Support Vector Regression And Its Application

Support Vector Machine (SVM) is new data mining technique developed from the middle of 1990s. SVM based on the foundations of Statistical Learning Theory (SLT), which is a small-sample statistics and concerns mainly the statistic principles when samples are limited. SVM is developing promisingly either in theory or applications.There are mainly two focuses of SVM including support vector classification (SVC) and support vector regression (SVR), while the research of SVR is not a patch on SVC either in theory or in applications. This paper focuses on SVM in several aspects, including theory foundation and application.1. A generalized weighted model of SVR is proposed, in the optimization problem of which a flexible convex function and the weighted coefficient are included. The different choices of the function and the coefficient will equate the balance between the generalization and the risk, while will derive some kinds of existing algorithms of SVR. The model is more effective in the selection of the kernel function and the confirmation of the real risk.2. A multidimensional output support vector regression model is designed that introduces the solution of the multi-outputs training. The present SVM is only useful for the pattern with multi-input and single output, and there are not enough discussions in the topic of multi-output. Regression hyperplane based matrix is structured, which can be implemented by solving the iterative procedures of the basic support vector regression model. Instead of the vector coefficient, a weighted matrix coefficient is gained as the result. This model is useful in estimating the movement trend of the multi-outputs, and the experiments show that the model is feasible and effective.3. Different support vector classification and regression predict models are constructed and applied to the solution of the customer classification, credit scoring, business prediction and so on. Some examination for the support vector regression models are presented in the paper, as well as the particular analysis for the different results with contrasting experiments.4. Unfixed the kernel function parameter is introduced while training. The choice of the kernel function and parameter will be according to the characterization of the date, rather than lying on the experimenter's experience. The experiment with kernel function according to the information gains shows more accuracy and generalization, and there are also result tables, which indicate the impact when different parameters are chose.