The Research On Multi-view Clustering Based On Constraint Representation Learning

With the rapid development of network information technology and sensor technology,data collection methods are increasing,and data objects in the real world are usually able to be obtained from multiple data sources or present different representations,and they are often referred to as multi-view data.In multi-view data,data from different views often have different degrees of correlation and heterogeneity,and the feature information among different views plays a complementary role.Multi-view learning has attracted the widespread attention of researchers and has become a hot topic in the fields of artificial intelligence,data mining and other scientific research.How to model the description of data objects from each view and the relationship between different views,and design strategies that integrate the characteristics of different views,so as to provide more reasonable guidance for inference and decision-making in practical application,has important application value.Multi-view clustering is also a main research direction in multi-view learning.It aims to explore the information within the view and the correlation information between the views by integrating the data characteristics of multiple views,making full use of various constraints,and dividing the samples into the corresponding clusters.With the rapid development of Internet technology,the data in the real world is becoming more and more complex,usually high-dimensional and nonlinear,and also contains a variety of complex hierarchical attributes.Traditional clustering algorithms usually belong to shallow models,often ignore the nonlinear structure and complex hierarchical information contained in multi-view data,and gradually cannot meet the production and application in real life.Deep learning can capture the corresponding attribute features through different levels of networks,project the data from the original high-dimensional space to the low-dimensional space and then perform the clustering task.Therefore,deep learning-based clustering methods have become the current hot spot in the field of clustering research.However,most of the current deep learning-based approaches are supervised work,i.e.,training the network by labels so that the representation of the last layer matches the features of the corresponding task,and less work has been done in the direction of multiview clustering.Although some unsupervised deep multi-view clustering models have been designed and developed,they are not able to impose sufficient constraints to adequately capture the hierarchical information,the nearest neighbor information between samples,and the correlation information of views in multi-view data.In addition,some of the existing deep multi-view clustering methods map the input data to a lowdimensional space and then directly use existing clustering methods to obtain clustering results on the low-dimensional representation,which separates the relationship between representation learning and clustering,and results in the learned low-dimensional representation to obtain the suboptimal clustering results to a certain extent.In this paper,we conduct the research on constrained representation learning for multi-view clustering based on deep learning framework to enhance the performance of multi-view clustering by modeling the information within each view and the correlation information among views,which contains the following three aspects:(1)We propose a multi-view clustering method based on the deep autoencoder(MVC-MAE).This method uses multiple deep autoencoders to model each view separately,capturing the hierarchical semantics of multi-view data in a layer-wise manner,enforcing the samples to approach layer by layer in the low-dimensional space,and then obtaining high-level semantic information.And to protect the local geometric structure of each view,by constructing the nearest neighbor graph corresponding to each view,we use the nearest neighbor graph to guide the representation learning of the corresponding samples in the low-dimensional space by combining the manifold learning theory.In addition,inspired by the cross-entropy,a regularization term is also designed which can enforce the representation of the same sample in different views to be as similar as possible and the representation of different samples in different views to be as different as possible.Finally,the clustering embedding layer is also introduced to integrate representation learning and clustering learning into a unified step.The corresponding experimental results verify that the proposed method can effectively capture multi-view high-level semantic information and improve multi-view clustering performance.(2)We propose multi-view subspace clustering with multi-level representations and adversarial regularization(Mv SC-MRAR).The method first captures the lowdimensional representations of attributes at different levels through a multilayer nonlinear fully connected network,and then constructs the corresponding self-expression matrices based on the representations at different levels.A diversity canonical is also designed to maximize the complementarity between different perspectives as well as different levels of subspace representations.In addition,an adversarial training canon is introduced to guide the learning of low-dimensional representations by adversarial training,forcing them to obey the specified prior distribution.The experimental results show that the method can capture the information of multi-view data in a better and more comprehensive way.(3)We proposed a nearest-neighbor-aware multi-view clustering method based on graph convolutional network(NMv C-GCN).This method adopts graph convolutional network to aggregate the attribute features of central samples and their neighboring samples in each view and then optimizes the graph convolutional network for each view using mutual information maximization theory to make the representation of each view more discriminative.A consensus regularization is also designed to be able to generate a consensus representation with the joint help of low-dimensional representations of the original samples and low-dimensional representations of the corrupted samples.In addition,NMv C-GCN also introduces the clustering embedding layer to integrate the representation learning and clustering tasks into a unified step,making NMv C-GCN an end-to-end multi-view clustering method.Finally,extensive experiments on real datasets show that NMv C-GCN can effectively aggregate information about surrounding neighbors and improve multi-view clustering performance.