In my multi-label classification task I'd like to provide label ranking for each prediction, i.e. the following classification result:
(0.8, 'class A'), (0.2, 'class B')
would mean that this observation most probably belongs to the class A, but it also has something in common with the class B.
At the moment I am using OneVsRestClassifer from scikit-learn to assign multiple labels. This class has the predict_proba method but it returns the prediction confidence for each underlying binary classifier which, to me, is slightly different from providing a measure describing to which extent an observation belongs to a certain class.
Is there any method that can be used to provide such a label ranking in the case of multi-label classification?