Enterprise AI Analysis

Explainable AI Framework for Credit Card Fraud Detection Using XGBoost, Deep Neural Networks and SHAP-Based Interpretability

Financial fraud continues to threaten the stability of global financial ecosystems, particularly with the increasing growth of digital transactions. Traditional rule-based fraud detection methods struggle to adapt to evolving fraud patterns, severe class imbalance, and increasing regulatory demands for transparency. This study proposes an explainable artificial intelligence (XAI) framework for credit card fraud detection that integrates an optimised XGBoost classifier and a deep neural network (DNN), supported by SHAP-based interpretability. Using a widely recognised real-world European credit card fraud dataset, two models were developed and rigorously evaluated. To mitigate extreme class imbalance, random undersampling was applied, and model performance was assessed using precision, recall, F1-score, ROC AUC, PR AUC, and Matthews Correlation Coefficient (MCC). The DNN achieved superior performance with an accuracy of 0.95, PR AUC of 0.987 and MCC of 0.909, while XGBoost demonstrated strong stability and interpretability with PR AUC of 0.980 and MCC of 0.865. SHAP was used to provide both global and local explanations, ensuring transparent and auditable decision support. The results demonstrate that high-performance fraud detection models can be effectively combined with explainability to support trustworthy, regulation-aligned financial analytics. The study highlights the need for explainable AI integration in fraud detection pipelines and outlines future opportunities for real-time deployment and adaptive fraud learning.