Utilizing Explainable Machine Learning for Robust and Accurate Fraud Detection in E-Commerce

In the rapidly growing field of e-commerce, fraudulent actions are responsible for significant financial losses, and the challenge of detecting financial fraud remains critical. In this study, we utilize a real dataset, specifically the IEEE-CIS Fraud Detection dataset provided by Vesta Corporation, which contains a diverse mix of real-world e-commerce transactions. We implement supervised machine learning models to identify fraudulent transactions while addressing the common problem of class imbalance using SMOTE and Random Over Sampling methods. Our evaluation employs machine learning algorithms such as Random Forest, XGBoost, and Support Vector Machines (SVM) to assess how various models and oversampling techniques impact detection performance. We also apply explainable AI methods to enhance the clarity of model behaviors and feature significance. Local interpretable methods, including Local Interpretable Model-agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP), and global interpretation methods, such as permutation feature importance and univariate partial dependencies, are implemented to determine features that are crucial for robust prediction. Performance is evaluated using crucial metrics including accuracy, precision, recall, F1-score, and the area under the ROC curve (AUC-ROC). The results demonstrate that both the oversampling strategy and the machine learning model used significantly influence detection effectiveness and provide insights on effective combinations to enhance fraud detection. Overall, this research lays the foundation for designing more effective and interpretable fraud detection systems to help financial institutions mitigate losses and reduce the incidence of fraudulent transactions.