UPI Fraud Detection Using Hybrid Machine Learning Models with Explainable Risk Scoring and Real-Time Monitoring

Abstract: Unified Payments Interface has become a major transaction channel in day-to-day digital payments, which makes transaction security an important technical concern. Fraud detection in such environments is difficult because suspicious transactions form only a very small portion of the total data, while legitimate user activity can vary widely in amount, timing, and transaction type. This work presents a hybrid machine learning framework for UPI fraud detection that combines a baseline Random Forest model with an improved XGBoost-based detection model, supported by imbalance handling, engineered transaction features, explainable prediction analysis, and a real-time monitoring dashboard. The system is designed as a complete end-to-end pipeline consisting of dataset preparation, preprocessing, feature transformation, model training, fraud scoring, API-ready prediction flow, and interactive visualization. In the implementation, transaction attributes are transformed into a compact feature set that includes temporal behaviour, transaction-category indicators, and fraud-rate information by transaction type. Synthetic Minority Over-sampling Technique is used to reduce the effect of class imbalance during training. The trained model produces a fraud probability score, and a decision threshold of 0.6 is used for final classification. To improve transparency, SHAP-based feature explanations are integrated so that important factors behind a prediction can be viewed instead of treating the model as a complete black box. A Streamlit dashboard was developed to support manual fraud checking, risk-gauge visualization, live transaction simulation, feature-importance display, and monitoring analytics. Experimental results show strong discrimination between legitimate and suspicious transactions. From the recorded confusion matrix, the system correctly identifies 314224 legitimate transactions and 168 fraudulent transactions, with zero false positives and 181 false negatives. These results indicate that the proposed framework is highly reliable for safe-transaction confirmation and practically useful for risk-aware UPI monitoring applications. The work demonstrates that combining ensemble learning, explainability, and dashboard-based monitoring can provide a more usable fraud detection system than rule-based screening alone.

Keywords: UPI fraud detection, XGBoost, Random Forest, Explainable AI, SHAP, SMOTE, Streamlit dashboard, financial transaction security