Boosting Attributed Graph Anomaly Detection via Negative Sample Awareness

Anomaly detection is widely used in domains like spam identification and financial fraud detection. The Contrastive Learning-based graph abnormal detection approach obtains sufficient self-supervised information from well-designed pretext tasks. However, arbitrary selection of negative samples in existing constrastive learning frameworks introduces sampling bias. To tackle this issue, we propose a novelty attributed Graph anomaLy dEtection framework via negAtive sample awareNess(GLEAN). Our method is sensitive to hard negative samples to alleviate sampling bias. These hard negative samples, obtained by shuffling features, are employed for node-node contrast. Additionally, we restrain sampling bias in node-subgraph contrasts by reconstructing the original attributes from the feature masking graph and the original graph. By combining the generation and contrastive modules, we calculate an anomaly score for each node. Extensive experiments on five benchmark datasets show our model outperforms current state-of-the-art models.