Predicting the Probability of Correct Classification ; CU-CS-974-04
Gregory Z. Grudić · CU Scholar (University of Colorado Boulder) · 2004
We propose a formulation for binary classification, called the Probabilistic CDF algorithm, that both makes a classification prediction, and estimates the probability that the classification is correct.Our model space consists of the widely used basis function models (which includes support vector machines and other kernel classifiers).Our formulation is based on using existing algorithms (such as SVM) to build the classifiers, and therefore achieves state of the art classification accuracy, while at the same time giving accurate estimates point specific probabilities of correct classification.We further demonstrate that the Probabilistic CDF algorithm can improve the overall classification accuracy of an existing basis function classifier by appropriate local modification the decision boundary.Our framework is theoretically justified and empirical evaluations show promising results.