Multi--view Learning Based On Graph Neural Networks

With the increasing development of the Internet,the sources and collection meth-ods of data become diversified and complicated.The same data object is often obtained from multiple sources or has many different characteristics and expressions.Data in the form of multiple feature representations is called multi-view data,and learning using multi-view data is called multi-view learning.Due to the rise of deep learn-ing,researchers have proposed multi-view learning algorithms based on different deep learning components.However,these multi-view learning algorithms based on deep neural networks do not consider important graph information in the data.The graph structure information in the data can transfer information between multiple views.The graph neural network techniques also make it possible for neural networks to use graph structure information by embedding graph structure information in the hidden layer representation of the neural network.This thesis introduces graph neural network to multi-view learning to achieve structural information interaction among views,and has mainly achieved the following research results:1.Multi-view representation learning based on graph attention mechanism.Traditional multi-view representation learning uses CCA(Canonical Correlation Analysis)representations learned under unsupervised settings affect learning per-formance due to insufficient discriminativeness.We propose Mv DGAT,a semi-supervised multi-view representation learning method that enhance the discrim-inativeness and learns the appropriate weights of edges in the graph structure by introducing an attention mechanism.Experimental results show that the perfor-mance of the proposed method Mv DGAT is superior to the classical multi-view representation learning method.2.Multi-view classification based on weighted Laplacian.The traditional multi-view classification method based on co-training has excellent performance.How-ever,the method of exchanging confidence pseudo-labels to improve the per-formance of learners in each view has the problems of introducing noise and excessive training cost.Starting from the theoretical analysis of graph-based co-training,we propose Co-GCN,a multi-view classification method that can use the combination of weighted Laplacian matrix to exchange structural information between views.Co-GCN also consider the importance of each view by using adaptive weights in the weighted Laplacian matrix.Experimental results show that the proposed Co-GCN is superior to the classical multi-view classification method.