Multiclass Classification Method Research With SVM Arithmetic

Support Vector Machine (SVM) is a new statistical learning method which proposed by Vapnik. The learning principle of SVM is to minimize the structural risk, which gives SVM better generalization. Great progress has been made in theoretical study and algorithmic realization of SVM recently, which has become a new technique of data mining to overcome traditional difficulties, such as dimension disaster or over-fitting , etc.The way SVM solves the non-linear separable problem is that: a map impliedly defined by kernel function is used to transfer the samples in the original feature space into a higher dimensional feature space. Therefore, the non-linear separable problem becomes a linear separable one. In the course of solving decision-making function, the computation can be conducted in the original feature space, thus greatly decreasing computational complexity in the higher dimensional feature space.For practical problems, there are too many data indexes, and besides that, the method of kernel function adopted always magnifies the dimensions of the feature space. All these below cost too much computation, and bring about much difficulty. We often study two-class classification problem, but it doesn't mean we can find a solution to classifying multi-class classification problem, even if we can classify two classes correctly.This paper discussed support vector machine theory and kernel function's property. It also considers feature selection method based on the sequential minimization technique. As a result, it can improve SVM application ability in classification problem. Firstly, this paper studies and compares the performance of several multi-class classification methods based on support vector machine theory. Secondly, it suggests a practical method to solve fade area problem in the multi-class classification problem which uses fade support vector machine method. Thirdly, it gives a feasible method to settle sample unbalanced problem which often appears in one to many multi-class classification method. All these improvements have excellent practical value in solving multi-class classification problem. Lastly, this paper illustrates the application of SVM arithmetic in multi-class classification problem by classifying multi-class letter images recognition problem.