<![CDATA[E-commerce has increased the productivity of international trade, causing an increase in credit card fraud that has damaged finances and weakened public confidence in digital payment systems. This study aims to improve the sensitivity and reliability of fraud detection on highly imbalanced transaction data through the design and evaluation of the DXCL model compared to conventional individual and ensemble models. This methodology uses resampling approaches such as random undersampling and SMOTE oversampling during training to reduce class imbalance. DXCL's performance is evaluated against six benchmark models Random Forest, standalone DNN, XGBoost, LightGBM, CatBoost, and a dummy classifier utilizing accuracy, precision, recall, F1-score, ROC-AUC, TPR, and FPR metrics on a 2013 European credit card transaction dataset. The results prove that DXCL outperforms individual models and Random Forest in effective rate while eliminating false positive rate with 99.98% accuracy, precision, recall, F1-score, and ROCAUC of 1.00. Deep feature extraction and ensemble enhancement significantly improve fraud examination of class imbalanced transaction datasets. DXCL supports the application of a more dependable approach for detecting credit card fraud with low false positive rates in highly imbalanced digital transaction environments]]>
Copyrights © 2025
