PolySHAP: Extending KernelSHAP with Interaction-Informed Polynomial Regression

Shapley values are computationally expensive but important for AI transparency: The paper addresses a fundamental challenge in explainable AI—computing Shapley values (a method for determining which features most influence a model's decisions) normally requires an impractical number of calculations that grows exponentially with the number of features.

PolySHAP improves upon existing approximation methods by capturing feature interactions: While the current leading approach (KernelSHAP) treats features as independent, PolySHAP uses polynomial regression to account for how features interact with each other, resulting in more accurate explanations of AI model behavior across various benchmark datasets.

A popular practical technique now has theoretical backing: The paper provides the first rigorous mathematical explanation for why "paired sampling"—a commonly-used modification to KernelSHAP that practitioners have observed works well—actually produces accurate results, validating an approach that was previously used without strong theoretical justification.

Better explanations are achievable without proportionally higher computational costs: PolySHAP achieves improved accuracy in explaining AI decisions while maintaining computational efficiency, making more reliable model interpretability accessible for real-world applications.