Community detection and anomaly prediction in dynamic networks

Abstract

Anomaly detection is an essential task in the analysis of dynamic networks, as it can provide early warning of potential threats or abnormal behavior. We present a principled approach to detect anomalies in dynamic networks that integrates community structure as a foundational model for regular behavior. Our model identifies anomalies as irregular edges while capturing structural changes. Leveraging a Markovian approach for temporal transitions and incorporating structural information via latent variables for communities and anomaly detection, our model infers these hidden parameters to pinpoint abnormal interactions within the network. Our approach is evaluated on both synthetic and real-world datasets. Real-world network analysis shows strong anomaly detection across diverse scenarios. In a more specific study of transfers of professional male football players, we observe various types of unexpected patterns and investigate how the country and wealth of clubs influence interactions. Additionally, we identify anomalies between clubs with incompatible community memberships, but also instances of anomalous transactions between clubs with similar memberships. The latter is due in particular to the dynamic nature of the transactions, as we find that the frequency of transfers results in anomalous behaviors that are otherwise expected to interact as they belong to similar communities.