Similarity preserving overlapping community detection in signed networks

Community detection in signed networks is a challenging research problem, and is of great importance to understanding the structural and functional properties of signed networks. It aims at dividing nodes into different clusters with more intra-cluster and less inter-cluster links. Meanwhile, most positive links should lie within clusters and most negative links should lie between clusters. In recent years, some methods for community detection in signed networks have been proposed, but few of them focus on overlapping community detection. Moreover, most of them directly exploit the sparse link topology to detect communities, which often makes them perform poorly. In view of this, in this paper we propose a similarity preserving overlapping community detection (SPOCD) method. SPOCD firstly extracts node similarity information and geometric structure information from the link topology, and then uses a graph regularized binary semi-nonnegative matrix factorization (GRBSNMF) model to fuse these two sources of information to detect communities. Through this mechanism, nodes with high similarity can be well preserved in the same community. Besides, SPOCD devises a special discretization strategy to obtain the binary community indicator matrix, which is very convenient for directly identifying overlapping communities in signed networks. We conduct extensive experiments on synthetic and real-world signed networks, and the results demonstrate that our method outperforms state-of-the-art methods.