#platt-scaling

Question 1: Single choice

How does scikit-learn's SVC(probability=True) obtain the class probabilities returned by predict_proba?

It fits a logistic regression on the SVM scores, using extra cross-validation during training1
It applies a sigmoid to the distance from the boundary at prediction time, with no extra fitting2
It counts the fraction of support vectors of each class among the nearest training points3
The SVM margin directly defines exact posterior probabilities, so nothing has to be fitted4