Enhancing anomaly detection using graph neural networks: a case of fraud in mobile money

Abstract

Anomaly detection remains a major challenge in sectors such as data quality management, healthcare, and especially finance, where fraud poses serious risks to institutions and consumers. The surge of mobile money over 1.6 billion accounts and $1 trillion in annual transactions has enhanced financial inclusion but also expanded fraud exposure, with global losses estimated at $485 billion yearly. Sub-Saharan Africa, handling 70% of these transactions, faces adaptive schemes such as social engineering and agent collusion. In Uganda, identity fraud represents 90.4% of reported cases. Although 94.12% of service providers have fraud management systems, 54.9% question their effectiveness, as most rely on rule-based models from decision trees that overlook interaction effects among variables, reducing detection accuracy. Capturing these interacrtons between observations, could substantially improve fraud identification. This study introduces a statistically rigorous anomaly detection framework using Graph Neural Networks (GNNs), specifically Graph Convolutional Networks (GCNs), to model complex relational structures in transaction data. Secondary, simulated data modeled on Ugandan mobile money transactions comprising 1.7 million records and 10 variables was used. The client transaction network was represented as a directed graph and analyzed using metrics such as degree, betweenness, closeness, eigenvector centrality, clustering coefficient, and assortativity. The GNN model was benchmarked against rule-based and decision tree models using balanced accuracy, snsitivity, specificity, F1-score, and ROC-AUC. The network was sparse (average degree 6.25), weakly clustered (0.02), and exhibited low assortativity, reflecting opportunistic, individualistic fraud behavior. High-betweenness nodes were flagged as potential hubs for fraud propagation. The GNN achieved a balanced accuracy of 91.9%, F1-score of 91.2%, and ROC-AUC of 97.6%, outperforming the decision tree’s 69.2%, 52.4%, and 94.9%. Compared with prior studies applying XGBoost ,random forest and decision trees (F1-scores of 85% ,84% and 81%), the GNN demonstrates a clear performance advantage. However, reliance on simulated data and limited variables constrains generalizability. The study concludes that GNN-based models provide scalable, dynamic solutions for mobile money fraud detection and recommends validation using real transaction data, integration of richer contextual features.

Subject Keywords: Anomaly detection; graph neural networks; fraud; mobile money