It might be better to do the opposite. In multilabel classification, there is the jaccard similarity score that is also called accuracy and is another type of averaging to compute the accuracy.

I thought the zero_one_loss was more general as we can get the unnormalized version.
As these are only internals, we can always change them. I was just a bit confused that there are two basically identical implementations that are >20 lines that only differ by a sign as far as I can tell.

I will submit a patch. I would like to add the normalize option to the accuracy_score.
Is it ok for you?

sure! thanks :)

Fix in #1750