Temporal Clustering in Dynamic Networks with Tensor Decomposition

Abstract: Dynamic networks are increasingly being usedd to model real world datasets. A challenging task in their analysis is to detect and characterize clusters. It is useful for analyzing real-world data such as detecting evolving communities in networks. We propose a temporal clustering framework based on a set of network generative models to address this problem. We use PARAFAC decomposition to learn network models from datasets.We then use $K$-means for clustering, the Silhouette criterion to determine the number of clusters, and a similarity score to order the clusters and retain the significant ones. In order to address the time-dependent aspect of these clusters, we propose a segmentation algorithm to detect their formations, dissolutions and lifetimes. Synthetic networks with ground truth and real-world datasets are used to test our method against state-of-the-art, and the results show that our method has better performance in clustering and lifetime detection than previous methods.