A Graph Autoencoder-based Anomaly Detection Method for Attributed Networks

Anomaly detection in attributed networks aims to find anomalous nodes in the network that differ from the behavior pattern of most nodes, and graph neural network provide a way to use fused structural and attribute information. However, existing methods based on Graph Convolutional Network (GCN) detection do not consider the over-smoothing phenomenon of GCN due to the stacks of network layers, which causes significant performance deterioration. To address the above problems, we propose a graph autoencoder-based anomaly detection method for attributed networks: Residual Graph Autoencoder (Res-GAE), by which the performance is effectively improved. Res-GAE contains an encoder and two decoders. More specifically, the encoder consists of a GCN and a residual network is utilized to learn the network representation. The decoders are designed to reconstruct the network structure and node attributes respectively. After that, the objective function is used to analyze the reconstruction error to generate the anomaly score ranking, to realize anomaly detection. Extensive experiments on the three datasets (BlogCatalog, Flickr, ACM) demonstrate that the proposed method has the significant improvement compared with other baseline methods.