PolySHAP Improves Shapley Value Estimates via Polynomial Regression
A new method called PolySHAP extends KernelSHAP by using higher-degree polynomials to approximate Shapley values, capturing non-linear interactions between features. The approach yields empirically better estimates on benchmark datasets and is proven consistent. PolySHAP also connects to paired sampling (antithetic sampling), a common modification that improves KernelSHAP's accuracy. The work is published on arXiv (2601.18608) and addresses the exponential cost of exact Shapley value computation, which requires 2^d evaluations for d features.
Key facts
- PolySHAP extends KernelSHAP using higher-degree polynomials.
- It captures non-linear interactions between features.
- Empirically better Shapley value estimates on benchmark datasets.
- Proven consistent estimates.
- Connects to paired sampling (antithetic sampling).
- Exact Shapley value computation requires 2^d evaluations.
- KernelSHAP approximates Shapley values via linear function.
- Published on arXiv with ID 2601.18608.
Entities
Institutions
- arXiv