(A) bayesian approach to universum support vector machine

Abstract

Support Vector Machine(SVM) have been developed in many ways after the soft-margin SVM had been suggested in 1995. Bayesian SVM and Universum SVM are two methods for developing SVM. Bayesian SVM interpreting in probabilistic frameworks, samples classifiers from the particular posterior distribution. Bayesian SVM classifies new data by averaging the results of the sampled classifiers, and it have better performance than the soft-margin SVM for some data. Universum SVM which is trained by not only training data but also additory data called universum data, has better performance than soft-margin SVM. In this paper, we suggest a probabilistic interpretation of Universum SVM, and we use bayesian method to Universum SVM. In addition, we compare the performance of bayesian Universum SVM and Universum SVM. As it is denoted that the importance of selecting the universum data in the paper which suggested Universum SVM, solving the problem selecting the universum data under the probabilistic model will be done in our future work.