Overlapping community detection via self-constrained symmetric non-negative matrix factorization

A number of approaches based on symmetric nonnegative matrix factorization (SNMF) have been proposed to improve the performance and the interpretability of community detection. Due to the nature of NMF, the partition results obtained by conventional NMF without post processing are soft assignments of nodes w.r.t. communities, which demonstrates overlapping of communities. Based on the traditional SNMF method, we propose a self-constrained symmetric non-negative matrix factorization (SC-SNMF) with tuning ability to control the degree of community overlapping, which controls if the community partition result is "most overlapping", "nearly overlapping" or "nearly non-overlapping". We use both traditional and overlapping version of modularity and partition density to investigate community overlapping on five real-world social network datasets. The experimental results show that SCSNMF has the ability of interpretation for overlapping degree of communities.