Make it possible to specify interaction_cst and monotonic_cst with feature names.

[ogrisel]

Describe the workflow you want to enable

Instead of passing an array of monotonicity constraints (-1 for a decrease constraint, +1 for an increase constraint or 0 for no constraint) specified by feature positions in the training set, it would be more convenient to pass a dict to pass constraints spec only for the required feature names (when those are available as str values in the dataset columns). For instance

from sklearn.datasets import load_diabetes
from sklearn.ensemble import HistGradientBoostingRegressor

X, y = load_diabetes(return_X_y=True, as_frame=True)

reg = HistGradientBoostingRegressor(
    monotonic_cst={"bmi": +1, "s3": -1}
)
reg.fit(X, y)

Not that here X has column names because it is a pandas dataframe.

See #24845 for a similar feature for interaction_cst by passing a list of tuple of str names instead.

Describe your proposed solution

This requires updating the fit method, docstring and examples to accept a dict of constraints with feature names as keys.

If feature_names_in_ is not defined in fit, then a value error with a helpful error message must be raised.

Describe alternatives you've considered, if relevant

No response

Additional context

Once #13649 is finalized and merged, a similar treatment should be adapted to it for the sake of consistency.

[ogrisel]

I will give it a quick shot.

[ogrisel]

I am working on a similar UX improvement for categorical_features.

[lorentzenchr]

If feature_names_in_ is defined in fit, then a value error with a helpful error message must be raised.

Did you mean „if … is NOT defined, …“?

[ogrisel]

monotonic_cst was implemented in:

• ENH Add support for feature names in monotonic_cst #24855

and something similar for categorical_features in:

• ENH Specify categorical features with feature names in HGBDT #24889

but we still need to do the same for interaction_cst, maybe after #24849 has been reviewed and merged.

[lorentzenchr]

I guess this can be closed with #24855, #24889 and #24849 merged.