Abstract
The COVID-19 pandemic has limited people's mobility to a certain extent, making it difficult to purchase goods and services offline, which has led the creation of a culture of increased dependence on online services. One of the crucial issues with using credit cards is fraud, which is a serious challenge in the realm of online transactions. Consequently, there is a huge need to develop the best approach possible to using machine learning in order to prevent almost all fraudulent credit card transactions. This paper studies a total of 66 machine learning models based on two stages of evaluation. A real-world credit card fraud detection dataset of European cardholders is used in each model along with stratified K-fold cross-validation. In the first stage, nine machine learning algorithms are tested to detect fraudulent transactions. The best three algorithms are nominated to be used again in the second stage, with 19 resampling techniques used with each one of the best three algorithms. Out of 330 evaluation metric values that took nearly one month to obtain, the All K-Nearest Neighbors (AllKNN) undersampling technique along with CatBoost (AllKNN-CatBoost) is considered to be the best proposed model. Accordingly, the AllKNN-CatBoost model is compared with related works. The results indicate that the proposed model outperforms previous models with an AUC value of 97.94%, a Recall value of 95.91%, and an F1-Score value of 87.40%.