EGNN-AD: An Effective Graph Neural Network-Based Approach for Anomaly Detection on Edge-Attributed Graphs

The emergence of Graph Neural Networks (GNNs) has led to the development of several GNN-based anomaly detection models that detect anomalies in attributed graphs using graph structural and attribute information. However, most existing solutions focus on detecting anomalous nodes, while some applications require detecting anomalous edges, such as fraudulent financial transactions or product reviews. In this work, we present a method for detecting anomalous edges from a node classification perspective in a transformed node-attributed graph, where each edge in the original graph is converted into an attributed node. We propose an effective solution, \(\mathtt {EGNN\text {-}AD}\) , which leverages two GNN models, \(\texttt{EGNN}\) and \(\texttt{NGNN}\) , to incorporate information from both edges and nodes for detecting edge anomalies. \(\texttt{EGNN}\) uses top-k Personalized PageRank (PPR) to establish different edge-to-edge relations in the new graph, while \(\texttt{NGNN}\) employs trainable node embeddings to enhance performance. Our experiments on real edge-attributed datasets demonstrate that \(\mathtt {EGNN\text {-}AD}\) consistently outperforms several baselines in terms of anomaly detection accuracy.