Research On Fast Training Method Base On Core Vector Machine And Support Vector Machine

Classification problem is an important issue in machine learning and data mining, whichhas been widely used. Support vector machine (SVM) base on statistical learning theory, is aneffective classification algorithm proposed in recent decades. SVM is based on structural riskminimization, it has a solid theoretical foundation and it is do well in solving nonlinear andhigh dimensional pattern recognition problem. But the scalability aspect and training speed ofSVM to handle large data sets still need much of exploration. In order to solve the large scaletraining problem, some approximation algorithms have appeared. Core Vector Machine(CVM) proposed by W.Tsang is one of them. CVM transfer a two-class SVM problem into aMinimum enclosing ball(MEB) problem and then CVM solve the MEB problem by using thecomputational geometry approximate algorithm. CVM has a time complexity that is linear inm and a space complexity that is independent of m, where m is the size of the training set.However, it is computationally infeasible to use CVM to deal with the data set with masssupport vectors (SVs), as its training time is related to the number of SVs.In this paper, a fast training algorithm combining CVM with SVM (CCS) is proposed. Inthe first step, CCS reduces the scale of data sets by using CVM, and then used labelingmethod to rapidly reconstruct training set, which aim is to reduce the scale of training set. Inthe second step, CCS use SVM to deal with the new data set. Experimental results on6datasets indicate the proposed algorithm can decreased training time by about30%in averagewithout reducing the classification accuracy, it is an efficient method for large-scaleclassification problem.