Multi-view Clustering Based On Deep Graph Regularized Non-negative Matrix Factorization

With the development of Internet information technology,data becomes an important part of human's production activities and social life.The Internet era produces huge high-dimensional and polymorphic data which have posed big challenge for data analysis.Therefore,there is an urgent need to find efficient data analysis tools to help us understand the meaning of data and generate valuable information.However,traditional data analysis methods are difficult to achieve efficient analysis and processing for these complex data.Research on cognitive physiology points out that high-dimensional data often has an inherent low-dimensional structure.In the meanwhile,Non-negative Matrix Factorization(NMF)is able to discover this low-dimensional structure in original data and at the same time preserve local features of data.This technology helps people understand the implications of data and it has been widely applied in many fields such as machine learning,pattern recognition and signal processing.However,most existing NMF techniques only focus on a single dataset(view)and is hard to achieve satisfactory performance in some applications when multi-view analysis is crucial.Since multi-view learning combines the compatibility of different views data,it breaks the bottleneck of single-view.The researchers point out observing an object from different aspects can provide complementary information that is not available in single view.Multi-view learning,an effective extension of single-view learning,aims to synthesize the information across all views to make up for the shortcomings of single-view and improves the performance of high-level applications such as clustering.This thesis is devoted to build new methods for multi-view clustering based on NMF.Firstly,this paper expounds the significance of NMF and Multi-view Clustering and makes a comprehensive analysis of review on their research status.Besides,a briefly introduction of several common popular variants of NMF and their corresponding principles are presented.Furthermore,this paper introduces the Multi-view Clustering algorithm based on NMF and the fusion style after representation learning.Secondly,this paper focuses on the application of NMF-based multi-view learning,and developed new methods and algorithms for Multi-view Clustering based on Deep Non-negative Matrix Factorization(Mv DNMF)and Multi-view Clustering based on Deep Graph Regularized Non-negative Matrix Factorization(Mv DGNMF)methods.Both Mv DNMF and Mv DGNMF not only can extract hierarchical attributes of data by imposing the deep structure but also fuse the complementary information across all views and hence improves the clustering performance of the data as much as possible.In addition,Mv DGNMF effectively alleviates the problems of poor interpretability and over-abstract representation in the procedure of linear combination of multi-layer non-negative matrices(or semi-negative).Besides,Mv DGNMF can discover hierarchical local representations by constructing graph Laplace regularizations.In summary,Mv DGNMF is able to exploit the similarity between instances to improve the efficiency of information transmission layer by layer and drive the representations of the last layer of all views to a consensus.Efficient algorithms based on the alternating multiplicative update rules are developed for Mv DNMF and Mv DGNMF.Finally,in order to demonstrate the effectiveness of the proposed Mv DNMF and Mv DGNMF methods,we employ several open datasets(including image and text datasets)to evaluate their clustering performances.By comparing with the performances of other state-of-art methods,we demonstrate observed that Mv DNMF and Mv DGNMF methods are efficient and often achieved better performance than the other compared methods.