Deep Dual Support Vector Data description for anomaly detection on attributed networks

Networks are ubiquitous in the real world such as social networks and communication networks, anomaly detection on aims at finding nodes whose structural or attributed patterns deviate significantly from majority of reference nodes. However, most traditional methods neglect relation structure information among data points therefore cannot effectively generalize to graph data. In this paper, we propose an end-to-end model Deep Dual Support Vector Data description based Autoencoder (Dual-SVDAE) for which considers both attribute networks. Specifically, Dual-SVDAE consists a autoencoder learn latent representation node space space, respectively. Then, dual-hypersphere learning mechanism is imposed them two hyperspheres normal perspectives, Moreover, achieve joint between network, fuse embedding final input feature decoder generate attribute. Finally, abnormal can be detected by measuring distance learned center each hypersphere Extensive experiments real-world show that consistently outperforms state-of-the-arts, demonstrates effectiveness proposed method.