Research On Support Vector Machine Leaning Algorithms

Support Vector Machine(SVM) is a new method of Machine Learning which is proposed by Vapnik and his group based on the statistical learning theory.It can solve small samples learning problems better by using structural risk minimization in place of experiential risk minimization.Because SVM adopts the thought of kernel function,it can change the nonlinear problems into linear problems to reduce the complexity of algorithm.It also has some advantages in the generality,robustness,effectiveness,simple calculation.So it has been widely used in face recognition,handwriting recognition, pattern classification and other fields.In this thesis,the main learning algorithms of SVM are deeply investigated;In order to overcome the poor classification performance and the levity result,some improved methods for SVM are proposed.The results are as follows:(1) Researches on the basis algorithms of SVM:quadratic programming algorithms, classification algorithms,incremental algorithms,and the analysis of the performance.(2) Base on the k-means clustering algorithm,a criterion of screening informative samples is designed.Then the steps for the incremental training algorithm are given by extracting useful information in accordance with that criterion.It can greatly reduce the training time and further improve the ability of dealing with large-scale leaning samples. And the performance of the improved incremental algorithm is analyzed from various aspects.(3) The Sequential Minimal Optimization(SMO) Algorithm applied to the classification of the unbalanced datasets not only leads to a poor classification performance but also makes the result levity.In order to overcome those disadvantages, an improved SMO algorithm which uses different error costs for different class is proposed.Besides,the formula and the steps of the improved SMO algorithm are given.(4) The Keerthi's SMO Algorithm and the improved SMO algorithm are separately trained by the UCI standard data,the results show that the improved algorithm's ability of dealing with unbalanced datasets Can be improved and its stability can also be intensified.Finally,the work of the paper is summarized,and future research is proposed.