A Novel Graph-level Anomaly Detection Model

Graph-level anomaly detection, which seeks to find unusual graphs exhibiting abnormal graph structures or node features within a collection of graphs, is enormously impactful in the real world. Despite this, much of the past research has predominantly targeted the identification of anomalies at the node, edge, or subgraph levels within individual graphs. A principal challenge in this field is developing a thorough comprehension of ordinary patterns within graphs, in order to enable effective detection of graphs that stray from established norms, either locally or globally. To tackle this challenge, we introduce a novel method, which learns both local and global normal and anomalous pattern information through the contrastive distillation of node and graph representations. Our model adheres to a tri-network configuration optimized to fully utilize supervisory information. In particular, the contrast network successfully augments the range of anomalous patterns. The model efficiently employs the representation error produced by the tri-network as a scoring mechanism for anomalous graph detection. Empirical results from six different indicate that our model attains superior performance in anomaly detection, surpassing established leading methods.