In order to speed up the face recognition algorithm under the condition of keeping high accuracy, a discriminative feature selection algorithm is proposed to handle multi-class face recognition problems. It is supported by vector machine to select features and employ total probability rule to integrate feature selection and multi-class classification into an integral frame. The experiment on the face databases of UMIST and FERET displays that the proposed algorithm can effectively select the features which have obvious physical meanings, thus speeding up the response of classifier without degrading the generalization performance.