Explainable and Cost-Sensitive Machine Learning for Fraud Prevention in U.S. Banking: Threshold Optimization, SHAP Explainability, and Fairness Analysis.

Abstract

Credit card fraud is a major challenge for U.S. financial institutions. Banks need fraud detection systems that can identify fraudulent transactions, reduce false alarms for legitimate customers, and meet regulatory expectations. This study develops an explainable and cost-sensitive machine learning model for detecting credit card fraud using the publicly available Sparkov synthetic credit card transaction dataset from Kaggle. The dataset contains 1,296,675 transactions for training and 555,719 transactions for testing. The research creates and tests four kinds of machine learning models: XGBoost, LightGBM, random forest, and logistic regression. Instead of selecting the model based on accuracy, the study applies decision-oriented threshold optimization to minimize the total economic cost of classification errors. The cost assumes $1,000 for each false negative that misses a fraud case and $50 for each false positive that blocks a legitimate transaction. The results show that XGBoost has the lowest total economic cost of error with a threshold of 0.75 at $355,700 and catches 91.2% of fraud cases. The study uses SHAP values to figure out what the most important things are that lead to fraud prediction and improve model transparency for banking decision-makers. Furthermore, gender-based fairness analysis is conducted for using false negative rates, false positive rates, and true positive rates. The findings show that cost-sensitive optimization, explainable AI, and fairness monitoring can help create more responsible fraud detection systems in the banking sector.