Co-regularized nonnegative matrix factorization for evolving community detection in dynamic networks

Abstract

Community detection in static networks solely emphasizes the clustering accuracy, while evolving community detection in dynamic networks simultaneously takes into account both the clustering accuracy and clustering drift. The available evolutionary clustering algorithms are criticized for failing to fully characterize dynamics of networks and to accurately balance the clustering accuracy and clustering drift. To solve these problems, we propose a co-regularized evolutionary nonnegative matrix factorization for evolving communities in dynamic networks (Cr-ENMF). Specifically, both the network and communities at the previous time step are utilized to characterize the clustering drift, which are incorporated into the objective function of Cr-ENMF by regularization. We show that the well-known temporal smoothness framework for evolutionary clustering is a special case of the proposed framework, and prove the equivalence between Cr-ENMF and evolutionary clustering. Thereafter, an iterative strategy is presented to optimize the objective function. The experimental results over both artificial and real world dynamic networks illustrate that Cr-ENMF outperforms state-of-the-art approaches.