The Best ML Classifier(s): An empirical study on the learning of imbalanced and resampled credit card data

Machine Learning-based fraud detection systems are more effective at detecting financial fraud. Credit card fraud detection is one of them. The datasets used in learning by the machine learning algorithms faced many difficulties due to class-imbalance issues. The aim of this research is to discuss the latest developments in credit card fraud detection and also assess the predictive performance of eight popular machine learning classifiers, such as Logistic Regression, K-Nearest Neighbour, Decision Tree, Random Forest, AdaBoost, Bagging, CatBoost, and LightGBM, on learning from an imbalanced credit card fraud detection dataset. Here, the synthetic minority oversampling technique with the edited nearest neighbor, i.e., SMOTE-ENN, has been used to deal with class imbalance issues. The performance evaluation of each classifier is performed on learning in two situations using a raw dataset and a synthetically resampled, class-balanced dataset. When learning from a raw dataset, Random Forest and CatBoost outperformed all the classifiers. The RF attained the highest recall, G-mean, and AUC as 0.785, 0.886, and 0.892, respectively, while CatBoost attained the highest precision and F1-score as 0.921 and 0.843, respectively. In contrast, when learning from a resampled class-balanced dataset, Random Forest, AdaBoost, and Bagging attained the highest precision of 0.999, while K-Nearest Neighbor, Random Forest, and AdaBoost attained the highest recall of 1.000. On other performance metrics, all classifiers performed well except Logistic Regression and Decision Tree. The findings illustrate that the Random Forest classifier is more effective in imbalanced data as well as resampled data learning.