Research On Selection Strategy Of Cluster Members In Clustering Fusion Algorithm

Without any prior information of data distribution,clustering often needs to make the assumption of the data,and then uses certain rules to divide the data,but the assumption is not sure to fit in with the real distribution of the data.Clustering consensus is a good way to solve this problem,which uses consensus function to combine various clustering members,achieved by certain clustering algorithms,and then can obtain more robust,stable and consistent final clustering result.However,the accuracy of the final clustering result may be reduced,when the quality of certain clustering members participated in the consensus(i.e.consensus members)are low.So,some high quality consensus members should be selected to be participated in the consensus,which is called the selective clustering consensus algorithm,where selection strategy and consensus function are two important research areas.The thesis analyzed the existing selection strategies of selective clustering consensus algorithms.One kind of selection strategies mainly focus on selecting the high quality consensus members.However,there are various evaluation criteria of clustering quality,each of which is often only suitable for certain data distributions,in addition,high quality clustering members often have little difference between each other.The other kind of selection strategies first select the highest quality cluster member as the reference member,and then select consensus members,by considering the influence of quality and difference into account.However,the kind of selection strategies depend on the selection of this single reference member,and there is the above problems in evaluating the quality of the reference member.To solve these problems,this thesis proposed a novel selection strategy based on multiple reference members.The selection strategy first groups the initial clustering members into various clusters,so that there are relatively large difference between the clusters.Then,the highest quality clustering member is selected as reference member in each cluster,based on which consensus members are selected in each cluster,by considering the influence of quality and difference into account.Unlike the existing selection strategies,the selection strategy doesn't depend on one certain referencemember,because it selects a number of reference members with relatively large difference between each other.By this way,consensus members can be selected with high quality and relatively large difference between each other,and then better final clustering result can be obtained.Experiment results show that the selective clustering consensus algorithm proposed in this these is effective,and can obtain better clustering results than the existing algorithms in most cases.