Web-Based Explainable Credit Card Fraud Detection Using SMOTE, Ensemble Feature Selection, and XGBoost

Abstract: The rapid growth of digital has enhanced the risk of fraud using the credit cards, resulting into the massive financial losses. This paper describes an explainable fraud detector system based on machine learning and Explainable Artificial Intelligence (XAI) as a web-based system. Synthetic Minority Oversampling Technique (SMOTE) is employed in the management of class imbalance in transaction data. A voting-based feature selection strategy that combines the importance of Random Forest with L1-regularized Logistic Regression (LASSO), and Chi-Square test is used to select the most significant features. An XGBoost classifier is trained on the selected features in order to predict fraud effectively. The system is implemented in the form of Flask web application, which allows real-time entry of transactions and uploading of CSV file. All transactions are categorized as fraudulent or normal with a probability score attached to it. Many features of SHAP are used to guarantee transparency including global and local feature importance, and LIME generates explanations on an instance level. The system presented has attained a high detection performance with interpretability and real life application which renders it appropriate in a financial fraud monitoring and decision support in the real world.

Keywords: credit card fraud detection, XGBoost, SMOTE, explainable AI, SHAP, LIME, web-based system.