Hybrid Machine Learning-Based Data Mining Framework for Early Fraud Detection in Digital Transactions

Abstract

The exponential growth of global electronic payment systems has simultaneously catalysed a dramatic surge in sophisticated digital transaction fraud. Traditional rule-based engines and standalone machine learning classifiers face critical bottlenecks: they either struggle with high false-alarm rates or lack real-time computational scalability when handling massively imbalanced streaming datasets. This paper presents a novel, two-stage HybridMachine Learning-Based Data Mining Framework designed specifically for the ultra-early isolation of fraudulent patterns. The first stage employs unsupervised Autoencoders to conduct low-latency feature extraction and isolate non-linear transaction anomalies from heavily skewed datasets. The second stage feeds these optimal feature mappings into an optimized ensemble pipeline featuring Light Gradient Boosting Machine (LightGBM) to execute precise fraud classification. Tested against the standard benchmark PaySim mobiletransaction dataset, our hybrid architecture achieves a fraud classification accuracy of 98.4% and an overall recall of 97.8%, while cutting transaction assessment latency to just 2.6milliseconds. These outcomes systematically outperform legacy standalone classifiers, presenting an adaptive, risk-aware security layer ideal for high-throughput digital banking ecosystems.