A fast training algorithm for unbiased proximal SVM

This paper presents a new algorithm for fast training of unbiased proximal support vector machines. PSVM was first introduced as an alternative to SVM classifiers that usually require a large amount of computation time for training. Unfortunately PSVM may present poor performance, especially for low values of a regularization parameter C, due to biased optimal hyperplanes. The proposed algorithm, named UPSVM (unbiased proximal support vector machines), uses a slightly different approach to circumvent this problem, such that an unbiased optimal hyperplane is always obtained. Simulations show that the proposed algorithm performs better than PSVM and sequential minimal optimization (SMO) with respect to training time with similar probability of correct pattern classification.