Research On Fuzzy Twin Support Vector Machine Classification Algorithm And Its Application

Classification problem is a common problem in practical applications, as is one of the most important research topic in machine learning domain. The twin support vector machines (TSVM) is a rapid classification algorithm for resolving discriminating problems by using a pair of optimization objectives in the corresponding quadratic programming. It does not only have the advantages of conventional support vector machine, but also have strong data processing capabilities. This thesis focuses on the classification problem, particularly concerns of the researches on the classifier model and the dual optimization problem solving algorithm of twin support vector machine, the main works of this thesis are as the follows:Chapter One briefly introduces the research background and the recent research situation of twin support vector machine, including statistical learning theory, support vector machines, fuzzy support vector machines, twin support vector machine and optimization algorithm of its quadratic programming problems.Chapter Two presents fuzzy margin twin support vector machine (FMTSVM) and L2-FMTSVM by adding fuzzy membership and margin in the primal problems of TSVM for binary classification, in order to enhance the noise immunity of the outlier data and im-prove the classification prediction performance of TSVM. The algorithm implements different penalty parameter for different sample and optimize the classification model based on the principle of structural risk minimization. At the same time, a novel dual shrinking coordinate descent method is proposed to solve FMTSVM dual problems which lead to very fast training. Average prediction accuracy and training speed of our FMTSVM algorithm are6%and1%higher than the TSVM respectively on UCI datasets. Experiments results on artificial and UCI datasets indicate our algorithm not only obtains faster learning speed, but also shows better generalization performance.Chapter Three proposes three new multi-class classification FMTSVM algorithms based on one-against-rest SVM、one-against-one SVM and partial binary tree SVM. FMTSVM is extended to multi-classification from binary classification and the performance of the new method is validated by its application in face recognition and traditional Chinese painting classification problem. Experiments on face database (ORL、Yale、Face94) show that av-erage prediction accuracy and training speed of our algorithm are1%and9%higher than the TSVM respectively. Experiments on Chinese painting images datasets indicate that pre-diction accuracy and training speed of the proposed algorithm are7%and74%higher than the TSVM respectively. This experiment demonstrates that the one-against-one FMTSVM is more effective and feasible than one-against-one SVM, one-against-one FSVM and one-against-one TSVM.Chapter Four is the conclusion part, in which the major findings and implications of the study will be made clear. In addition, some tentative suggestions will be provided for further study.