Optimizing Multilayer Perceptron with Cost-Sensitive Learning for Addressing Class Imbalance in Credit Card Fraud Detection

The increasing use of credit cards in global financial transactions offers significant convenience for consumers and businesses. However, credit card fraud remains a major challenge due to its potential to cause substantial financial losses. Detecting credit card fraud is a top priority, but the primary challenge lies in class imbalance, where fraudulent transactions are significantly fewer than non-fraudulent ones. This imbalance often leads to machine learning algorithms overlooking fraudulent transactions, resulting in suboptimal performance. This study aims to enhance the performance of Multilayer Perceptron (MLP) in addressing class imbalance by employing cost-sensitive learning strategies. The research utilizes a credit card transaction dataset obtained from Kaggle, with additional validation using an e-commerce transaction dataset to strengthen the robustness of the findings. The dataset undergoes preprocessing with RUS and SMOTE techniques to balance the data before comparing the performance of baseline MLP models to those optimized with cost-sensitive learning. Evaluation metrics such as accuracy, recall, F1 score, and AUC indicate that the optimized MLP model significantly outperforms the baseline, achieving an AUC of 0.99 and a recall of 0.6. The model's superior performance is further validated through statistical tests, including Friedman and T-tests. These results underscore the practical implications of implementing cost-sensitive learning in MLPs, highlighting its potential to significantly enhance fraud detection accuracy and offer substantial benefits to financial institutions.