CatBoost returns SHAP values with one more entry than the input feature count. Should the final entry be plotted as…
Classic MLLevel: IntermediateCatBoost · SHAP values
CatBoost returns SHAP values with one more entry than the input feature count. Should the final entry be plotted as another feature?
Hints
- An additive decomposition needs an intercept-like term.
- Separate the reference level from input contributions.
Answer
No. It is the expected prediction baseline, not an input feature. Add it to the feature contributions to recover the model output in the explained output space.
Source and licence: CatBoost · SHAP values · Apache-2.0 · answer: Quantoros editorial