Identifying Anomalies in Graph Streams Using Change Detection

Anomaly detection in graph streams requires both the discovery of normative subgraph patterns in the stream and then the identification of subgraphs that are unexpected deviations from the normative patterns. Both of these processes are computationally complex, and therefore we should only update them when necessary. We present an approach based on a change detection metric used for graph sampling that selectively updates the normative patterns only when significant change has occurred. Using a batch processing method on the graph stream, we show that the change detection approach significantly reduces running times while maintaining the accuracy of more exhaustive approaches. CCS Concepts • Information systems~Data stream mining • Computing methodologies~Anomaly detection.