Bootstrap Based Pattern Selection for Support Vector Regression

Dongil Kim, Sungzoon Cho

Abstract: Support Vector Machine (SVM) results in a good generalization performance by employing the Structural Risk Minimization (SRM) principle. However, one drawback is O(n 3) training time complexity. In this paper, we propose a pattern selection method designed specifically for Support Vector Regression (SVR). In SVR training, only a few patterns called support vectors are used to construct the regression model while other patterns are not used at all. The proposed method tries to select patterns which are likely to become support vectors. With multiple bootstrap samples, we estimate the likelihood of each pattern to become a support vector. The proposed method automatically determines the appropriate number of patterns selected by estimating the expected number of support vectors. Through the experiments involving twenty datasets, the proposed method resulted in the best accuracy among the competing methods.