Research On Improved Method Of Fuzzy Twin Support Vector Machine Algorithm

Twin support vector machine algorithm cannot effectively apply data sets with noises or outliers,which can easily lead to the degradation of classifier performance.Therefore,the fuzzy theory is brought into FTSVM algorithm to reduce the influence of noises or outliers on classification hyperplane by calculating the membership of each sample point.Additionally,FTSVM algorithm can achieve an effective,straightforward and a satisfying solution.But the algorithm still has some limitations,in response to these limitations,the main work of this paper is to improve FTSVM algorithm.The research content is as follows:(1)FTSVM algorithm may possess poor classification performance,and needs to calculate complex inversion operations on the matrix when solving dual problems,an improved fuzzy twin support vector machine algorithm based on Universum data(UIFTSVM)is provided.Firstly,a regularization term and Universum data are added into the optimization problem to minimize the structural risk and improve the generalization performance of the model.Then,a new Lagrange function is introduced to avoid the computation of inverse matrix.Furthermore,the kernel technique can be directly applied to UIFTSVM algorithm.Finally,UIFTSVM algorithm is assessed on UCI data sets and NDC data sets.The experimental results show that the optimization problem with Universum data and regularization term can improve the classification performance of UIFTSVM algorithm.(2)In view of the issues that FTSVM algorithm is still sensitive to noises and cannot distinguish support vectors and outliers effectively,a fuzzy twin support vector machine based on pinball loss(PFTSVM)is proposed.Firstly,a new kind of mixed membership function is constructed by combining the intra-class hyperplane membership function and the improved k-nearest neighbor membership function.Then,the hinge loss function is replaced by the pinball loss function to reduce noise sensitivity.Finally,like UIFTSVM algorithm,PFTSVM algorithm realizes the structural risk minimization and avoids the computation of inverse matrix.And the proposed algorithm is assessed on Ripley data set and some UCI data sets.The experimental results show that PFTSVM algorithm has strong robustness.(3)The proposed PFTSVM algorithm cannot be used to deal with the multi-classification data effectively,a pinball loss based fuzzy multiple birth support vector machine algorithm(PFMBSVM)is present.Firstly,based on the multiple birth support vector machine(MBSVM),PFTSVM algorithm is extended to multi-classification method.Then,PFMBSVM algorithm is compared with OVO-PFTSVM,OVR-PFTSVM,OVO-TWSVM,OVR-TWSVM and MBSVM algorithms on some UCI data sets.The experimental results indicate that the multi-classification strategy is useful,and the comprehensive classification performance of PFMBSVM algorithm is superior to that of the comparison algorithms.