I think decision function should be sorted just as we do for predict proba, where we sort by class label as in probas[:,np.argsort(self.label_)]