A community-aware approach for identifying node anomalies in complex networks

The overwhelming amount of network data that is nowadays available, leads to an increased demand for techniques that automatically identify anomalous nodes. Examples are network intruders in physical networks or spammers spreading unwanted advertisements in online social networks. Existing methods typically identify network anomalies from a local perspective, only considering metrics related to a node and connections in its direct neighborhood. However, such methods often miss anomalies as they overlook crucial distortions of the network structure that are only visible at the macro level. To solve these problems, in this paper, the CADA algorithm is proposed, which identifies irregular nodes from a global perspective. It does so by measuring the extent to which a node connects to man y different communities while not obviously belonging to one community itself. Results on synthetic and real-world data show that the incorporation of the community aspect is of critical importance, as our algorithm significantly outperforms previously suggested techniques. In addition, it scales well to larger networks of hundreds of thousands of nodes and millions of links. Moreover, the proposed method is parameter-free, enabling the hassle-free identification of anomalies in a wide variety of application domains.