Research And Application Of Cluster Integration

Cluster analysis is a active branch in data mining, has been widely used in pattern recognition, information retrieval, machine learning and other areas. Different algorithms get different result on the same data, no single clustering algorithm that can be the best solution for a wide range of data sets. The emerging of cluster ensemble gives a prominent solution for this question.Clustering ensemble is made up of two steps: generating diversity base clusterings and use consensus function to get the final result. To generate base clusterings, one can use different parameter of one algorithm, different subset of the data set or different subspace. consensus function via co- association matrix, cluster association matrix or other method obtain the ensemble relationship, and then algorithm like hierarchical cluster or hyper graph partitioning could be used to get the final clustering result.First the thesis use Auto Code and ClusterDP generated cluster member,Then use the consus method of WOMC algorithm get the final result. This method show good performance Second the theses research the voting based method.The method has a drawback since it has a strict restrict on all base clusterings must generate the same number of clusters. This pose the very similar element in mutual information matrix, and lead to the weight of base clustering is very close to each other which degenerate the effect of weighting. so we propose a new method to embed the weight of base clustering in LCE algorithm which can let base clustering generate different number of clusters. At last the thesis research the image classification method based on k-means. For the k-means is sensitive to the initial cluster centre and easily get in local optimal. For this reason the thesis use the cluster ensemble method to feature transform, get the better represent of the data. Through experiment obtain the higher accuracy than the original method.