Multi-Relation Graph Transformer and Edge Aware Learning: A Scalable Approach for Detecting Fraud in Heterogeneous Graphs

Abstract

Detecting fraud in multi-relational graphs is a significant difficulty due to the complex nature of fraudulent activities and the inadequacies of conventional Graph Neural Networks (GNNs) in managing heterophilic relationships and long-range dependencies. Current GNN-based models experience over-smoothing and over-squashing, which constrains their capacity to differentiate between fraudulent and legitimate interactions. To tackle these difficulties, we offer MR-GT, a Graph Transformer-based model that integrates multi-relational attention and edge-aware techniques to enhance fraud detection. MR-GT implements edge-aware message passing and edge-based attention bias, enabling the model to concentrate on significant transactional links while disregarding less pertinent interactions. The approach utilizes relation-aware aggregation to accurately identify fraud-related patterns in diverse interactions. Experimental findings on the Yelp and Amazon datasets indicate that MR-GT significantly surpasses state-of-the-art approaches, attaining an AVC of 92.33 on Yelp and 98.25 on Amazon. The ablation study further illustrates the significance of each component, indicating that MR-GT provides a scalable and efficient approach for identifying fraud in extensive, multi-relational graphs.