An improved multi-class algorithm for SVMs

A multi-class algorithm based on the posterior probability outputs for SVMs is presented in this paper. Our algorithm reduces a multi-class classification problem to multiple binary ones, which are solved by the binary SVMs. The binary posterior probabilities for the binary SVMs are obtained for computing the total posterior probabilities and making the final decision. If the estimation of the posterior probability outputs of the binary SVMs were sufficiently exact, our algorithm could approximate to the optimal Bayesian classifier. However it has some difficulty to do this in fact. Our algorithm is comparable to the other algorithms on the recognition performance. Moreover, the number of binary problems is decreased greatly, so the training and test complexity are decreased, too. Our algorithm can give not only classification results but also the posterior probability of classification, which is usable for solving many practical problems. Experimental results confirm the feasibility and the validation of our algorithm.