Multi-view Data Classification Research

With the continuous development of data collection technology,complex multi-view data with massive expression forms is constantly emerging in the various fields of scientific ex-ploration and daily life.Therefore,studying the multi-view data analysis has very important practical value and significance.Multi-view classification is the basis of multi-view data analysis.Most of the existing multi-view classification methods are based on linear dis-criminant analysis.They require the input data to obey the global Gaussian distribution and cannot capture the potential local structure of the data,resulting in poor applicability.While some locality-aware methods can capture the local structure of data,they involve KNN pro-cesses,which need to preset neighbor number or local feature matrices,resulting in low flexibility.At the same time,most of the existing deep multi-view classification methods do not fully consider the inherent structural characteristics of multi-view data,and can not effectively extract the view-invariant representations.In view of these problems,this paper studies the multi-view locality adaptive discriminant analysis(MvLADA),multi-view clas-sification method based on deep adversarial network(MvDAN)and enhanced multi-view classification method based on deep adversarial network(MvDANE),The main research contents of the paper are as follows:1,Most of the existing multi-view classification methods based on linear discriminant anal-ysis depend on the assumption that the input data obeys the global Gaussian distribution,but the data in real life is usually nonlinear,resulting in poor applicability.In addition,some locality-aware multi-view classification methods involve KNN processes,which need to preset neighbor number or local feature matrices,leading to low flexibility.In view of the above problem,this paper proposes the multi-view local adaptive discriminant analysis(MvLADA)method.MvLADA integrates linear transforms learning and weighted matrix learning into a uniform framework,and the adaptively learned weighted matrix is shared by all views.MvLADA does not limit the distribution of input data-It captures the intrinsic local structure of the data through a weighted matrix obtained by adaptive learning,and has strong applicability.In addition,MvLADA adaptively learns the weighted matrix without performing the KNN process,mote reliably capturing the potential local structure of the data,and has strong flexibility.MvLADA is nearly free of parameters.Experimental results on the multi-view databases verify the effectiveness of the proposed algorithm.2.MvLADA is a shallow multi-view classification method,which can not extract the deep features of multi-view data.However,most of the existing deep multi-view classification methods can not effectively extract the view-invariant representations.In view of these shortcomings,this paper proposes a multi-view classification method based on deep adver-sarial network(MvDAN),which utilizes the interaction between the feature projector and the view classifier to obtain the view-invariant representations more effectively*Feature projector captures the potential shared properties of multi-view data,which uses structure preservation to force the samples of the same label from different views to be close,and the samples of different label from different views to be far away.The view classifier captures the specific properties of each view and attempts to distinguish the sample's view based on the differences between the different views.In addition,MvDAN also introduces the label prediction structure to obtain the predictive labels of samples by minimizing cross entropy loss.However,MvDAN does not measure the distance of the extracted deep features in the same view,so it cannot ensure that the ideal classification margin has been learned.In view of this problem,this paper further proposes an enhanced multi-view classification method based on deep adversarial network(MvDANE).It introduces the intra-view metric loss on the basis of MvDAN,so that the deep features of the same class in the same view are close,and the deep features of the different class are far apart,sufficiently extracting the discrimi-nant information of the multi-view data.This paper has carried out experiments on several commonly used multi-view databases to prove the effectiveness of the proposed algorithm.