A Novel Label Propagation Method for Community Detection Based on Game Theory

Abstract

Community is a mesoscopic feature of the multi-scale phenomenon of complex networks, which is the bridge to revealing the formation and evolution of complex networks. Due to high computational efficiency, label propagation becomes a topic of considerable interest within community detection, but its randomness yet produces serious fluctuations. Facing the inherent flaws of label propagation, this paper proposes a series of solutions. Firstly, this paper presents a heuristic label propagation algorithm named Label Propagation Algorithm use Cliques and Weight (LPA-CW). In this algorithm, labels are expanded from seeds and propagated based on node linkage index. Seeds are produced from complete subgraph, and node linkage index is related to neighboring nodes. This method can produce competitive modularity Q but not Normalized Mutual Information (NMI), and compensate with existing methods, such as Stepping Community Detection Algorithm based on Label Propagation and Similarity (LPA-S). Secondly, in order to combine the advantages of different algorithms, this paper introduces a game theory framework, design the profit function of the participant algorithms to attain Nash equilibrium, and build an algorithm integration model for community detection (IA-GT). Thirdly, based on the above model, this presents an algorithm, named Label Propagation Algorithm based on IA-GT model (LPA-CW-S), which integrates LPA-CW and LPA-S and solves the incompatibility between modularity and NMI. Fully tested on both computer-generated and real-world networks, this method gives better results in indicators such as modularity and NMI than existing methods, effectively resolving the contradiction between the theoretical community and the real community. Moreover, this method significantly reduces the randomness and runs faster.