A v-twin projection SVR with automatic accuracy adjustment

Abstract

Taking motivation from v-insensitive twin support vector regression (v-TSVR) and the projection idea, this paper proposes a novel v-twin projection support vector regression model, called v-TPSVR. This v-TPSVR, based on v-TSVR, determines the regression function through a pair of nonparallel hyperplanes solved by two smaller sized quadratic programming problems (QPPs). The proposed v-TPSVR model also can automatically optimize the parameters $\varepsilon_{1}$ and $\varepsilon_{2}$ via the user specified parameters $v_{1}$ and $v_{2}$, which is the same as v-TSVR. But different from v-TSVR, v-TPSVR seeks a projection axis in each QPP such that the variance of the projected points is minimized, so the empirical correlation coefficient between each hyperplane and the projected inputs is maximized. Although the training speed of the proposed algorithm is similar to that of other compared algorithms, it is obvious that the introduction of the projection axis makes the number of SV less than that of v -TSVR under the same values of $v_{1}$ and $v_{2}$, it would lead to faster testing speed. In addition, The experimental results indicate that the proposed v-TPSVR obtains the better prediction performance than the popular ɛ-TSVR and v-TSVR.