A Hybrid Ensemble Learning Framework for Real-Time Fraud Detection and Risk Scoring in UPI Transactions

Abstract

With the rapid growth of India's Unified Payments Interface (UPI) as the country's dominant real-time digital payment rail, there has been a corresponding rise in sophisticated financial fraud, including phishing campaigns, social engineering attacks, unauthorized collect requests and QR code manipulation. Conventional rule-based detection systems are inherently limited when confronted with new and evolving fraud typologies, which motivates the adoption of adaptive, data-driven countermeasures. This paper proposes a Hybrid Ensemble Learning Framework (HELF) that unites four complementary machine learning algorithms, Logistic Regression, Random Forest, XGBoost and Isolation Forest within a dynamic, weighted risk-score fusion architecture. A continuous composite threat score, derived jointly from supervised probabilistic outputs and an unsupervised anomaly signal, drives a graduated three-tier verdict system that categorizes transactions as Legitimate, Suspicious or Fraudulent, thereby supporting both automated blocking and human analyst escalation. HELF was evaluated on a curated UPI transaction dataset of 147,382 records and achieved an accuracy of 98.6%, precision of 97.9%, recall of 99.1% and F1-score of 98.5%, outperforming all single-model and uniform-voting ensemble baselines. The framework records zero false negatives at the Fraudulent verdict tier, a property of substantial operational value in financial security settings, and its risk-score fusion formula remains fully auditable, satisfying the interpretability expectations of regulatory frameworks that govern digital payment systems.