Robust Graph based Deep Anomaly Detection on Attributed networks

Abstract

Anomalous users' identification on attributed social networks involves finding users whose profile characteristics go amiss fundamentally from the greater part of reference profiles both in terms of network structure and node attributes as well. Most of the existing methods uses graph convolutional networks (GCN) to generate latent representation of nodes for various tasks like node classification, link prediction and anomaly detection. This method primarily represents every node as the aggregate of its neighbouring node's features. But it has a problem that (i) the representation of normal node is affected by the presence anomalous neighbour nodes and as a result, even normal nodes are considered as anomalous and (ii) anomalous nodes go undetected as their representation is flattened by aggregate operation. To overcome this problem, we propose a robust anomaly detection(RAD) method to better handle the anomaly detection task. weighted aggregate mechanism is employed to distinguish between node's self features and its neighbourhood. Experiments on twitter,enron and amazon datasets give results which shows that the proposed method is robust in detection of anomalies based on weighted average of self and neighbouring node's features.