Abstract: Due to disadvantage of the affinity propagation algorithm of which the number of clusters can not be pre-specified, an improved method including three phases is proposed in this paper. The proposed method uses affinity propagation algorithm to obtain the representation center points of the dataset. Then K-means is applied to the clustering of the center points and produces the initial training set. Moreover, the modified K nearest neighbor algorithm is applied to the procedure of clustering analysis. Artificial data and UCI datasets are used in experiment to compare the new algorithm with other clustering menthes. The results demonstrate that the new clustering algorithm is outperforms the affinity propagation algorithm and traditional clustering algorithms.