Why does explaining log loss require labels even though ordinary prediction explanations do not?
Classic MLLevel: IntermediateSHAP · TreeExplainer
Why does explaining log loss require labels even though ordinary prediction explanations do not?
Hints
- Prediction and error are different functions.
- One of them takes the true outcome as an argument.
Answer
Loss depends on both prediction and observed outcome. Supply aligned labels for every explained row; changing the label can change loss attribution even when features and prediction remain fixed.
Source and licence: SHAP · TreeExplainer · MIT · answer: Quantoros editorial