Effect of Sampling Techniques on Fraud Detection Performance: An Empirical Study with XGBoost

Fraud detection in financial transactions is critical yet challenging due to extreme class imbalance, where fraudulent activities constitute a very small fraction of overall data. This imbalance tends to bias classifiers towards the majority class, resulting in poor detection of fraud cases. This study investigates the impact of several sampling techniques—Random Under Sampling (RUS), Synthetic Minority Oversampling Technique (SMOTE), Adaptive Synthetic Sampling (ADASYN), Tomek Links, and SMOTE combined with Tomek Links—on the performance of the XGBoost classifier for credit card fraud detection. Using the publicly available Kaggle credit card fraud dataset, the models were evaluated based on precision, recall, F1-score, and area under the ROC curve (AUC). Results show that hybrid oversampling and cleaning methods such as SMOTE+Tomek offer a balanced trade-off between recall and precision, while undersampling achieves high precision but lower recall. This work provides practical insights for selecting sampling strategies to improve fraud detection systems.