic communities), even local methods might cause significant memory overhead
(this is further discussed in Sect. 5.2).
4.2.2 Experimental study
In order to gain insight into the performance aspects of community detection, we

have conducted a comparative experimental study including eight well-known meth-
ods, namely10 : (1) the greedy modularity maximization scheme by Clauset, Newman,
Moore (CNM) (Clauset et al. 2004), (2) the Walktrap method (WALKTRAP) by Pons
and Latapy (2005), (3) the “leading eigenvector” method (LDEIGEN) by Newman
(2006), (4) the spin configuration method (SPIN) by Reichardt and Bornholdt (2006),
(5) the label propagation method (LPROP) by Raghavan et al. (2007), (6) the Markov
Cluster Algorithm (MCL) by Van Dongen (2000), (7) the heuristic modularity maximi-
zation scheme (LOUVAIN) by Blondel et al. (2008), and (8) the information theoretic
method (INFOMAP) by Rosvall and Bergstrom (2008).
10 For ease of reference, we provide for each method a short name in parentheses.
123
In order to test the method performance under a wide range of structural net-
work features, we used the synthetic benchmark graphs by Lancichinetti et al. (2008).
The graph generation process is based on five parameters: (a) number of nodes, N , (b)
average degree of nodes, k̂,11 (c) mixing parameter, μ, indicating the “conspicuous-
ness” of communities (the higher its value, the fuzzier the communities), (d) power
law exponent for node degree distribution, γ , (e) power law exponent for community
size distribution, β. For testing the effect of each parameter on the performance of the
methods, we created synthetic graphs with varying values for each one of the parame-
ters while keeping the rest of the parameters fixed. For each parameter set, we created
10 graph realizations in order to obtain more reliable performance measurements by
averaging over the respective measurements. Our performance measurements include
execution time, Normalized Mutual Information (NMI),12 and peak memory consump-
tion for each method. Thus, apart from measuring the computational requirements of
the methods, we also benchmarked by means of NMI the quality of the community
structure that they produce.
Figure 5 provides an overview of the benchmark results by depicting the relation of
method performance to three graph generation parameters, namely N , k̂, and μ. For
the sake of brevity and due to the fact that there was no significant influence of per-
formance on parameters γ and β, the respective diagrams were omitted. Figure 5a–c
depicts the relation of the method execution time in relation to the three graph genera-
tion parameters. Figures 5d–f illustrates the execution time in relation to the achieved
community structure quality in terms of NMI. Finally, Fig. 5g, h presents the peak
memory requirements of the competing methods in relation to the number of nodes
and average degree (i.e. number of edges) of the network.
In terms of execution time, there is a remarkable performance difference between
the fastest methods, namely LPROP and LOUVAIN, and the slowest ones, i.e. SPIN
and MCL. For instance, for a network of 10 thousands of nodes, LPROP and LOU-
VAIN take approximately 0.5 and 0.6 s to complete, while MCL takes 75 s and SPIN
almost 19 min. In terms of scalability, all methods appear to scale almost linearly to
the number of nodes in the network. An exception is the WALKTRAP method that
exhibits an abrupt increase in the execution time after the network grows larger than
10 thousands of nodes. This is actually not an inherent complexity characteristic of
the method but it is due to the fact that the method consumes all available memory,
as becomes obvious from Fig. 5g, and it thus needs to frequently access the disk.
Another method that scales a little worse than linear is the LPROP method, which
for small size networks is the fastest one, but for larger size networks comes second
after the LOUVAIN method. The execution time of some methods is also affected
by the average degree of the network. SPIN, MCL, CNM and WALKTRAP are the
methods affected most by the average network degree, while INFOMAP, LOUVAIN
and LPROP do not seem to be significantly influenced. Finally, changing the network
mixing parameter, i.e. how much the different communities “blend” with each other,
11 This is equivalent to the number of edges of a network.

12 This measure is typically used in community detection studies, e.g. the comparative study by Danon et
al. (2005), to quantify the extent to which the detected community structure matches the “true” one, which
in the case of synthetic graphs is known.
123
6
10
6
EXEC TIME (msec)
EXEC TIME (msec)
EXEC TIME (msec)

10
5
6 10
10
4
10
4
4 10
10 3
10
2
2 10
10 2
1
10 1
10 10
1K 5K 10K 50K 100K 10 25 50 100 150 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
(a) (b) (c)

1 1 1
0.9 0.8 0.8

0.8
0.6 0.6
NMI
NMI
NMI
0.7
0.4 0.4
0.6
0.5 0.2 0.2
0.4 0 0
1K 5K 10K 50K 100K 10 25 50 100 150 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
(d) (e) (f)

4
14 x 10
7
10
memory usage (KB)
memory usage (KB)
12 LOUVAIN
6
10 CNM
10
INFOMAP
5
10 8 LPROP
6 LDEIGEN
4
10 MCL
4
SPIN
3 2
10 WALKTRAP
0
1K 5K 10K 50K 100K 10 25 50 100 150
(g) (h)
Fig. 5 Benchmark results for eight community detection methods
does not seem to have significant influence on the execution time of methods with the
exception of INFOMAP, MCL and SPIN.
In terms of memory consumption, WALKTRAP was found to be the most memory
hungry method, consuming all available 8 GB of the test machine when networks of
50 thousands of nodes or larger were analyzed. This was also the reason for the steep
decrease in its performance in terms of execution time, as described above. MCL was
the second most memory hungry method needing more than 1.5 GB of memory for
a network of 100 thousands of nodes. In contrast, the LOUVAIN method exhibited
the best memory performance, consuming only 18 MB of memory for a network of
100 thousands of nodes. This is due to the fact that the LOUVAIN implementation is
based on an internal binary network representation that is much more compact than
the adjacency lists used by the majority of its competing methods. An additional use-
ful observation of how the memory requirements of methods scale with the average
degree (i.e. number of edges) in the network can be drawn from Fig. 5h: LOUVAIN
and MCL appear relatively insensitive to k̂, while the memory requirements for the
rest of the methods are heavily dependent on k̂.
Our final observations pertain to the precision of the detected community structure,
quantified by use of the NMI measure. The number of nodes in the network has only
123
limited effect on the precision of the detected community structure. Only the LOU-
VAIN and SPIN methods appear to suffer from the increase in the number of nodes;
specifically, their NMI performance drops approximately 10% as the number of nodes
in the network grows from one to 100 thousands of nodes. Furthermore, the average
node degree appears to have a strong impact on the precision of the detected commu-
nity structure. All algorithms sustain considerable loss in NMI, which is particularly
pronounced for MCL (NMI drops below zero after k̂ exceeds 50) and LPROP (NMI
gets zero once k̂ gets 100). The most resilient methods in that respect are WALKTRAP,
LOUVAIN and SPIN. Finally, the effect of the mixing parameter on the precision of the
detected community structure is in accordance with the observations of Lancichinetti
and Fortunato (2009), where a sharp drop in NMI is observed as μ increases from
0.5 to 0.8. The most sensitive algorithms to μ are LPROP and CNM, while MCL and
WALKTRAP appear to be the most resilient.
4.3 Dynamic community detection
So far, the discussion on community detection has progressed under the silent assump-
tion that the network under consideration is static. In fact, it is only recent works in
this area that take the evolving nature of network data into account. Due to the highly
dynamic nature of Social Media data, this is a significant aspect of community detec-
tion that is worth further attention. In general, time-awareness can be incorporated in
community detection approaches in a number of ways: (a) longitudinal application on
successive snapshots, (b) vertex-centric time-awareness, (c) incremental application.
4.3.1 Longitudinal application on successive snapshots
This is a simple approach for extending any community detection method for the
analysis of dynamic networks. It consists of two basic steps: (a) application of com-
munity detection on a sequence of static network snapshots, and (b) identification
of correspondence between communities found in successive network snapshots.
For instance, this is the approach adopted by Palla et al. (2007), where the Clique
Percolation Method (Palla et al. 2005) is applied on successive graph snapshots, G(t)
and G(t + 1), as well as on their union G(t) ∪ G(t + 1) in order to facilitate the
community correspondence identification. An important result of this study pertains
to the possible changes that communities undergo throughout their life, e.g. growth,
contraction, merging, splitting, birth and death.
Similar community evolution operations are defined in Asur et al. (2007), where
the MCL graph clustering algorithm (Van Dongen 2000) is used to discover commu-
nities on successive graph snapshots. A more recent approach (Kim and Han 2009)
relies on the results of SCAN (Xu et al. 2007) on individual graph snapshots and
identifies the correspondence between communities in different times by imposing
temporal constraints on successive graph snapshots. This is achieved by considering
vertex similarity links across graph snapshots at different times and identifying dense
t-partite quasi-cliques on the t-partite graphs that are formed by linking similar ver-
tices of successive graph snapshots. Alternatively, edges between vertices have been
123
used as community features and their correlations across different time snapshots have
been used to match and track communities in time (Lin et al. 2007).
4.3.2 Vertex-centric time-awareness
According to this approach, the dynamic character of community structure is attributed

to their vertices. For instance, Fenn et al. (2009) apply the spin model of Reichardt and
Bornholdt (2006) on successive graph snapshots to identify communities. Then, for
each vertex it quantifies the extent to which the vertex remains in the same community
or switches from one community to another. An alternative approach is introduced
by Wang et al. (2008): community evolution is described in terms of a small set of
core vertices. For instance, when two successive communities share a common core
vertex, then the second community (from a temporal point of view) is considered as
an evolved version of the first. Due to the fact that vertex-centric approaches are also
based on successive graph snapshots, they are quite similar to the previous category
of dynamic community detection methods.
4.3.3 Incremental application
This is a more sophisticated means of considering the dynamic character of networks.

The community detection at time t + 1 is initialized with information derived from the
community structure discovered at time t. The incremental nature of these methods
has two advantages: (a) it is considerably more efficient in terms of computations, (b) it
leads to more consistent community detection results, since the clustering at time t + 1
is not dramatically different than the one at time t, i.e. a balance is achieved between
temporal noise and concept drift (Chakrabarti et al. 2006). Incremental methods can
be further divided in iterative approximation schemes (Sun et al. 2007; Yang and Liu
2006; Lin et al. 2008) and graph modification oriented reclusterings (Falkowski et al.
2007; Franke and Geyer-Schulz 2009).
Iterative approximation schemes are inherently incremental since they rely on
some iterative scheme that progressively approximates the target community struc-
ture. Therefore, in case they are initialized with the community structure found for
the previous graph snapshot, they tend to converge to the new community structure
in much fewer iterations. For instance, in GraphScope (Sun et al. 2007), community
detection is carried out by minimizing the information required to encode a graph and
its community structure (Minimum Description Length). Initializing the community
structure with some rough estimate (the structure found in the previous timestep) leads
to faster minimization of the Minimum Description Length. Similarly, the incremental
scheme of Yang and Liu (2006) uses an iterative approximation technique for estimat-
ing repulsion and attraction forces among vertices that are responsible for separating
the network into communities. Finally, FacetNet (Lin et al. 2008) presents another
important paradigm of incremental community detection formulated on the basis of
evolutionary clustering (Chakrabarti et al. 2006). According to it, community detec-
tion is formulated as a cost minimization problem that is expressed as the combination
of two costs: the snapshot cost, which is related to the current network structure,
and the historic cost, which is derived from the divergence of the current community
123
structure from the previous one. This cost is minimized through an iterative rule update
scheme.
The second approach to incremental community detection is based on the idea that
the incremental method should only take into account the graph modification opera-
tions (e.g. new vertex, new edge, etc.) in order to derive a new community structure,
which would be the same or equivalent with the one that would have been derived in
case a complete reclustering had taken place. For example, the DenGraph framework
(Falkowski et al. 2007), which uses a notion of subgraph density equivalent to the
concept of (μ, )-cores (Xu et al. 2007), translates each modification taking place on
the graph into some update operation (or no operation in some cases) on the graph
cores and their associated communities. A more sophisticated scheme is presented
by Franke and Geyer-Schulz (2009) based on an incremental version of a restricted
random walk clustering scheme, which takes into account only the changes that have
taken place on the graph since the last clustering.
5 Strategies for scaling community detection on Social Media scale
The remarkable growth of Social Media data has brought attention to the problem
of scaling community detection methods to real-world networks that contain many
billions of vertices and edges. Here, we present five strategies that can be employed
in order to scale community discovery methods to networks of such magnitude.
5.1 Sampling techniques
One possibility for reducing the complexity of community detection methods is to

employ a sampling strategy in the computationally expensive part of the method.
Previous works (Leskovec and Faloutsos 2006; Hübler et al. 2008 have demonstrated
that it is possible to sample a small subgraph from a large network so that the induced
subgraph has the same network properties (e.g. distribution of degree and cluster-
ing coefficient) with the original large graph. However, in the context of community
detection, vertex (or edge) sampling may have different goals.
For instance, Tyler et al. (2003) make use of sampling in the computation of
edge betweeness in large graphs, i.e. their sampling targets at approximating the true
betweeness of edges. In their betweeness computation, they take into account only
the contribution from a limited number of randomly selected vertices, which saves
significant computational overhead at a modest statistical error that is further reduced
by repeating the sample-based computation several times and aggregating the results.
A crude means of approximating community detection is by use of high-degree vertex
sampling (Java et al. 2008b). Subsequently, the unsampled vertices are assigned to the
community to which they have more connections. Following a more refined approach
(Maiya and Berger-Wolf 2010), vertices are sampled based on their expansion proper-
ties in order to derive a subgraph, of which the community structure is representative
and can thus be used to infer the community assignments of unsampled vertices by
use of statistical relational learning.
123
Aside the computational efficiency aspects, sampling network data is important for
Social Media data for an additional reason: Any Social Media network under study
is essentially the result of a network sampling process that is implemented by some
crawling algorithm (Gjoka et al. 2009; Ye et al. 2010). For that reason, the proper-
ties of the sampled network and the extent to which these are representative of the
original network depend on the employed crawling/sampling principle. In the end,
the results of community detection will implicitly depend on the adopted network
sampling process.
5.2 Local graph processing
Local community detection methods (Flake et al. 2000; Clauset 2005; Luo et al. 2006;
Papadopoulos et al. 2009a; Chen et al. 2009a) provide a means to alleviate scalabili-
ty challenges by focusing on a portion of the network under study. In practice, such
methods start a network exploration process from a seed vertex or vertex set and start
attaching adjacent vertices to the community as long as these attachments lead to the
increase of some local community quality measure. This principle is illustrated in
Fig. 6. According to it, the network is divided in four parts: (a) the set B of border
vertices that are adjacent to at least one vertex not belonging to the community, (b)
the set C of core community members that have no connection to vertices outside the
community, and (c) the set U of vertices that are adjacent to at least one vertex of the
community border, (d) the rest of the network. A local method proceeds by examining
which vertex (if any) from set U is appropriate for attachment to the community. Once
such a vertex is identified and attached to the community, sets B, C, and U are updated
and the process continues until it is not possible to identify any other vertex that should
be attached to the community.
Since local methods are limited to a portion of the network under study, it is expected
that they can circumvent the memory bottleneck faced by global methods. In most
local community quality measures, such as local modularity (Clauset 2005), subgraph
Fig. 6 Basic principle of local community detection: Starting from some seed vertex, one progressively
explores the neighborhood around it. Vertices are distinguished between Core (C), Boundary (B) and
Un-visited (U), with respect to their relation with the local community
123
Fig. 7 Basic principle of streaming/iterative schemes: an iterative process visits the vertex adjacency lists
and fetches part of them in main-memory, where the community computation is performed by use of the
local adjacency information and some additional state variables
modularity (Luo et al. 2006) or node outwardness (Bagrow 2008), it is necessary to

maintain in memory only the subgraph G L = (VL , E L ) ⊆ G, where VL = C ∪ B ∪ U
and E L = {(u, v) ∈ E|u, v ∈ VL }. However, in case such a local method leads to
the formation of a gigantic community, violating the assumption that |G L | |G|,
even local methods are not sufficient to address scalability concerns. Such cases are
expected to be rare and dependent on the network topology under study as well as the
local community quality measure employed by the method.
Local methods have the added advantage that they are targeted at some specific
topic. This is especially important for the Web in general and Social Media appli-
cations in particular, since it is not feasible to study the community structure of the
whole Web. Usually, a community of interest can be defined by means of some seed
items (e.g. web pages, online photos, videos) and then the community containing them
can be progressively discovered by use of either their hyperlinks (conventional Web
communities (Flake et al. 2000; Ino et al. 2005) or by use of the associated metadata
(e.g. tags, comments, etc.), which create implicit links between content items.
5.3 Streaming/iterative schemes
Streaming and iterative schemes are a very promising means of scaling community
detection. According to them, an iterative process examines each vertex along with its
neighbors in a given order and performs a computation, of which the result is associ-
ated with the processed vertex. This vertex iteration process is repeated multiple times
until convergence. The same principle can be applied by use of edge iteration. In that
way, it is possible to perform community detection on networks that do not fit in main
memory under the assumption that the relevant parts of the graph (e.g. vertex neigh-
borhoods) are streamed in main memory and state information (e.g. values associated
with vertices) resides in main memory. This principle is illustrated in Fig. 7.
Label propagation (Raghavan et al. 2007) is a prominent example of the iterative
paradigm. According to it, each vertex is visited and assigned a label as a result of
majority voting by its neighbors. This vertex iteration is repeated multiple times until
123
convergence.13 A more complex example of iterative community detection is provided

by the (μ, )-core community detection scheme (Xu et al. 2007) that consists of a
single edge and a single vertex iteration: each edge is visited once in order to compute
the structural similarity between the vertices it connects. Subsequently, a vertex vis-
iting scheme is carried out that first identifies (μ, )-cores and then attaches to them
those vertices that are structure reachable from them (in total each vertex is visited
once).
5.4 Multi-level approaches
Multi-level approaches rely on the following principle: First, they try to find a rough
partition of the network into communities by use of a fast process (at the expense of
accuracy). Subsequently, the create a meta-network, where vertices represent commu-
nities and edges stand for inter-community associations. This meta-network is much
smaller in size than the original, therefore more accurate partitioning techniques can be
devised. This principle was popularized by the METIS graph partitioning framework
(Karypis and Kumar 1998) and applied within a community detection framework by
Djidjev (2008), where modularity maximization is cast as a minimum weighted cut
problem that is solved by recursive bisection and refinement.
In addition, a community “folding” or “contraction” process can be integrated in
various community detection schemes (Ino et al. 2005; Blondel et al. 2008) with the
goal of uncovering community organization at multiple resolutions. Ino et al. (2005)
devise an efficient scheme for enumerating certain types of subgraphs (satisfying the
definition of the so-called IKN-community) and apply it recursively at multiple levels
in order to uncover hierarchical community structure. Similarly, a heuristic scheme
for modularity maximization on large scale networks is presented by Blondel et al.
(2008). Once an optimal partition is identified at the lowest resolution, the identified
communities are “folded” into vertices and the same scheme is applied at the meta-net-
work comprising communities as vertices. In such methods, the main computational
burden falls on the first level of the multi-level scheme. For that reason, heuristics are
commonly devised to speed up computations at this level of community detection.
An alternative multi-level scheme was presented by Gibson et al. (2005), where
large dense bi-cliques of vertices are identified on a directed network. A shingling
operation is proposed that maps an input set of size n to a set of much smaller size
(a so-called “fingerprint”). This operation is used to create fingerprints of the neigh-
borhoods of vertices (first level shingles) and then re-applied to create second-level
shingles. Then, the extraction of dense bi-cliques is translated to a problem of con-
nected component identification on the network of first-level shingles G S . This prob-
lem is much easier than the original due to the fact that it only needs memory access
to the vertices of G S , since its edges are encoded in the second-level shingles. In that
way, the authors could analyze the community structure of a network consisting of
two billion vertices and 11 billion edges.
13 Sometimes there may be oscillation phenomena that can be addressed by extensions of the method
(Leung et al. 2009).
123
5.5 Distribution and parallelization
The constant increase in the size of the networks that need to be analyzed will even-
tually bring the aforementioned strategies for scaling community detection to their
limits assuming they are applied on a single machine. For that reason, an essential
strategy for making community detection scalable to peta-scale networks is the distri-
bution of the network into multiple processing nodes and parallel execution of many
community detection processes.
Hui et al. (2007) present an interesting approach to distributed community detection.
Community detection is seen as a local process taking place in each processing node
in a network of devices so that each device is “aware” of its community. Although this
perspective does not address at all the scalability aspects of the problem, it is interest-
ing for two reasons: (a) it introduces the paradigm of distributed community detection
by use of local techniques (Sect. 5.2), (b) it demonstrates the practical issues arising
when a local community detection method (such as the one based on optimizing the
local modularity by Clauset (2005)) is applied in an environment where only partial
knowledge of the network is possible.
The Bulk Synchronous Model (BSP) is used by Zhang et al. (2009) to detect com-
munities in large networks by means of distribution. The processing model is vertex-
oriented (similar to the one by Hui et al. (2007)) and involves a set of local updates
on a local structural property termed “propinquity” and message passing to neigh-
boring vertices. It is demonstrated by authors that this dynamic process leads to the
emergence of communities in the form of connected components. The performance
of the technique is heavily dependent on the vertex degree distribution, which led
the authors to the adoption of some ad hoc degree-based filtering on the network. In
terms of parallelism, the authors achieve substantial gains with the addition of new
processing nodes up to a meaningful number of nodes given the network size (i.e. it
is of no use to partition a relatively small network in many nodes due to the overhead
that will be incurred by the excessive message passing).
Alternatively, Yang et al. (2009) implement a parallel community detection scheme
on top of the MapReduce (Dean and Ghemawat 2004) framework. A special mapping
and reduction scheme is employed with the goal of distributing the problem of enu-
merating all maximal cliques in large networks. Subsequently, a clique merging step
is performed that results in a clique-based community structure similar to the one
discovered by the Clique Percolation Method (Palla et al. 2005), but with no need to
specify the k parameter. However, their experiments indicate that after distributing the
problem to more than 30 processors, the total processing throughput levels off indi-
cating increased computational overhead and thus calling for improved parallelization
mechanisms.
6 Exploitation of community detection results
Community detection has attracted extensive research interest for some time now.
Despite the proliferation of methods developed for analyzing different types of com-
munities, there have been relatively less works that actually exploit the results of
123
community detection in some application. Traditionally, social network analysis and

biological systems have benefited from the results of community detection (Girvan
and Newman 2002). In addition, research on the structure and processes taking place
on the Web (Kumar et al. 1999; Flake et al. 2000) constitute an additional applica-
tion domain. It is only recently that some works have started exploiting the results of
community detection research on Social Media data. Yet, Social Media present unique
opportunities and challenges for community detection research due to the following
properties: (a) the unprecedented growth and magnitude of networked data generated
by Social Media applications, (b) the high rate of spam and the large variance in the
quality of Social Media content and metadata, (c) the highly interactive and dynamic
character of Social Media systems.
Here we collect the main applications that community detection has seen to date in
the context of Social Media. We differentiate between observational studies that are
focused on the structure and dynamics of communities in Social Media networks and
application-oriented studies that attempt to integrate community detection results in
the context of some information retrieval or web mining task.
6.1 Observational studies
6.1.1 Statistical analysis of Social Media community structure
Several studies focus on the structure of communities in Social Media networks. Most
of them consider some specific Social Media application and examine the commu-
nity structure of the network derived from it. For instance, a study on a collaborative
tagging system (delicious) is presented by Cattuto et al. (2008b), where an online book-
mark network is created from the tagging activities of delicious users14 and a spectral
technique is applied for separating the resources into communities. Another study
(Zakharov 2006) focuses on the community structure of the users of the LiveJournal
blogging platform. Through a thermodynamic diffusion process on the network, the
authors could identify several prominent communities on the network (e.g. a Russian-
speaking blogging community and two communities focused on two RPGs). Another
study on blog communities is presented by Java et al. (2008a); in that, a co-clustering
framework is employed in order to exploit both the hyperlink structure of blog articles
as well as the tag information of articles.
One of the most extensive and systematic studies on the community structure of
networks, including several Social Media ones, is presented by Leskovec et al. (2008).
The examined Social Media networks are mostly user affiliation networks from dif-
ferent Social Media applications (Delicious, Epinions, Flickr, LinkedIn, LiveJournal,
Yahoo! Answers), but there is also an interesting network comprising Netflix users
and movies as vertices and ratings as edges (bipartite). The authors employ both a
hybrid graph partitioning technique based on METIS (Karypis and Kumar 1998) and
MQI (Gallo et al. 1989), and a local spectral partitioning technique (Andersen et al.
2006), both of which identify partitions with minimum conductance (Kannan et al.
14 Two bookmarks are linked by a weighted edge depending on the number of tags they share.
123
Fig. 8 Core-periphery
structure, a characteristic of
many Social Media networks
(Leskovec et al. 2008)
2004).15 In order to express the community structure quality in different scales, the
authors employ the Network Community Plot, i.e. a plot of the minimum conductance
value in relation to the community of size.
From the analysis across a wide range of networks and by use of different graph
partition algorithms, it appears that the communities of the networks of interest are
integrated in a core-periphery structure, where the periphery consists of the so-called
“whiskers”, densely connected subgraphs (of size in the order of 100 vertices) that are
connected to the rest of the network by only few (typically one or two) edges. A pic-
torial representation of this structural pattern is depicted in Fig. 8. The study revealed
that such peripheral communities can collectively constitute even more than half the
size of the network and are responsible for the “best” conductance values achieved by
graph partitioning schemes. As communities get larger, they tend to blend in with the
core of the network and it is harder to separate them by means of graph partitioning.
6.1.2 Temporal analysis of Social Media communities
Thanks to the recent development of numerous methods for discovering and dynami-
cally tracking the community structure of networks (Sect. 4.3), the focus of community
detection studies on Social Media has begun to shift from static community structure
analysis to their temporal evolution and dynamics. Consequently, there have been
several efforts to understand the evolving phenomena that characterize Social Media
communities. For instance, Lin et al. study the evolution of blog communities form-
ing around real-world events by analyzing the temporal correlations of interactions
between bloggers of the same community through observation of their mutual links
Lin et al. (2007). Schlitter and Falkowski (2009) study the community dynamics of
an online music listening application (last.fm) by use of their dynamic community
detection framework.
The aforementioned works observed the formation of new communities, the evolu-
tion (growth or contraction) of persistent ones, as well as more complicated community
dynamics (merging, splitting). Comparing these studies with the seminal work by Palla
et al. (2007), it appears that some consensus is formed regarding the possible transfor-
15 The reason for using two methods is that the global one focuses on minimizing the conductance over
the whole graph, while the second also makes sure that local community properties, such as compactness,
are preserved even at the expense of conductance minimization.
123
(a) (b)
(c)
Fig. 9 Basic primitives of community structure evolution: (a) growth and contraction, (b) merge and split,
(c) emergence and extinction
mations that communities may undergo. Figure 9 illustrates six basic transformations
that have been identified in a number of studies (Lin et al. 2007; Palla et al. 2007;
Schlitter and Falkowski 2009). Basically, there are three types of transformations:
(a) one-to-one (Fig. 9a), which involves community growth or contraction, (b) one-to-
many (Fig. 9b), which involves one community splitting to many or many communities
merging to one, and (c) one-to-zero (Fig. 9c), which involves the emergence or the
extinction of a community.
6.2 Applications
6.2.1 Topic detection in collaborative tagging systems
The huge amount of tags attached to online content by users of Social Media applica-
tions creates the need for imposing organization on the flat tag spaces of collaborative
tagging applications. This can be directly achieved by grouping tags based on the
topic they are associated with. There have been several recent works that attempt
to derive meaningful clusterings of tags that correspond to topics of social interest
(Begelman et al. 2003; Papadopoulos et al. 2009a,b, 2010a,b; Simpson 2008). For
instance, Begelman et al. (2003) were among the first to apply community detection
methods, namely spectral modularity maximization, to identify interesting tag clusters.
Similarly, tag clustering is pursued by means of a variant of the modularity maximi-
zation method of Newman and Grivan (2004) on enterprise folksonomies (Simpson
2008). A problem that the latter study pinpointed is that the employed modularity max-
imization method led to the discovery of disproportionately large communities when
123
the underlying tag co-occurrence network was dense (e.g. the tag network formed
from delicious data).
Alternative tag community detection schemes were employed by Papadopoulos et
al. (2009a,b, 2010a). In the first work Papadopoulos et al. (2009a), an efficient local
community detection scheme was devised that could discover the community around
a seed tag. Figure 10 presents several examples of tag communities discovered on a
tag network created from the LYCOS iQ community question answering application
(similar to Yahoo! Answers). Such communities can contribute to understanding the
interests of the users of Social Media applications. For instance, by observing the
community structure for the topics Music (Fig. 9c) and Science (Fig. 9d), it becomes
obvious that in LYCOS iQ there is more interest and a more sophisticated topic struc-
ture with respect to the topic of Science than to Music.
An alternative tag community detection scheme was presented in Papadopoulos
et al. (2009b). The proposed scheme relied on a two-step process that involved the
selection of densely connected groups of tags by use of SCAN (Xu et al. 2007) and
the expansion of these groups with the goal of maximizing the subgraph modularity
of Luo et al. (2006), which could be efficiently computed in an incremental fashion.
A refinement of this method appeared in Papadopoulos et al. (2010a) solving the prob-
lem of parameter selection introduced by the first step of the algorithm. A limitation
of this approach was the topic blending problem, i.e. the situation in which some tag
communities contained tags related to different topics due to some polysemous tag
that acted as bridge between them. This has been addressed by tag disambiguation
methods that are discussed below.
6.2.2 Tag disambiguation
Due to the unrestricted and informal nature of tagging, there are numerous cases where
the use of a single tag in isolation is not sufficient to convey the intended semantics.
For that reason, tags need to be considered in context in order to disambiguate their
meaning. Recently, several research efforts (Au Yeung et al. 2009; Specia and Motta
2007) attempted to address the problem of tag disambiguation by use of community
detection. Starting from a particular tag, Au Yeung et al. (2009) derive several Social
Media networks, e.g. a network of documents that have been tagged with the partic-
ular tag by the same user, and the community detection method of Newman (2004b)
is applied to extract communities of tags or documents (that eventually lead to tags)
that correspond to the different senses of a tag. This approach was demonstrated to
yield superior performance compared to consulting some static external source of
information such as WordNet.
Specia and Motta (2007) employ tag clustering within a tag-to-ontology mapping
framework. They achieve this by first clustering tags through a heuristic scheme sim-
ilar to seed-based local community expansion (as described in Sect. 5.2) in order to
identify tag communities around pairs of tags that co-occur frequently with each other.
This tag clustering scheme is extended by a cluster post-processing step to remove
redundant clusters and the resulting clusters are used for mapping tags to ontology con-
cepts and relations, as well as to Wikipedia entries. Experimental results on Flickr and
123
(c)
(a) (d)
(e)
(b) (f)
Fig. 10 Examples of tag communities discovered by the local method of Papadopoulos et al. (2009a).
The presented communities were created using “computers”, “history”, “music”, “science”, “film” and
“animals” as seed nodes. For Fig. 9a, b, the tags are depicted. For the sake of brevity, only the network
structure is illustrated for the rest of the cases. Edges are weighted by co-occurrence frequency and only
those exceeding a certain threshold are drawn for clarity
123
delicious demonstrated the feasibility of this approach to semantify user-contributed

content.
6.2.3 User profiling
Personalized search and recommendation constitute an additional information retrieval

problem that can benefit from the use of community detection. More specifically,
clusters of tags have been demonstrated to act as effective proxies of users’ interests.
Gemmell et al. (2008) base the ranking performance of a personalized search mech-
anism on tag clusters outperforming conventional ranking schemes. The tag clusters
were extracted by use of a variant of hierarchical agglomerative clustering. However,
since this scheme requires manual parameter tuning that may have significant impact
on performance (as remarked by the authors), a viable alternative would be to use
some community detection to identify tag clusters.
Tsatsou et al. (2010) integrate the results of tag community detection in a personal-
ized ad recommendation system and compared against conventional nearest-neighbor
tag expansion schemes. More specifically, tags belonging to the same community are
used as a terminological description of semantic concepts within a domain ontology.
It was found that the use of tag communities discovered by a local community detec-
tion algorithm (Papadopoulos et al. 2009b) led to increased recall performance and
faster convergence of the personalized profiles. This is due to the fact that the rich
associations among tags that emerge through the identified community structure may
partly alleviate the vocabulary impedance problem that is known to trouble online
advertising (Ribeiro-Neto et al. 2005).
6.2.4 Photo clustering
The huge increase in the amounts of user contributed content in Social Media sharing
applications such as Flickr creates the need for automatically organizing content. For
instance, clustering the photos in such an application may help users to navigate larger
parts of the photo collection more efficiently (i.e. by looking at one representative
photo per cluster instead of all cluster photos). Moëllic et al. (2008) pursue photo
clustering by use of a shared nearest neighbors approach on two graphs of photos
where edges between photos are considered either by use of shared tags (tag-based
graph) or due to visual similarity (visual graph). The employed clustering technique
is shown to achieve improved clustering performance compared with conventional
clustering algorithms (k-means and one of its speeded-up variants). Also, comparing
the results of their methods with the clusters available from Flickr, the authors noted
similar clustering quality.
A more sophisticated application of graph clustering is presented by Li et al. (2008).
Their goal is to collect different representative (iconic) photos for popular landmarks
and use the massive visual content that is associated with them in order to create 3D
landmark models. They devise a multi-stage photo processing framework, in which
an important task is to group iconic photos together in order to reduce the amount
of photos that are processed by the computationally intensive 3D reconstruction step.
They achieve photo grouping by creating an iconic photo graph where photos are
123
Fig. 11 Five sample landmarks and events in Barcelona as identified by the cluster-based landmark and
event detection framework of Papadopoulos et al. (2010c). Thanks to the geotagging information of the
photos, the landmarks and events can be localized with high precision
connected by edges when they are visually similar (by use of SIFT descriptors) and by
applying the N -cut graph clustering algorithm by Shi and Malik (2000) on this graph.
In that way, they managed to reconstruct the major views of three famous landmarks
(Statue of Liberty, Notre Dame and San Marco).
Papadopoulos et al. (2010c) identify real-world landmarks and events in large
tagged photo collections by use of photo cluster classification. They apply the SCAN
algorithm (Xu et al. 2007) on a hybrid photo similarity graph that encodes both visual
and tag similarity between photos. Subsequently, the derived photo and tag communi-
ties are classified as landmarks or events based on cluster features similar to the ones
used by Quack et al. (2008). Figure 11 illustrates five sample landmarks and event
found for the city of Barcelona by use of this method. Manual inspection of the results
reveals that most of the clusters correspond either to famous landmarks of the city or
to real events (e.g. music concerts). Furthermore, the automatically selected cluster
tags provide meaningful descriptions for them.
6.2.5 Event detection
Events constitute an important unit of organization for Social Media content, since a
large part of user contributed content revolves around real-world events. Community
detection has found applications in the detection and tracking of events from social
text streams. For instance, the framework presented by Zhao et al. (2007) incorporates
textual, social and temporal aspects of blog feeds with the goal of tracking events.
The N -cut graph partitioning method of Shi and Malik (2000) is used twice in this
framework: once to cluster a graph of blog posts connected by their textual similar-
ity into topics, and at a second level, to cluster a graph of temporal activity profiles
123
among users (created by their comments) into communities that correspond to real-
world events. They tested their approach on a sample of 250 thousand posts from the
Dailykos political blog, from which they could extract well-recognizable real-world
events.
An even simpler approach for the extraction of events from blog streams is described
by Sayyadi et al. (2009). A keyword graph is built by extracting important keywords
(named entities, noun phrases) from text documents and associating them through
their co-occurrence. Then, the authors employ the divisive community detection ap-
proach of Girvan and Newman (2002) with a slight variation that enables overlap
between communities: keywords that are estimated to belong to many communities
are split into multiple vertices. They also use the efficient sample-based version of the
algorithm (Tyler et al. 2003) and observe that the derived communities are the same.
Subsequently, they use the extracted keyword communities as prototype vectors for
clustering the news articles. With the help of some additional cluster filtering and
merging, they end up with document clusters corresponding to real-world events.
7 Conclusions
7.1 Summary of main findings
The paper provided an overview of the emerging topic of community detection on

Social Media with a focus on scalability aspects and applications. Due to the huge
number of related works on the topic, the discussion focused on the most established
notions of community on networks and widely used methodological paradigms of
community detection, including their performance characteristics in terms of com-
plexity and memory usage. Furthermore, several strategies were presented that can be
used by community detection frameworks for scaling their applicability to real-world
datasets. Finally, the paper presented a series of observational studies and applications,
where the results of community detection have been exploited.
A first conclusion pertains to the very concept of community. There are numer-
ous network-based definitions and perceptions of community. In the context of Social
Media, communities comprise entities, such as online users, content items and meta-
data associated with content, that revolve around particular topics or events of social
interest. The association of such entities through their contextual co-occurrence leads
to the formation of Social Media networks. Analyzing the structure of such networks
leads to the discovery of Social Media communities.
Due to the graph partitioning heritage of community detection, many of the available
definitions are based on a global network perspective, i.e. they consider community
structure as a property of the whole network. However, a local perspective on the prob-
lem may be more meaningful in the context of Social Media, since one can only have
partial knowledge of the network under study. In addition, the possibility of overlap
among Social Media communities and weighted membership to them are especially
important attributes. Finally, the support for leaving noisy data out of the discov-
ered community structure and considering multiple levels of community organization
constitute significant aspects of the problem in this domain.
123
In addition, this survey acknowledged the remarkable proliferation of community

detection methods developed in the last years. Due to the diversity of research disci-
plines spawning the development of these methods, namely social network analysis,
computer science and statistical physics, there is no widely accepted terminology and
methodological classification of existing methods. For that reason, we attempted to
provide such a classification in Sect. 4.1. Despite the large number of community
detection methods abounding in literature, in practice most applications are limited to
a few seminal ones (or variants of them): the divisive scheme by Girvan and Newman
(2002), the modularity maximization method of Newman and Grivan (2004) and its
computationally efficient implementation (Clauset et al. 2004), the N -cut graph parti-
tioning scheme by Shi and Malik (2000), and spectral partitioning by use of the graph
Laplacian (Von Luxburg 2006). However, as it became clear from the performance
analysis of Sect. 4.2, these methods are not the most efficient, neither in terms of
complexity nor in terms of memory usage.
We believe that there are some under-exploited categories of community detection
methods. For instance, local density-based schemes, such as the SCAN algorithm
(Xu et al. 2007), are particularly important in the context of Social Media due to their
support for leaving spuriously connected vertices (i.e. noise) out of the detected com-
munity structure. In addition, iterative dynamic processes, such as the label propagation
algorithm by Raghavan et al. (2007) are also promising due to their computational effi-
ciency and their conceptual simplicity, which facilitates the development of method
extensions or adaptations that are catered for particular problems.
A final note concerns the relation of community detection to the emergence of core-
periphery structure in Social Media networks. While acknowledging the existence of
such structure from a macroscopic perspective, one should understand its implica-
tions on the exploitation of community detection in Social Media mining. Established
community detection methods that rely on the optimization of some quality measure
(modularity maximization, conductance/N -cut minimization) result in the extraction
of disproportionately large communities from the core of Social Media networks.
Such gigantic communities are of limited use within information retrieval tasks, such
as online content recommendation and retrieval. Therefore, we deem that appropriate
methods capable of extracting fine-grained community structure within such gigantic
components are the most relevant in the context of Social Media mining applications.
7.2 Outlook
Given the growing interest in the problem of community detection and the paramount
importance of Social Media for business and society in whole, it is only natural to
expect that exciting developments will take place on this field in the years to come.
We foresee significant progress on several methodological issues faced by existing
methods, but most importantly we expect that community detection results will be
further exploited in several online scenarios and will attain a sufficiently mature state
in order to be deployed in real-world Social Media applications.
On a methodological level, one should expect further refinement of methods to
cope with the increasing scale of Social Media data. Method parallelization and
123
fingerprinting or summarization techniques are deemed as promising strategies to

achieve scalability. Moreover, the development of new methods that will enable the
study of k-partite, multi-relational and hyper networks is likely to attract considerable
interest. Currently, there is a clear preference towards applying community detec-
tion on simple one-mode networks (even when it is necessary to employ some lossy
transformation on the original data) due to the conceptual and computational simplic-
ity of these methods. It remains to be seen whether such methods will be replaced
by more sophisticated that are applicable directly to the “raw” (implying k-partite,
multi-relational) form of Social Media networks.
In terms of application, one should expect that recent community detection meth-
ods will displace conventional clustering and partitioning schemes, such as k-means
and hierarchical agglomerative clustering, which are still very popular in a series of
domains, e.g text and photo clustering. In addition, as noted above, even within the
space of community detection methods, one should expect a shift in their adoption:
the currently established global optimization algorithms will give their place to local
density-based methods and iterative approximation schemes. Finally, one may look
forward to the wide adoption of dynamic community detection and tracking methods
in real-world contexts and applications. In conclusion, the forthcoming developments
in the field of community detection are expected to have a transformational impact on
mining and understanding the realm of Social Media.
Acknowledgment This work was supported by the WeKnowIt and GLOCAL projects, partially funded
by the European Commission, under contract numbers FP7-215453 and FP7-248984 respectively.
References
Agichtein E, Castillo C, Donato D, Gionis A, Mishne G (2008) Finding high-quality content in social media.
In: Proceedings of WSDM ’08: the international conference on Web Search and Web Data Mining,
Palo Alto, CA, USA, 11–12 Feb 2008. ACM, New York, pp 183–194
Andersen R, Chung FRK, Lang K (2006) Local graph partitioning using PageRank vectors. In: FOCS’06:
Proceedings of the 47th annual IEEE symposium on foundations of computer science, pp 475–486
Arenas A, Díaz-Guilera A, Pérez-Vicente CJ (2006) Synchronization reveals topological scales in complex
networks. Phys Rev Lett 96(11):114102
Arenas A, Duch J, Fernaández A, Gómez S (2007) Size reduction of complex networks preserving modu-
larity. New J Phys 9:176
Asur S, Parthasarathy S, Ucar D (2007) An event-based framework for characterizing the evolutionary
behavior of interaction graphs. In: Proceedings of the 13th ACM SIGKDD international conference
on knowledge discovery and data mining, San Jose, CA, USA, 12–15 Aug 2007. KDD ’07. ACM,
New York, pp 913–921
Au Yeung CM, Gibbins N, Shadbolt N (2009) Contextualising tags in collaborative tagging systems. In:
Proceedings of ACM conference on hypertext and hypermedia, pp 251–260
Baeza-Yates R (2007) Graphs from search engine queries. Theory and Practice of Computer Science
(SOFSEM), LNCS 4362. Springer, Harrachov1–8
Bagrow JP (2008) Evaluating local community methods in networks. J Stat Mech 5:P05001
Barber MJ (2007) Modularity and community detection in bipartite networks. Phys Rev E 76:066102
Batagelj V, Zaversnik M (2003) An O(m) algorithm for cores decomposition of networks. Eprint
arXiv:cs/0310049
Begelman G, Keller P, Smadja F (2006) Automated tag clustering: improving search and exploration in the
tag space. http://www.pui.ch/phred/automated_tag_clustering
Blondel VD, Guillaume J-L, Lambiotte R, Lefebvre E (2008) Fast unfolding of communities in large
networks. Eprint arXiv:0803.0476
123
Borgatti S, Everett M, Shirey P (1990) LS sets, lambda sets, and other cohesive subsets. Soc Netw 12:337–
358
Breiger R, Boorman S, Arabie P (1975) An algorithm for clustering relational data with applications to
social network analysis and comparison with multidimensional scaling. J Mathl Psychol 12:328–383
Bron C, Kerbosch J (1973) Algorithm 457: finding all cliques of an undirected graph. Commun ACM
16(9):575–577
Cattuto C, Benz D, Hotho A, Stumme G (2008a) Semantic grounding of tag relatedness in social book-
marking systems. In: Proceedings of ISWC 2008, Karlsruhe, Germany
Cattuto C, Baldassarri A, Servedio VDP, Loreto V (2008b) Emergent community structure in social tagging
systems. Adv Complex Syst (ACS) 11(4):597–608
Chakrabarti D (2004) Autopart: parameter-free graph partitioning and outlier detection. Lecture notes in
computer science 3202. Springer, pp 112–124
Chakrabarti D, Kumar R, Tomkins A (2006) Evolutionary clustering. In: Proceedings of the 12th ACM
SIGKDD international conference on knowledge discovery and data mining, Philadelphia, PA, USA,
20–23 Aug 2006. KDD ’06. ACM, New York, pp 554–560
Chen J, Zaiane OR, Goebel R (2009a) Local community identification in social networks. In: International
conference on advances in social networks analysis and mining (ASONAM), Athens, Greece
Chen J, Zaiane OR, Goebel R (2009b) A visual data mining approach to find overlapping communities in net-
works. In: International conference on advances in social networks analysis and mining (ASONAM),
Athens, Greece
Chi Y, Zhu S, Hino K, Gong Y, Zhang Y (2009) iOLAP: a framework for analyzing the internet, social
networks, and other networked data. Trans Multimed 11(3):372–382
Clauset A (2005) Finding local community structure in networks. Phys Rev E 72 026132
Clauset A, Newman MEJ, Moore C (2004) Finding community structure in very large networks. Phys Rev
E 70:066111
Danon L, Diaz-Guilera A, Duch J, Arenas A (2005) Comparing community structure identification. J Stat
Mech P09008. http://iopscience.iop.org/1742-5468/2005/09/P09008/
Dean J, Ghemawat S (2004) Mapreduce: simplified data processing on large clusters. In: Proceedings of
OSDI, 04, pp 137–150
Dhillon IS, Guan Y, Kulis B (2007) Weighted graph cuts without eigenvectors: a multilevel approach. IEEE
Trans Pattern Anal Mach Intell 29(11):1944–1957
Djidjev HN (2008) A scalable multilevel algorithm for graph clustering and community structure detection.
Lecture notes in computer science, vol 4936. Springer-Verlag, Berlin, pp 117–128
Donetti L, Munoz MA (2004) Detecting network communities: a new systematic and efficient algorithm.
J Stat Mech P10012. doi:10.1088/1742-5468/2004/10/P10012
Duch J, Arenas A (2005) Community detection in complex networks using extremal optimization. Phys
Rev E 72:027104
Falkowski T, Barth A, Spiliopoulou M (2007) DENGRAPH: a density-based community detection algo-
rithm. In: Proceedings of web intelligence 2007, pp 112–115
Fenn D, Porter M, McDonald M, Williams S, Johnson N, Jones N (2009) Dynamic communities in multi-
channel data: an application to the foreign exchange market during the 2007–2008 credit crisis. Eprint
arXiv:0811.3988
Flake GW, Lawrence S, Giles CL (2000) Efficient identification of Web communities. In: Proceedings of
KDD ’00, ACM, pp 150–160
Fortunato S (2009) Community detection in graphs. Eprint arXiv:0906.0612
Fortunato S (2010) Community detection in graphs. Phys Rep 486:75–174
Fortunato S, Castellano C (2007) Community structure in graphs. Eprint arXiv:0712.2716
Fortunato S, Latora V, Marchiori M (2004) Method to find community structures based on information
centrality. Phys Rev E 70:056104
Franke M, Geyer-Schulz A (2009) An update algorithm for restricted random walk clustering for dynamic
data sets. Adv Data Anal Classif 3(1):63–92
Gallo G, Grigoriadis MD, Tarjan RE (1989) A fast parametric maximum flow algorithm and applications.
SIAM J Comput 18(1):30–55
Gemmell J, Shepitsen A, Mobasher B, Burke R (2008) Personalizing navigation in folksonomies using
hierarchical tag clustering. In: Proceedings of DaWaK 2008, LNCS 5182, pp 196–205
123
Gibson D, Kumar R, Tomkins A (2005) Discovering large dense subgraphs in massive graphs. In: Proceed-
ings of the 31st international conference on very large data bases, Trondheim, Norway, Aug 30–Sept
2, 2005. Very Large Data Bases. VLDB Endowment, pp 721–732
Girvan M, Newman MEJ (2002) Community structure in social and biological networks. Proc Natl Acad
Sci USA 99(12):7821–7826
Gjoka M, Kurant M, Butts CT, Markopoulou A (2009) A walk in facebook: uniform sampling of users in
online social networks. Eprint arXiv:0906.0060
Gregory S (2009) Finding overlapping communities in networks by label propagation. Eprint arXiv:
0910.5516
Hastings MB (2006) Community detection as an inference problem. Phys Rev E 74:035102
Hübler C, Kriegel H, Borgwardt K, Ghahramani Z (2008) Metropolis algorithms for representative sub-
graph sampling. In: Proceedings of the 2008 eighth IEEE international conference on data mining,
Dec 15–19, 2008. ICDM. IEEE Computer Society, Washington, DC, pp 283–292
Hui P, Yoneki E, Chan SY, Crowcroft J (2007) Distributed community detection in delay tolerant net-
works. In: Proceedings of 2nd ACM/IEEE international workshop on mobility in the evolving internet
architecture, MobiArch ’07. ACM, pp 1–8
Ino H, Kudo M, Nakamura A (2005) Partitioning of Web graphs by community topology. In: Proceedings
of the 14th international conference on World Wide Web, Chiba, Japan 10–14 May 2005. WWW ’05.
ACM, New York, pp 661–669
Java A, Joshi A, Finin T (2008a) Detecting communities via simultaneous clustering of graphs and folkso-
nomies. In: Proceedings of WebKDD 2008, KDD workshop on web mining and web usage analysis,
Las Vegas, NV
Java A, Joshi A, Finin T (2008b) Approximating the community structure of the long tail. In: Proceedings
of the international conference on weblogs and social media
Kannan R, Vempala S, Vetta A (2004) On clusterings: good, bad and spectral. J ACM 51(3):497–515
Karypis G, Kumar V (1998) A fast and high quality multilevel scheme for partitioning irregular graphs.
SIAM J Sci Comput 20(1):359–392
Kim M, Han J (2009) A particle-and-density based evolutionary clustering method for dynamic networks.
Proc VLDB Endow 2(1):622–633
Kovács IA, Palotai R, Szalay MS, Csermely P (2010) Community landscapes: an integrative approach to
determine overlapping network module hierarchy, identify key nodes and predict network dynamics.
PLoS ONE 5(9):e12528
Kumar R, Raghavan P, Rajagopalan S, Tomkins A (1999) Trawling the Web for emerging cyber-commu-
nities. Comput Netw 31(11–16):1481–1493
Kumar SR, Raghavan P, Rajagopalan S, Sivakumar D, Tomkins A, Upfal E (2000) The web as a graph. In:
ACM symposium on principles of database systems, Dallas, Texas
Lancichinetti A, Fortunato S (2009) Community detection algorithms: a comparative analysis. Phys Rev E
80:056117
Lancichinetti A, Fortunato S, Radicchi F (2008) Benchmark graphs for testing community detection
algorithms. Phys Rev E 78:046110
Leskovec J, Faloutsos C (2006) Sampling from large graphs. In: Proceedings of the 12th ACM SIGKDD
international conference on knowledge discovery and data mining, Philadelphia, PA, USA, 20–23 Aug
2006. KDD ’06. ACM, New York, pp 631–636
Leskovec J, Lang K, Dasgupta A, Mahoney M (2008) Community structure in large networks: natural
cluster sizes and the absence of large well-defined clusters. Eprint arXiv:0810.1355
Leung IXY, Hui P, Lio P, Crowcroft J (2009) Towards real-time community detection in large networks.
Phys Rev E 79:066107
Li X, Wu C, Zach C, Lazebnik S, Frahm J (2008) Modeling and recognition of landmark image collections
using iconic scene graphs. Lecture notes in computer science, vol 5302. Springer-Verlag, Berlin, pp
427–440
Lin Y, Sundaram H, Chi Y, Tatemura J, Tseng BL (2007) Blog community discovery and evolution based
on mutual awareness expansion. In: Proceedings of the IEEE/WIC/ACM international conference on
web intelligence. IEEE Computer Society, Washington, DC, pp 48–56
Lin Y, Chi Y, Zhu S, Sundaram H, Tseng BL (2008) Facetnet: a framework for analyzing communities and
their evolutions in dynamic networks. In: Proceeding of the 17th international conference on World
Wide Web, Beijing, China, 21–25 April 2008. WWW ’08. ACM, New York, pp 685–694
123
Lin Y, Sun J, Castro P, Konuru R, Sundaram H, Kelliher A (2009) MetaFac: community discovery via
relational hypergraph factorization. In: Proceedings of KDD ’09. ACM, pp 527–536
Lorrain F, White H (1971) Structural equivalence of individuals in social networks. J Math Sociol 1:49–80
Luo F, Wang JZ, Promislow E (2006) Exploring local community structures in large networks. In: Proceed-
ings of web intelligence 2006. IEEE Computer Society, pp 233–239
Maiya AS, Berger-Wolf TY (2010) Sampling community structure. In: Proceedings of the 19th international
conference on World Wide Web, Raleigh, North Carolina, USA, 26–30 April 2010. WWW ’10. ACM,
New York, pp 701–710
Massen CP, Doye JPK (2005) Identifying “communities” within energy landscapes. Phys Rev E 71:046101
Mika P (2005) Ontologies are us: a unified model of social networks and semantics. In: Proceedings of
ISWC 2005. Springer, Berlin, pp 522–536
Moëllic P, Haugeard J, Pitel G (2008) Image clustering based on a shared nearest neighbors approach for
tagged collections. In: Proceedings of CIVR ’08, Niagara Falls, Canada, 7–9 July. ACM, New York,
pp 269–278
Newman MEJ (2004a) Fast algorithm for detecting community structure in networks. Phys Rev E 69:066133
Newman MEJ (2004b) Analysis of weighted networks. Phys Rev E 70:056131
Newman MEJ (2006) Finding community structure in networks using the eigenvectors of matrices. Phys
Rev E 74:036104
Newman MEJ, Girvan M (2004) Finding and evaluating community structure in networks. Phys Rev E
69:026113
Palla G, Derenyi I, Farkas I, Vicsek T (2005) Uncovering the overlapping community structure of complex
networks in nature and society. Nature 435(7043):814–818
Palla G, Barabasi A-L, Vicsek T (2007) Quantifying social group evolution. Nature 446:664–667
Papadopoulos S, Skusa A, Vakali A, Kompatsiaris Y, Wagner N (2009a) Bridge bounding: a local approach
for efficient community discovery in complex networks. Eprint arXiv: 0902.0871
Papadopoulos S, Kompatsiaris Y, Vakali A (2009b) Leveraging collective intelligence through community
detection in tag networks. In: Proceedings of CKCaR’09 workshop on collective knowledge capturing
and representation, Redondo Beach, California, USA
Papadopoulos S, Kompatsiaris Y, Vakali A (2010a) A graph-based clustering scheme for identifying related
tags in folksonomies. In: Proceedings of DaWaK’10, Bilbao, Spain. Springer-Verlag, pp 65–76
Papadopoulos S, Vakali A, Kompatsiaris Y (2010b) Community detection in collaborative tagging systems.
In: Pardede E (ed) Book community-built database: research and development. Springer, New York
Papadopoulos S, Zigkolis C, Kompatsiaris Y, Vakali A (2010c) Cluster-based landmark and event detection
on tagged photo collections. IEEE Multimed Mag 18(1):52–63
Pons P, Latapy M (2005) Computing communities in large networks using random walks. Computer and
Information Sciences—ISCIS 2005
Porter MA, Onnela JP, Mucha PJ (2009) Communities in networks. Not Am Math Soc 56(9):1082–1097
Quack T, Leibe B, Van Gool L (2008) World-scale mining of objects and events from community photo
collections. In: Proceedings of the 2008 international conference on content-based image and video
retrieval, Niagara Falls, Canada, 07–09 July 2008. CIVR ’08. ACM, New York, pp 47–56
Radicchi F, Castellano C, Cecconi F, Loreto V, Parisi D (2004) Defining and identifying communities in
networks. Proc Natl Acad Sci USA 101:2658–2663
Raghavan UN, Albert R, Kumara S (2007) Near linear time algorithm to detect community structures in
large-scale networks. Phys Rev E 76:036106
Reichardt J, Bornholdt S (2006) Statistical mechanics of community detection. Phys Rev E 74:016110
Ribeiro-Neto B, Cristo M, Golgher PB, Silva de Moura E (2005) Impedance coupling in content-targeted
advertising. In: Proceedings of the 28th annual international ACM SIGIR conference, Salvador, Brazil,
15–19 Aug. SIGIR ’05. ACM, New York, pp 496–503
Rosvall M, Bergstrom CT (2008) Maps of random walks on complex networks reveal community structure.
Proc Natl Acad Sci USA 105:1118–1123
Sayyadi H, Hurst M, Maykov A (2009) Event detection and tracking in social streams. In: Proceedings of
international AAAI conference on weblogs and social media. AAAI Press
Schaeffer SE (2007) Graph clustering. Comput Sci Rev 1(1):27–64
Schlitter N, Falkowski T (2009) Mining the dynamics of music preferences from a social networking site.
In: Proceedings of the international conference on advances in social network analysis and mining,
Athens, Greece
123
Schmitz C, Hotho A, Jäschke R, Stumme G (2006) Mining association rules in folksonomies. In: Proceed-
ings of the 10th IFCS conference on data science and classification, pp 261–270
Scott J (2000) Social network analysis: a handbook. Sage Publications Ltd, London
Scripps J, Tan P, Esfahanian A (2007) Node roles and community structure in networks. In: Proceedings
of the 9th WebKDD and 1st SNA-KDD 2007 workshop on web mining and social network analysis,
San Jose, CA, 12–12 Aug 2007. WebKDD/SNA-KDD ’07. ACM, New York, pp 26–35
Shi J, Malik J (2000) Normalized cuts and image segmentation. IEEE Trans Pattern Anal Mach Intell
22(8):888–905
Šíma J, Schaeffer SE (2006) On the NP-completeness of some graph cluster measures. In: Proceedings of
SOFSEM 2006: theory and practice of computer science, pp 530–537
Simpson E (2008) Clustering tags in enterprise and web folksonomies. Technical report HPL-2008-18
Specia L, Motta E (2007) Integrating folksonomies with the semantic web. Lecture notes in computer
science, vol 4519. Springer-Verlag, Berlin, pp 624–639
Sun J, Faloutsos C, Papadimitriou S, Yu PS (2007) GraphScope: parameter-free mining of large time-evolv-
ing graphs. In: Proceedings of KDD ’07. ACM, pp 687–696
Tang L, Liu H (2010) Graph mining applications to social network analysis. In: Aggarwal C, Wang H
Managing and mining graph data. Springer, New York
Tsatsou D, Papadopoulos S, Kompatsiaris I, Davis PC (2010) Distributed technologies for personalized
advertisement delivery. In: Hua X–S, Mei T, Hanjalic A (eds) Online multimedia advertising: tech-
niques and technologies. IGI Global, pp 233–261. http://www.igi-global.com/bookstore/chapter.aspx?
titleid=51963
Tyler JR, Wilkinson DM, Huberman BA (2003) Email as spectroscopy: automated discovery of community
structure within organizations. In: Huysman M, Wenger E, Wulf V (eds) Communities and technolo-
gies. Kluwer B.V., Deventer, pp 81–96
Van Dongen S (2000) Graph clustering by flow simulation. Ph.D. Thesis, Dutch National Research Institute
for Mathematics and Computer Science, Utrecht, Netherlands
Von Luxburg U (2006) A tutorial on spectral clustering. Technical report 149. Max Planck Institute for
Biological Cybernetics, August 2006
Vragović I, Louis E (2006) Network community structure and loop coefficient method. Phys Rev E
74:016105
Wang Y, Wu B, Du N (2008) Community evolution of social network: feature, algorithm and model. Eprint
arXiv: 0804.4356
Wasserman S, Faust K (1994) Social network analysis: methods and applications. Cambridge University
Press, Cambridge
Xu X, Yuruk N, Feng Z, Schweiger TA (2007) SCAN: a structural clustering algorithm for networks. In:
Proceedings of KDD ’07. ACM, pp 824–833
Yang B, Liu D-Y (2006) Force-based incremental algorithm for mining community structure in dynamic
network. J Comput Sci Technol 21(3):393–400
Yang S, Wang B, Zhao H, Wu B (2009) Efficient dense structure mining using MapReduce. In: Proceedings
of international conference on data mining workshops, pp 332–337
Ye S, Lang J, Wu F (2010) Crawling online social graphs. In: Proceedings of 12th international Asia-Pacific
web conference, APWeb 2010
Zakharov P (2006) Thermodynamic approach for community discovering within the complex networks:
LiveJournal study. Eprint arXiv:physics/0602063
Zhang Y, Wang J, Wang Y, Zhou L (2009) Parallel community detection on large networks with propinquity
dynamics. In: Proceedings of KDD ’09. ACM, pp 997–1006
Zhao Q, Mitra P, Chen B (2007) Temporal and information flow based event detection from social text
streams. In: Proceedings of the 22nd national conference on artificial intelligence, Vancouver, BC,
Canada, July 2007. AAAI Press, pp 1501–1506
123

Community Detection in Social Media: Performance and Application Considerations

Uploaded by

Copyright:

Available Formats

Community Detection in Social Media: Performance and Application Considerations

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Community Detection in Social Media: Performance and Application Considerations

Uploaded by

Copyright:

Available Formats

Data Min Knowl Disc (2012) 24:515–554

Community detection in Social Media

Symeon Papadopoulos · Yiannis Kompatsiaris ·

S. Papadopoulos (B) · Y. Kompatsiaris

S. Papadopoulos · A. Vakali · P. Spyridonos

requirements. It presents both a theoretical and an experimental comparative discus-

Keywords Community detection · Large-scale networks · Social Media

– Discuss the interpretation and exploitation of community detection results by

2.1 Elements of a Social Media network

2.2 Social Media network creation

of vertices and three-way relations among them. Subsequently, simpler (two-mode

3 Social Media communities

3.1 Classes of implicitly defined communities

3.1.1 Local definitions

3.1.2 Global definitions

Global community definitions consider community structure as a property of the whole

3.1.3 Process-based definitions

An alternative means of defining communities is by considering some community

3.2 Community attributes

Across all community definitions presented above, a set-based view of communities

(a) (b) (c)

4 Community detection methods

Community detection and graph partitioning. Graph partitioning is a well-specified

Community detection and graph clustering. Community detection is almost inter-

4.1 Methodological classification

Depending on the underlying methodological principle as well as the adopted def-

4.1.1 Cohesive subgraph discovery

4.1.2 Vertex clustering

4.1.3 Community quality optimization

4.2 Performance comparison

4.2.1 Theoretical discussion

Computational complexity. It is customary to quantify the computational complexity

Table 1 Comparison of community detection complexity

Method Complex-A Complex-B Scale

Cohesive substructure detection

the complexity of the modularity maximization method based on simulated anneal-

4.2.2 Experimental study

In order to gain insight into the performance aspects of community detection, we

11 This is equivalent to the number of edges of a network.

EXEC TIME (msec)

EXEC TIME (msec)

(a) (b) (c)

0.9 0.8 0.8

0.5 0.2 0.2

(d) (e) (f)

memory usage (KB)

4.3 Dynamic community detection

4.3.1 Longitudinal application on successive snapshots

4.3.2 Vertex-centric time-awareness

According to this approach, the dynamic character of community structure is attributed

4.3.3 Incremental application

This is a more sophisticated means of considering the dynamic character of networks.

5 Strategies for scaling community detection on Social Media scale

5.1 Sampling techniques

One possibility for reducing the complexity of community detection methods is to

5.2 Local graph processing

modularity (Luo et al. 2006) or node outwardness (Bagrow 2008), it is necessary to

5.3 Streaming/iterative schemes

convergence.13 A more complex example of iterative community detection is provided