Why can the same category have different ordered target-statistic values for different training rows in CatBoost?
Classic MLLevel: IntermediateCatBoost · Categorical statistics
Why can the same category have different ordered target-statistic values for different training rows in CatBoost?
Hints
- The statistic is indexed by a row history as well as a category.
- Full-data aggregation is a different computation.
Answer
Each row uses its preceding history in a permutation, not one full-data category average. Early occurrences have less evidence and more prior influence; replacing this with a global mean changes leakage behavior.
Source and licence: CatBoost · Categorical statistics · Apache-2.0 · answer: Quantoros editorial