Bug in utils/multiclass.py/_ovr_decision_function

Describe the workflow you want to enable

Dear scikit learn developpers,

I think the implementation of _ovr_decision_function in utils
/multiclass.py doesn't work properly when the parameter confidences is probability. While as the documentation suggests, it can be a probability .

confidences : array-like of shape (n_samples, n_classifiers)
        Decision functions or predicted probabilities for positive class
        for each binary classifier.

The problem is the following two lines of codes

sum_of_confidences[:, i] -= confidences[:, k]
sum_of_confidences[:, j] += confidences[:, k]

In this context, there is a binary classifier for class i vs j. And j is the positive class.

If confidences is "decision_function", then it works. Because if "decision funtion" is negative, it means the classifier thinks the negatve class i is more possible. And the -= will increase the sum_of_confidences of i, and decrease the sum_of_confidences of j.

However, if confidences is "probability", it doesn't work. Because probability is always greater than zero. So the sum_of_confidences of i will always decrease, even when i is more likely to happend (prob of j < 0.5).

Describe your proposed solution

"decision_function" is centered at 0, while "probability" is centerd at 0.5. These two cases should be handled seperately.

Describe alternatives you've considered, if relevant

No response

Additional context

No response

Start in utils/multiclass.py at _ovr_decision_function and inspect its documentation and callers to understand how decision functions and probabilities are passed. Check the existing multiclass utility tests, then define coverage for both input forms; done means the aggregation handles probability inputs consistently with the documented behavior without regressing decision-function inputs.