A performance nondestructive pruning method for reduced robust weighted Least Squares Support Vector Regression

摘要

In this paper, a method is proposed to get a sparse solution of Least Squares Support Vector Regression (LSSVR) without damaging the performance while can also deal with outliers and keep robustness. The method uses the weighted LSSVR as the primal robust model, and reduces the redundancy part properly by considering both the support values and the density of train samples. And then the samples are divided into different parts to monitor theirs performance changes. Only if the performance change of any part is larger than the fluctuation threshold, the pruning process will cease while a simple and reliable model will be received. The experiments result more intuitively displays the efficacy and feasibility of proposed method, and the final sparseness and robustness solution can be seen as a performance nondestructive model which are favorable for our viewpoints.