Adaptive Class-Imbalance Learning for High-Sensitivity Fraud Detection in Large-Scale Digital Transaction Networks

Abstract

This study addresses the persistent problem of severe class imbalance in large-scale digital transaction networks, where fraudulent transactions represent only a small minority of total observations and conventional accuracy-focused classifiers can overlook genuine fraud cases, increase false negatives, and weaken minority-class sensitivity. The purpose of the research is to quantitatively evaluate how adaptive class-imbalance learning capabilities influence High-Sensitivity Fraud Detection Performance across banking, fintech, electronic payment processing, digital-wallet, e-commerce, and related enterprise transaction environments. A quantitative, cross-sectional, case-study-based design was adopted, using purposive sampling and a five-point Likert-scale questionnaire administered to professionals in fraud analytics, machine learning, cybersecurity, transaction monitoring, payment risk, and digital transaction management. Of 320 questionnaires distributed, 296 were returned and 288 valid responses were retained, producing a 90.0% usable response rate. The key independent variables were Adaptive Minority-Class Resampling, Cost-Sensitive Learning, Dynamic Classification-Threshold Optimization, and Adaptive Ensemble Learning, while High-Sensitivity Fraud Detection Performance was the dependent variable. Analysis included descriptive statistics, Cronbach’s alpha, KMO and Bartlett’s tests, Pearson correlation, multiple regression, ANOVA, tolerance, VIF, and Durbin-Watson diagnostics. High-Sensitivity Fraud Detection Performance Recorded M = 4.19, SD = 0.54. Dynamic Classification-Threshold Optimization produced the strongest relationship, r = .74, p < .001, and the strongest regression effect, β = .32, p < .001. The overall model achieved R = .842, R² = .709, adjusted R² = .705, F (4, 283) = 172.45, p < .001, explaining 70.9% of the variance and supporting all five hypotheses. The findings imply that transaction-network operators should integrate adaptive threshold optimization, cost-sensitive learning, resampling, and ensemble methods to improve fraud sensitivity while controlling false-positive burdens.