Aimed at the problems that he traditional linear Support Vector Machine (SVM) is equal to each dimension of input features in training data sets, and the classification in the original space directly leads to the low prediction accuracy, a method for combining lowdegree polynomial mappings and feature weighted is proposed to improve the classification performance of linear SVM. First of all, each sample is mapped into the twodegree explicit feature space by using the polynomial trick to increase the implicit information of samples. Then, the feature weights of each dimension are calculated by using the fuzzy entropy feature weighted algorithm. By feature weighting, the magnitude of contribution to the result of classification can be measured. To verify the robust of the proposed method, the totally seven data sets from different database are tested. Making a comparison between Kernel SVM and other improved Linear SVM algorithms in training time and prediction accuracy, the results show that the training time reduces an order of magnitude with the expansion of data set, and the prediction accuracy can keep up with few training samples even close to Kernel SVM. The running results show that the proposed method can effectively improve the overall performance of the linear vector machine.