Research On Multi-view Clustering Algorithm Based On Multi-level Information Fusion

As a critical technique in data mining,multi-view clustering has received widespread attention.And it has been successfully applied in the fields of biology,medicine,natural language processing and graphics and image processing,and has high research and application value.Multi-view clustering obtains a comprehensive clustering partition by fusing the data information of multiple views.To use the heterogeneous information of different views for clustering,the key lies in how to efficiently merge the information of all views.Most of the existing multi-view clustering algorithms first extract the latent representations of the multiview data,and then use the data-level fusion method or the partition-level fusion method to fuse the extracted latent representations.The representation learning,data-level information fusion and partition-level information fusion of these algorithms are disconnected.In order to solve the above problems and improve the clustering performance of the multi-view clustering algorithm,this paper proposes a multi-view clustering algorithm based on multi-level information fusion.The main research contents of this paper are as follows.First of all,based on the existing MvDSCN method that combines representation learning and data-level information fusion,a Multi-view Clustering algorithm based on Two-level fusion is proposed,which is called MvC-T.The basic idea is to add the multi-view information fusion of the partition level,and form a two-level fusion structure with the existing data-level information fusion.Among them,the multi-view information fusion operation of the partition level is to combine the graph structure and use the idea of updating the graph to perform the basic partitions fusion.The experimental results show that compared to the basic model and some other existing multi-view clustering algorithms,MvC-T has better clustering performance on four evaluation indicators including accuracy.This shows that MvC-T is an effective multiview clustering algorithm.Secondly,because the data-level information fusion and the partition-level information fusion in MvC-T cannot achieve joint fusion,and only the common self-representation matrix is used to extract the edge relations.The clustering result is still easily affected by views with a lot of noisy data.In order to further improve the clustering performance of the algorithm,a Multi-view Clustering algorithm based on Joint learning is proposed,which is called MvC-J.The basic idea is as follows: First extract new edge relations for better data-level information fusion,and then add the clustering loss to guide the generation of the self-representation matrix and the update of the basic partitions.So as to realize the joint learning of representation learning,data-level information fusion and partition-level information fusion.In the experiment,by comparing with MvC-T and some other existing multi-view clustering algorithms,the results show that MvC-J performs better in four evaluation indicators including accuracy.