Abstract: To reduce large-scale training sample set, the concept of k-nearest vectors is proposed, and a new account method for dissimilarity is given accordingly. Then, the paper proposes and proves the methods of noise identification and boundaries distance description. Based on these methods, an efficient sample reduction algorithm is proposed. The algorithm removes noise samples according to the dissimilarity at first step, then according to the similarity of samples, and the dissimilarity which describes the distance between sample and classification boundary, the algorithm removes minor training samples from the original sample space directly. Experiments indicate that the reduction algorithm can effectively reduce the sample, and improve the training efficiency.