Change Point Detection in Dynamic Networks Based on Community Identification

Detecting or recognizing event related change points in dynamic networks becomes an increasingly important task, as a change in network's structure may associate with a change in function of the networked system. However, general change point detection methods either fail to extract effective features or do not scale well. In this work, we introduce the probability distribution of nodes' importance to characterize a network, the profile that allows for comparison between two networks and clustering on snapshots of dynamic networks. Based on this, we develop summarization scheme to detect change points on dynamical networks by segmenting the snapshots into disjoint clusters, which can guarantee the scalability on large dynamical networks. Specifically, we construct a network whose nodes represent the dynamic network snapshots. Then we do community detection on the constructed network and serialize the community detection results in chronological order. The resultant sequence naturally indicates the potential changes. Experiments on both synthetic and real-world networks show the outperformance of our framework compared to the state-of-the-art methods.