Dual Perspective Contrastive Learning Based Subgraph Anomaly Detection on Attributed Networks

Network anomaly detection is widely used to discover the anomalies of complex attributed networks in reality. Existing approaches can detect independent abnormal nodes by comparing the attribute differences between nodes and their neighbors. However, in real attributed networks, some abnormal nodes are concentrated in a local subgraph, so it is difficult to find out by comparing neighbor nodes because the features within the subgraph are similar. Furthermore, most of these methods use GCN for feature extraction, which means that each node will indiscriminately aggregate its neighbors, causing the value of normal nodes to be severely affected by the surrounding abnormal nodes. In this paper, we propose an improved unsupervised contrastive learning method that is universally applicable to multiple anomaly forms. It will comprehensively compare the inside and outside of the subgraph as two perspectives and use the knowledge of the trained teacher model to adjust the sampling probability for the selectively aggregating of neighbor nodes. Experimental results show that our proposed framework is not limited by the distribution of abnormal nodes and outperforms the state-of-the-art baseline methods on all four benchmark datasets.