An improved weighted posterior probability reconstruction strategy for multi-class support vector machine

Abstract

After analysis of the existing one-versus-one (OVO) decomposition methods for multi-class support vector machine (SVM), the improved weighted posterior probability (IWPP) reconstruction strategy is presented to combine binary SVM-based classifiers to multi-class one. The new strategy can resolve the unclassifiable region problems in the conventional max-win-voting (MWV) method and increase recognition accuracy. Firstly, the different prior probabilities of these binary SVM-based classifiers in OVO decomposition are considered. Then, the improved weight coefficients for combination of the probability output among these binary classifiers are re-derived based on conditional probability theory. To validate the proposed strategy, four examples on UCI database are used to illustrate the logical correctness of IWPP. Tests show that the IWPP strategy has less classification errors, better classification ability and more stable probability output than the original ones.