Research on a new membership functions based on fuzzy SVM
COMPUTER MODELLING & NEW TECHNOLOGIES 2013 17(5C) 164-166
Zhoukou vocational college of science and technology, Henan, 466000, China
Traditional membership functions in fuzzy SVM (FSVM) were designed based on the distance between a sample and its cluster center, which are irrational for dataset with non-spherical-shape distribution. A new membership function was proposed based on the distance between a sample and a hyper plane within the class. It overcomes disadvantages of traditional designing methods and improves the generalization ability of FSVM, while reducing the dependence of membership function on the geometric shape of sample data. Numerical experiments show that, compared with the traditional SVM and three FSVM with different membership functions, FSVM with new membership function has better classification accuracy. The New method is simple and its computation time is less.