Answers are grounded in this paper and cite their sources.
Financial fraud detection is a challenging data science problem because fraudulent transactions are rare, adaptive, and costly when not detected early. Machine learning models can identify suspicious transaction patterns, but many high-performing models remain difficult to interpret, limiting their use in audit-oriented financial decision support. This study proposes an explainable fraud detection framework combining ensemble learning and SHAP-based feature interpretation. Random Forest, XGBoost, and LightGBM were compared with a soft-voting ensemble model using a controlled synthetic financial transaction dataset of 20,000 records, including 400 fraudulent and 19,600 legitimate transactions. Performance was evaluated using accuracy, precision, recall, F1-score, and AUC-ROC, while SHAP was used to explain global feature importance and local prediction outcomes. The voting ensemble achieved the strongest performance, with accuracy of 0.991, precision of 0.729, recall of 0.875, F1-score of 0.795, and AUC-ROC of 0.973. SHAP identified merchant risk score, customer velocity, transaction amount, location distance, and failed attempt count as the most influential fraud indicators. The findings show that explainable ensemble learning can improve fraud detection performance while supporting transparent and auditable fraud analytics
No takes yet. Share an insight, caveat, or question.
Khadafi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: