Motif-Level Anomaly Detection in Dynamic Graphs

As many real-world networks evolve over time, such as social networks, user-item networks, and IP-IP networks, anomaly detection for dynamic graphs has attracted growing attention. Most existing studies focus on detecting anomalous nodes or edges but fail to detect anomalous motif instances. In this paper, we propose MADG, a general Motif-level Anomaly Detection framework for dynamic Graphs, which can identify the anomaly in different motifs. Motifs are specific subgraph structures that frequently occur in a network and have been widely used in network analysis. In order to learn discriminative motif-level representations and leverage the temporal information from the dynamic graph, we design motif-augmented GCN and temporal self-attention. We first use motif-augmented GCN to model the topological structure among nodes and motif instances to learn their representations at each snapshot. Then, we feed the representations of multiple snapshots into the self-attention layer with relative temporal encoding in order to capture the evolutionary patterns. Extensive experiments on real-world dynamic graph datasets demonstrate the effectiveness of our proposed MADG framework.