Aiming at the traditional method and its precision which used as evaluation index can not meet the requirements of practical classification. Introduce cost sensitive method into SVM, put forward a new learning algorithm CSSVM (cost-sensitive SVM), and acquire projection sub-gradient solving method which is similar as Pegasos to deal with large scale data. The Pegasos process includes initialization, iteration, ascertaining step lengths and direction of sub-gradient descent, update, projection and the end. The test results show that this algorithm can effectively improve identifying rate and identifying precision and it is competitive.