A Double Non-Negative Matrix Factorization Model for Signed Network Analysis

Abstract

As one of the most important characteristics of complex networks, community structure have been studied extensively in real cases. However, in signed networks that include both positive and negative edges, the development of community discovery is still limited now. Because the sign information on edges poses a challenge to modeling the signed network. Most existing methods are based on heuristics, so these methods tend to have high computational complexity and ignore the generation of the networks. Here, we propose a Double Non-negative Matrix Factorization (DouNMF) model from the perspective of generative model to detect communities in the signed network. This algorithm skillfully applies the Non-negative matrix Factorization algorithm to the signed network. In addition, the algorithm integrates indegree information into the process of matrix factorization. Large amounts of experiments on several artificial and real-world signed networks validate that the effectiveness and accuracy of our proposed approaches both in community discovery and link prediction.