In order to overcome the limitation of single kernel in Support Vector Machine(SVM) model, hybrid kernel is usually used in forecasting. However, the weight of functions in the hybrid kernel is hard to calculate. To solve this problem, we propose a new method based on feature-distance. This method firstly gets an optimization function based on SVM's geometric meaning and a principle, which is the feature-distance of the same kind should be minimized and the different should be maximized, and then analyzes the optimization function to work out the weight. Experimental results show that compared with the cross validation method and PSO algorithm, this method reduces the computing time nearly by 70% with the accuracy kept unchanged.
分享
引用
