Research On Multi-view Dimension Reduction Methods Based On Matrix Decomposition

With the rapid development of Internet,data acquisition technology and information retrieval,we are now facing with the explosive growth of related data for research.How to find the intrinsic structure of complicated data,namely dimension reduction has been widely used in many fields,such as computer vision,data mining technology,pattern recognition,machine learning and so on.With the continuous improvement of feature exaction algorithms,data of single view cannot describe complicated data and people are faced with more and more multi-view data.However different views have different statistic properties,traditional machine learning algorithms cannot deal with these data.Therefore,how to explore the common information and heterogeneous features simultaneously becomes the core of dealing with multi-view data.Based on depth study of the existing dimension reduction algorithms,first,a new dimensionality reduction framework with sparse graph embedding is proposed.O n the one hand,the characteristic of sparse representation is used to find compact discriminative latent subspaces of the data;through "sampling and projection" strategy,the linear dimensionality reduction methods using sparse graph embedding,on the other hand,can be employed to effectively deal with large scale data.Quantitative and qualitative experimental results verify that the performance of original algorithms can be improved by the proposed methods;simultaneously applicatio n of "sampling and projection" strategy greatly reduces the time complexity and space complexity of the algorithm,and makes it suitable for large scale data.Second,a robust multi-view dimensionality reduction algorithm based on Markov chain is proposed.It not only takes advantage of the properties of the stationary distribution of Markov chain to effectively eliminate potential noise in transition probability matrix by low rank and sparse decomposition but also constructs linear map matrix via similarity or reconstruction which makes high-dimensional data can easily get their corresponding low-dimensional representations.By compared with some classical multi-view dimension reduction algorithms the proposed method is proved to get more accurate low-dimensional embeddings and achieve high precision in classification tasks.Finally,a multi-view dimension reduction method based on matrix decomposition is proposed.The proposed method explores and integrates common information and heterogeneous features among different views via matrix decomposition when compared with existing algorithms.Moreover,the weight of each view can be learned automatically in the dimensionality reduction process,which would be used to adjust the contribution of each view in the process of constructing a low-dimensional consensus representation.Experimental results show that the proposed can explore shared features and different statistic properties of multiple views,so better classification performance can be gotten.