PR to #4816

Please add tests!

Please adhere to PEP8. Using flake8 might help

From the definition of the StackingClassifier here (http://machine-learning.martinsewell.com/ensembles/stacking/), self.stage_two_clfs should beself.stage_two_clf and there should be no weights argument.

I can understand the use case, but you should be applying a VotingClassifer on top of the StackingClassifier and StackingClassifier should not be doing it internally. This should be the API for that use case.

X, y = make_classification()
estimators1 = [Classifier1(), Classifier2()]
estimator2 = LogisticRegression()
estimator3 = RandomForestClassifier()
sc1 = StackingClassifier(estimators1, estimator2)
sc2 = StackingClassifier(estimators1, estimator3)
vc = VotingClassifier([("sc1", sc1), ("sc2", sc2)], weights=[0.3, 0.7])
vc.fit(X, y)

Also we should rename self.stage_one_clfs to something that has estimators in it. Also we should have a list of tuples where each tuple is of two length, (string, estimator) as done in VotingClassifier for GridSearch support. See (https://github.com/scikit-learn/scikit-learn/blob/master/sklearn/ensemble/voting_classifier.py#L227)

@tsterbak Thanks for the effort!

Please do consider adding a bit of documentation, tests and PEP8 compliant.

@MechCoder Thanks for your comments!! I will change the code asap.

@MechCoder about the API:
I think you are right in some ways, but using the voting internally and compute first and second stage in one object is computionally more efficient. So you have to fit and predict the first stage models only once to predict with all second stage models.

More suggestions are welcome! :)

All right, I can have a closer look after the code is cleaned up.

@MechCoder Hi, is there anyone working on it currently? If not, I want to do a PR

Is anybody planning to work on this? Otherwise I would be happy to look into it.

ping @jnothman

Sure, please go ahead. There were some questions with regards to the API, but those can be discussed during the course of the pull request.

I would like to participate in this project. In fact, I almost win the BNP competition (22nd place) with my own implementation, which I would like to adapt to sklearn... If possible, I would like to work into it!!

@ivallesp I'm currently working on a stacking (aka Super Learner) implementation too - would you like to collaborate?

@JivanRoquet I have already had a PR #7427 to stacking classifier, now just waiting for #7674 merged after which I can use the new _BaseComposition class. If you have any suggestions please let me know.

@yl565 awesome, thanks! I can see you provided a StackingClassifier.

My work is related to a StackingRegressor. Your PR is very helpful in that way, so that both codes can be consistent. I'll try to adapt what I've already done with the conventions you follow.

@JivanRoquet Sure! I would love to colaborate on this.

@jnothman should this PR be closed with #7427 as the more up to date?

EDIT: the even more up to date version is #8960