Comparison of four variants of RE-GCN.

Community detection is a classical problem for analyzing the structures of various graph-structured data. An efficient approach is to expand the community structure from a few structure centers based on the graph topology. Considering them as pseudo-labeled nodes, graph convolutional network (GCN) is recently exploited to realize unsupervised community detection. However, the results are highly dependent on initial structure centers. Moreover, a shallow GCN cannot effectively propagate a limited amount of label information to the entire graph, since the graph convolution is a localized filter. In this paper, we develop a GCN-based unsupervised community detection method with structure center Refinement and pseudo-labeled set Expansion (RE-GCN), considering both the network topology and node attributes. To reduce the adverse effect of inappropriate structure centers, we iteratively refine them by alternating between two steps: obtaining a temporary graph partition by a GCN trained with the current structure centers; updating each structure center to the node with the highest structure importance in the corresponding induced subgraph. To improve the label propagation ability of shallow GCN, we expand the pseudo-labeled set by selecting a few nodes whose affiliation strengths to a community are similar to that of its structure center. The final GCN is trained with the expanded pseudo-labeled set to realize community detection. Extensive experiments demonstrate the effectiveness of the proposed approach on both attributed and non-attributed networks. The refinement process yields a set of more representative structure centers, and the community detection performance of GCN improves as the number of pseudo-labeled nodes increase.