Research On Non-negative Matrix Factorization Algorithm For Multiple Views

With the advancement of metadata technology globalization,metadata displays the traits of high dimension,multi-morphology,multi-source and complex structure.Facing such a large amount of data,the traditional data processing methods often have little effect,even disappointing,and a lot of information contained in the data can’t be obtained and perceived.Therefore,how to deal with high-dimensional data efficiently and how to obtain key information effectively has become an important topic in information science and technology.Non-negative matrix factorization,as a representative representation learning method,can decompose a high-dimensional nonnegative matrix into two non-negative factor matrices,which are interpreted as basis matrix and coefficient matrix respectively,and effectively mine and learn the low dimensional structure and local representation of data.The nonnegative matrix factorization algorithm,however,is mainly aimed at single-view data.In the era of big data,the data is huge and diverse,and multi-view data is the norm.Multi-view data may be collected from different sources or represented by different features for different tasks.Multi-view data is rich in information,and multi-view learning method is superior to single-view learning method in various tasks.Traditional single-view data clustering can’t integrate comprehensive information,so its clustering performance and robustness will be limited when facing multi-view data.Multi-view clustering aims to eliminate this defect and can integrate the differences and complementary information of multi-view data,which has fascinated extensive interest.The principal works of this dissertation are as follows:(1)A general multi-view non-negative matrix factorization clustering method based on structural regularization is proposed.The proposed algorithm can achieve local and global consistency,not only capture the global and local structure of data for discriminant learning,but also learn the representation by using the diversity and differences of multi-view data.In addition,the optimal factor matrix can be obtained by optimizing the objective function with efficient alternating iterative algorithm.Most of those results on several data sets indicated that the proposed model is superior to relational clustering algorithms.(2)A novel multi-view nonnegative matrix factorization algorithm based on adaptive learning is proposed.The model combines the information complementarity and compatibility of multi-view data,and integrates adaptive structural awareness learning to enhance the acquisition of local structural information and different types of discriminant information.In addition,an effective update algorithm is proposed for the optimization problem,which supervises the tendency of the method.The experimental results on benchmark data sets are compared with other clustering methods,and the robustness and efficiency of this proposed method are confirmed.Generally speaking,based on the similarity,complementarity and difference of multi-view data,this dissertation proposes two non-negative matrix factorization algorithms based on multi-view,and the theoretical analysis and experimental verification show that the proposed two models can effectively improve the clustering accuracy of multi-view data.