How does CatBoost PerTreeLevel resampling differ from PerTree for split comparisons?
Classic MLLevel: IntermediateCatBoost · Bootstrap
How does CatBoost PerTreeLevel resampling differ from PerTree for split comparisons?
Hints
- Separate the frequency of stages from candidates within a stage.
- Fair split comparison uses the same sampled evidence at that stage.
Answer
PerTreeLevel refreshes sampling or weights before each new tree level/split stage, whereas PerTree refreshes before a new tree. Neither means drawing unrelated samples separately for every candidate being compared within one stage.
Source and licence: CatBoost · Bootstrap · Apache-2.0 · answer: Quantoros editorial