AI-based credit card fraud detection: a machine learning approach with model explainability on real-world data

Abstract Fraud detection is a critical challenge in the financial industry, as the growing volume of transactions increases the risk of fraudulent activities, leading to significant financial and reputational harm. This study aims to address this challenge by evaluating advanced machine learning and deep learning models designed to enhance the accuracy, efficiency, and scalability of fraud detection systems, ultimately contributing to more secure financial operations. The study evaluates nine machine and deep learning models encompassing both supervised and unsupervised learning techniques across four sampling techniques: imbalanced data, random undersampling, synthetic minority oversampling technique (SMOTE), and majority voting with random undersampling on a private and a public dataset. Each model’s effectiveness is compared and analyzed, highlighting their suitability for fraud detection tasks. The study also integrates explainability techniques to uncover the most influential features contributing to fraud detection, by using feature importance, Shapley additive explanations (SHAP), Local Interpretable Model-Agnostic Explanations (LIME), and deep learning important features (DeepLIFT). The results demonstrate the efficacy of the evaluated models in detecting fraudulent activities, with Random Forests achieving the best performance among machine learning models and TabTransformer leading among deep learning models. While hyperparameter tuning enabled XGBoost to achieve a statistically superior F1-score, Random Forest demonstrated remarkable inherent robustness and stability, maintaining high performance even without optimization. Consequently, Random Forest was selected over XGBoost as the representative model for in-depth interpretability analysis. This research contributes to the financial industry’s efforts in combating fraud by presenting a comprehensive evaluation of models, identifying optimal strategies for data sampling, feature engineering, and enhancing model interpretability to support informed decision-making.