Raw Tree SHAP contributions sum to a logit. Is applying sigmoid to each contribution and then adding a valid…
Classic MLLevel: IntermediateSHAP · TreeExplainer
Raw Tree SHAP contributions sum to a logit. Is applying sigmoid to each contribution and then adding a valid probability explanation?
Hints
- A nonlinear transform does not distribute over sums.
- Choose the output space before computing attributions.
Answer
No. Sigmoid is nonlinear, so transforming each term does not preserve additivity. Transform the complete raw sum for the prediction, or use a supported probability-space explanation with its own baseline and contributions.
Source and licence: SHAP · TreeExplainer · MIT · answer: Quantoros editorial