Abstract: A new initializing algorithm based on data distribution is proposed for K-Means in this paper.First we partition data space into grid and find the local-maximum cell which counts more data points than its neighborhood cells.Then we use distance optimization method to choose the seed clusters from local-maximum cells globally.Benchmark experiments evaluate the proposed method and five other typical initialization methods on both synthetic and real-life data sets,and the results demonstrated that our proposed algorithm gives faster convergence speed without descending in clustering performance.