Efficient relation-aware heterogeneous graph neural network for fraud detection

Enxia Li, Jin Ouyang, Sheng Xiang, Lu Qin, Ling Chen

Abstract: Fraud detection is a vital application of data-mining techniques in both financial sectors and social platforms. Graph Neural Networks (GNNs) have demonstrated exceptional performance in fraud detection tasks on graph-structured data. However, the complexity and heterogeneity of social networks pose significant challenges to both model efficiency and scalability. This paper presents a novel Relation-Aware Heterogeneous GNN framework for fraud detection in heterogeneous graphs. Our approach incorporates two key innovations: a relation-aware computation graph preprocessing step and a heterogeneous aggregation that integrates both feature-based and topological information. Specifically, we first preprocess the input graph using relation-aware node mapping techniques, which optimizes neighborhood aggregation by alternating between 1-hop and 2-hop neighbors. This hybrid propagation strategy reduces computational complexity without compromising fraud detection performance. Additionally, we employ stochastic projection reduction to manage feature dimensionality effectively. To enhance the efficiency of heterogeneous GNN, we integrate a distilled method (DRHGNN) that enables the transfer of heterogeneous patterns from high-dimensional features to compact representations, while maintaining predictive accuracy. The experimental results on benchmark datasets demonstrate that our model achieves superior accuracy compared to existing methods while maintaining high efficiency when detecting fraud in large graphs.

External IDs:dblp:journals/www/LiOXQC25