Excellent work. Just to clarify: Will the new options allow to both

1. keep np.nan as they are and
2. map new unknown categories to np.nan?

If yes, this will be super good news for fitting boosted trees!

this PR supports 2 but 1 is still not supported. An error is raised when nans are present in the training data: it's unclear where to map them, as the output of OrdinalEncoder is supposed to be interpreted as ordered quantities.

If yes, this will be super good news for fitting boosted trees

HistGradientBoostingClassifier and the correspoding regressor natively support both missing values (as nans) and categorical data now :)

https://scikit-learn.org/stable/modules/ensemble.html#histogram-based-gradient-boosting

It's great to see all the progress in handling categorical data with native scikit-learn components!

Is what @mayer79 asked in 1. what's proposed in #17123 ?

yes, with 1 and 2 being inverted

@NicolasHug : Thx for clarifying. From a practical perspective, it is not desirable that remaining nans would raise an error. If my subsequent model algorithm cannot natively deal with nans, we can simply add an imputer after the encoder and voila.

@mayer79 Couldn’t you just run the encoder for not nans only to get the desired behavior?