Design and Investigation of Various ML Techniques for Credit Card Fraud Detection

Abstract

This study evaluates the effectiveness of various machine learning models for detecting fraudulent online credit card transactions. It identifies that traditional methods are insufficient due to the high volume of legitimate transactions and evolving attack techniques. The research involved gathering and pre-processing transaction records to minimize outliers. Three predictive modeling methods were developed: an ensemble approach utilizing bagging for accuracy, a Support Vector Machine (SVM) employing radial basis functions, and an entropy-based decision tree. Each model's flexibility and effectiveness were rigorously tested against unseen transaction data. The decision tree achieved a 99.91% correctness rate, missing 40 fraudulent transactions while incorrectly flagging 33 legitimate ones. The SVM model showed a slight decline at 99.84%. In contrast, the ensemble bagging method garnered the highest success rate at 99.95%, correctly identifying 109 genuine positives and 85,293 negatives, while missing 38 frauds and misclassifying only two real transactions. This approach outperformed previous benchmarks like KNN and Random Forest, demonstrating superior speed and reliability for real-time fraud detection in digital payments through model stacking.