Background: SHapley Additive exPlanations (SHAP) methods are widely used to interpret machine learning models, yet most implementations assume feature independence. This assumption rarely holds in practice, especially when features are correlated, leading to biased and unstable attributions.

Objective: We introduce Corr-SHAP, a correlation-aware SHAP approach that produces more faithful and stable feature attributions by explicitly modeling feature dependencies. Our aim is to enhance the accuracy, robustness, and scalability of SHAP explanations for models trained on correlated data.

Methods: Corr-SHAP models feature correlations via a multivariate Gaussian approximation with a Ledoit–Wolf covariance estimator. We design a correlation-aware sampling distribution that penalizes redundant coalitions, improving computational efficiency in higher dimensions. To correct the induced bias, we employ a Self-Normalized Importance Sampling estimator, which re-weights samples by the ratio of the true Shapley kernel to the sampling probability. Our analysis establishes high probability error bounds in terms of Effective Sample Size, extending convergence guarantees to correlated feature spaces.

Results: Across synthetic and real-world datasets, Corr-SHAP achieves Shapley value estimates that closely align with Kernel SHAP, while exhibiting substantially lower variance and more stable feature rankings. In correlated clusters, Corr-SHAP systematically down-weights redundant features, improving ranking fidelity without introducing bias. To further support scalability, we demonstrate that combining Corr-SHAP with Leverage-SHAP reduces variance in higher-dimensional settings.

Conclusion: Corr-SHAP provides a statistically grounded and computationally efficient framework for SHAP value estimation under feature correlation. By integrating correlation modeling, bias correction, and variance reduction, it scales beyond small toy problems and delivers explanations that are both accurate and reliable, making it a valuable tool for practitioners analyzing complex real-world datasets.

Keywords: Explainable artificial intelligence; XAI; SHapley Additive exPlanations; Feature correlation; Model interpretability; Importance sampling; Variance reduction.