Hierarchical Group-Based Shapley Explanation for Interpretable Financial Fraud Detection

Shapley additive explanations (SHAP) are widely used in explainable AI (XAI) due to their solid theoretical basis. However, their applications in financial fraud detection systems (FDS) are limited by independence assumptions, sampling variance, and poor scalability in correlated, high-dimensional data. To overcome these challenges, we propose a hierarchical explainability framework for financial fraud detection, extending group SHAP with hierarchical Shapley attribution. Features are grouped using pre-trained language model embeddings, and a two-stage attribution process estimates group-level contributions followed by fine-grained feature-level attributions. Our approach enhances stability and interpretability for stakeholders in financial fraud detection, while substantially reducing computational cost. It also enables multi-level explanations, allowing users with different expertise to interpret explanations at suitable levels of detail. The framework is evaluated on large-scale online and ATM transfer datasets from real-world FDS pipelines. Experimental results demonstrate improved accuracy, stronger redundancy mitigation, interpretability, and runtime improvement compared to Kernel SHAP.