Reconstruction Strategy for Multi-Class SVM Based on Posterior Probability

Abstract

After analysis and comparison of the problems of the existing one-versus-one (OVO) and one-versus-rest (OVR) decomposition methods of multi-class support vector machine (SVM), the novel strategy based on posterior probability is presented to reconstruct a multi-class classifier from binary SVM-based classifiers. The new reconstruction strategy can increase recognition accuracy and resolve the unclassifiable region problems in the conventional ones. Firstly, the geometric distance of test sample to the optimal classification hyperplane is used as the criterion of estimating the class probabilities to decrease the incomparability existing in different binary SVM-based classifiers. Then based on the Bayesian posterior probability theory, the combination strategy of the probability output among these binary SVM-based classifiers in OVO decomposition is given and the different prior probabilities of them are considered. Lastly, the prior probabilities are evaluated by OVR decomposition. In order to verify the effectiveness of this strategy, experiments have been made on UCI database; the experiment results show that the reconstruction strategy presented is effective over conventional ones.