Application And Extension Of SVM In Multi-class Classification Problems

In this paper, we study a new SVM based on hidden information and derive a newone-class SVM based on hidden information. First, we introduce the basic theoryknowledge, such as, support vector machine, kernel, the reproducing kernel Hilbert spaceand the dimension reduction. Then, three kinds of decision tree multi-class classifiersbased on SVM are presented by means of three clustering methods, which are respectivelyclustering with minimum distance of class means, maximum distance of class means andmaximum margin criteria. The experiments with AVIRIS remote sensing image are madefor testing the validity and advantage of our proposed algorithms. The experimental resultsdemonstrate that our methods are significantly better than minimum distance classification,linear discriminant classification, decision tree classification, OAR-SVM and OAO-SVM.Then, a SVM based method called SVM+was developed under both Learning UsingPrivileged Information (LUPI) setting and Learning With Structured Data (LWSD) setting.SVM+introduces several new concepts such as correcting function and projecting trainingdata into two different spaces. These new ideas have lead to the SVM+based Multi-TaskLearning (MTL) method for classification. We empirically compared SVM+MTLclassification with several methods such as SVM.At last, this paper is denoted to study the effect of the group information of data inone-class kernel support vector machines (OC-KSVMs) for classification accuracy andtime consumed of multi-class classification data. Two new classification methods based onOC-KSVMs are presented. One is OC-KSVM with maximum margin from the origin andgroup information of data (briefly, MMOC-KSVM+). And another is OC-KSVM withhyper-sphere and group information of data (briefly, HSOC-KSVM+). We provedtheoretically that MMOC-KSVM and HSOC-KSVM are equivalent for Gaussian RBFkernels. Experiments on three real-words data sets are performed in order to test andevaluate the efficacy of the proposed methods. Experimental results indicate that the groupinformation of data can improve the classification accuracy of data and meanwhile increasethe time consumed of algorithms.