Causal Feature Selection for Explainable AI (XAI) in Fraud Detection: A Systematic Literature Review

Fraud detection systems increasingly rely on machine learning models that offer high predictive accuracy but suffer from limited transparency and interpretability, raising concerns about trust, accountability, and regulatory compliance. In sensitive domains such as financial fraud detection, explainability is not only desirable but necessary to meet legal and ethical requirements, including data protection and fairness regulations such as the General Data Protection Regulation (GDPR). This systematic literature review examines the integration of Causal Feature Selection (CFS) with Explainable Artificial Intelligence (XAI) techniques to address the trade-off between predictive performance and interpretability in fraud detection models. The review synthesizes recent studies that employ Structural Causal Models (SCMs) alongside post-hoc explanation methods such as SHAP and LIME to enhance transparency, robustness, and causal interpretability. Findings indicate that causal-driven feature selection improves model trustworthiness by identifying stable and meaningful relationships rather than spurious correlations. The study further highlights key application domains, existing challenges in causal discovery—particularly scalability and data quality—and the lack of standardized causal benchmarking datasets. Additionally, emerging approaches such as Causal Graph Networks and Counterfactual Reasoning are identified as promising future directions for advancing explainable and regulation-aware fraud detection systems. This review provides researchers and practitioners with structured insights into current methods, limitations, and research opportunities at the intersection of causality and explainable AI in fraud detection.