A KERNEL-BASED WEIGHT-SETTING METHOD IN ROBUST WEIGHTED LEAST SQUARES SUPPORT VECTOR REGRESSION

摘要

By combining the basic idea of weighted least squares support vector machines (WLS-SVM) and fuzzy support vector machines (FSVM), a weight-setting strategy based on 2-norm distance and neighborhood density (WLS-SVM Ⅰ) is presented in this paper. Then the relationship between the 2-norm distance and RBF kernel is revealed. Consequently, an equivalent weight setting strategy (WLS-SVM Ⅱ) using information from RBF kernel is put forward. Numerical experiments show both the 2-norm distance-based strategy and the Kernel-based strategy produce robust LS-SVM estimators of noisy data. And when satisfying some conditions,WLS-SVM Ⅰ can be substituted by WLS- SVM Ⅱ, which may provide an efficiency-enhancing strategy for online LS-SVM.