Probabilistic SVM outputs for pattern recognition using analytical geometry

Abstract

We present an alternative way of interpreting and modifying the outputs of the support vector machine (SVM) classifiers. Stemming from the geometrical interpretation of the SVM outputs as a distance of individual patterns from the hyperplane, allows us to calculate its posterior probability, i.e. to construct a probability-based measure of belonging to one of the classes, depending on the vector's relative distance from the hyperplane. We illustrate the results by providing suitable analysis of three classification problems and comparing them with an already published method for modifying SVM outputs.