Research On Multi-view Learning Algorithm Based On Tensor Correlation

With the development of data collection technology,the ways of acquiring data are becoming more and more diverse,and the obtained data usually has multiple views,thereby forming multi-view data.How to efficiently use the information contained in multi-view data for learning is the research goal of multi-view learning.In order to make better use of multi-view data and promote the application of multi-view learning in practice,we must consider the multi variable correlation.To solve the problem of multivariate correlation of single view data,based on the matrix information channel of single input and single output,this paper proposes the tensor information channel of multivariate input,which can study how to share information among multiple variables.Compared with the matrix information channel,the tensor information channel is more effective in solving practical problems.In addition,this paper extends the information bottleneck algorithm to tensor information bottleneck algorithm,and the corresponding experiments show that the method has good performance in multi variable compression,and solves the problem of low efficiency of matrix information channel method,especially when the output variable depends on multiple input variables.As a classical algorithm in multi-view learning,multi-view dimensionality reduction algorithm has attracted many researchers' attention.However,some existing algorithms ignore the high-order statistics(related information)which can only be found by exploring all features at the same time.In addition,the current typical correlation analysis method only considers the linear relationship between variables,and only kernel function can be used for the nonlinear relationship between variables.Therefore,in order to solve the above problems,this paper proposes a mutual information tensor analysis(MITA)method based on mutual information matrix analysis and tensor canonical correlation analysis(TCCA).This method deals with the linear or non-linear interaction of any number of view data by analyzing the mutual information tensor of different views.The goal of MITA is to directly maximize the typical correlation of multiple views.This problem is equivalent to finding the best rank-1 approximation of data mutual information tensor,which can be effectively solved by alternating least squares(ALS).Through the multi-view dimensionality reduction experiments of attack network traffic prediction,advertisement classification and biometric structure prediction,the effectiveness of this method is proved,especially in the case of low dimensional subspace.