Why should switching to Bayesian bootstrap in CatBoost not be presented as row-subsampling speedup?
Classic MLLevel: IntermediateCatBoost · Bootstrap
Why should switching to Bayesian bootstrap in CatBoost not be presented as row-subsampling speedup?
Hints
- Changing weights need not reduce the number of evaluated rows.
- Regularization and computational subsampling have different goals.
Answer
It randomizes example weights for regularization but does not perform the documented row-removal speedup. If computation is the bottleneck, evaluate a compatible sampling scheme such as Bernoulli or MVS separately.
Source and licence: CatBoost · Bootstrap · Apache-2.0 · answer: Quantoros editorial