Data augmentation based multi-view contrastive learning graph anomaly detection

Graph anomaly detection is valuable in preventing harmful events such as financial fraud and network intrusion. Although contrast-based anomaly detection methods could effectively mine anomaly information based on the inconsistency of anomalous node instance pairs, avoiding the drawback of using self-coding architecture that led to the need for full graph training for the model. However, most existing contrast-based graph anomaly detection methods focused only on node-subgraph contrast patterns, ignoring the fact that the sampled node-subgraph instance pairs contained only the local information of the target node, and at the same time did not take into account the importance of each subgraph to the target node, which led to the lack of global information about the node and the emergence of the problem that the contrast patterns were too generalized. In order to solve the problems mentioned above and to improve the accuracy of graph anomaly detection, a graph anomaly detection by data augmentation and multi-view contrastive learning (DAMC-GAD) was proposed. Specifically, a graph data augmentation method for anomaly detection was proposed, in which the relative local structure of target nodes and their own attributes were used to correlate distant nodes in order to construct an augmented view that was rich in global information. Layer-by-layer sampling combined with node-subgraph contrast was introduced, and optimization strategies for the contrast model were developed based on the importance of the subgraph. A multi-view contrastive learning model with data augmentation was constructed through a complementary fusion strategy, and extensive experiments were conducted on synthetic anomaly and real anomaly datasets, which show that DAMC-GAD outperforms the current state-of-the-art baseline model on both types of datasets.