An Improved Twin Support Vector Regression with Automatic Margin Determination

Abstract

In this paper, a novel regression algorithm named ν-twin support vector regression (ν-TSVR) is presented, improving upon the recently proposed twin support vector regression (TSVR). It also tries to seek two nonparallel down- and up-bounds for the unknown function. By treating the size of one-sided -insensitive tube as optimization variables with corresponding parameters s, we reformulate the original TSVR as a more sensible model. To this end, ν-TSVR has the advantage that s are learned simultaneously with regressor. Meantime, we give a theoretical result concerning the meaning of s. Moreover, by introducing structural risk minimization principle, the over-fitting phenomenon in TSVR can be avoided. We analyze the algorithm theoretically and demonstrate its effectiveness via the experimental results on several artificial and benchmark datasets.

DOI: http://dx.doi.org/10.11591/telkomnika.v11i1.1895