A Graph Based Approach For Clustering Ensemble Of Fuzzy Partitions

Abstract

Fuzzy clustering and Cluster Ensemble are important subjects in data mining. In recent years, fuzzy clustering algorithms have been growing rapidly, but fuzzy Clustering ensemble techniques have not grown much and most of them have been created by converting them to a fuzzy version of Consensus Function. In this paper, a fuzzy cluster ensemble method based on graph is introduced. Proposed approach uses membership matrixes obtained from multiple fuzzy partitions resulted by various fuzzy methods, and then creates fuzzy co-association matrixes for each partition which their entries present degree of correlation between related data points. Finally all of these matrixes summarize in another matrix called strength matrix and the final result is specified by an iterative decreasing process until one gets the desired number of clusters. Also a few data sets and some UCI datasets data set are used for evaluation of proposed methods. The proposed approach shows this could be more effective than base clustering algorithms same of FCM, K-means and spectral method and in comparison with various cluster ensemble methods, the proposed methods consist of results that are more reliable and less error rates than other methods.