Community Enhanced Graph Convolutional Networks

Abstract Graph representation learning is a key technology for processing graph-structured data. Graph convolutional networks (GCNs), as a type of currently emerging and commonly used model for graph representation learning, have achieved significant performance improvement. However, GCNs acquire node representations mainly through aggregating their neighbor information, largely ignoring the community structure which is one of the most important feature of the graph. In this paper, we propose a novel method called Community Enhanced Graph Convolutional Networks (CE-GCN), which integrates both neighborhood and community information to learn node representations. Specifically, the neighborhood information of nodes is aggregated by a graph convolutional network. The community information of nodes is calculated by a modularity constraint. Finally, we incorporate the modularity constraint into the graph convolutional network, and then form a unified model framework. Experimental results on five real-world network datasets demonstrate that CE-GCN significantly outperforms state-of-the-art methods.