Dynamic explainability-constrained hyperparameter tuning for transparent credit card fraud risk scoring

We propose a Dynamic Explainability-Constrained Hyperparameter Tuning Framework (DXHTF) for credit card fraud risk scoring, which addresses the critical trade-off between predictive performance and interpretability in high-stakes financial applications. Conventional fraud detection systems frequently emphasize precision while sacrificing clarity, which hinders analysts’ ability to trust and respond to the model outputs. The proposed method embeds real-time explainability metrics obtained from SHAP (SHapley Additive exPlanations) within the hyperparameter optimization loop so that model modifications uphold both detection performance and interpretability. At its core, the DXHTF dynamically adjusts the regularization parameters based on the stability of feature attributions, measured through a novel SHAP variance monitor, while a multi-objective controller balances these constraints against predictive loss. Furthermore, the framework introduces an explainability-aware validation tier that supplements conventional performance metrics by including robustness assessments for feature importance consistency. The DXHTF employs a gradient boosting machine and is refined through Bayesian optimization, dynamically adjusting to changing fraud behaviors while upholding the transparency standards required by regulations. The key advancements consist of assigning weights to domain-essential features and a closed-loop feedback system that updates the explainability constraints almost instantaneously. Experimental testing on actual transaction data shows that the framework attains a comparable fraud detection performance while preserving the stability and clarity of its explanatory outputs. This study bridges a longstanding gap in financial machine learning, where interpretability is often an afterthought, by embedding it as a first-class constraint in the model development lifecycle.