Evolutionary clustering via graph regularized nonnegative matrix factorization for exploring temporal networks

Abstract

Evolutionary clustering is a classic and helpful framework for modeling dynamic data and has been devoted to analyzing the temporal networks recently. However, all methods based on evolutionary clustering either does not directly model the evolution, predict the varying of dynamic communities, deal with the case of the time-varying number of communities or have a high computational complexity and are not easily extended, which limits the applications for exploring the dynamic networks. In this paper, we propose a new novel framework of Evolutionary Clustering based on Graph regularized Nonnegative Matrix Factorization (ECGNMF), to detect dynamic communities and the evolution patterns and predict the varying structure across the temporal networks. More concretely, we construct a generated model to fit the observed snapshot networks based on nonnegative matrix factorization (NMF). Any one of series intuitive and interpretable penalty items as we denoted could be integrated into the proposed framework via a graph regularization way, which can be optimized efficiently and model changes of the number of communities across different snapshots of temporal networks. We also detect the evolution patterns of the dynamic communities in the temporal networks in a principled manner. Experimental results show that our framework has better performance on community detection in temporal networks compared to some widely used models based on evolutionary clustering and heuristic methods.