Finding overlapped communities in online social networks with Nonnegative Matrix Factorization

In this work, we introduce two approaches, namely iSNMF and iANMF, for effectively identifying social communities using Nonnegative Matrix Factorization (NMF) with I-divergence as the cost function. Our approaches work by iteratively factorizing the nonnegative input matrix through derived multiplicative update rules. By doing so, we can not only extract meaningful overlapping communities via soft community assignments produced by NMF, but also nicely handle all directed and undirected networks with or without weights. To validate the performance of our approaches, we extensively conduct experiments on both synthesized networks and real-world datasets in comparison with other NMF methods. Experimental results show that iSNMF is among the best efficient detection methods on reciprocity networks while iANMF outperforms current available methods on directed networks, especially in terms of detection quality.