Robust Cross-Validation Score Functions with Application to Weighted Least Squares Support Vector Machine Function Estimation

In this paper new robust methods for tuning regularization parameters or other tuning parameters of a learning process for non-linear function estimation are proposed: repeated robust cross-validation score functions (repeated-CV Robust V −fold) and a robust generalized cross-validation score function (GCVRobust). Both methods are effective for dealing with outliers and non-Gaussian noise distributions on the data. The robust procedures are based both on a robust cross-validation estimate and a robust function estimator. Simulation results for weighted Least Squares Support Vector Machine (weighted LS-SVM) function estimation are given to illustrate that the proposed robust methods outperform other cross-validation procedures and methods based on a number of other complexity criteria.