Credit card fraud classification using applied machine learning – a comparative study of 24 machine learning algorithms

Abstract

This paper presents a comprehensive study on credit card fraud detection, addressing the escalating issue of fraudulent activities that significantly impact both financial institutions and consumers. We introduce a novel framework for evaluating the collective performance of diverse machine learning (ML) models—including Logistic Regression, Decision Trees, Random Forests, Support Vector Machines, and Neural Networks—using a synthetic dataset carefully constructed to mirror real-world transaction features and behavioral patterns. By applying various sampling strategies to this highly imbalanced dataset and leveraging domain knowledge for feature selection, this study aims to enhance both the accuracy and stability of fraud detection models, while identifying the minimum feature set required for optimal detection speed and efficiency. Our results reveal that algorithms such as Gaussian Naive Bayes, Kernel Naive Bayes, Cubic SVM, and Trilayered Neural Networks each provide strong, balanced performance. Building on these findings, we propose that ensembling these top-performing models could further improve detection rates and reliablity, harnessing their complementary strengths to achieve superior overall performance. This paper underscores the necessity of advanced and integrated ML techniques for robust, timely fraud detection, offering valuable insights for real-time implementation and presenting a comprehensive solution to a pressing financial security challenge.