Multi-layer Graph Neural Network-Based Random Anomalous Behavior Detection

Abstract

Random anomalous behavior is a false or redundant behavior that randomly appears in the network structure, affecting the analysis result of the network. Current methods mainly capture the relationship between entities in a complex network, using structure features to distinguish this behavior, which is known as graph-based anomaly detection. As a representation learning method that abstracts network entities into feature vectors, the graph embedding method is widely used in abnormal node detection and has been proven to have a relatively accurate detection effect. But it is difficult to apply in anomaly detection of numerous edges. Therefore, we combine the GCN and GAT based on node characteristics, proposing an anomaly detection model for random anomalous edges. It compares with other methods in 6 real networks and proves its effectiveness.