[BUG] Rename transform_train to resample.

[anopsy]

The method in preprocessing.Outlier_Remover transform_train should be changed to fit_transform or transform.

[koaning]

@FBruzzesi so this came up during the code spring. It indeed seems that our OutlierRemover doesn't have a transform method. And I also just noticed that we also only allow X as input.

I really forgot what we had in mind when we designed this one. But I'm wondering what we might want to do with it. @MBrouns do you remember? Is there a reason why we called it transform_train? Maybe related to the fact that we cannot really use it in a pipeline because it changes the shape of X?

[FBruzzesi]

It should be related to #342.
The TL;DR is that the scikit learn Pipeline would not filter y and this would not work with supervised learning

[MBrouns]

Yea I think that's indeed it. I'm up for calling it resample to make it work with the imblearn folk

[koaning]

That does feel better, yeah. Let's do that and update the docs accordingly.

[FBruzzesi]

I am certainly not a user of imblearn. I tried to play around with it and it seems not to be so straightforward.
Curiously enough, one of the user guide explaining how to create a custom sampler implements an outlier detection.

To add more details, when adding resample and fit_resample methods, then I end up with the issue of having both fit_resample and transform implemented (the latter due to inheritance) which the imblearn Pipeline seems to not like.

[koaning]

I think we can drop the TransformerMixin here if we want it to be more like a resampler.

Then again, part of me would also be "ok" with dropping this feature from this library. Doing sampling stuff in a pipeline really requires imblearn and I'm not sure if I like the idea of adding imblearn as a dependency for scikit-lego.

[FBruzzesi]

I definitely agree in not adding imblearn as a dependency. I honestly like the idea of having such feature, but maybe it is the wrong place for it

[koaning]

We can also just make a utility function that just removes outliers. Something like:

X_new, y_new = remove_outliers(estimator, X, y)

Wouldn't this be the simplest/cleanest?