SmartFraud-Lite: A Lightweight, Cost-Aware Hybrid Framework for Credit Card Fraud

Abstract

Credit card fraud detection remains a major challenge for financial institutions due to severe class imbalance, evolving fraud strategies, and the need for rapid decisionmaking. Many existing approaches primarily focus on improving overall predictive accuracy, often overlooking business costs and performance across different risk segments. To address these gaps, this study introduces SmartFraud-Lite (SFL), a lightweight hybrid framework designed for costsensitive fraud detection. The framework incorporates three key components: (1) SHAP-Guided Pruning (SGP), which leverages SHAP values to identify and retain the most relevant features; (2) Cost-Aware Threshold Tuning (CATT), which adjusts classification thresholds based on business cost considerations; and (3) Risk-Tier Stratified Evaluation (RTSE), which evaluates model performance across distinct risk categories. Experiments conducted on the ULB dataset (284,807 transactions with 0.17% fraud cases) demonstrate that SmartFraud-Lite achieves a ROC-AUC of 0.9773 and a PR-AUC of 0.7729, outperforming established models such as XGBoost (0.9614) and LightGBM (0.9589). In addition, the proposed framework yields the lowest overall business cost (690) while maintaining minimal latency overhead.