Analytical Study of Graph Neural Networks for Fraud Detection in Nigerian Mobile Payment Systems

Fraud in Nigerian mobile payment systems causes significant financial losses and erodes consumer confidence in digital financial services, and the relational patterns of fraudulent transaction networks are inadequately captured by conventional tabular machine learning models. This study analytically evaluates graph neural network (GNN) approaches for fraud detection in mobile payment transaction data from Nigeria. A dataset of 680,000 transactions (3.8 percent labelled fraudulent) was obtained from a Nigerian payment service provider operating in Lagos and Abuja, with transactional nodes and relationships modelled as a heterogeneous financial graph. Three GNN architectures were implemented: GraphSAGE, Graph Attention Network (GAT), and a heterogeneous graph transformer. Conventional baseline models (XGBoost and Logistic Regression) were included for comparison. Evaluation used AUC-ROC, average precision, and F1 score on a temporal train-test split to prevent data leakage. GAT achieved the highest AUC-ROC of 0.943 and average precision of 0.812, outperforming XGBoost (AUC-ROC 0.887) by a statistically significant margin. The graph-based features, particularly transaction velocity within device-account subgraphs, showed the highest contribution to fraud prediction. The study fills a research gap in GNN application to African fintech fraud detection and provides an analytically validated approach for Nigerian payment service providers. Implementation using a streaming graph update architecture for real-time fraud detection is recommended as a future development.

Keywords: graph neural networks, fraud detection, mobile payments, Nigeria, fintech security