Beyond Additivity: Sparse Isotonic Shapley Regression toward Nonlinear Explainability

Shapley values, a widely-used tool for explaining AI decisions, often produce inaccurate results when real-world data violates their core assumption of additivity—a mathematical property that rarely holds in practice due to non-Gaussian distributions, heavy-tailed data, feature dependencies, and domain-specific factors.

The paper introduces SISR (Sparse Isotonic Shapley Regression), which automatically learns how to adjust explanations to account for non-additive real-world conditions, eliminating the need for manual specification and making Shapley-based explanations more reliable for practitioners.

SISR addresses a practical efficiency problem: current methods require computing explanations for all features before removing unimportant ones, which is computationally expensive—SISR instead identifies important features directly, reducing computational costs in high-dimensional settings where thousands of features exist.

The research demonstrates that seemingly unrelated factors like irrelevant features and dependencies between features can fundamentally distort how much credit each feature deserves, suggesting that standard Shapley explanations may systematically misattribute importance in ways practitioners haven't previously recognized.

The proposed method comes with mathematical guarantees for finding optimal solutions and recovering true feature importance, providing more rigorous foundations for using these explanations in high-stakes decisions where accuracy of feature attribution matters.