Research On Structure Support Vector Machine Classification Models

In the field of machine learning,support vector machine(SVM)has been widely applied as a novel and effective classification or regression method.Compared with traditional machine learning methods,SVM has good generalization performance.However,there still exists a great space for improvements in computational complexity.In order to reduce the computational complexity,twin support vector machine(TSVM)is proposed.The learning speed of TSVM is faster than that of SVM.Meanwhile,TSVM can improve generalization performance.Therefore,TSVM has become the most popular classification method at this stage.Although SVM and TSVM have been very successful,it is found that both SVM and TSVM ignore structural information.In this paper,more effective classification models with potential structural information are proposed.The main contents are summarized as follows:Firstly,the local structural information of samples compared with global structural information is important to the generalization performance.Thus,inspired by neighborhood margin Fisher discriminant analysis(NMFDA)and Fisher discriminant analysis(FDA),on the basis of SVM,a support vector machine with local structural information(LSI-SVM)is proposed in this paper.The algorithm needs three steps to obtain the final classification hyperplane.Firstly,based on FDA algorithm,the global structure of data is excavated.Secondly,based on NMFDA algorithm,the local within-class and between-class scatters are captured.Thirdly,the obtained structural information is introduced into the SVM to build a new classification model.Finally,the experimental results on UCI datasets show that our LSI-SVM can improve the classification accuracy.Secondly,the local between-class margin compared with global margin really reflects the separability of between-class samples.So,inspired by NMFDA algorithm,on the basis of TSVM,a twin support vector machine with local structural information(LSI-TSVM)is proposed in this paper.Firstly,based on NMFDA algorithm,the local within-class scatter and the local between-class scatter are excavated for each class.Then,the obtained local structural informationa is introduced into the TSVM to obtain a novel classification model.Finally,all experiments show that our LSI-TSVM is superior to the state-of-the-art algorithms in generalization performance.Thirdly,the margin distribution compared with margin theory is crucial to the generalization performance.Therefore,in order to improve the generalization performance of TSVM,a twin support vector machine based on asjustable large margin distribution(ALD-TSVM)is proposed in this paper.The new method firstly reconstructs the margin distribution for each class.Then,the reconstructed margin distribution is introduced into the TSVM to obtain a new classification model.The numerical experimental results show that our ALD-TSVM has a significant improvement in generalization performance.