Research Of Support Vector Machine Model

Support Vector Machine (SVM) is a new developed machine learning method. This technique has its roots in statistical learning theory, could be expressed in mathmatical form, has visual geometrical significance and has shown promising empirical results in many practical applications. Moreover, SVM can serve better in the processing of small samples, high dimension and local minima, which exist in most of learning methods. Currently, SVM is attracting more and more researchers and becoming a hotspot in the fields of artificial intelligent and machine learning.Starting with an introduction to the foundational theory ,an analysis of the SVM model training and SVM models selection, and aiming at the samples reduction and kernel function parameters selection, the thesis proposes new methods to improve training efficiency and speed up classification process of SVM. Experiments on real data sets indicate that the proposed methods are effective. The main research in this thesis can be classed as follows:1. A novel samples reduction method based on similarties is proposed. The training processing of SVM model in fact is a convex quadratic programming. The approach remove data from the training set before model training, and thus the number of support vectors becomes small and also the quadratic programming becomes simple. Here, the paper uses samples reduction method to deal with imbalanced data classification learning. Experimental results show that SVM with the selected patterns was trained without compromising the generalization capability.2. This paper proposes a new method to choose the kernel parameter. Radial Basis kernel function has been successfully used for a wide variety. The function could map the nonlinear classification problem to linear problem if the kernel parameter was suitable. We proposed a quick kernel parameter selection algorithm which contains the samples'label and samples'distance in feature space. Also this method doesn't need train and test the SVM model. Experiment results show that the proposed method is feasible and effective.