Abstract: K-Means is one of classical and heuristic clustering algorithm,which is sensitive to the model's initial state.This makes the initialization of the model deterministic to the clustering solution,and the process usually can obtain the local optima result.The study on supplying a initial seeds set that can reflect the characteristics about the distribution of the data is of great value for clustering research.They would be selected in a denser region as far as possible and would be dispersive as much as possible.And then Hybrid Distance Density Based Seeking (HYDD) stategy is offered:first of all,this method rearranges the original data on the principle of decreasing density,then the date higher density and longer distance than the diameter wer interted into a candidate set.Secondly,k seeds is selected from the candidate set based on the theory that the sum of distance is decreasing.Finally,K-Means is run with the initial seed set.Experiments results on 5 synthesis and 3 real datasets show that HYDD K-Means could obtain clusters which have maximal intra-cluster homogeneity and maximal inter-cluster separation.

Key words: clustering, initial seeds, heuristic searching, K-Means algorithm