Anomaly Detection in Evolving Heterogeneous Graphs

For dynamic graphs, which are used to model temporal-relational data such as social networks, anomaly detection is often leveraged to identify vertices exhibiting unusual spikes in activity (e.g., posts of a user). As an indicator for the activity of a vertex, the incident edges and their evolution over time hold valuable information to characterize anomalous behavior. However, many existing approaches do not exploit this information fully by either not considering edges at all or reducing them to quantified numerical attributes. We propose a sliding window analysis for edges to assess their evolving behavior and incorporate it into an existing anomaly detection approach. Initial experiments show how our approach is able to detect anomalies that are statistically more significant than existing methods in our exemplary use-case featuring urban anomaly detection.