There are several multiclass variants of the hinge loss. The one in the article you link is the variant by Crammer & Singer. We should either accept a parameter to let the user choose or use one function per variant.

We can also implement the "one-vs-rest" loss, which is simply to sum the hinge losses of each binary classification problem.

Thanks @mblondel for the precision.

I am new to Scikit-learn. I am trying to fix this issue.

Hi, I am reading this paper
http://www.ttic.edu/sigml/symposium2011/papers/Moore+DeNero_Regularization.pdf
in order to understand the notation of multi class hinge loss equation at wikipedia(w subscript y and w subscript t). Can anyone please confirm if w subscript f and l are functional margins? Are they product of w evaulated at some feature and the corresponding product?

Also, is there any way to find w from est?

Fixed by #3607