The Application Of Multi-category Support Vector Machine In Credit Rating And Study Of Kernel Parameter Selection

Support Vector Machine is a new method of Data Mining, and also a tool of small sample statistics, which performs better in dealing with small sample, non-linear and high-dimension pattern recognition problem than any other machine learning method. On the foundation of two-category Support Vector Machine, the algorithm and the application of multi-category is a problem worthy to research.BASEL NEW CAPITAL ACCORD is encouraging the banking developing their inner credit rating system, so the rating problem is becoming more important. An objective and effective rating method will be a feasible way to deal with this problem. Credit rating makes qualitative and quantitative analysis to the credit level on the view of science, which is also a non-linear classification problem. Support Vector Machine will perform well in this problem.This dissertation is focusing on following 3 points.First, basing on the BASEL NEW CAPITAL ACCORD, builds the model of credit-rating, and standardizes and hierarchical the process of credit rating. The dissertation analysis the model level, implements the different rating methods by improving the main algorithm of model level.Second, this dissertation based on the sample of bank, improves and optimizer which are suitable for multi-category, including the Hierarchy Support Vector Machine and correcting coding Support Vector Machine algorithms, implements the machine learning and data rating process with the help the tool box, which gets the better classification result, compares the performance, classification result of these algorithms.Finally in the dissertation, the effect factor of kernel function is analyzed; the experiments in different parameters by adjusting the parameter are implemented. Besides these, the performance and features of several common kernel functions are also analyzed, including their structure and related theories. The kernel function of this dissertation is Gauss Radial Basis is selected according to the analyses based on its own features; it studies the two parameters of Gauss Radial Basis kernel function, by adjustingσand C, looking for the relationship the parameter and the learning and populating ability, and summarize the method of parameter adjutancy.