Research On Clustering Ensemble Method For Fusing Structural Information

Cluster analysis methods are widely used by researchers in different fields.However,when clustering analysis is applied to practical tasks,due to the complexity of data in the real world,the scope,reliability and stability of a single clustering method are restricted.Therefore,researchers have proposed a new method called clustering ensemble method.The goal of clustering ensemble method is to integrate multiple existing base clusters through a certain strategy to obtain a consistent,more accurate,and more robust clustering result.But the existing clustering ensemble methods still have certain limitations,that is,when generating consistent clustering results based on the clustering members,usually only the correlation between the data point pairs is calculated,and the correlation between the clusterings in the clustering members are ignored.so this method can not fully explore the potential relationship between data objects.For the reasons given above,how to mine the potential association relationships between data objects based on the clustering members,and to fuse their association relationship and characteristics is a key problem that needs to be solved in the research of clustering ensemble methods.Aiming at these problems,this paper mainly focuses on the following contents:(1)In view of the problem that there are a large number of potential correlations between data objects in the real world,a global structure information extraction method for data objects is proposed.In the process of extracting structural information,the information entropy is used to extract the correlation between data objects and the correlation between clusterings according to the results of base clustering methods.(2)In order to get the results of consistent clustering ensemble,we fused the global structure information and features of the data objects.The graph neural network is used to fuse the global structure information and features of the data objects to obtain the low-dimensional representation.Then a self-supervised clustering integration model is constructed to integrate the low dimensional representation of data objects.Thus,an integrated optimization model is obtained.(3)Experiments were performed on the real data set.the correlation matrix obtained by the global structure information extraction method and the correlation matrix obtained by the classical data objects correlation calculation methods were input into the spectral clustering as a adjacency matrix.The results show that methods in this paper can obtain better consistent integration result than other methods.Then,comparing the clustering ensemble algorithm that fuses structural information with the classic clustering ensemble algorithms,the results show that the algorithm in this paper is compared with HGPA,CSPA and MCLA can further improve the accuracy of clustering ensemble results.