Inconsistent behavior in averaging choices micro and samples.

[amueller]

For the multi-class case, precision, recall and f-score with micro all produce accuracy, while with samples they produce an error.
That seem inconsistent. Using the definitions in the docs, they should also all be accuracy, I think.

I think I'd propose to deprecate micro averaging for multiclass.

The docs actually give an example of micro-average recall for multiclass, which is really weird imho.

[amueller]

The docs at the top actually recommend micro average for multi-class:

Micro-averaging may be preferred in multilabel settings,
including multiclass classification where a majority class is to be ignored.

That seems weird to me, given that it's just accuracy.

[jnothman]

Wording could be clearer, but the intention there is that using multiclass with a majority class ignored (labels=np.setdiff1d(classes_, 'default class') will return something other than accuracy. I think it is hard to deprecate because of how it is used in classification_report and elsewhere.