Unsupervised Anomaly Detection on Node Attributed Networks: A Deep Learning Approach

Anomaly detection has been one of the important issues in social network analysis in recent years due to the crucial role it plays in different applications such as fraud and spammer detection. Using both graph and node characteristics leads to more accurate results in detecting anomalous nodes of node attributed networks. Most of the research works in this field are concentrated on supervised methods for anomaly detection. However, in real-world problems, there is not enough labeled data to use supervised methods for anomaly detection. This paper proposes an unsupervised method for detecting anomalous nodes in node attributed networks. The methods used a two-step deep learning approach. In the first step, structural features of the network are extracted using node2vec; in the next step, Variational AutoEncoder (VAE) is used to detect the anomalies considering both structural and node attributes. The anomalous nodes are recognized by their higher reconstruction loss. Our experimental results on two datasets, BlogCatalog and Flickr, show that the suggested method can compete with the state-of-the-art approaches of anomaly detection in attributed networks. Our method (Deep2NAD) outperforms the state-of-the-art result on the Flickr dataset based on AUC. Moreover, it receives an acceptable AUC over the BlogCatalog dataset in comparison to the state-of-the-art methods.