Research And Improvement On The Variant Of Support Vector Machine

Support Vector Machine comes down to solving a quadratic programming problem,. Newton's method, interior point method and other classical optimization algorithms are good solutions. But when dealing with large-scale data, there are some shortcomings such as a memory limitation, too long training time and other issues. After years of research, it appears many excellent algorithms which have good generalization ability. But when faced with a complex form of data, how to improve the algorithm's better performance, expand its applications are still issues to be examined.On the basis of such in-depth study in the design of algorithm, classification criteria, classification performance of some variant SVMs, this paper complete the following work:First, we research several variant of SVM algorithms, including the least squares support vector machine(LSSVM),proximal support vector machine(PSVM), proximal support vector machine via generalized eigenvalue(GEPSVM) and twin support vector machine(TWSVM). These variations are designed to improve the learning speed classification accuracy of SVM. Comparing the performance of the algorithm on artificial data sets and UCI data sets, it is proved that these algorithms have obvious advantages to the variant of SVM in classification accuracy and speed.Second, integrated the advantages of SVM and LSSVM, this paper presents a new fast train classification based on LSSVM (FTLSVM).Experiments show that FTLSVM has considerable accuracy with LSSVM and PSVM, but is faster than LSSVM and PSVM.Last, propose a novel algorithm in this thesis:WMPSVM aiming to overcome the singular problems appearing in GEPSVM. Each classification surface of the method is obtained by solving a simple eigenvalue problem, avoiding the singularity of the GEPSVM. Its learning speed is considerable to GEPSVM but with better classification accuracy.