Explainable UPI Fraud Detection Using XGBoost and SHAP-Based Interpretable AI

Abstract

The rapid growth of Unified Payments Interface (UPI) transactions in India has created an urgent need for intelligent fraud detection systems capable of operating at scale in real time. This paper presents an explainable UPI fraud detection system combining XGBoost gradient boosting with SHAP (SHapley Additive exPlanations)-based interpretable AI. Raw transaction data from the PaySim synthetic dataset (6.36 million transactions, 0.13% fraud rate) is transformed into 25+ engineered features spanning temporal patterns, amount behaviour, account balance anomalies, and user behavioural profiles. A temporal train/validation/test split preserves chronological integrity. Cost-sensitive training via XGBoost’s scale_pos_weight parameter compensates for the extreme class imbalance. The trained model achieves Precision 97.6%, Recall 91.5%, F1-Score 94.5%, and AUC-ROC 0.99 on the held-out test set. A Streamlit-based web dashboard provides both single-transaction real-time analysis with SHAP explanations and bulk batch processing via CSV upload, delivering sub-200 ms inference latency suitable for payment gateway integration.