Optimal Weight-Tuning for Unbalanced Data in Credit Card Fraud Detection

With the proliferation of technology and the expansion of e-commerce services, credit cards have emerged as one of the foremost payment methods, leading to a surge in banking transactions. However, this growth has also brought about a corresponding increase in fraudulent activities, necessitating heightened vigilance and increased banking transaction costs. In this study, we explore the efficacy of class weight-tuning hyperparameters in regulating the balance between legitimate and fraudulent transactions. Specifically, we employ Bayesian optimization to fine-tune these hyperparameters, taking into consideration practical challenges such as imbalanced data. Our approach involves utilizing weight-tuning as a preprocessing step for addressing unbalanced datasets, alongside leveraging Cat Boost and XG Boost to augment the efficiency of the LightGBM method by integrating a voting mechanism. Furthermore, we incorporate deep learning to further refine the hyperparameters, particularly focusing on our proposed weight-tuning technique. Real-world data is utilized in conducting experiments to evaluate the efficacy of the proposed methods. In addition to traditional metrics like ROC-AUC, we employ recall-precision metrics to provide a more comprehensive assessment of performance, particularly for unbalanced datasets. Through 5-fold cross-validation, we individually evaluate Cat Boost, LightGBM, XGBoost, and logistic regression. Moreover, we employ a majority voting ensemble learning technique to assess the combined performance of these algorithms. Our results demonstrate that the proposed methods outperform state-of-the-art approaches, yielding significant improvements in fraud detection performance.