Generation of prepared transforms after feature selection

[d-a-bunin]

Description

After fit in feature selection transforms we get a list of features to be filtered during transform. The problem is that during inference we don't really need the transforms that aren't passing the filter, but we are still computing them and wasting resources.

In theory, we could somehow regenerate the list of transforms after feature selection in a way that only necessary transforms are present.

Thoughts

Even if the pipeline should be refit after such regeneration it is still a useful feature.

Related

• More granular top features selection #325