This study presents a comparative evaluation of multiple machine learning models for credit card fraud detection, including Logistic Regression, Support Vector Machine (SVM), Random Forest, XGBoost, Multi-Layer Perceptron (MLP) Neural Network, and a Neural Network built using TensorFlow/Keras. Each model underwent a comprehensive methodology involving data preprocessing, feature scaling, and undersampling to address class imbalance. Model performance was evaluated using a variety of metrics accuracy, precision, recall, F1 score, and AUC-ROC to capture their effectiveness in distinguishing fraudulent transactions. Results indicated that XGBoost achieved the highest accuracy at 96.6% with a perfect precision score, while MLP and TensorFlow/Keras neural networks demonstrated excellent AUC scores (0.982 and 0.977, respectively), reflecting strong discrimination between classes. The SVM, Logistic Regression, and Random Forest models also showed competitive performance, with high precision and recall values. This analysis highlights the effectiveness of ensemble and neural network models for fraud detection in imbalanced datasets and emphasizes the importance of metric selection based on application-specific priorities, particularly for balancing false positive and false negative outcomes.
Shi et al. (Tue,) studied this question.