A Systematic Review of Explainable Hybrid AI Approaches for Suspicious UPI Transaction Detection

Abstract

The rapid adoption of Unified Payments Interface has significantly transformed digital payment systems by enabling fast, convenient, and cashless financial transactions. However, the increasing volume of online transactions has also led to a rise in fraudulent activities, creating serious challenges for financial institutions in detecting both known and emerging fraud patterns. Conventional fraud detection approaches primarily rely on individual supervised or unsupervised learning models, which often struggle to effectively identify evolving fraud patterns while maintaining high detection accuracy. To address this challenge, this project proposes an enhanced Hybrid Learning Framework for detecting fraud in UPI transactions by integrating supervised and unsupervised learning models through a weighted risk score fusion strategy. The framework performs data cleaning, behavioral and temporal feature engineering, feature preprocessing using StandardScaler and One-Hot Encoding, class balancing using SMOTE, parallel prediction using a Stacking Classifier and a Deep Autoencoder, weighted risk score fusion, and decision-based fraud classification. The developed system is further enhanced with Explainable Artificial Intelligence techniques using LIME, SHAP, Partial Dependence Plot, and Individual Conditional Expectation improve prediction transparency and model interpretability. A FastAPI-based web application is developed to provide an interactive user interface that performs fraud prediction, displays fraud and legitimate probabilities, risk level, explanation visualizations, and personalized fraud prevention recommendations. Experimental evaluation demonstrates that the proposed Hybrid Stacking–Autoencoder framework achieves an accuracy of 96.8%, precision of 98.2%, recall of 95.3%, F1-score of 96.8%, and a ROC-AUC score of 0.996, outperforming the individual baseline and extended models. The developed framework provides an accurate, reliable, and interpretable solution for intelligent UPI fraud detection and effectively supports secure digital payment systems by combining high predictive performance with transparent decision-making and personalized security guidance. Index Terms: Unified Payments Interface , Fraud Detection, Machine Learning, Hybrid Learning, Anomaly Detection, Explainable Artificial Intelligence, Digital Payments.