Machine Learning-Based Fraud Detection and Risk Mitigation Framework for High-Volume Digital Transactions

Abstract

Increased levels of digital transactions have created the need for more intelligent approaches to fraud detection and risk mitigation in enterprise digital payment applications. Rule-based approaches in fraud detection usually suffer from high levels of false positives, delayed responses and little adaptability to evolving fraud patterns. In this paper, an Adaptive Intelligent Risk and Fraud Detection (AIR-FD) Framework is proposed, which makes use of prediction, anomaly detection, adaptive risk scoring and automated decision-making response for enterprise digital transactions. A Design Science Research Methodology (DSRM) approach with experimental evaluation was used to design and test the AIR-FD Framework. For the experimental testing, the dataset comprised of 500,000 digital transaction records gathered from banking, eCommerce, mobile payments, authentication and transaction logs environment. The dataset processing involved duplication removal, missing values handling, normalizing the timestamp and behavioral features extraction. Classification and detection of fraud were consistently performed using a unified architecture comprising Random Forest, XGBoost, Isolation Forest, and LSTM models. Machine learning based model obtained 94.0% of accuracy in fraud detection, 92.8% of precision, 91.6% of recall, and 92.2% of F1-score, which is 18.7 percentage points higher than that of rule-based approach. False positive rate decreased from 11.2% to 6.1%. In addition, detection latency was decreased from 42.6 seconds to 18.4 seconds. Meanwhile, automated risk mitigation process has decreased response time from 76 seconds to 14 seconds. The results showed that the combination of machine learning with anomaly detection and adaptive risk scoring greatly improves the security, decision-making processes, operation speed and financial risks mitigation of enterprises.