Great explanations!

A noticed a small mistake in the Kernel SHAP section. When converting z' to the original feature space, we use h⁻¹ (the inverse of h mapping function, as in the paper). Same error in the loss function definition.

Cheers