Incorporating Network Structure with Node Information for Semi-supervised Anomaly Detection on Attributed Graphs

Anomaly detection on attributed graphs has attracted lots of research attention recently. A great deal of existing work focuses on unsupervised anomaly detection. However, in practical applications, we can obtain some labeled instances by experts, and it remains unexplored how to utilize limited labeled instances for improving the accuracy of anomaly detection. In this paper, we propose a semi-supervised anomaly detection method by considering both structure anomalies and attribute anomalies. Firstly, based on graph convolutional networks (GCNs), we learn a hypersphere that encloses normal nodes and forces anomalous nodes to lie outside, we detect structure anomalies by calculating the distances between the node embeddings and the hypersphere center. Since the trained GCNs always fail to learn better representations for anomaly detection due to noise edges are mixed into neighborhood aggregation, we use deep neural networks (DNNs) to detect attribute anomalies. Besides, to make full use of the labeled data, we incorporate the semi-supervised learning into anomaly detection, which can propagate limited label information to a large number of unlabeled instances and learn accurate nodes embeddings. Extensive experiments on five real-world datasets validate the superiority of our method and the significance of each module.