Fraud detection in Nigerian financial card transactions using a graph attention network

The growth of the digital payment ecosystem presents significant challenges to economic stability and consumer trust because of financial fraud. Traditional rule-based and machine-learning approaches often fail to capture the complex relational patterns present in fraudulent transaction networks. This study develops and evaluates a Graph Attention Network (GAT) model for detecting fraudulent card transactions in the Nigerian financial sector by using graph-based representations to capture relationships among entities. We constructed a heterogeneous graph representation of transaction data from a Nigerian bank, in which nodes represent cards, merchants, account holders, and transactions, while edges represent transaction relationships. A GAT architecture was implemented to learn node embeddings, and an ablation study was conducted to evaluate the contribution of graph structure by comparing the proposed GAT with a no-edge GAT and a randomized-edge GAT. The model was also benchmarked against GraphSAGE, a Gated Graph Recurrent Network (GRNN), a Graph Convolutional Network (GCN), and non-graph baselines including XGBoost, a feedforward neural network, and LightGBM. All models were evaluated on a real-world Nigerian banking dataset using stratified train--test splits, normalized numerical features, and median imputation for missing age values. The results show that graph structure substantially improves fraud detection: the proposed GAT achieved an F1-score of 0.9612, outperforming the no-edge GAT (0.2650) and the randomized-edge GAT (0.8769). Although the GAT achieved a higher F1-score than the GCN (0.9612 versus 0.8998), GraphSAGE achieved a higher AUC than the GAT (0.9961 versus 0.9929), indicating a trade-off between threshold-dependent performance and ranking performance.