Research On Multi-view Learning Based On Subspace Learning

In recent years,the rapid development of information technology has greatly improved the ability of data collection and storage.Data can be representated variously.The feature information for the same object obtained from different angles or different sensors is called multi-view data.Since data from different views may lie in completely separate spaces,it is impossible to utilize the traditional single-view machine learning method to deal with multi-view data,thus,multi-view learning model arises at the historic moment.Multi-view learning studies the relationship between different views and fuses the features of the data from each view to improve the generalization ability of the multi-view learning model.At the moment,although the theoretical research on multi-view learning has made some progress,more and more practical application requires the identification of samples from different views,which is known as the problem of cross-view classification in multi-view learning.Due to the view discrepancy,most of the existing multi-view learning methods use the idea of maximum correlation to explore the relationship between views,but it is difficult to further improve the generalization performance.This thesis proposes a Deep Cross-view Autoencoder Network(DCVAE)model to simultaneously handle view-specific,view-correlation,and consistency problems,which is a novel multi-view subspace learning framework.The model is based on the Autoencoder,including the newly designed cross-view reconstruction module,self-reconstruction module,and the encoding consistency constraints module.Self-reconstruction ensures the view-specific,cross-view reconstruction transfers the information from one view to another view and further fuses them,and encoding consistency term constrains the learned encodings to be more consistent.The model explores the relationship among views from a new perspective,and finally produces a multi-view common representation subspace.Compared with numerous classical multi-view learning algorithms,the accuracy of the model has been significantly improved on several databases.In addition,the 2D embeddings of the learned common representation subspace further demonstrate the proposed model is valid and favorable.In order to further improve the generalization performance of the model,this thesis introduces supervised information and proposes a deep cross-view discriminant analysis network model based on DCVAE.The discriminant regularization is added in the process of learning the common representation subspace,which makes the distance within-class closer while the distance between-class farther,namely the subspace is more discriminant.A significant improvement of the corss-view classification results and the ablation experiment prove the wisdom of introducing label information and the effectiveness of the further proposed model.