Abstract:RSKNN is an improved kNN algorithm based on variable parameter rough set model. The algorithm guarantees under the premise of a certain classification accuracy, effectively reduces the computation burden of the classified samples, and improves the computation efficiency and precision of classification. But in this algorithm,the instances of each class are simply classified into core and boundary areas. It has the limitation that it isn't classified according the features of datasets. An efficient algorithm aiming at learning multi-representatives for RSKNN is proposed. Using the theory of structural risk minimization, a few factors that determine the expected risk of new classification model are analyzed. And an unsupervised algorithm for partial clustering is used to build an optimal set of representatives. Experimental results on UCI public datasets demonstrate that the proposed method significantly improves the accuracy of the classification.