Research On Multi-view Clustering Methods Based On Non-negative Matrix Factorization

With the rapid development of Internet,digital multimedia,and information acquisition technology,data gradually presents the characteristics of mass,high dimensional,multi-source and heterogeneous,multi-source or heterogeneous data constitutes multi-view data.Therefore,how to extract useful information from high dimensional multi-view data becomes a pivotal task for data processing and analysis.Clustering is a basic method in the field of machine learning and data mining.The purpose of clustering is to divide the target data set according to the similarity between the data.Because of its interpretability and physical meaning,non-negative matrix factorization has become a significant method in clustering problem.In recent years,several multi-view clustering methods based on non-negative matrix factorization have been proposed.Although these multi-view clustering methods have achieved promising clustering performance,non-negative matrix factorization-based methods still have drawbacks requiring further in-depth research for multi-view clustering problems.Existing multi-view clustering methods based on non-negative matrix factorization pay more attention to the shared information contained in the multi-view data,neglect the complementary information contained in the multi-view data.These methods do not take the similarity between inter-view into consideration.In addition,these multi-view clustering methods based on non-negative matrix factorization separate representation learning and clustering analysis into two steps,which affects the effectiveness of multi-view clustering.In this paper,on the basis of some related theories,such as non-negative matrix factorization,pairwise co-regularization,manifold regularization,and Hilbert-Schmidt independence criterion,we propose two novel multi-view clustering algorithms based on non-negative matrix factorization to promote the clustering performance.The main contributions are summarized as follows.Firstly,a multi-view clustering method based on non-negative matrix factorization and pairwise co-regularization is proposed.Most existing non-negative matrix factorization-based multi-view clustering methods ignore the similarity between inter-view.Based on the nonnegative matrix factorization framework,pairwise co-regularization method and manifold regularization method are used to construct similarity constraints between inter-view and intra-view,respectively.Then,the low-dimensional representations corresponding to multiview data are more discriminative.Experimental results show that the proposed method has higher clustering accuracy for multi-view clustering task.Secondly,a diversity-consistency multi-view clustering based on non-negative matrix factorization is proposed.Most existing non-negative matrix factorization-based multi-view clustering methods have not made full use of the complementary information in multi-view data,and separate representation learning and clustering analysis into two steps.To address these challenges,we propose a novel multi-view clustering method,named diversityconsistency multi-view clustering method based on non-negative matrix factorization,which conducts representational learning and clustering analysis simultaneously.This method uses the Hilbert-Schmidt independence criterion to model multiple representations to capture the complementary information of multi-view data.With the help of class-label consistency constraints,each data point under different feature representations can obtain the same clustering results.Experimental results show that the proposed method significantly improves the clustering accuracy and normalized mutual information.In summary,this thesis proposes two multi-view clustering methods based on non-negative matrix factorization to improve the clustering performance from the aspects of similarity,diversity and consistency of multi-view data.Theoretical analysis and validation experiments show that the superior clustering performance of the proposed methods over other state-of-the-art multi-view clustering methods based on non-negative matrix factorization.