FIX accept input_features parameter in TableVectorizer.get_feature_names_out

[glemaitre]

closes #1256

This PR makes sure to expose an input_features parameter in the get_feature_names_out to be compatible with scikit-learn.

Right now, the parameter is ignored. One question would be: do we want to allow overwriting the column names. In this case, we would need to change more code because we would need to overwrite the name of the columns of the output dataframe in transform to have a consistency between get_feature_names_out() and the feature_names_in_ of the subsequent steps in a scikit-learn pipeline.

[GaelVaroquaux]

LGTM, but could we have a tiny changelog entry. it's useful for users

[glemaitre]

LGTM, but could we have a tiny changelog entry. it's useful for users

Indeed :)

[jeromedockes]

Thanks a lot @glemaitre !!

Note we have a bunch of estimators defining get_feature_names_out and none of them accept input_features so it might be worth creating another issue with a TODO list

do we want to allow overwriting the column names. In this case, we would need to change more code because we would need to overwrite the name of the columns of the output dataframe in transform to have a consistency between get_feature_names_out() and the feature_names_in_ of the subsequent steps in a scikit-learn pipeline.

I would say no, in any case a preprocessing check in the tablevectorizer raises an exception if the columns of the input are different in transform than in fit. but you know more about get_feature_names_out and how it is used in scikit-learn

[GaelVaroquaux]

Note we have a bunch of estimators defining get_feature_names_out and none of them accept input_features so it might be worth creating another issue with a TODO list

+1

Can we use scikit-learn's common tests to detect this?

[glemaitre]

Can we use scikit-learn's common tests to detect this?

We should check if the common test is covering this particular feature but I would hope so. I think that we are still bypassing the common test. I should give it another look in skrub.