Beyond Homophily: Attributed Graph Anomaly Detection via Heterophily-Aware Contrastive Learning Network

Wangyu Jin, Huifang Ma, Yingyue Zhang, Zhixin Li, Liang Chang

Abstract: Attributed graph anomaly detection aims to identify abnormal nodes that significantly differ from most nodes in terms of their attribute or structure. Recent graph contrastive learning methods, which follow an augmenting-contrasting learning scheme, have shown great potential. However, these methods still have two key limitations. Firstly, the contrastive views generated through random perturbation may not conform to abnormal patterns. Secondly, the inability to effectively discriminate heterophilic edges in anomalous networks can adversely impact the detection performance. To this end, we propose a Heterophily-aware Contrastive Learning network (HCL for abbreviation) for attributed graph anomaly detection. Our approach addresses the aforementioned limitations by integrating human knowledge of anomaly patterns for constructing contrastive views and identifying edge homophily/heterophily using an unsupervised edge discriminator. Additionally, a dual-channel encoder is designed to capture representative node representations from discriminated edges. Extensive experiments on four public benchmark datasets demonstrate that our method is competitive with the most advanced baseline.

External IDs:dblp:conf/icann/JinMZLC24