The two methods used for predicting good probabilities in Supervised Learning are

☛ a) Platt Calibration

☛ b) Isotonic Regression

These methods are designed for binary classification, and it is not trivial.