Community Detection via Joint Graph Convolutional Network Embedding in Attribute Network

Community detection is a foundational task in network analysis. Besides the topology information, in recent years, there have been many methods utilizing network attribute information for community detection. The key of introducing the attribute information is how to integrate these two sources of information for better community detection. Graph Convolutional Networks (GCN) is an effective way to integrate network topologies and node attributes. However, when GCN is used in community detection, it often suffers from two problems: (1) the embedding derived from GCN is not community-oriented, and (2) this model is semi-supervised rather than unsupervised. To address these problems, we propose an unsupervised model for community detection via joint GCN embedding, i.e. JGE-CD. We employ GCN as the basic structure of encoder to match the above two sources of information, and use a dual encoder to derive two different embeddings by using an attribute network and its variant with random transformation. We further introduce a community detection module considering the community properties into the joint learning process. It derives two community detection results for a relative-entropy minimization which work together with a topology reconstruction module in order to make the model discover community structure in an unsupervised way. Extensive experiments on seven real-world networks show a superior performance of our model over some state-of-the-art methods.