A global support vector regression based on sorted K-fold method

Support vector regression (SVR), as one of the most popular metamodels, has been widely used to replace expensive and time-consuming simulation models in complex engineering design process. In this essay, a global support vector regression metamodeling approach based on the sorted K-fold method (SKM-SVR) is proposed, which combines the K-fold cross-validation technique with SVR. Firstly, all sample points are sorted and divided into K groups according to their absolute prediction errors from traditional SVR metamodel. Secondly, K number of SVR metamodels are constructed based on cross-validation technique. The weighting factor of each SVR metamodel depends on their prediction accuracy. Finally, K number of SVR metamodels are integrated to form the SKM-SVR metamodel. The effectiveness of SKM-SVR method is demonstrated by a mathematical example and a pressure vessel design problem. Results indicate that SKM-SVR approach has better prediction ability than that of SVR.