A Comparative Analysis of Imbalanced Learning Techniques for Optimizing Credit Card Fraud Detection

Abstract

Credit card fraud is a growing concern for financial institutions and consumers, leading to significant financial losses and increased security risks. One of the main challenges in fraud detection is the extreme class imbalance, where fraudulent transactions make up only a tiny fraction of all transactions. This imbalance makes it difficult for machine learning models to correctly identify fraud, as they tend to be biased toward the majority class. This paper explores and compares the implementation of various imbalanced learning techniques, including SMOTE, ROS, Borderline-SMOTE, ADASYN, K-means SMOTE, SMOTE-ENN, SMOTETomek, CT-GAN, and CT-GAN Synthesizer. The goal is to assist in the selection of highperformance imbalanced learning techniques for fraud detection, ensuring its applicability and robustness across imbalanced fraud datasets. Empirical results of extensive experiments with 5 datasets show that traditional oversampling methods like ROS and SMOTE variants, consistently improved model performance when combined with strong classifiers like Random Forest and XGBoost. These methods not only increased recall, ensuring a higher detection rate of fraudulent transactions, but also maintained a favorable balance with precision, reducing the risk of flagging legitimate transactions as fraudulent. In contrast, more advanced techniques, including GAN’s and K-means SMOTE, did not demonstrate the expected improvements. Instead, these methods occasionally introduced variability that did not translate into overall performance gains when compared to the traditional oversampling strategies.