Research On Fuzzy Multi-classification Support Vector Machine Algorithm Based On Structural Information

Support vector machine(SVM)is an effective machine learning method.It shows many unique advantages in solving small sample,nonlinear and high-dimensional pattern recognition problems,making it one of the most widely used classification techniques in data mining,and it has been widely used in many fields.Because SVM has defects such as high computational complexity,researchers have proposed different twin support vector machines(TWSVM)based on support vector machine.However,no matter whether SVM or TWSVM,they are only used for binary classification problems.In order to solve the multi-classification problem of support vector machines,people have conducted research on it within different support vector machine algorithms have been proposed.These algorithms are mainly divided into three types: The first category is to directly construct methods to solve multiclassification problems,for example,the simplified multi-class support vector machine(Sim MSVM);the second category is a multi-class strategy based on two classifications,including one-versus-all twin support vector machine(OVA TWSVM),one-versus-one Twin Support Vector Machine(OVO TWSVM)and directed acyclic graph twin support vector machine(DAG TWSVM);the third category is an evolution of multi-classification strategy based on binary classifications,include one-versus-one-versus-all twin support vector machines(Twin KSVC)and multiple birth support vector machines(MBSVM).It can be seen that for the existing multi-classification support vector machines,they are mainly based on the hinge loss function,which makes the obtained support vector machines more sensitive to noise and unstable resampled data.In order to further improve the performance of the multiclass support vector machine algorithm,this paper studies the structure-fuzzy multi-class support vector machine based on pinball loss and L₁ loss,constructs the corresponding support vector machine model,and proposes some improved support vector machines or twin support vectors machine algorithm.The specific research content is as follows:1.Based on the pinball loss function and L₁ loss function,the existing six multiclassification support vector machine algorithms are improved,including Sim MSVM,OVA TWSVM,OVO TWSVM,DAG TWSVM,Twin KSVC and MBSVM.Multi-classification support vector machine models based on pinball loss function and L₁ loss function are proposed and built.Use the pinball loss function to calculate the error of the sample points,not only to punish the incorrectly classified samples,but also to give additional punishment to the correctly classified samples,so that the support vector machine is insensitive to noise and the resampling data is stable.The L₁ loss function is used in algorithms,and at the same time considering the absolute loss of one type of sample points based on the L₁ norm and the hinge loss of other types of sample points based on the L₁ norm,so that the algorithm has better stability and robustness.2.Research on structural fuzzy multi-class support vector machine based on pinball loss function and L₁ loss function.In view of the existing multi-class support vector machine algorithms,it does not take into account the different roles of different samples in constructing the classification surface,and does not consider the potential structural information between the samples,which may contain some important prior knowledge and other issues.The sample fuzzy membership and sample structural information are introduced into the algorithm,and a structural fuzzy multi-classification support vector machine based on pinball loss and a structural fuzzy multi-classification support vector machine based on L₁ loss are proposed.3.Select UCI standard data set and synthetic data set,use ten-fold cross-validation method to conduct experiments on the proposed algorithms,and compare them with the existing multi-classification algorithms.At the same time,in order to study the impact of different samples with different weights on the classification results.Use three different methods including class center distance method,fuzzy C-means method and S-type to obtain sample weights,and use intra-class dispersion and clustering technology-based fuzzy Cmeans,hierarchical clustering and K-means clustering to obtain samples structural information and experimental comparison.