<![CDATA[Detecting financial fraud is a complex and evolving challenge, particularly because of the relational nature of transaction data, graph sparsity, and severe class imbalance. To the best of our knowledge, this study repre-sents one of the first systematic benchmarks of five prominent Graph Neural Network (GNN) architectures, GCN, GAT, GraphSAGE, GIN, and SGCN, for fraud detection under balanced and imbalanced conditions across multiple public datasets. We explicitly evaluate the impact of the Synthetic Minority Oversampling Technique (SMOTE) on graph-based fraud detection performance, an aspect that has rarely been addressed in prior research. The comparative analysis considers predictive performance (AUC, F1-Score, Precision, Re-call) and computational efficiency to provide actionable guidance for real-world development. The experimental results show that GraphSAGE offers the best trade-off between accuracy and execution time for laten-cy-sensitive environments, while GAT’s attention mechanism supports offline, interpretability-driven analysis. These findings provide empirical evidence to inform GNN selection strategies for scalable and effective fraud detection systems.]]>
Copyrights © 2026
