Two analysts explain the same customer using different background populations. Why can both obtain valid but different…
Classic MLLevel: IntermediateSHAP · TreeExplainer
Two analysts explain the same customer using different background populations. Why can both obtain valid but different SHAP values?
Hints
- An explanation is relative to a reference, not only to a model.
- Changing the reference can redistribute attribution.
Answer
They answer different reference-population questions: baseline expectations and missing-feature integration differ. Align the background and dependence assumptions before comparing values as if they measured the same quantity.
Source and licence: SHAP · TreeExplainer · MIT · answer: Quantoros editorial