One VS Rest Classifier Native Implementation for Multiclass Problems

[f-gibellini]

Hello,

First of all thank you for building this amazing product, I use it all the time in my job and find it very useful and performant.

Recently when dealing with a complex multi-class (8 classes) problem with few high dimensional categorical features, I noticed that wrapping the CatBoostClassifier in a sklearn.multiclass.OneVsRestClassifier improves my performance on predictions by ~10%.
This improvement is obtained at even model size so for example training 1 single CatBoostClassifier to predict 8 classes having 4000 predictors is outperformed by 8 binary (CatBoostClassifier) classifiers each having 500 predictors [all other hyper params being the same].

Given this performance difference and the fact that the sklearn wrapper is quite slow and inefficient I was wondering whether a native implementation of a one vs rest multi-class approach could be implemented directly in the CatBoostClassifier as this would probably be much more performant and optimized.

I could find no reference to this feature in the docs and hence I think it is not implemented, if it is already please let me know.

Thanks,
Kind regards,
Francesco