FewGAD: Few-Shot Enhanced Graph Anomaly Detection via Generative Contrastive Learning

Nannan Wu, hongdoudong, Yiming Zhao, Yazheng Zhao

Keywords: Anomaly Detection; Graph Neural Network; Few-shot Learning;

Abstract: Graph anomaly detection (GAD) is critical in domains such as fraud detection, cybersecurity, and social network monitoring. However, existing approaches face two major challenges: the inherent scarcity of labeled anomalies in practical scenarios, and the widespread reliance on graph augmentation, which often distorts anomaly semantics and undermines model robustness. To address these issues, we propose FewGAD, a framework that leverages limited anomaly labels to enhance contrastive discrimination through high-order subgraph sampling without augmentation. By avoiding augmentation-induced distortion, this design fundamentally improves the robustness and semantic validity of learned representations, thereby enabling clearer separation between normal and anomalous nodes. Furthermore, a kernel density estimation mechanism expands the utility of scarce labels, enhancing data efficiency and strengthening anomaly discrimination under few-shot settings. Extensive experiments on five benchmark datasets demonstrate that FewGAD consistently surpasses state-of-the-art unsupervised and few-shot GAD methods, achieving an average AUC gain of 6.2\%.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 5520