KernelSHAP explains one prediction by approximating each feature’s Shapley value with a weighted linear surrogate. It samples feature coalitions, fills missing features with realistic values from other examples, and weights the resulting surrogate examples with the Shapley kernel. section 9.4.2
A practical KernelSHAP workflow is:
- Represent each coalition by included and excluded features, using realistic replacement values for excluded features. section 9.4.2
- Weight each example with the Shapley kernel, then fit linear regression; its coefficients approximate the feature Shapley values. section 9.4.2
Coalitions as a diagnostic
Treat features like members of a team: compare predictions from coalitions with and without a feature, then use weighted differences to estimate its contribution. section 9.4.2
