I guess so. But surely we would need similar checks in other ensembles?

Shall we impose that any estimators (classifiers + regressors): an estimator should predict something of the same type than y. It could be check in predict and tested in common test. Is there any estimator not allowing for such behaviour?

travis is not happy

maybe there is an update to do in what's new?

So common tests check that the output classes are exactly the same types as the output classes, i.e. do a label encoder or equivalent inside. That means that if you input float, it will output float.

We actually allow floats that encode integers as targets (see type_of_target), so you can run a classifier on something like np.array([1, 2, 3], dtype='float'), and then you will run into this error.

I don't think we need this test because a Classifier that internally doesn't to a LabelEncoder scheme is not a valid sklearn estimator and won't pass check_estimator.

[Theoretically an estimator could satisfy check_estimator by remapping strings correctly but always casting ints to floats in the output, but that seems like a weird edge case. We could add an additional check to check_estimators, but I think that's overkill]

@agramfort @amueller I don't think there is a need for a what's new.
waiting for green lights otherwise should be good

LGTM

Thanks @glemaitre 🎉