Using Hard Classifiers to Estimate Conditional Class Probabilities

In many classification problems, it is desirable to have estimates of conditional class probabilities rather than just "hard" class predictions. Many algorithms specifically designed for this purpose exist; here, we present a way in which hard classification algorithms may be applied to this problem without modification. The main idea is that by stochastically changing the class labels in the training data in a simple way, a classification algorithm may be used for estimating any contour of the conditional class probability function. The method has been tested on a toy problem and a problem with real-world data; both experiments yielded encouraging results.