Multi-view discriminative edge heterophily contrastive learning network for attributed graph anomaly detection

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 promising results. 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 Multi-view discriminative Edge heterophiLy cOntrastive learning Network (MELON) for attributed graph anomaly detection. Our approach addresses the aforementioned limitations by integrating human knowledge of different anomaly patterns for data augmentation and designing an edge discriminator to identify edge homophily/heterophily in an unsupervised manner. In addition, a dual-channel encoder is also designed to capture representative representations of the nodes from the discriminated homophilic/heterophilic edges. We then use a decoder to reconstruct the original network from the node representations learned by the dual-channel encoder and rank nodes according to their reconstruction error as the anomaly metric. Extensive experiments on four public benchmark datasets demonstrate that our proposed method outperforms state-of-the-art baselines.