A Graph Convolutional Network-Based Approach for Community Detection in Attributed Networks

Community detection has attracted widespread attention since it helps reveal geometric structures and latent functions of complex networks. Recently community detection has been revisited with the development of network representation learning, many approaches have been presented, including graph convolutional network (GCN) based methods. Existing GCN-based community detection methods usually rely on a considerable number of prior labels to infer unknown nodes. To address this problem, we propose a new GCN-based method for community detection in attributed networks without any label information. Based on the local self-organization characteristics of the communities, we integrate a label sampling model and the shallow GCN architecture into an unsupervised learning framework, the former helps construct a balanced training set via a local expansion strategy to train GCN. Moreover, we reveal the underlying community structures by fusing topology and attribute information. Experimental results on several real-world networks indicate our method is effective compared with the state-of-the-art community detection algorithms.