Overlapping Community Detection in Weighted Temporal Text Networks

Abstract

Network is a powerful language to represent relational data. One way to understand network is to analyze groups of nodes which share same properties or functions. The task of discovering such groups is known as community detection. The community detection in real-life networks, the majority of which are weighted temporal text networks, is confronted with two main problems - how to model the weight of edges and how to exploit the temporal information. Existing works either ignore the edge weight or utilize it in graph measures like modularity, which lacks scalability. And currently the common-used method involving temporal information is to discretize the time, which leads to series of problems. We are thus motivated to present a new method to encode the edge weight and temporal information. A probabilistic generative model, named Custom Temporal Community Detection (CTCD) is introduced, which views the link between two nodes as a weighted edge with several time stamps. Our model utilizes network, semantic and temporal information simultaneously to extract temporal community affiliations for individual user, influence strength across communities and temporal interested topic in each community. An efficient inference method, which scales linearly, and corresponding parallel implementation are proposed to adapt to large datasets. Through the knowledge extracted by CTCD, we are able to spot the community shift of the individual user, to which little attention has been given, and employ it to track the development of the communities over time. Moreover, experiments on two large-scale weighted temporal text networks show that CTCD gains significant improvement over state-of-the-art methods on a series of tasks.