A novel method which selects the approximate bases of high dimensional feature space based on the scale of linear independency is proposed；and after combining the presented method with the partial reduction strategy

SLS-SVRM(Sparse Least Squares Support Vector Regression Machine) is built. In addition

the recursive trick is used to accelerate the establishment of SLS-SVRM. SLS-SVRM obviously decreases the number of support vector without loss of the predicted accuracy. Finally

three UCI (university of California at irvine) datasets confirm the effectiveness of the proposed model.