Dual Graph Neural Network for Overlapping Community Detection

Community detection has long been designed to find communities with different structures in various networks. It is now widely believed that these communities often overlap each other. However, due to the complexity and diversity of the network, it is often difficult to accurately identify the overlapping community structure in many real networks. Considering the above problem, we introduce a dual graph neural network for overlapping community detection (DGOCD) under the framework of the extended Bernoulli-Poisson (EBP). First, we build two graphs to model information of different orders between nodes respectively, and use a set of GCNs as a backbone to learn semantic representations of the above graphs in parallel. Then we introduce the concept of topological potential matrix to aggregate the embedding representations of the two channel graphs. Moreover, for learning the affiliations between nodes and communities, we carry out network reconstruction based on the former information. Finally, the reconstructed network is sent into the GCN to get the final community division. Experimental results on real network datasets demonstrate that the proposed DGOCD consistently outperforms existing methods.