Interpretable Machine Learning for Fraud Detection in Large-Scale Financial Transactions

Financial fraud detection is characterized by extreme class imbalance and the need for high explainability. This study implements an interpretable framework using a dataset of 18 million transactions. By employing Logistic Regression with balanced class weights and extensive behavioral feature engineering, we achieved a ROC-AUC of 98.93% and a PR-AUC of 55.73%. The results demonstrate that transparent models, when supported by logical behavioral features, can effectively isolate fraudulent activities while providing clear risk scores for financial auditing.