Research On Multi-Scale Kernel Support Vector Regression Model

SVR is mainly used to solve the problem of the regression, the learning effect depends largely on the choice of kernel. For the study of nuclear functions, from initial single kernel to multiply kernel learning methods in recent years, it has been a hot issue in support vector machine theory.Most of the currently research on multi-scale kernel support vector regression machine are limited to study a single Gaussian kernel function or wavelet multi-scale kernel function of. In fact, the Gaussian kernel function has a good ability to learn, but its generalization ability is not high, and multi-scale wavelet kernel function support vector regression method with good generalization ability, so we will combine them to construct a new kernel, so that it can fully reflect the characteristics of each function.In order to improve the result of the support vector regression model, in addition to select the appropriate kernel, it is also required to determine the optimal parameters of the model. Parameters for the proposed model, we use density-based clustering algorithm to determine the scale of kernel, respectively, with coefficient of variation and the upper quartile of the parameters to determine the width of the Gaussian kernel function and wavelet kernel function, the remaining model parameters namely punishment coefficient kernel weighting function coefficients and loss function parameters are obtained by chaos genetic algorithm search. Finally, in order to verify the performance of the model, compared with the other three models- mononuclear SVR model, multiscale Gauss kernel SVR model and multi-scale wavelet nuclear SVR model. Do the experiments on the three UCI datasets, the obtained results may prove the best support vector regression model is the one based on new multi-scale kernel.