Analytical Study of the Research Gap in Graph Neural Networks for Fraud Detection in Nigerian Banking

Graph neural networks capture relational patterns in transaction networks that conventional tabular machine learning models cannot, and their potential for improving fraud detection in Nigerian banking represents an important but unstudied research frontier. This study examined the research gap in graph neural network fraud detection for Nigerian banking through systematic review and original empirical prototyping. A systematic scoping review of nine databases identified 58 publications from 2020 to 2024 on GNN-based fraud detection in financial services, of which zero addressed Nigerian banking contexts. Original empirical prototyping constructed a transaction graph from 18 months of anonymised interbank transfer records provided by one commercial bank in Lagos. Nodes represented accounts and edges represented transactions weighted by amount and frequency. Three GNN architectures (GraphSAGE, Graph Attention Network, and Graph Convolutional Network) were trained to identify fraudulent account communities and individual high-risk nodes, compared against XGBoost as a non-graph baseline. GraphSAGE achieved AUC of 0.91 on the test set compared to XGBoost's 0.84, confirming the graph structural advantage. False negative rates for coordinated fraud rings were 47 percent lower under GraphSAGE. The study identifies dataset availability, computational infrastructure, and legal data sharing as the three primary barriers to GNN adoption in Nigerian banking fraud detection. Recommendations include CBN establishing a financial crime graph data consortium, and academic-bank partnerships to develop Nigerian-specific GNN fraud benchmarks.

Keywords: graph neural networks, fraud detection, Nigerian banking, GraphSAGE, research gap