SVM classification result on imbalance data set is partial to majority class. It makes the F1_measure value of minority class inadequate. This paper presents a method which improves the classification accuracy of minority class. The method generates weights to optimize parameters b of the classification hyper plane without changing the number of samples, and combines the total number of samples, the number of support vector, the accuracy of minority class and majority class. Finally, the effectiveness of the method is proved by experiments.