A Non-Cooperative Game Model for Overlapping Community Detection in Social Networks

Due to its broad real-life application, overlapping community detection (in the realm of a social network) has attracted considerable interests from many researchers. However, current methods fail to reveal the full community structure and its formation process. Thus, here we present an effective and scalable overlapping community detection algorithm, which formulates the target problem as a non-cooperative game played by multiple social actors (players). Specifically, we allow each player to join multiple communities, and the strategy of each player is denoted by a community membership vector. The adopted utility function in our game integrates both the high-order proximity and the similarity between membership vectors of social actors. By properly weighting and rewiring the original social network, our approach can nicely enhance the global community structure by incorporating higher-order node proximities. Moreover, we formally prove that the proposed game resembles and matches how a potential game works (in the classical sense in game theory), indicating that the Nash equilibrium point must exist. To find such Nash equilibrium point, we use a stochastic gradientascent method to update the community membership vectors of players. Extensive experiments are conducted on both synthetic and real-world social networks. After comparing our method with six baseline methods, we obtain convincing results in terms of how well the methods reveal communities, as well as its scalability.