Anomaly Detection in Evolutionary Social Networks leveraging Community Structure

Abstract

Social networks are temporally evolving by nature and in general, the evolution takes place gradually with time. But occasionally, the nodes may exhibit anomalous behavior that is to be detected and reported in real time. Usually, the social networks are represented as a temporal stream of weighted graphs, where edge weights stand for frequency or power of actions. Here, for identification of anomalies in evolutionary social networks, we follow an action based approach leveraging the community structure of the network, where it is shown that only the Community Boundary Nodes are enough to detect anomaly in a temporally evolving social network. We introduce a new definition of anomaly score for nodes based on its action history. To the best of our knowledge, this is the first work to exploit the community boundary nodes of an evolutionary social network to identify the graph snapshots containing anomalous events. Experimental studies on real-world as well as synthetic evolutionary networks show that the proposed technique achieves almost 13% improvement in F-Score compared to the state-of-the-art algorithms. Moreover, parallel implementation of the proposed technique results nearly 6 x speedup making it suitable for real time computation of anomaly in evolutionary social networks.