StratifiedShuffleSplit requires three copies of a lower class, rather than 2 about scikit-learn

[glemaitre]

I think this is not a bug but rather a known implementation detail: looking at the code, we use _approximate_mode that is know to be an approximate estimate that can be off by 1 (the value that you observed). I assume that this approximation is done for some computation reasons.

However, since the behaviour is surprising, I think that we could document it in a "Note" section to mention this corner case.

[jeremycg]

Fair enough, the code in the function:

        if n_train < n_classes:
            raise ValueError(
                "The train_size = %d should be greater or "
                "equal to the number of classes = %d" % (n_train, n_classes)
            )
        if n_test < n_classes:
            raise ValueError(
                "The test_size = %d should be greater or "
                "equal to the number of classes = %d" % (n_test, n_classes)
            )

does a check I'd expect to catch the issue here: I'd suggest that should be modified too.