GRAPH-Guard: A Framework for Heterogeneous Graph Anomaly Detection using Supervised and Unsupervised Techniques

Recently, graph data has received a lot of attention as it is used to represent other data types including social networks, banking, security, financial, medical, and textual data. Therefore, anomaly detection in such data has become an important research area due to its ability to prevent adverse events like financial fraud, network intrusion, and social spam. Anomalies in graph data can occur at the node, edge, subgraph or graph level. Node anomalies can be due to unusual structure or attributes. Edge anomalies aim to identify unusual connections which are often unexpected or unusual relationships between real-world entities. The main challenge is detecting and categorizing such anomalies. In recent years, many computational methods have been proposed for anomaly detection in graphs using statistical analysis and machine learning approaches. In this research, we propose a framework called "GRAPH-Guard" for node-level anomaly detection using deep and ensemble learning algorithms. The framework is evaluated on heterogeneous attributed graphs and compared to other node anomaly detection algorithms quantitatively and qualitatively. The quantitative evaluation shows the proposed framework improves the area under the curve (AUC) by 4 points compared to the average AUC of previous works. It also improves the F1 score by 1 point compared to the best previous F1 score.