Applying explainable artificial intelligence to interpret supervised ensemble learning models for robust credit card...

As the usage of digital financial transactions continue to swell, it becomes all the more crucial to employ classifiers with machine learning techniques in order to process credit card fraud detection. While complex ensemble models can reach incredible levels of predictive accuracy, the black-box nature of these algorithms often leaves us at a loss, and there hasn't been much research on how interpretable these high-performing models are across different settings. To contribute to this research gap, this study assesses four supervised learning algorithms (Logistic Regression, Random Forest, XGBoost and LightGBM) based on their predictive performance and applicability with Explainable Artificial Intelligence (XAI) techniques. To optimize the generalizability of their findings, the models were extensively tested and compared against three disparate public credit card transaction datasets. The performance, as measured by different metrics such as accuracy, precision, recall, F1-score and ROC-AUC gives the best results to tree-based algorithms ensembles (especially XGBoost) with linear methods also providing decent improvement over baseline models. Data used as input for training consisted of the top-performing models, and then SHAP (SHapley Additive exPlanations) framework was applied to help identify leading feature importance and interpret complicated predictive output. This study provides a comprehensive outline linking predictive performance to explainability in every combination of models, yielding impactful results for developing effective, transparent and accountable financial security systems.