Research On Multi-view Subspace Learning Based On Deep Learning And NMF

Multi-view data is usually defined as comprehensive representation data composed of different representations.The two key factors in the process of multi-view data processing are consistency and complementarity.Based on these two characteristics,multi-view learning can cover all features of data samples more comprehensively.However,most algorithms for multi-view data can only be processed for a single view and ignore these two factors.In order to make more complete use of various information between multi-view data,many emerging algorithms have been derived to process the feature data in each view.The multi-view subspace learning method learns the unified representation of multiple subspaces or hidden spaces of all view data,and uses the unified representation for classification or clustering to complete tasks more easily and efficiently.In this study,multi-view subspace learning methods are divided into two main types,namely,subspace self-representation learning methods and non-negative matrix factorization(NMF)methods,and a new multi-view subspace learning algorithm is proposed according to the basic type.In order to further improve the utilization rate of information between multi-view data and take into account the two factors,a semi-supervised multi-view learning algorithm based on NMF is proposed,which called Multi-view classification via Multi-view Partially Common Feature Latent Factor Learning(MVPCFLF).The MVPCFLF algorithm is an extended learning form based on partially shared latent factor learning.It can make full use of public information and special information to obtain latent representations.The key idea of the MVPCFLF algorithm is to increase the constraints on the common feature matrix to maintain the consistency of the common features.The experimental results show that the MPCPLFF algorithm is more effective than the existing multi-view learning algorithm on the classification problem of 6 data sets.Next,this paper proposes a multi-view subspace adaptive learning via Autoencoder and Attention(MSALAA)based on attention mechanism and autoencoder.The traditional subspace clustering method clusters the affinity matrix of a single view,ignoring the problem of fusion between views.The MSALAA algorithm combines the autoencoder algorithm,the attention mechanism algorithm and the multi-view low-rank sparse subspace clustering method,which can not only improve the nonlinear fitting ability,but also meet the consistency and complementarity characteristics.The experimental results show that the MSALAA algorithm can significantly outperform existing benchmark methods on the clustering problem of 8 data sets.