Research On Svm Based On Large-Scale Training Set

Support Vector Machine based on structural risk minimization is determined by Vapnik and others on the basis of statistical learning theory put forward. Since SVM has good generalization performance and promotion of capacity, be better able to solve the small practical problems, which has been widely used. With the advent of the information age, information production and dissemination of the rapid rate of increase is also increasing the size of the database, how mining useful information from a massive dataset, has attracted growing attention. On basis of the comprehensive analysis for present research situation at home and abroad, this paper has further deep research to the svm on a large-scale training set.Firstly, it analyses the transformation between support vectors and normal vectors during new samples added to support vector set. Aimed at the inefficient removing method, an improved sifting algorithm for incremental SVM learning–Twice Removing Algorithm is proposed. In this algorithm, the useless samples are discarded by two useful removing methods, leads to new incremental training choose removing effective dataset instead of using the whole dataset they can not deal easily with very large dadasets, it can reduce subsequence training time marked ness.Secondly, based on the least square support vector machine. Then a novel compound kernel function based on least square support vector machine is proposed by taking advantage of local kernel function and global kernel function. These algorithms are realized through simulating experiment.Finally, the new multi-class classification method——improved multi-class classification method based binary treeis proposed based on already existing multi-class classification method based binary tree. It is a difficult and important problem that to determine the class division order of the multi-class classification method based binary tree, this algorithm use the clara clustering algorithm can solve this problem well.