Overlapping community detection on complex networks with Graph Convolutional Networks

Abstract

Discovering the community structure within networks is of significance with respect to many realistic applications, like recommendation systems and cyberattack detection. In this study, we propose an overlapping community detection method CDMG based on Graph Convolutional Network (GCN) from the perspective of maximizing the Markov stability of community structure, which is defined in terms of the clustered autocovariance of a Markov process taking place on the network. Extensive experiments on both the attributed networks and the normal networks with different scales demonstrate the superiority of CDMG compared to other established community detection algorithms. Additionally, the Markov stability of the community structure relies on a time parameter, Markov time, and we observe that the performance of CDMG can be further improved by utilizing the optimal Markov time. According to the variation curve that demonstrates the influence of Markov time t on the performance of CDMG, we propose a trichotomy-based method to search for the optimal Markov time for our method.