Fraud targeting digital payments across Africa has surged alongside the continent’s rapid shift to mobile and card-based transactions, yet most detection research focuses on data-rich Western markets. African financial institutions face a compounding challenge: they must stop fraud in real time while working with far less labelled data than their counterparts in Europe or North America. We tackle this gap head-on by comparing three machine learning paradigms on the Vesta Corporation (IEEE-CIS) dataset, which contains 590 540 real e-commerce transactions at a 3.5% fraud rate. A gradient-boosted tree (LightGBM) serves as the supervised baseline, an Isolation Forest provides an unsupervised anomaly scoring, and a deep neural network trained with Focal Loss bridges the two philosophies. Each model goes through a disciplined pipeline of correlation filtering, SVM-RFE feature selection, and Bayesian hyperparameter search via Optuna. Our results tell a clear story. LightGBM dominates across metrics: AUPRC of 0.5974, AUROC of 0.9309, and MCC of 0.5738, finishing training on CPU in under five seconds. Even when we strip away 90% of the training data to mimic a newly launched African payment platform, LightGBM holds an AUPRC of 0.5172, comfortably above every other model at full data. The Isolation Forest, though weaker in absolute terms, proves its worth as a label-free fallback for coldstart scenarios. We integrate TreeSHAP and LIME so that every flagged transaction carries an auditable, feature-level explanation aligned with South Africa’s POPIA and FICA requirements. Taken together, these findings show that careful feature engineering and tuning outperform complex hybrid architectures run with default settings, offering practical, explainable fraud detection suited to Africa’s data-scarce reality.

credit card fraud detection, Africa, data scarcity, LightGBM, Isolation Forest, neural networks, focal loss, explainable AI, SHAP, POPIA