Contrastive self-supervised learning for graph anomaly detection based on local structural pattern
Ching Yao Kao, Yi‐Hsuan Lee · IET conference proceedings. · 2025
Numerous studies have highlighted the growing importance of anomaly detection in attributed networks for uncovering malicious activities, data inconsistencies, and emerging trends. Contrastive learning is distinguished in this task by learning discriminative representations without the need for extensive labelled data. In this study, we proposed a modified Contrastive self-supervised Learning framework for Anomaly detection on attributed networks (CoLA), adjusting its instance-pair sampling strategy to handle graphs with high mean degrees more effectively. Evaluated the performance with injected synthetic anomalies, our method outperforms the original CoLA, showing improved detection of local structural patterns in various datasets.