The effects of class rebalancing techniques on ensemble classifiers on credit card fraud detection: An empirical study

Millions of dollars in financial fraud losses can be minimized, and in some cases, completely avoided by implementing the appropriate fraud prediction model. Combining a suitable rebalancing strategy with data mining techniques on a large dataset can enhance the prediction model for credit card fraud. The objective of this study is to investigate the impact of sampling techniques on ensemble classifiers for constructing credit card default prediction models. To decide which combination of rebalancing technique and ensemble classifier works best on skewed datasets for credit card fraud detection, in this paper, we investigate and assess the performance of no sampling, random under sampling, Tomek link removal, random oversampling, SMOTE, and a combination of SMOTE and Tomek link removal using ensemble classifiers including XGBoost, LightGBM, and Random Forest. For evaluating the best combination of rebalancing technique and ensemble classifier, we have used precision, recall, f1 score, mcc, PR-AUC curve and ROC-AUC curve as evaluation metrics. Based on overall evaluation matrics Random Forest, XGBoost perform best when paired with Tomek link removal, and LightGBM performs best when paired with random oversampling. All evaluation metrics of our empirical study indicate that Tomek link removal with Random Forest works best among all the different combinations of rebalancing techniques and ensemble classifiers for predicting fraudulent credit card transactions.