Is KernelShap not Baseline Shapley?

[conorosully]

Is KernelShap not Baseline Shapley?

Question

Is Captum's KernelShap actually implementing Baseline Shapley (BShap) rather than the original marginal KernelSHAP from Lundberg & Lee (2017)?

Key Difference

Original KernelSHAP (marginal):
Samples missing features from training data:
$v(S) = E_{X_{\bar{S}}}[f(x_S, X_{\bar{S}})]$

Makes independence assumption to enable sampling from marginals.
Involves multiple model evaluations per coalition.

Captum's KernelShap:
Uses fixed baseline (default: zeros):
$v(S) = f(x_S, b_{\bar{S}})$
where b is the baseline value

No sampling - just direct evaluation with baseline substitution
One model evaluation per coalition
Equivalent to Baseline Shapley with Shapley kernel weighting

So, the Captum implementation makes no independence assumption. Something I think is important to understand when interpreting the output.

Lundberg & Lee (2017): A Unified Approach to Interpreting Model Predictions
Sundararajan et al. (2020): The Many Shapley Values for Model Explanation