Model performance optimisation in credit card fraud detection using class imbalance techniques, feature engineering and feature selection techniques

Abstract

Fraud detection in financial datasets, particularly in credit card transactions, presents a significant challenge due to the prevalence of irrelevant features and class imbalances. Addressing these issues is crucial for optimizing model performance and accurately identifying fraudulent activities. This research focuses on the application of feature engineering, class imbalance handling techniques alongside a comparative analysis of feature selection techniques such as Chi-Square, ANOVA, (Recursive Feature Elimination) RFE, and (Information Gain) IG all in bid to find the best combination of techniques that enhance model accuracy in (Credit Card Fraud Detection) CCFD. To mitigate class imbalances, Synthetic Minority Oversampling Technique) SMOTE , (Synthetic Minority Oversampling Technique With Edited Nearest-Neighbours) SMOTE-EEN, and simple oversampling were employed. These methods aimed to balance the class distribution, improving the models’ ability to detect fraud. Popular classification models, including Decision Trees, KNN, AdaBoost, and XGBoost, were trained on datasets that had undergone feature engineering, class imbalance techniques and feature selection all in bid to produce optimized model performances. The study utilized evaluation metrics like F1-score, Balanced Accuracy and ROC-AUC to assess model performance and the results demonstrated how feature engineering, combined with specifically SMOTE-EEN as the class imbalance handling technique alongside strategic ANOVA or Chi-Square Test, significantly improved the accuracy and robustness of the fraud detection models with accuracy scores of over 90% across the four classifiers on the four datasets. These findings will thus help provide valuable insights for industry researchers in selecting the most effective techniques for optimizing model performance in fraud detection studies.