GANAN: GAN-based for Anomaly Detection on Attributed Networks

Anomaly detection on attributed networks is crucial for many applications such as social networks security and fraud detection. Existing models utilize deep autoencoder to reconstruct structure and attribute and measure reconstruction errors to obtain anomaly nodes. However, there are problems such as over-smoothing and node’s representation learning. In order to solve these problems, in this paper, we propose a GAN-based for Anomaly Detection on Attributed Networks (GANAN). The fake graph network is generated by autoencoder. We use two encoders map both fake and real graph to latent space. The structure and attribute are used as the input of the discriminator. The game method can alleviate the over-smoothing and help encoder learn node representation. With the decoder output, the measurement considers the reconstruction error of nodes from both structure and attribute perspectives. Extensive experiments on real-world datasets demonstrate that our proposed model has better performance than various existing mothed.