There is no incremental learning ability for the traditional support vector machine (SVM) and there are all kind of merits and flaws for usually used incremental learning method. Normal SVM is unable to train in large-scale samples, while the computer’s memory is limited. In order to resolve this problem and improve training speed of the SVM, we analyze essential characteristic of SVM and bring up the incremental learning algorithm of SVM based on regression of SVM related to SV (support vectors). The algorithm increases the speed of training and can be able to learning with large-scale samples while its regressive precision loses fewer. The experiments show that SVM performs effectively and practically. Its application to prediction of the transition temperature (Tg) for high molecular polymers show that this model (R2=0.9427) proved to be considerably more accurate compared to a ANNs regression model (R2=0.9269).