AttentionAE: Autoencoder for Anomaly Detection in Attributed Networks

Abstract

Anomaly detection in attributed networks is to find nodes that deviate from the behavior patterns of most nodes, which is widely used in social network false account detection or network intrusion detection and so on. However, most existing methods only focus on one aspect of network either network structure or node attributes, ignoring the interaction between network structure and node attributes. Meanwhile, they regard reconstruction error as node anomaly score, which lacks considering the other factors for computing score. Therefore, in this paper, we propose a method for attributed network anomaly detection based on an autoencoder considering the node attention mechanism and an anomaly score generator. The autoencoder considers not only network structure but also node attributes to obtain a higher-quality embedded representation. Meanwhile, the decoder reconstructs the adjacency matrix and calculates the reconstruction error. The anomaly score generator uses a multi-layer perceptron (MLP) as the basic framework. In addition, in order to better consider the calculation of the node anomaly score by the reconstruction error, we concatenate the reconstruction error, reconstruct the residual direction vector and embedding vector to construct input vector for training MLP. At last, the final output is node anomaly score, and anomaly detection achieved. Experiments on three real-world datasets prove the effectiveness of our method.