Skip to content
Quantoros

Raising CatBoost onehotmaxsize makes more features one-hot. Why might that change learning beyond merely expanding…

Classic MLLevel: IntermediateCatBoost · Categorical features

Raising CatBoost one_hot_max_size makes more features one-hot. Why might that change learning beyond merely expanding representation size?

Hints
  • Check which transformations are skipped for one-hot features.
  • This threshold switches mechanisms.

Answer

Features treated as one-hot do not receive the documented CTR transformations. The parameter changes which categorical-statistic mechanism is available, so measure quality and resource usage rather than assuming equivalent encodings.

Source and licence: CatBoost · Categorical features · Apache-2.0 · answer: Quantoros editorial

Related questions