Multiclassification Method Research Based On Fuzzy Support Vector Machines

Statistical Learning Theory is for small samples machine learning, and SVMs (Support Vector Machines) proposed by Vapnik are based on this theory. SVMs have great generalizability and pursue the best result through limited information. Because of the replacement of ERM (Empirical risk minimization) by SRM (Structural Risk Minimization), SVMs have a firm theoretic base; because of the use of kernel functions, SVMs transform non-linear problem to linear problem, so reduce the complexity of algorithm. Now SVMs behave well in comparison with ANN, GA, and this method is studied in many fields such as handwritten numerals recognition, face recognition, texture classification.Multiclassification is a branch of machine learning. SVMs were originally proposed for 2-classification, so people extended this method to meet multiclassification. Furthermore in some practical issues, the margin between classes is not clear, so FSVMs (Fuzzy Suppport Vector Machines) are given. For different purposes, there is mainly two methods based on FSVMs. In this paper, the author analyses these two methods in detail, gives advanced algorithms based on FSVMs and does some experiments to approve them. The two methods are listed as follow:One was given to resolve unclassifiable regions in one-aganist-one and one-against-all SVMs by Takuga and Shigeo. In this paper, An advanced model is proposed to keep coherence of all multiclassification methods.The other is given to incarnate different influence of samples and reduce the influence of noisy data by Chun-Fu Liu, Sheng-De Wang, Han-Pang Huang. In this paper. The author proposes a combination of this method and one-against-all rule to get a more generalizable algorithm.Through expeiments, the algorithms proposed in this paper get more remarkable results.