A larger foldpermutationblock speeds up CatBoost. Why is it not a free optimization?
Classic MLLevel: IntermediateCatBoost · Training parameters
A larger fold_permutation_block speeds up CatBoost. Why is it not a free optimization?
Hints
- The parameter changes the permutation unit.
- Less fine-grained randomization may change the learned model.
Answer
It permutes larger groups of objects rather than individual small blocks, altering the randomization used in training. The documented speed benefit can trade off quality, so compare on the same held-out split and seed protocol.
Source and licence: CatBoost · Training parameters · Apache-2.0 · answer: Quantoros editorial