This PR fixes a bug for multi-label targets in StratifiedShuffleSplit. The solution being used now is the "label powerset" method: each sequence of labels is mapped to a string with str(row), which transforms a multi-label problem into a multi-class problem.

To see the source of the problem, note that len(str(np.arange(1000))) returns 4056 while len(str(np.arange(1001))) returns 36. The reason is that arrays with > 1000 elements are truncated with an ellipsis: str(np.arange(1001)) gives '[ 0 1 2 ..., 998 999 1000]'. Thus, for multi-label targets with > 1000 labels, samples are mapped onto the same short string whenever their first three values and last three values are the same, which is not the intended behavior.

The solution proposed here, discussed with @vene in this comment thread, is to use ' '.join(row.astype('str')) to convert each target to a string. We are guaranteed that we can do call .astype('str') on row because y = check_array(y, ensure_2d=False, dtype=None) converts y to a numpy array.

As an added benefit, this approach ends up being several faster than str(row) when len(row) < 1000:

In [1]: import numpy as np
In [2]: row = np.random.randint(0, 2, size=500)
In [3]: %timeit ' '.join(row.astype('str'))
169 µs ± 2.39 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)
In [4]: %timeit str(row)
738 µs ± 23.1 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

I also considered using sklearn.utils.murmurhash3_32 (suggested by @vene as another option) but concluded it was tricky to coerce all the kinds of labels people might use into an int32 data type.