GATAE: Graph Attention-based Anomaly Detection on Attributed Networks

Abstract

Anomaly detection on attributed network has broad applications in many practical scenarios. Most of existing methods figure out the anomaly detection task by using graph convolution networks to embed the attributed networks. However, these methods will inevitably suffer over-smoothing problems. To approach this problem, in this paper, we propose a graph attention-based autoencoder model. Firstly, we encode the attributed network with a graph attention network. The attention mechanism not only alleviate the over-smoothing problem, but also help encoder learn nodes' representation better. Secondly, we use two decoders to reconstruct the original network and obtain reconstruction errors subsequently. Thus, we are able to detect anomalies by measuring the reconstruction errors. Experiments on real-word datasets show that our proposed model has better performance than other baseline methods in the area under a receiver operating characteristic curve (AUC).