Explainable AI in Financial Fraud Detection: Evidence from the IEEE-CIS Fraud Dataset

Financial fraud detection systems increasingly rely on advanced machine learning models, whose decisions should remain easily interpretable, in order to make transparent decisions and to align with regulatory requirements. In this study, we explore explainable AI methods for fraud detection, using the IEEE-CIS Fraud detection database. Thus, we compare the results obtained using tabular models (XGBoost, CatBoost) with those generated by graph-based models (GraphSAGE) and observe that although gradient boosting models have a better performance - measured in terms of ROC-AUC -, graph-based models capture relational structures, which cannot be described by tabular models. We propose a multi-level explainability framework, which combines methods of explaining model variables (SHAP values) with graph analysis methods, such as edge and node ablation. The results highlight a trade-off between discriminatory power and the interpretation of interactions between network nodes. Thus, although tabular models are better at identifying risk at feature level, graph-based models allow for a better interpretation of transactional fraud patterns. These results demonstrate that explainable AI in fraud detection should not be based on a single mpdel, but it should combine several types of models that capture both the relational structure and the predictive power at feature level.