Research On Subspace Learning Method For Multi-view Data Scene

In the era of big data,the data is growing exponentially.People are usually surrounded by the massive amounts of uncertain and fuzzy data from different views,which reflects the characteristics of multi-modality and high dimensionality.This type of data appears frequently,and impacts daily life.When facing massive multi-view high-dimensional data,analyzing and processing multi-view high-dimensional data efficiently and quickly is a hot topic in the field of machine learning and data mining.Although the corresponding proposals have been provided to solve some issues in recent years,they still cannot satisfy the increasing demands in the fields of industry,agriculture,production and life.This thesis mainly focuses on the issues of highdimensional data analysis in multi-view scenarios,using many technologies,such as subspace learning methods,information fusion methods and multi-view learning methods,to study various aspects of the subject.This thesis mainly focuses on dimensionality reduction(DR),information fusion(IF)and general learning framework(GLF)in multi-view high-dimensional data scenarios.First,two multi-view dimensionality reduction algorithms are proposed to reduce the dimensionality of high-dimensional data and improve the efficiency of multi-view high-dimensional data analysis tasks.Then,as to the information fusion problem of multi-view high-dimensional data,this thesis proposes a novel multi-view non-negative matrix factorization method to further improve the performance of information fusion between different views.Finally,two general learning frameworks are proposed for constructing a universal multi-view subspace learning framework,which directly extend most of the existing subspace learning methods to the field of multi-view learning.Therefore,this thesis has conducted extensive research on subspace learning methods in multi-view high-dimensional data scenarios and achieves the following results:(1)When facing the issues of dimensionality reduction of high-dimensional data,subspace learning methods based on low-rank representations are widely used.However,such methods usually ignore the local topological structure of the sample space and cannot be directly used to solve multi-view problems.To tackle these issues,this thesis takes full advantage of the lowrank subspace learning algorithm and introduces the local structured information in the sample space into the construction of the low-rank relationship matrix among samples,which proposes two multi-view dimension reduction algorithms based on local low-rank structure,including Multi-view Local Low-rank Embedding and Multi-view Low-rank Preserving Embedding.The former constructs a suitable manifold structure by using the local low-rank relationship between samples,and proposes a central view-based scheme to make full use of the correlation among the characteristics of different views;The latter provides three different embedding methods to fully maintain the low-rank reconstruction structure in each view,and integrates all views into one common latent space by minimizing the inconsistency between it and each view.(2)The information between different views is not only independent with each other but also complementary to each other.Thus,how to properly integrate information from different views is still a difficult problem.To fuse of multi-view data information,this thesis proposes a novel graph-agreement non-negative matrix factorization method,which fully considers the potential structural information in each perspective and the complementary information between the perspectives.The proposed algorithm uses non-negative matrix factorization technology to explore the potential structural information under each view,and uses the graph structure information between and within the views to maintain consistency between each other.Therefore,the proposed algorithm fuses the multi-view information consistently to improve the performance of multi-view high-dimensional data processing.(3)In recent years,subspace learning has been widely used in the field of high-dimensional data processing,but most of the current subspace learning methods cannot be directly applied to multi-view scenarios.Meanwhile,the model principles of most subspace learning methods are not con sistent,and the hypothesis biases also vary obviously.Against how to construct a universal multi-view subspace learning framework,this thesis proposes two multi-view subspace learning frameworks,which base on the similarity-regularization term and the graph-structure consensus term,respectively.Our goal is to directly extend most of the existing subspace learning methods into multi-view scenarios,which can facilitate related research work in the multiview field.Meanwhile,the proposed GLF aims to not only fully discover Rich information of multi-view high-dimensional features,but also further make full use of the application fields and performance advantages of existing subspace learning works.The former proposes a regular term based on the consistency of the similarity between samples to discover the complementary information among views,and provides two strategies based on the pairwise views and the central view to extend single-view subspace learning algorithms(including manifold learning,linear subspace learning and kernel subspace learning)into multi-view scenes;the latter proposes a graph-structure consensus term based on heterogeneous graphs to explore supplementary information between different perspectives and explore the correlation among different views,and maintains the graph structure information in each view to discover the diversity information among multiple views.In summary,this thesis studies a series of problems in the process of high-dimensional data processing in multi-view scenarios,and introduces related research results and innovations in detail,which hopes that the research results proposed in this thesis can open up a new idea and solution for the field of multi-view high-dimensional data analysis.