Temporal subgraph contrastive learning for anomaly detection on dynamic attributed graphs

A dynamic attributed graph exists in which features and structures evolve. Some researchers have focused on the study of anomaly detection methods under such complex evolution patterns. However, they cannot address the discrepancy problem of coupled evolution of multitemporal features, i.e., how to portray and capture the anomaly patterns under coupled evolution is a key problem that needs to be solved. Therefore, in this paper, we propose the Temporal Subgraph Contrastive Learning (TSCL) method for anomaly detection on dynamic attributed graphs, which learns node representations by sampling and comparing temporal subgraphs and uses the statistical results of multiround comparison scores to predict node anomalies. In particular, the Temporal Features Evolving module and the Temporal Subgraph Sampling module capture the coupled evolutionary patterns of features and structures, and the combination of the Temporal Contrastive Learning module and the Statistical Anomaly Estimator module implements an end-to-end working approach between representation learning and anomaly detection. Finally, extensive comparative experiments and analyses on real datasets demonstrate the effectiveness of our proposed TSCL approach for anomaly detection on dynamic attributed graphs.