The determination of fuzzy membership function in the fuzzy support vector machine (FSVM) is a difficult problem. To solve the problem of being sensitive to the noises and outliers in support vector machine, by the inspiration of Bayesian decision theory, combining with sample density characteristics, sample points relation between same class and other class is researched, and the tightness on each sample points is described. Based on that, method of posterior probability and sample density weight are given to each sample, and new fuzzy membership function is proposed. The detection of the noises and outliers is avoided by this method. Numerical simulation shows that the improved fuzzy membership function method is effective.