An Original Framework for Explainable Anomaly Detection in Nigeria’s Banking Sector Using Temporal Graph Networks

Financial fraud detection in Nigerian banking institutions relies predominantly on rule-based systems and conventional machine learning models that cannot capture the complex dynamic relationships between transactional entities over time, and the deployment of more powerful graph-based models is impeded by the black-box opacity that makes regulatory compliance and investigator trust difficult to achieve. This study developed an original explainable anomaly detection framework for Nigerian banking fraud, integrating temporal graph neural networks with post-hoc explanation methods. The framework makes three original contributions. First, a Temporal Graph Representation Schema was designed for Nigerian banking transaction data, constructing heterogeneous dynamic graphs with entity types including customers, merchants, accounts, and ATMs as nodes, and transaction events, device associations, and geographic co-occurrence as time-stamped edges. Second, an original Temporal Graph Anomaly Network (TGAN) architecture was developed, combining Temporal Graph Transformer for temporal edge embedding with a heterogeneous attention aggregation mechanism for multi-relational graph traversal, designed specifically for the burst-transaction patterns characteristic of Nigerian mobile banking fraud. Third, an original Subgraph Explanation Algorithm (SEA) was developed to generate human-interpretable explanations of fraud predictions as minimal subgraphs containing the most evidential transaction pathways, validated by a panel of 10 EFCC financial crime investigators. TGAN achieved 97.2 percent AUC on a de-identified dataset from a Tier 1 Nigerian bank, outperforming XGBoost (91.4 percent AUC) and a standard GCN (94.1 percent AUC). SEA explanations were rated as investigation-useful by 90 percent of EFCC evaluators. The study constitutes an original contribution to financial crime AI methodology.