Towards Decoupled Contrastive Learning with Neighborhood Enhancement for Graph Anomaly Detection
Abstract
Unsupervised graph anomaly detection (UGAD) is crucial to identify anomalous behavior in graph-structured data. Most existing methods rely on contrastive learning, typically adopting node-subgraph and node-node contrastive paradigms. However, anomalous nodes can interfere with local neighborhood modeling, compromising node-subgraph contrast. Moreover, the training of these two paradigms is often coupled, which limits their ability to fully capture node-level anomaly cues. To address these limitations, we propose DeSEN, a novel two-stage UGAD learning framework that decouples the two contrastive learning paradigms. In Stage 1, we introduce a structural-semantic neighborhood enhancement strategy based on biased views to suppress anomalous interference during neighborhood modeling. In Stage 2, we leverage the confident node sets constructed by affinity ranking in Stage 1 to guide node-node contrast, thereby providing effective global node-level anomaly signals. Extensive experiments on eight real-world benchmark datasets, including a large-scale Open Graph Benchmark (OGB) dataset, against ten state-of-the-art baselines show that DeSEN significantly improves both AUROC and AUPRC, while the training and inference pipeline completes within a few minutes. These results validate the effectiveness and efficiency of DeSEN.