Figure 2
From: Deep graph level anomaly detection with contrastive learning
The model architecture of GLADC is constructed by the following three modules. For dual-graph encoder module, we first use a graph convolution autoencoder to learn node level representation as \({\mathbf{Z}}_{node}\) by encoding and decoding the input real graph. Then, we introduce a disturbed graph encoder to get another node level representation as \({\hat{\mathbf{Z}}}_{node}\). We design a contrastive learning paradigm to enhance graph level representation based on initial graph level representations \({\hat{\mathbf{Z}}}_{G}\) and \({\mathbf{Z}}_{G}\) learned by shared projection head. For graph encoder module, we use a shared graph encoder to encode the reconstruction graph to get latent node level and graph level representations as \({\acute{\mathbf{Z}}}_{node}\) and \({\acute{\mathbf{Z}}}_{G}\) respectively. For graph anomaly detection module, we first train GLADC model in normal graphs, and test model by test dataset containing normal and abnormal graphs. Then, anomalous graphs are recognized according to the error score of corresponding graph representations from the input graph and the reconstruction graph.
