Regression Modeling And Parameter Estimation Based On Kernel Methods And Its Application

The goal of data modeling and analyzing is to forecast the unknown data by extracting the rule of the observed ones.The multi-correlation among variables and built-in complexity of data are two main factors which lead to the weakness forecast performance of general algorithms.Partial Least Squares(PLS) is a new regression method.By means of screening and synthesizing data information,it can reduce the multi-correlation and construct mathematical model effectively even when the samples are less than the variables. Exploiting the excellent modeling capability of PLS and combing with variable selection techniques,we put forward a backward method based on PLS.Experiment results show that the ability to predict and stability are effectively enhanced in comparison with stepwise regression.Kernel methods is an advanced strategy of data processing,which maps the low dimension data into a higher feature space,meanwhile transforms the complicated nonlinear problem into a linear one in the feature space.Support Vector Machine(SVM) is the successful application apotheosis of Kernel Methods.We proposed a new method based on Uniform Design to overcome the difficulty of selecting the model parameters of Support Vector Regression.The results of experiment show its validity.Theoretic analysis and Experiment results demonstrate that the model parameter selected by the proposed method can yield a regression model with good forecasting ability.