Research On Feature Selection And Multi-Class Classification Methods Based On Twin Support Vector Machine

As an effective learning algorithm, Support Vector Machine(SVM) has been widely used in areas such as pattern recognition, Twin Support Vector Machine(TWSVM) is a new machine learning algorithm developed on the basis of SVM, it gets one pair of non-parallel hyperplanes by solving two smaller sized QPPs, which not only has a faster computing speed, but also has a strong generalization ability. Given its excellent learning performance, it has received extensive attention of scholars. In this paper, we focus on the expansion methods in feature selection and multi-class classification problem based on the TWSVM classification algorithm. Specific contents are as follows:Firstly, this paper introduces the present situation of the TWSVM algorithm, meanwhile, the classification principle of TWSVM is discussed detaily in binary classification problem, then we carry on research to many kinds of improved methods of TWSVM that existed at present, and analyze their performance and range of application.Secondly, in feature selection problem, taking into account that there are two feature weight vectors in TWSVM, where each feature corresponds to two weights, we can not directly use these two weights for feature selection. In response to this situation, we introduce a new feature weight item to give the corresponding weights for different features, to solve the above weights selection problem, and then a new feature selection method based on linear twin support vector machine is proposed. It selects features during classifier construction by introducing weight matrix of features in the primal formulation of TWSVM. In the solving process, it utilizes alternating iterative optimization method to decompose the problem of solving the model into two sub-problems, and solves the sub-problems effectively. The feature selection method is analyzed and compared on UCI datasets, simulation results verify the proposed method is effective.Finally, in multi-class classification problem, based on the SVM multi-class classification algorithm, one-against-rest TWSVM, one-against-one TWSVM and binary tree TWSVM are researched in depth, meanwhile, we compare their advantages and performance, which is analyzed and verified on UCI datasets. Next, the TWSVM multi-class classification algorithm is applied to the bearing fault diagnosis problem and a method of fault diagnosis based on TWSVM is proposed. In the process of fault identification, we use wavelet packet analysis techniques to extract fault features efficiently and use one-against-one TWSVM to classify the bearing fault patterns, it achieves good recognition effect.