pipeline.get_feature_names()

[jackwellsxyz]

Describe the workflow you want to enable

This is possibly a duplicate of #6424. I'd love to do a pipeline.fit(X) with a pandas dataframe with named columns, then a pipeline.get_feature_names() as input into an eli5 explainer, with scikit-learn being smart enough to call get_feature_names() for those transformers it makes sense to do so (OneHotEncoder, SelectFromModel, etc.)

Describe your proposed solution

I'm not sure what a good solution might be -- one start might be to implement get_feature_names() for all transformers and return just the input column names if it doesn't change them, as would be the case for a Binarizer, for example.

Describe alternatives you've considered, if relevant

Additional context

[rth]

Thanks! The end goal of SLEP14 and proposed implementation in #16772 is to allow this.

Closing this issue to avoid duplicates, please comment in one of those exiting ones.

[jnothman]

eli5 already handles this case. But you shouldn't be passing the output of pipeline.get_feature_names() as these are the names of features output by the Pipeline's transform method.