Add a warning when one-class vector is passed to LabelBinarizer::fit()

[ysk24ok]

This change is

[TomDLT]

'negative'

[ysk24ok]

Oops, sorry, I have fixed the typo.

[amueller]

hm... do we want this? What was the motivation @ysk24ok? I guess it can lead to counter-intuitive results, but it's a normal use-case, isn't it?

[ysk24ok]

I had encounterd the same situation as #4546.
When only one class is in y_true, LabelBinarizer recognizes the label as negative one
and the return value of log_loss gets weired.
I thought it is necessary to make it clear that LabelBinarizer recognizes the label as a negative label when only one label in y_true, and that's why I created this PR.

But I didn't know #7239 is merged and this bug is now already fixed.

scikit-learn 0.17:

>>> y_true = np.array(['class1', 'class1', 'class1'])
>>> y_predict = np.array([[0.1],[0.7],[0.5]])
>>> log_loss(y_true, y_predict)
0.66749350018123588

current master:

>>> y_true = np.array(['class1', 'class1', 'class1'])
>>> y_pred = np.array([[0.1],[0.7],[0.5]])
>>> log_loss(y_true, y_pred)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/Users/yusuke-nishioka/Documents/Github/scikit-learn/sklearn/metrics/classification.py", line 1620, in log_loss
    'labels argument.'.format(lb.classes_[0]))
ValueError: y_true contains only one label (class1). Please provide the true labels explicitly through the labels argument.

LabelBinarizer doesn't permit one-class 'y_true'.
So this PR is outdated. Closing is good.

[amueller]

ok thanks for the feedback @ysk24ok