A Game-Theoretic Framework to Identify Overlapping

Communities in Social Networks

Wei Chen1 , Zhenming Liu2 , Xiaorui Sun3 , and Yajun Wang1

1
Microsoft Research Asia, Beijing, China. weic,[email protected]
2
Harvard School of Engineering and Applied Sciences. [email protected]
3
Department of Computer Science, Shanghai Jiao Tong University.
[email protected]

Abstract. In this paper, we introduce a game-theoretic framework to address

the community detection problem based on the structures of social networks. We
formulate the dynamics of community formation as a strategic game called com-
munity formation game: Given an underlying social graph, we assume that each
node is a selfish agent who selects communities to join or leave based on her own
utility measurement. A community structure can be interpreted as an equilibrium
of this game.
We formulate the agents’ utility by the combination of a gain function and a loss
function. We allow each agents to select multiple communities, which naturally
captures the concept of “overlapping communities”. We propose a gain function
based on the modularity concept introduced by Newman (2006), and a simple loss
function that reflects the intrinsic costs incurred when people join the communi-
ties. We conduct extensive experiments under this framework, and our results
show that our algorithm is effective in identifying overlapping communities, and
are often better then other algorithms we evaluated especially when many people
belong to multiple communities.
To the best of our knowledge, this is the first time the community detection prob-
lem is addressed by a game-theoretic framework that considers community for-
mation as the result of individual agents’ rational behaviors.

1 Introduction

Understanding the formation and evolution of communities is a long-standing research

topic in sociology because in part of its important relations with the studies of urban de-
velopment [9], criminology [22], social marketing [13, 10], and many other areas. With
recent increasing popularity of online social network services like Facebook and Twit-
ter, the studies of community structures become more important than ever. Among these
studies, identifying and detecting the communities are not only of particular prominence
but have immediate applications. For example, for effective online marketing, such as
placing online ads or deploying viral marketing strategies, identifying communities in
the social network could often lead to more accurate targeting and achieve better mar-
keting results. Albeit online user profiles or other semantic information is certainly
helpful to identify user segments and communities, this kind of information is often at a
coarse-grained level and overlooks information on communities at a more fine-grained
level. On the other hand, the structure of social networks reveals rich information over
the community structure, especially at the fine-grained level. In this paper, we study
community detection based on the network structure.
There exists extensive research work on detecting communities sheerly based on
network structures, for which we provide a more detailed introduction in the next sec-
tion. From these researches, however, we find a couple of issues that we feel are less
satisfactory. First, many community detection algorithms either set a global optimiza-
tion goal for detection, such as optimizing for modularity [17], betweenness [8], or
conductance [3], or has no optimization goal and just looking for sub-components with
certain predetermined structures (e.g. [21]). Neither of them is grounded with a system-
atic theory for the emergence of communities over a social network. Communities in
real social networks are certainly formed to serve a purpose, but they are also formed or-
ganically from bottom up, without a global authority trying to enforce a global optimal
objective. Second, a majority of the researches in community detection focus on parti-
tioning graphs into disjoint components and only a few of them accommodate overlap-
ping communities. In social networks, every individual typically belongs to more than
one communities, such as the community of her family members, that of her friends and
classmates, that of her co-workers, etc. Therefore, a satisfactory community detection
algorithm shall incorporate overlapping communities, and preferably do so in a natural
way.
In this paper, we address the above two issues by borrowing game-theoretic con-
cepts often used in economic researches. We model each node in a social network as
a rational agent trying to optimize its own utility by joining or leaving communities,
which we define as a community formation game. A Nash equilibrium of the game can
be readily interpreted into a community structure of the network: the communities every
node belongs to in a Nash equilibrium become the output of our community detection
algorithm. Due to computational constraints, we use local equilibria [1] instead of Nash
equilibria as our solutions. We believe that this game-theoretic framework reflects real-
world organic community formations, and thus in principle is a more systematic frame-
work than others that rely on global optimization objectives or have no objectives. This
framework also naturally incorporates overlapping communities, because we allow each
individual agent to select multiple communities, as long as it could improve her utility.
In our framework, the utility of each individual consists of two parts: a gain function
and a loss function. This is to match the real-world scenario, in which each individual
not only receives benefit from the communities she belongs to but also needs to pay cer-
tain cost to maintain her membership in the communities. To materialize our framework
into actual community detection algorithms, we address a number of issues, in partic-
ular the existence and the computation efficiency of Nash equilibria in our community
formation game.
For the existence of Nash equilibria, we show that in general Nash equilibria may
not exist, but we provide a sufficient condition to permit the existence of Nash equilib-
ria. More specifically, we define a family of functions called locally linear functions and
show that if all gain and loss functions of all the agents are locally linear, the game is
guaranteed to have a Nash equilibrium. For computation efficiency, we show that some
locally linear gain and loss functions will result in games for which the computation of
Nash equilibria belongs to hard computation classes such as NP-hard or PLS-complete
classes.
To deal with these computational issues, rather than computing a Nash equilibrium
of a game, we focus on finding local equilibria [1], which are the states such that no
agent can improve her own utility with a “small” deviation from her current strategy.
Notice that the Nash equilibria are local equilibria by definition. We are able to define
a set of gain and loss functions that are both locally linear and allow efficient computa-
tions of the local equilibria, when the agents’ strategy space is locally constrained.
Our gain function is based on a “personalized” version of the modularity [17], and
thus reflects the individual desire of forming a close-knit community structure. Our
proposed loss functions are based on the number of communities one need to maintain.
We conduct experiments on both synthetic and real networks to evaluate our com-
munity detection algorithm. It is not easy to find social networks with an associated
underlying communities as a ground truth, and thus we adopt the recently proposed syn-
thetic benchmark graphs of [11] to quantitatively evaluate our algorithm and compare
with two recently proposed detection algorithms that allow overlapping communities:
the clique percolation method [21] and the CONGA algorithm [8]. Our results show
that our algorithms performs quite well overall, and are usually better than the other
algorithms especially when more nodes belong to multiple communities. We also run
our algorithms on a couple of real networks, and show visually the results of our detec-
tion. Finally, as a proof of concept application, our algorithm is used to disambiguate
Chinese authors with the same English names in DBLP and our results look promising.
As a summary, our contribution is to propose the use of a game-theoretic framework
for the community detection problem and to provide an actual algorithm based on this
framework. Our experimental results demonstrate that our algorithm is very effective,
especially when dealing with the case where a large portion of people belong to multiple
communities. To the best of our knowledge, this is the first work that adopts a game-
theoretic approach in community detection problem. We believe that this approach more
naturally reflects real-world community formation in social networks and is a promising
direction worth further investigation and expansion.

1.1 Related work

The earliest result related to community detection in computer science is probably the
graph partition problem, which can date back to the design of VLSI [6] and modeling
roles and positions in social structure [24]. While these approaches are relevant to the
community detection, Newman pointed out a few facts that make the approaches unsuit-
able for detecting communities [17]. For instance, in a typical graph partitioning prob-
lem, the number of nodes to be partitioned and the number of groups to be partitioned
are usually known in advance, which is of great difference from the community detec-
tion problem we are considering. A second serious drawback of these graph-partition-
based methods is that all of them are essentially variations of finding partitions of a
graph so that the number of crossing edges between partitions are minimized. However,
small number of crossing edges alone may not be a good indication for communities
without considering the intrinsic connections among the nodes in the graph.
 Newman’s revolutionary notion of modularity is the first successful attempt to re-
solve the drawbacks specified above [17]. The modularity is defined on a partition of
the nodes in a graph. Let G = (V, E) be an undirected graph modeling a social network
with n nodes and m edges. Assume that each node v belongs to community cv . We
define the indicator function δ(cu , cv ) = 1 if and only if cu = cv , i.e., u, v are in the
same community, and otherwise δ(cu , cv ) = 0 for two nodes u, v ∈ V . The modularity
Q of this specific community partition is
( )
1 ∑ du dv
Q= Auv − δ(cu , cv ),
2m u,v 2m

where A is the adjacency matrix of G with Auv = 1 if (u, v) ∈ E and 0 otherwise,

and du , dv are degrees of u and v respectively. The above definition follows Clauset
et al.’s notation [4] and they also provided a detailed justification for defining modu-
larity in this way. The modularity measurement is basically calculating the number of
edges within the communities minus the expected number of such edges if we randomly
rewire the same number of edges. Clauset et al. believe that optimizing the modularity
corresponds to appropriately finding a family of communities. Although modularity
based algorithms have been proven to perform well in many real world data, this fam-
ily of approaches have several limitations. First, this approach implicitly assumes that
communities do not intersect with one another, which is usually not the case for real
world communities. Second, it is known that modularity based approaches suffer from
the resolution limits [7], i.e., they are favorable to merge small communities.
Researchers also start to pay attention to the model-based methods for community
detection. In this line of research, a generative model, depending on the underneath
community structure, for the social network is assumed. For example, in Copie et al.’s
work [5], they first assume that there is a fixed partition of the agents in the social
network, where each partition is a community. Two agents within the same community
is more likely to form an connection than those that are from different communities.
When the social network is presented, one may infer the most likely partition through
Bayesian law. One major question remains unsolved under this framework. Namely,
the choice of generative models adds one more level of uncertainty since it is always
difficult to validate the correctness of any specific model. Also, as we shall see, there is
a major distinction between model-based methods like Copie et al.’s work [5] and ours.
In model-based methods, the formation of the social network is usually the result of the
agents’ community structure while in our model, the agents choose to form community
structure after the social network is formed.
Recently, a noticeable body of work is focusing on discovering overlapping commu-
nity structures. In a pioneering work, Palla et al. [21] asserted that finding overlapping
communities is highly relevant to finding k-cliques in the social networks. Their algo-
rithm first finds all k-cliques in the network with a fixed constant k. Two k-cliques are
connected if they share k−1 nodes, and each community is formed by a maximum set
of connected k-cliques. Since one node may belong to multiple disconnected k-cliques,
the final communities are overlapping with each other. Palla et al. showed that their
algorithm has reasonable running time, despite the fact that computing cliques is hard.
However, this method requires the size of clique as an input, which is usually hard
to provide. Furthermore, if the social network is highly connected, the algorithm will
fail to uncover the underlying structure with any reasonable clique size. In fact, using
k-clique to find overlapping communities implicitly assumes that no two communities
share a k − 1 clique.
Lancichinetti et al. [11] proposed to iteratively compute a local community from a
node optimizing a fitness function, which is defined by the internal and external degrees
of the computed subgraph. By varying the parameter in the fitness function, they obtain
both overlapping and hierarchical community structures.
Another line of work in addressing the overlapping community is based on exten-
sions of the modularity. Nocosia et al. [18] introduces a vector of belonging factors for
each node in the graph, indicating the probability that this node belongs to a partic-
ular community. They then define a modularity measurement based on the belonging
factors. One particular drawback of their algorithm is that it requires the number of
communities as the input, which is not applicable in most cases. Furthermore, for each
node, the total belonging factors summarize to 1, which essentially denies the idea that
one node may just fit the communities as well as everyone else there.
The formation of communities was put in the game theoretic context by Athey and
Jha [2], in which they explicitly addressed the losses and gains of associating oneself
with a community. However, the social network underneath the agents in Athey and
Jha’s work is not in the picture. Their work is thus not comparable with ours.
The remaining of the paper is organized as follows. In Section 2, we formally define
our game-theoretic framework in community detection. We propose an algorithm to
simulate a local dynamic to reach local equilibria in Section 3. In Section 4, we conduct
experiments on various benchmark and real world networks. We discuss the benefits
and the limitation of our framework in Section 5.

2 A game-theoretic framework for community detection

As opposed to identifying the communities via attempting to globally optimize a certain

measurement defined, we propose that we shall interpret the social network naturally,
as a community formation game played by selfish agents. Every agent has her intrinsic
utility that associates with which communities she joins and which she does not. The
sheer goal every agent aims to achieve is to maximize her own utility. The formation
of communities is thus the joint result of each agent’s selfish decision. Interestingly,
the formation of communities in the social network is never studied from the game-
theoretic perspective.
The framework we are proposing is quite simple and natural. The social network is
given, i.e., each agent’s social interaction is fixed. There is an associated utility function
defined on each agent for the set of communities she participates. We decompose the
utility of the agent into two components. The first component is the gain, or pleasure,
of an agent joining the communities. The second component is the loss, or pain, associ-
ated with the agent’s action of joining the communities. It has been shown that joining a
community provides one with tremendous benefits, physical or emotional [23]. There-
fore, every agent shall have the incentive to join every single community unless there is
an associated loss to joint the communities. And indeed, it is not hard to see in the real
world that joining a community usually associates with certain loss, e.g., membership
fees. Finally, an agent’s utility is merely the difference between her gain function and
loss function.

2.1 Community formation game

In this section, we formally define our community formation game. We are given an un-
derlying static acquaintance graph G = (V, E), with n = |V | and m = |E|. We assume
G is undirected and unweighted though it is straightforward to generalize our results to
directed and weighted graphs. Depending on the context, an element in V may either
be called as an agent or a node. Each node chooses a collection of communities that
it wants to join. The set of all possible communities is denoted as [k] = {1, 2, . . . , k},
where k is polynomial in n (an exponential number of communities is both infeasible
to detect in theory and unrealistic in practice). Notice that our final community struc-
ture may have much smaller number of communities than k. The utility of the ith agent
vi ∈ V is measured by a gain function gi (·) and a loss function ℓi (·). Finally,
∑ for any
family of functions {f1 , f2 , ..., fn } defined over the agents, let f (·) = i∈[n] fi (·),
∑ ∑
e.g., g(·) = i∈[n] gi (·) and ℓ(·) = i∈[n] ℓi (·).

Strategy space and Nash equilibrium. In our community formation game, the strate-
gies of agent vi are subsets of communities that it wants to join, i.e, all subsets of [k].
We denote Li ⊆ [k] as a strategy of vi , which we also refer to as the community label
of vi . We allow Li = ∅, which means that i chooses to not belong to any community.
Define L = (L1 , L2 , . . . , Ln ) as a strategy profile, which is a vector of community
labels for all agents.
The utility of vi is measured by a gain function gi (·) and a loss function ℓi (·), which
map L to real numbers.4 Let the community labels of agents other than i to be L−i , and
we use (L−i , L′i ) to denote a strategy profile where the i-th entry of L is replaced by
L′i . We define the utility function for vi to be: ui (L) = gi (L) − ℓi (L).
In the community formation game, given the strategies of other agents L−i , the best
response strategy (or strategies) of agent vi is:
arg max
′
gi (L−i , L′i ) − ℓi (L−i , L′i ).
Li ⊆[k]

Definition 1 (Pure Nash equilibrium). Given graph G, the strategy profile L = (L1 , L2 ,
. . . , Ln ) forms a (pure) Nash equilibrium of the community formation game if all agents
are playing their best strategies, that is,
∀i and L′i ̸= Li , ui (L−i , L′i ) ≤ ui (L−i , Li ).
In other words, in a Nash equilibrium, no agent can improve her own utility by
changing her strategy unilaterally. One can interpret that each agent is satisfied with
her community selection at the state of a Nash equilibrium. Since each node may select
more than one community, the communities detected at the equilibrium naturally can
be overlapping with each other, which shall reflect what occurs in the real world.
4
A more appropriate notation should be giG (·) and ℓG i (·). However, since the underlying graph
G is static in our game, it is simpler to omit the superscript on G.
Existence and computation of Nash equilibria. In general, Nash equilibria may not
exist in a community formation game. To see this, one can easily formulate a “matching
pennies” game [20] in the community formation game, in which one node u always
prefer to be with another node v in the same community while v always prefer not to
be in the same community as u. It is thus interesting to know when a Nash equilibrium
exists in the community formation game.
Let us recall that potential games are a general class of games that permit pure Nash
equilibria [19]. In a potential game, there is an associated potential function Φ(·) defined
on the strategy profiles of the agents. A community formation game is a potential game
if Φ(L) − Φ(L−i , L′i ) = ui (L−i , L′i ) − ui (L) for every strategy profile L and every
strategy L′i of vi . In other words, if an agent changes her strategy to improve her own
utility, the potential function strictly decreases with the same amount as the increase
of the agent’s utility. In any potential game that contains a finite number of strategy
profiles, Nash equilibria always exist. Furthermore, every better response dynamic, in
which each agent sequentially changes her strategy to improve her own utility, will
converge to a Nash equilibrium.
Next, we provide a sufficient condition to make a community formation game poten-
tial, and thus address the existence of Nash equilibria for community detection purpose.

Definition 2. A set of functions {fi (·) : 1 ≤ i ≤ n} is locally linear with linear factor
ρ if for every strategy profile L and every strategy L′i of vi , the following relation holds:

∀i ∈ [n], fi (L−i , L′i ) − fi (L) = ρ (f (L−i , L′i ) − f (L)) .

We show that if the gain and loss functions in a community formation game are
locally linear, the game is a potential game.

Theorem 1. Let {gi (·) : i ∈ [n]} and {ℓi (·) : i ∈ [n]}) be the sets of gain and
loss functions of a community formation game. If {gi (·)} and {ℓi (·)}) are locally linear
functions with linear factor ρg and ρℓ , then the community formation game is a potential
game.

Proof. We define a potential function as Φ(L) = ρℓ · ℓ(L) − ρg · g(L). Now consider

agent vi who changes her strategy from Li to L′i . From the definitions of locally linear
functions and the utility functions ui (·), we have Φ(L) − Φ(L−i , L′i ) = ui (L−i , L′i ) −
ui (L). Therefore, the community formation game is a potential game. 2

With locally linear gain and loss functions, we can guarantee the existence of Nash
equilibria, but it is not necessarily true that finding a Nash equilibrium is easy. The
follow lemma indicates that computing a Nash equilibrium could be hard. Due to space
limitation, we omit the proof of the lemma.

Lemma 1. There exists a community formation game, in which the sets of gain and loss
functions are locally linear, such that both computing the best response for an individual
agent and computing a Nash Equilibrium in the game are NP-hard.
Gain and loss functions. We now propose a set of gain and loss functions. These gain
and loss functions have natural economic interpretations and they can be computed ef-
ficiently. Additionally, our experiments also demonstrate that the equilibria using these
gain and loss functions provide pretty accurate information regarding the community
formation.
The gain function we use here is a generalized version of the well accepted mod-
ularity function, which is adapted to fit the scenario that one node may participate in
multiple communities. We define δ̂(i, j) = 1 if |Li ∩Lj | ≥ 1 and δ̂(i, j) = 0 otherwise.
Let A be the adjacency matrix of graph G.

Definition 3 (Personalized modularity function). The personalized modularity func-

tion defined for the i-th agent is:
( )
1 ∑ di dj
Qi (L) = Aij δ̂(i, j) − · |Li ∩ Lj | .
2m 2m
j∈[n]

Similar to the original modularity, the personalized modularity function measures

the number of edges from this agent to the community comparing with such number of
edges if all edges within the community are randomly sampled according to the degree
constrains on the agents. In particular, it describes how well this agent fits the commu-
nity comparing with a random community. Note that the original modularity can be ob-
tained by summing over the personal modularity functions for all agents, under the orig-
inal non-overlapping condition of |Li | = 1 for all i ∈ [n]. We use |Li ∩ Lj | instead of
∑ ∑ ∑
δ̂(i, j) in the second term so that we have j∈[n] dj ·|Li ∩Lj | = c∈Li j∈[n],c∈Lj dj ,
which allows fast local computation in our algorithm (see Section 3).
We use a simple loss function to model the aspect that an agent may suffer by joining
new communities. This, in particular, reflects some fixed cost associated in joining a
new community. In real world, it may connect to the membership fee or some cost to
attend community events.

Definition 4 (Linear loss function). Let c > 0 be a constant. The loss of a node vi
with the linear loss function is (|Li | − 1) · c.

It is easy to verify that both the personalized modularity function and the linear
loss function are locally linear functions, with linear factor 1/2 and 1, respectively.
Therefore, we have the following result.

Theorem 2. Let gi (L) be the personalized modularity function and ℓi (L) be a linear
loss function. The community formation game has a Nash equilibrium.

3 Local equilibrium and a simple algorithm

We have shown that if the set of gain and loss functions are locally linear, there always
exists a Nash equilibrium in our community formation game (Theorem 1). However,
computing the best response might be hard in some simple cases from Lemma 1. It is
thus not reasonable to assume that individuals always make the best response. Instead,
Algorithm 1 LocalEquilibrium(G)
1: initialize each node to a singleton community
2: repeat the following process until no node can improve itself
3: randomly pick a node vi , and perform the best operation among join, leave and switch

we propose that an agent will only choose a strategy from a restricted space that de-
pends on her current state when she needs to respond to the other agents’ strategies. In
particular, an agent can only locally implement the following three operations,

1. join. Agent vi joins a new community on top of the communities she joins by adding
a new label in Li .
2. leave. Agent vi leaves a community she is in by removing a label from Li .
3. switch. Agent vi switches from one community to another by replacing a label in
Li .

In the restricted strategy spaces, an equilibrium is a state where no agent can deviate
from her current strategy within the locally allowed strategy space. Such kind of equi-
libria are referred as local equilibria [1] in the literature. In the community formation
game, the entire strategy space of agent i is S = 2[k] . For each agent i with the current
community label set Li , we use ls(Li ) to denote i’s local strategy space, which is the
set of possible label sets we could obtain by applying one of the operations join, leave
and switch once on Li . The local equilibrium is defined as follows:
Definition 5 (Local equilibrium). 5 Given G, the strategy profile L = (L1 , L2 , . . . , Ln )
forms a local equilibrium of the community formation game if all agents are playing
their local optimal strategies, that is,

∀i and L′i ∈ ls(Li ), ui (L−i , L′i ) ≤ ui (L−i , Li ).

Different from Nash equilibrium, at local equilibrium, the utility of each agent is
achieving a local maximum instead of a global one. This is useful when the local strat-
egy space is easy to explore, while computing a global optimal solution is not feasible.
Restricting strategy space to our local strategy space is further justified by the fact that
in real world individuals usually consider joining or leaving one community at a time.
In our setting, computing the local best response in the local strategy space is
polynomial-time, by simply enumerating all O(k) possible join, leave and switch op-
erations. We show in the following lemma that for the case of using our personalized
modularity gain function and linear loss function, the computation can be made efficient
by maintaining a quantity for each community, and by only checking new communities
to add or switch to from the set of communities of one’s neighbors.
Lemma 2. Let ∆i be the degree of the node vi and Ni be the set of neighbors of vi in
the graph G. Let gi (L) be the personalized modularity function and ℓi (L) be a linear
loss function. The time complexity to find the best local operation for agent i in join,
leave and switch is O(|Li | · |L(Ni )| · ∆i ), where L(Ni ) = ∪j∈Ni Li .
5
The local equilibrium in [1] is defined on Euclidean strategy spaces.
Proof. It is sufficient to show that the function Qi (L) can be efficiently updated. For
∑ dj
each community U , we maintain a quantity Q̂U = vj ∈U 2m . Notice that if only one
member leaves U or joins U in each step, Q̂U can be updated in constant time.
For the join operation, if vi is joining a community not in L(Ni ), Qi (L) as well
as ui (L) will be strictly decreasing. Therefore, vi only considers the communities in
L(Ni ) to join. For the switch operation, the same argument applies, except that vi may
have a possible gain if she switches to a brand new community U ′ , but in this case vi
can gain the same utility by simply removing the old community. Therefore vi only
needs to consider communities in L(Ni ).
In joining a new community, Qi (L) can be updated in O(∆i ) time. This is because
∑ di dj
the term − j∈U 2m can be computed in constant time given the maintained Q̂U . The
switch operation will consider leaving |Li | communities and joining |L(Ni )| commu-
nities. Therefore, the total running time is O(|Li | · |L(Ni )| · ∆i ). 2
The fast computation of the local best operation is important in our experiment. This
is in particular a reason we propose to use |Li ∩ Lj | in defining the personalized modu-
larity function. Next, we use the simple algorithm LocalEquilibrium(G) (Algorithm 1)
to compute a local equilibrium, which essentially simulates the best local response dy-
namic. Although in general the local response dynamic may take long time to converge,
we show below that for the case of the personalized modularity function and the linear
loss function, the convergence is fast if we set the parameter c in the linear loss function
appropriately.
Theorem 3. Let gi (L) be the personalized modularity function and ℓi (L) = c(|Li |−1)
be a linear loss function with constant c satisfying 4cm2 is an integer. LocalEquilib-
rium takes at most O(m2 ) steps to reach a local equilibrium.
∑
Proof. Define∑the potential function Φ(L) = ℓ(L) − g(L)/2, where ℓ(L) = i ℓi (L)
and g(L) = i gi (L). Notice that Φ(L) ≥ −1/2 since g(L) ≤ 1 and ℓ(L) ≥ 0.
Now consider 4m2 ·Φ(L). It is straightforward to verify that 4m2 ·Φ(L) is an integer
function. Notice that we have Φ(L) = 0 in the initial state of LocalEquilibrium, and
Φ(L) ≥ −1/2 during the algorithm execution. Since 4m2 · Φ(L) strictly decreases in
each round, Φ(L) will achieve minimum in at most 2m2 steps. 2
Our LocalEquilibrium algorithm uses the initial configuration in which every agent
has one community of her own. One reason we choose this starting point instead of
other possible ones, such as all nodes sharing one single community, is that with this
starting point most of the agents’ activities will be joining communities, which is likely
to be an individual decision. In contrast, when people share one community and the
community evolves by splitting, the splitting decision is usually a collective decision
made by a group of people, which is not modeled in our current game-theoretic frame-
work and it would be computationally expensive to incorporate group decisions. Our
experimental results show that this starting point indeed leads to the extraction of rea-
sonable community structures. Additionally, if we already have some partial knowledge
on who are in the same communities, we can choose it as our starting point and do not
allow the local dynamic to change it, and thus community learning can be naturally
incorporated in our framework as well. This is a future direction we will pursue next.
4 Experiments
It is very hard to obtain ground truth community information from real-world networks.
Therefore, detecting communities sometimes is more of an art than a science [5]. Never-
theless, we conduct experiments on some real-world networks as well as several bench-
mark graphs. In the experiment, our algorithm uses the personalized modularity func-
1
tions as the gain functions and the simple loss functions with the loss factor c = m .

4.1 Real world graphs

We run our algorithm on two classical graphs: The Dolphin Network [12] and The
Zachary’s Karate Club [25]. We present our decomposition of the communities in Fig-
ure 1. In particular, we believe that our algorithm finds richer overlapping structures
compared with previous studies [17, 18]. For example, Newman uses modularity to
partition the same karate club network into two components [17], which corresponds
to two upper overlapping communities and the four lower communities we discovered
in Figure 1 (b). Thus, our community structure is a strict refinement of the community
structure discovered in Newman’s work. Our refined structure allows overlapping com-
munities, which by visual inspection does provide meaningful information about the
community structures.
For some small communities detected by our algorithm (e.g. nodes 47 and 50 in the
upper-left corner of Figure 1 (a)) could be merged with their neighbor communities if
we consider joint decisions of all community members. Considering joint decisions is
more computational intensitive, however, and we defer it as a future research item.
Our conclusion from this experiment is that our algorithm is able to discover finer
grained communities than previous approaches. These finer grained communities re-
veals insightful community structures in the network and could be used in further in-
vestigation of community interconnections and larger community structures.

4.2 Benchmark graphs

Recently, Lancichinetti and Fortunato [11] proposed a set of benchmark graphs to eval-
uate the performance of the community detection algorithms that can discover overlap-
ping communities. The evaluation metric used is the “normalized mutual information”
between the recovered community structure and the underlying ground truth. Though it
could be rather hard to justify how realistic these benchmark graphs and the evaluation
metric are, we show that our method actually performs pretty well in the benchmark
graphs.
In this data set, we compare our algorithm with the clique percolation method [21]
and the CONGA algorithm [8]. The clique percolation algorithm requires the size of
the cliques as input. We run the algorithm with different clique sizes ranging from 2
to 6. The results presented in Figure 2 are obtained by taking the optimal value for
the clique sizes. The CONGA algorithm on the other hand, requires the number of
communities as input. Although this information is available here, in general, we do
not know the number of communities. Another disadvantage of CONGA is its slow
running time. This is due to the computation of the “betweenness” values at each node,
 (a) (b)

Fig. 1. The communities discovered for the dolphin network (left). The communities discovered
for the Zachary’s karate club (right).

which essentially has to compute all pairs of shortest paths. In particular, we are only
able to obtain results of CONGA for the graphs with 1,000 nodes.
In Figure 2, we show the results of different algorithms on the benchmark graphs
with overlapping communities. The networks used to produce the left figure consist of
1,000 nodes, whereas those of the right figure consist of 5,000 nodes. In each figure,
the community sizes of the two upper diagrams range between smin = 10 and smax =
50, and the community sizes of two bottom diagrams range between smin = 20 and
smax = 100. The mixing parameter, i.e., the portion of crossing edges, µ is 0.1 for
two left diagrams and 0.3 for two right diagrams. The other parameters are τ1 = 2,
τ2 = 1, kavg = 20, kmax = 50 and om = 2. The x-axis represents the portion of
nodes that belong to multiple communities. The graphs are generated in the following
steps: (1) generate the number of communities that each node will belong to; assign
the degrees to the nodes based on a power law distribution with exponent τ1 ; (2) assign
the community sizes from another power law distribution with exponent τ2 for a fixed
number of communities; (3) generate the bipartite graph between the nodes and the
communities with the configuration model [14]; (4) for each node, assign the cross-
community degree and internal degrees within each community based on µ; (5) build
the graphs for each community and the cross-community edges with the configuration
model. For more detailed steps on how the graphs are generated and refined, readers
can refer to [11].
As shown in Figure 2, our algorithm for the game-theoretic framework performs
better for the bottom two networks. In other words, our algorithm works better on
 1 1 1
0.99
0.9
0.98
0.5 0.5
Normalized mutual information

Normalized mutual information

0.97
smin=10 0.8 smin=10
µ=0.1 µ=0.3 smax=50 0.96 µ=0.1 µ=0.3 smax=50
0 0 0.95 0.7
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5

1 1 1 1

0.5 0.5 0.5 0.5

Game
Clique smin=20 Game smin=20
CONGA µ=0.1 µ=0.3 s =100 Clique µ=0.1 µ=0.3 s =100
max max
0 0 0 0
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
Fraction of Overlapping Nodes Fraction of Overlapping Nodes

Fig. 2. The network on the left consists of 1,000 nodes. The minimum degree and maximum
degree of the network are 10 and 50 respectively. The network on the right consists of 5,000
nodes. The minimum degree and maximum degree of the network are 10 and 50 respectively.

graphs with larger communities. Notice that the community sizes in the bottom two
graphs are within 20 and 100, while the upper two graphs have communities with size
10 to 50.
Compared with the clique percolation algorithm, both our algorithm and the clique
percolation algorithm perform very well on the two upper left networks, with mu-
tual information being above 90%. For the two upper right networks in Figure 2, the
clique percolation algorithm outperforms our algorithm when the fraction of overlap-
ping nodes is small. However, our algorithm is more stable than the clique percolation
algorithm over all instances. The performance of the clique percolation algorithm drops
significantly when the portion of overlapping nodes increases. In particular, when half
nodes belong to multiple communities (at the point 0.5 on the x-axis), the performance
of our algorithm actually is equally good to the clique percolation algorithm for graphs
with 1,000 nodes, and performs better on graphs with 5,000 nodes.
Compared with CONGA algorithm, our algorithm is better for µ = 0.1, and is not
as good as CONGA for µ = 0.3. Again the performance of our algorithm is more stable
than CONGA, and is better when the fraction of overlapping nodes is large. Also notice
that CONGA is not able to finish in reasonable time for graphs with 5000 nodes.

4.3 Identifying duplicated names in DBLP

Finally, we provide an application scenario of the game-theoretic based community
detection algorithms, which suggests our community-detection approach may extend
well beyond the notion of “communities” defined in traditional ways.
The study of co-authorship network in academic community has attracted much at-
tention recently [15, 16]. Examples of co-authorship databases include Microsoft Aca-
demic Search and the DBLP computer science bibliography. Extracting the co-authorship
graph from the existing databases like DBLP sometimes could be challenging. For ex-
ample, different scholars with the same name may be naively viewed as a single person.
While people using Indo-European languages do not often experience this problem,
people from countries like China, Japan, or Korea, whose names in Microsoft Academic
or DBLP are indeed Romanization of their original names, usually find the problem of
having duplicated names persistent.
 Therefore, it is interesting to find a way to distinguish different people with the
same name in academic community when relevant data, say, records from DBLP, are
presented. Let us now present one possible way that serves as the first step to tackle this
problem using the game-theoretic based algorithms we developed. First, we construct
a co-authorship graph based on DBLP entries. A node in the graph corresponds to a
name, and one node may represent more than one person in the real world. Two nodes
are linked by an undirected edge if the corresponding names of these two nodes ever
coauthored at least one paper. Next, in this co-authorship graph, each node is asked to
play the community-formation game using the personalized modularity gain functions
and linear cost functions until a local equilibrium is reached. One would naturally ex-
pect that even when people with the same name collapse into one node, the node will
join multiple communities in the game because these people with the same name shall
belong to different communities. Based on this observation, a node may be partitioned
into multiple entities and an individual represented by this node may correspond to one
or more entities.
We provide one instance of experiment which demonstrates our observation is rea-
sonably accurate. Our experiment searches for the node with name “Wei Chen” in the
co-authorship graph, which in fact represents more than 20 individuals that have pub-
lished in total more than 200 papers in computer science or relevant areas. We use only
a subgraph of the co-authorship graph that contains 20,000 nodes because processing
the whole graph would otherwise be too computationally intensive. The subgraph is
obtained by using breadth first search from the node “Wei Chen” until 20,000 nodes are
discovered. The data is for publications until the end of year 2008.We specifically focus
on two “Wei Chen”s: one is the first author of this paper from Microsoft Research Asia
(MSRA) and the other is a faculty member in Zhejiang University (ZJU).
Our algorithm discovers more than 40 communities containing “Wei Chen”. In most
communities, we find that the set of co-authors are relevant to a particular “Wei Chen”.
For example, we find four communities that are relevant to the two “Wei Chen”s we
mentioned above. Table 1 summarizes the four sets of co-authors of two “Wei Chen”s
in the four communities.

Wei Chen Jialin Zhang, Chao Jin, Zheng Zhang, Likun Liu, Shiding Lin, Ming Chen,
(MSRA) Shaomei Wu, Yu Chen, Qiao Lian, Ben Y. Zhao, Xuezheng Liu
Marcos Kawazoe Aguilera, Sam Toueg
William M. Andrews, Aidong Lu, David S. Ebert, Mario Costa Sousa,
Ross Maciejewski, Tobias Isenberg
Zhongding Jiang, Yi Gong, Yu Guan, Jin Wang, Yingchao Zhao, Chunxiao Liu,
Zi’ang Ding, Guofeng Zhang, Yingzhen Yang, Ling Zhuang, Hongxin Zhang,
Wei Chen Chengfang Song, Huafeng Liu, Huagen Wan, Luying Li, Hujun Bao, Xiao Liang,
(ZJU) Qunsheng Peng, Qifeng Tan, Pengcheng Shi, Yubo Zhang, Shang-Hua Teng,
Lincan Zou, Xiaobo An,Xueying Qin, Long Zhang, Yinan Fan, Dong Xu,
Yun Zeng, Wei Hua, Zhao Dong,
Table 1. The partition of the co-authors of two “Wei Chen”s
 The first “Wei Chen” is the first author of this paper. The co-authors of him are
splitted into two communities. One related to his research collaborators after he joins
Microsoft Research Asia, and the other are his collaborators back when he was at Cor-
nell.
The second “Wei Chen” is a faculty member in Zhejiang University. The first group
of his collaborators represents his connection in Purdue university. The second group of
the co-authors is his colleagues in Zhejiang University, with the exception of “Shang-
Hua Teng” and “Yingchao Zhao”. These two authors are actually the co-authors of “Wei
Chen (MSRA)”. A reason to explain the misclassification is that Teng and Zhao only
co-authored one paper with “Wei Chen (MSRA)” (by the end of 2008) and they did not
collaborate with any other “Wei Chen (MSRA)”s co-authors. On the other hand, Teng
had collaboration with “Harry Shum” that has strong connections with the graphics
researchers, one of whom is “Wei Chen (ZJU)”. In this respect, it is actually hard to say
that it was a “misclassification” since the two authors only collaborate with “Wei Chen
(MSRA)” once.

5 Conclusion
We propose for the first time a game-theoretic framework to detect community struc-
tures in social networks. This formulation intuitively matches the dynamic formation of
communities in real world scenarios. Furthermore, since we do not require each agent
to join exactly one community, the resulting community structure natually incorporates
overlapping communities.
Our experiment shows that, even with simple ulitity functions defined on the agents,
our method is effective in discovering overlapping communities in several benchmark
graphs and real world networks. Since the algorithm we use to find local equilibrium
only implements local operations, the running time is fast and the algorithm can fit into
a parallel framework.
There remain many interesting open problems under this framework. One direction
is to find more appropriate gain and loss functions. The proposed ones in this paper,
though simple and effective, are by no means the best choices for the community for-
mation games. In particular, we believe better gain and loss functions can be obtained by
deeper understanding of the community formation process in the real world networks.

6 Acknowledgement
The authors thank Prof. Wei Chen from State Key Lab of CAD&CG, Zhejiang Univer-
sity for commenting on the partition produced by our algorithm.

References
1. C. Alós-Ferrer and A. Ania. Local equilibria in economic games. Economics Letters,
70(2):165–173, 2001.
2. S. Athey and S. Jha. A theory of community formation and social hierarchy. working paper,
2006.
 3. U. Brandes and T. Erlebach. Network Analysis: methodological foundations. Springer Ver-
lag, 2005.
4. A. Clauset, M. E. J. Newman, and C. Moore. Finding Community Structure in Very Large
Networks. Phys. Rev. E, 70(6):066111, Dec 2004.
5. J. Copic, M. O. Jackson, and A. Kirman. Identifying Community Structures from Network
Data via Maximum Likelihood Methods. The B.E. Journal of Theoretical Economics, 9,
2009. working paper.
6. P. O. Fjällström. Algorithms for graph partitioning: A Survey. In Linkoping Electronic
Atricles in Computer and Information Science, 3., 1998.
7. S. Fortunato and M. Barthélemy. Resolution limit in community detection. Proceedings of
the National Academy of Sciences, 104(1):36, 2007.
8. S. Gregory. A fast algorithm to find overlapping communities in networks. In ECML/PKDD.
Springer, 2008.
9. J. D. Kasarda and M. Janowitz. Community Attachment in Mass Society. American Socio-
logical Review, 39(3):328–339, 1974.
10. P. Kotler and G. Zaltman. Social Marketing: An Approach to Planned Social Change. The
Journal of Marketing, 35(3):3–12, 1971.
11. A. Lancichinetti and S. Fortunato. Benchmarks for testing community detection algo-
rithms on directed and weighted graphs with overlapping communities. Physical Review
E, 80(1):16118, 2009.
12. D. Lusseau. The emergent properties of a dolphin social network. Proceedings: Biological
Sciences, 270:S186–S188, 2003.
13. D. McKenzie-Mohr and W. Smith. Fostering Sustainable Behavior: An Introduction to
Community-Based Social Marketing. New Society Publishers, 1999.
14. M. Molloy and B. Reed. A critical point for random graphs with a given degree sequence.
Random Structures and Algorithms, 6(2-3):161–180, 1995.
15. M. E. J. Newman. Coauthorship networks and patterns of scientific collaboration. Pro-
ceedings of the National Academy of Sciences of the United States of America, 101(Suppl
1):5200–5205, 2004.
16. M. E. J. Newman. Who Is the Best Connected Scientist?A Study of Scientific Coauthorship
Networks. Complex Networks, 650:337–370, 2004.
17. M. E. J. Newman. Modularity and community structure in networks. Proceedings of the
National Academy of Sciences, 103(23):8577–8582, 2006.
18. V. Nicosia, G. Mangioni, V. Carchiolo, and M. Malgeri. Extending the definition of modu-
larity to directed graphs with overlapping communities. J. Stat. Mech, 3024, 2009.
19. N. Nisan, T. Roughgarden, É. Tardos, and V. V. Vazirani. Algorithmic game theory. Cam-
bridge University Press, 2007.
20. M. J. Osborne and A. Rubinstein. A Course in Game Theory. MIT Press, 1994.
21. G. Palla, I. Derényi, I. Farkas, and T. Vicsek. Uncovering the overlapping community struc-
ture of complex networks in nature and society. Nature, 435(7043):814, 2005.
22. R. J. Sampson and W. B. Groves. Community Structure and Crime: Testing Social-
Disorganization Theory. American Journal of Sociology, 94(4):774, 1989.
23. S. Sarason. The Psychological Sense of Community. Jossey-Bass, 1974.
24. H. C. White, S. A. Boorman, and R. L. Breiger. Social Structure from Multiple Networks. I.
Blockmodels of Roles and Positions. American Journal of Sociology, 81(4):730, 1976.
25. W. W. Zachary. An information flow model for conflict and fission in small groups. Journal
of Anthropological Research, 33(4):452–473, 1977.