<![CDATA[Detecting fraudulent transactions in financial systems presents a major challenge due to the scarcity of fraud instances and the limited availability of labeled data. This study explores the use of few-shot learning techniques combined with Graph Neural Networks (GNNs) to address these constraints. We evaluate four GNN architectures—Graph Convolutional Network (GCN), GraphSAGE, Graph Attention Network (GAT), and Simplified Graph Convolutional Network (SGCN)—on four real-world fraud detection datasets: Bank Fraud, IEEE-CIS, PaySIM, and ECommerce. Graph-based representations are constructed for each dataset, and models are trained using only 0%, 1%, 5%, and 10% of labeled data to simulate few-shot conditions. Experimental results show that GNNs, particularly GAT and GraphSAGE, maintain strong performance even with minimal supervision. Notably, GAT and GCN achieved an F1-score of 0.88 on the PaySIM dataset with just 10% labeled data, and GraphSAGE reached 0.25 on the highly imbalanced IEEE-CIS dataset. ROC curve analysis further demonstrates the discriminative capabilities of each model under different label settings. These findings highlight the potential of GNNs for effective fraud detection in low-resource and imbalanced environments, offering a practical solution for financial institutions aiming to enhance security with minimal labeled data.]]>
Copyrights © 2025
