A Graph-Based Approach to Detecting Fraudulent Transactions in Nigerian Mobile Payment Systems

Mobile payment fraud in Nigeria imposes substantial financial losses on consumers and service providers, with the Nigeria Inter-Bank Settlement System reporting significant year-on-year increases in electronic payment fraud incidents. This study proposes and evaluates a graph-based approach to fraudulent transaction detection in Nigerian mobile payment systems, addressing limitations of conventional rule-based and tabular machine learning fraud detection approaches. A transaction dataset of 250,000 anonymised mobile payment records, obtained in collaboration with a licensed payment service provider, is modelled as a bipartite transaction graph connecting sender accounts, recipient accounts, and transaction metadata nodes. Graph Neural Network models, specifically GraphSAGE and Graph Attention Networks, are trained on transaction graph embeddings and compared against baseline XGBoost and random forest classifiers. The GNN-based approach achieves an AUC-ROC of 0.95 and a fraud recall of 0.88, outperforming the XGBoost baseline (AUC 0.89) and significantly improving detection of coordinated fraud ring patterns. Feature importance analysis identifies unusual transaction velocity, account age, and cross-device transaction inconsistency as the strongest fraud indicators in the Nigerian mobile payment context. Keywords: fraud detection, graph neural network, mobile payments, Nigeria, NIBSS