Research On Concept Factorization-based Multi-view Clustering Method

In the age of big data,there are a sharp increase in the number of sources and means to collect data.Data comes from different sources but describes the properties of the same thing is called multi-view data.How to effectively implement multiview data clustering has become an important research direction in the current data mining field.Non-negative matrix factorization(NMF)is widely used in multi-view clustering due to its operability and interpretability in mapping high-dimensional complex data to low-dimensional subspace.However,multi-view clustering based on NMF methods often only consider the clustering of multi-view data in latent space,and explore the consensus information of multi-view data around the lowdimensional representation matrix,while ignoring the complementary information among views.Concept factorization is a distortion of NMF,which solves the limitation that NMF is only useful on non-negative data,and increases the application range of clustering algorithm.Therefore,this paper focuses on concept factorization,fully considers the consensus information and complementary information of multi-view data,and makes use of the diversity of multi-view data in different views and different spaces for clustering.The main research contents are summarized as follows:1)Multi-view Clustering Method based on HSIC and Concept FactorizationConsidering that the current multi-view clustering method based on concept factorization focuses on the consensus information of multi-view data while ignoring the complementary information among the different views,this paper proposes a clustering method for multi-view data,called Multi-view Clustering Method based on HSIC and Concept Factorization(HSICMVCF).In this paper,the Hilbert-Schmidt independence criterion is utilized to measure the degree of dependence among different views,which enhances the diversity of different views and ensure the complementary information of multi-view data.In addition,a manifold regularization is introduced to keep the local geometry of the multi-view data approximately unchanged,which enhances the feature representation of the latent space.Finally,the results of the comparison experiment indicate that the proposed method can improve the clustering performance.2)Concept Factorization-based Collaborative Multi-view Clustering Method in Visible and Latent SpacesTo solve the problem that the current multi-view clustering methods treat the clustering of multi-view data in visible space and latent space as two independent processes,this paper proposes a multi-view clustering method,called Concept Factorization-based Collaborative Multi-view Clustering Method in Visible and Latent Spaces(Co MVCF).It fully exploits the consensus information of multi-view data in latent space and the discriminant information in visible space.In addition to introducing manifold regularization,this paper integrates the clustering of data samples in visible space and latent space for collaborative learning and optimization.At the same time,entropy weighting strategy is used to measure the importance of each view to realize the adaptive allocation of view weight.Finally,the proposed method is proved to be superior to other advanced baseline models in a comparison experiment.