With boostfromaverage, should CatBoost initialize every regression loss using the arithmetic mean target?
Classic MLLevel: IntermediateCatBoost · Training parameters
With boost_from_average, should CatBoost initialize every regression loss using the arithmetic mean target?
Hints
- Which constant minimizes the selected loss?
- The mean is specific to squared error.
Answer
No. It uses a best constant for the selected supported loss: squared error favors the mean, absolute error a median, and quantile loss the corresponding quantile. The optimal bias depends on the objective.
Source and licence: CatBoost · Training parameters · Apache-2.0 · answer: Quantoros editorial