最小二乘支持向量机的一个快速近似算法

A fast approximation algorithm for least squares support vector machines

最小二乘支持向量机(LeastSquares Support Vector Machine,LS-SVM)已成功地应用于许多分类问题上,但由于其解不具有稀疏性,该方法无法应用到大数据集的学习与测试上.提出了LS-SVM的一个快速近似HBILS-SVM算法.该算法结合Backfitting与Invfitting法则迭代增加或减少一个支持向量,这一过程重复直到满足给定的停止条件,从而构建出最终的分类决策函数.对比已有的稀疏化算法,HBILS-SVM算法的计算复杂度低,解更加稀疏,且支持向量更具有全局最优性.数值模拟试验表明:HBILS-SVM算法能在取得同样的泛化性能的前提下使得分类器更加稀疏.

Least squares support vector machines （LS-SVM） have been applied to many classification problems recently,but it is invalid to learn and test large data sets since the solutions of LS-SVM lack the sparseness.This paper presents a fast sparse approximation procedure for LS-SVM,called the HBILS-SVM algorithm.The HBILS-SVM algorithm iteratively adds or reduces a support vector by combining the Backfitting and Invfitting strategies,this process is terminated when some stopping criterion is satisfied,and then the decision function is built.Compared to the existing sparse approximation methods,the HBILS-SVM algorithm has three compelling features：low complexity,sparser solution,and the support vectors are nearly globally optimal.Numerical simulation experiments show that our HBILS-SVM algorithm obtains sparser classifiers without sacrificing the generalization performances.