Pipeline for ensemble methods

[glemaitre]

The current pipeline implementation does not allow to handle the ensemble method since that we return several balanced X and y.
We should probably discuss of a proper solution to solve this issue.

[glemaitre]

@chkoar @dvro any opinions on that?

[dvro]

We could clone the classifier (end of the pipeline) and train several classifiers, one with each subset of samples, and use a VotingClassifier to combine the results. But this would make the Pipeline less transparent and different from sklearn ... just something to think about.

[glemaitre]

It was the option that I was thinking of since that It will be easy to find 3D X and 2D y.

[prcastro]

How bad would be to suggest that sklearn.ensemble.VotingClassifier could fit 3D X by using the 1st dimension for each classifier? Therefore the shape of the X could be also [n_classifiers, n_samples, n_features]? Maybe we could delegate this for their side and keep imbalanced-learn focused in balancing methods and not in the classification task.

[glemaitre]

@prcastro I am not that optimistic for such change in scikit-learn.
It is more a specific case from our side than something generic in machine learning.
I think that changing the Pipeline should be the way as mention by @dvro even if this is less transparent at first.

[glemaitre]

Refers to #252 to how to solve the issue