A review on multi-class TWSVM

摘要

Twin support vector machines (TWSVM), a novel machine learning algorithm developing from traditional support vector machines (SVM), is one of the typical nonparallel support vector machines. Since the TWSVM has superiorities of the simple model, the high training speed and the good performance, it has drawn extensive attention. The initial TWSVM can only handle binary classification, however, the multi-class classification problems are also common in practice. How to extend TWSVM from binary classification to multi-class classification is an interesting issue. Many researchers have devoted to the study of multi-class TWSVM. Although the study of multi-class TWSVM has made great progress, there is little literature on the comparisons and summaries of different multi-class classifiers based on TWSVM, which not only makes it difficult for novices to understand the essential differences, but also leads to the problem that how to choose the suitable multi-class TWSVM for a practical multi-class classification problem. This paper aims to review the development of multi-class TWSVM in recent years. We group multi-classTWSVM reasonably and analyze them with the respect to the basic theories and geometric meaning. According to the structures of the multi-class TWSVM, we divide them to the following groups: “one-versus-rest” strategy based multi-classTWSVM, “one-versus-one” strategy based multi-class TWSVM, binary tree structure based multi-class TWSVM, “one-versus-one-versus-rest” strategy based multi-class TWSVM and “all-versus-one” strategy based multi-class TWSVM. Although the training processes of direct acyclic graph based multi-class TWSVM are much similar to that of “one-versus-one” multi-class TWSVM, the decision processes of direct acyclic graph based multi-class TWSVM have their own characteristics and disadvantages, so we divide them to a separate group. This paper analyzes and summarizes the basic thoughts, theories, applicability and complexities of different multi-class TWSVM of different groups and presents experimental results to compare the performances.