A New Pairing Support Vector Regression

Abstract

In this paper, the new pairing support vector regression (pair-SVR) algorithm is proposed to evaluate nonlinear regression models. In spirit of TSVR, the pair-SVR determines indirectly the regression function through a pair of nonparallel insensitive upper bound and lower bound functions solved by two smaller sized support vector machine (SVM)- type problems, which causes the pair-SVR not only have the faster learning speed than the classical SVR, but also be suitable for many cases, especially when the noise is heteroscedastic, that is, the noise strongly depends on the input value. Besides, the proposed approach improves the sparsity than that of TSVR by introducing an insensitive zone determined by a pair of nonparallel upper bound and lower bound function. Only points outside the insensitive zone are captured as SVs, and only those SVs determine the final regression model. In general, the number of SV is very few. This makes the prediction speed of pair-SVR is obviously faster than TSVR.