GrSrNMF: dynamic community detection with graph and symmetry bi-regularized non-negative matrix factorization

Community detection in dynamic networks has become an interesting and popular research direction in recent years, widely used in electronic commerce, social media, and other fields. Evolutionary clustering is a classical and effective framework for dynamic community detection. Most current evolutionary clustering frameworks do not directly model the evolutionary pattern of dynamic networks, but only discover their change points. Therefore, some researchers introduce graph-regularization to generalize the evolutionary clustering. However, the corresponding problem is that the effect of graph regularization depends too much on the quality of dynamic networks. If the dynamic networks have too much noise or their structural organization is not obvious, the graph-regularization may not improve the model effect, and it may lead to the problem of being too smooth. Consequently, the depiction of distinct node characteristics is too uniform and challenging to discern. To solve this problem, a dynamic community detection framework based on Graph and Symmetry Bi-regularized Non-negative Matrix Factorization (GrSrNMF) is proposed. GrSrNMF can successfully identify community structures and appropriately address variations in the number of communities within network snapshots. This is particularly crucial in dynamic networks. analysis, as the number and structure of communities can vary over time. GrSrNMF can not only learn the symmetric structure of an undirected network well but also can capture the local structure of the graph. It improves the over-smoothing problem caused by graph-regularization, to mine the evolution pattern of dynamic networks and explore their temporal changes. Our proposed GrSrNMF outperforms some state-of-the-art models, like those based on evolutionary clustering and graph regularization, as well as sophisticated methods in exploring community detection in dynamic networks, utilizing two synthetic networks and two real networks.