Research On Multi-view Clustering Based On Deep Learning

Multi-view clustering aims to use the consistency and complementarity of different view information to efficiently explore the latent value of multi-view data,which is one of the main research directions of big data analysis.Most of the traditional single-view clustering methods are not suitable for dealing with multi-view data,and the current multi-view clustering methods are either unable to effectively explore the latent links between views,or cannot fully consider the contribution of different views to the final clustering results.Therefore,how to propose an efficient multi-view clustering method has attracted the attention of many researchers.Due to the strong nonlinear fitting ability of deep neural network,the combination of multi-view clustering and deep learning model has become a research hotspot.In this thesis,the deep multi-view clustering method is studied based on the generative model variational autoencoder.The main innovations are as follows :(1)In order to make rational use of the consistency and complementarity of multi-view data,and considering the importance of different views,a deep multi-view clustering based on distribution aligned variational autoencoder is proposed.This method adds a distribution alignment strategy in the process of deep variational autoencoder learning the latent distribution of views,which can perform consistent learning on multi-view data,and at the same time use the reconstruction loss constraint of the decoder to retain the feature information with strong expressiveness of its own views.Then,in order to consider the influence of different views on the clustering results,a set of adaptive weight vectors are introduced to obtain a shared latent representation.Finally,the deep embedded clustering loss is combined for optimization learning.The proposed method has carried out extensive comparative experiments on five public multi-view datasets,and shows excellent clustering performance in multiple clustering evaluation indexes.(2)Inspired by the fact that supervised learning can often solve problems well with label information,we have made a beneficial attempt to combine the idea of self-supervised learning with the deep multi-view clustering methods,and proposed the deep self-supervised multi-view clustering based on cross-distribution alignment.First of all,in the experiment of the first method proposed in this thesis,it is found that the distributed alignment strategy may lead to a strong constraint problem of consistency learning.Therefore,we introduce cross-view alignment strategy to balance multi-view consistency and complementary learning,that is,in the decoding process,not only the view itself information is limited,but also the information of other views is added for cross-reconstruction training.In addition,by designing a classification network as a downstream task and supervising the clustering result label information as pseudo-labels,the network parameters and clustering results can be jointly optimized.The experimental results show that this method has a certain improvement on clustering performance.In summary,in order to effectively mine the potential value of multi-view data,this topic proposes two multi-view clustering methods with excellent performance,and conducts convergence analysis,running time comparison,parameter analysis and other experiments on the corresponding model.Both methods have achieved excellent clustering results on multiple public datasets,which verifies the effectiveness of the proposed method.