Standard C-support vector machine (C-SVM) algorithm has certain limitation when dealing with many factual pattern classification problems, especially in the extreme case such as the recognition error cost loss in great difference. A kind of generalized C-SVM algorithm is introduced. By estimating the cost of the recognition error, optimal separating hyperplane can be translated into the low cost passage， and leaves more space for the high lost cost to increase recognition rate, thus reducing the damage of recognition error. The new method improves the applicability of C-SVM and sample recognition correct rate. When applied to radar high resolution range profile′s recognition, experimental results show that the proposed method can achieve better recognition effect than the traditional method.