A refactor removes the background argument while leaving featureperturbation=auto. Why can explanations change with an…
Classic MLLevel: IntermediateSHAP · TreeExplainer
A refactor removes the background argument while leaving feature_perturbation=auto. Why can explanations change with an unchanged tree model?
Hints
- Auto is a conditional choice, not one fixed algorithm.
- Compare the explanation setup as well as the model checksum.
Answer
Auto selects interventional with background data and tree_path_dependent without it. The dependence/reference assumptions changed, so identical predictions need not imply identical attributions. Set the mode explicitly for reproducible comparisons.
Source and licence: SHAP · TreeExplainer · MIT · answer: Quantoros editorial