Research On Multi-view Subspace Clustering Algorithms

Cluster analysis is a commonly used data analysis and processing tool in the fields of machine learning and data mining.With the advent of the era of big data,the channels for people to collect data,the way to obtain data,the means of processing data,and the diversity of feature extractors make the data present different sources,different modalities,and different characteristics,etc.Traditional clustering methods are proposed and developed for singleview data scenarios,which are difficult to deal with multi-view data effectively.Multi-view clustering is a fundamental task for clustering multi-view data,which improves the clustering effect by exploiting the information increment brought by multi-view data.However,how to effectively integrate information from multiple different views and exploit the consistency and complementarity between views at the same time has become a huge challenge in the development of multi-view clustering.In recent years,multi-view subspace clustering has received extensive attention and research due to its good clustering performance and mathematical interpretability.Therefore,this thesis focuses on multi-view subspace clustering methods.Research on how to effectively fuse interaction information in multi-view data to improve clustering performance.In this thesis,relevant research is carried out to address this key issue.Specifically,the research work mainly includes the following three aspects:(1)Feature-level Fusion Multi-view Subspace ClusteringIn view of the potential noise and redundant information in the original multi-view data,in this thesis,we use the autoencoder network to perform effective feature extraction on the multiview data.Considering the potential complementary information between different views,this thesis obtains an intact space representation of multi-view data through intact space learning.In order to integrate autoencoder learning,intact space learning and self-representation learning into one optimization model,this thesis proposes a deep multi-view subspace clustering based on intact space learning method.The effectiveness of the proposed method is validated on several popular multi-view datasets.(2)Subspace-level Fusion Multi-view Subspace ClusteringThe self-representation model in multi-view subspace clustering usually obtains the global information of the data.In order to expand the local information of the self-representation model,this thesis adopts the graph learning method.Since the subspace representations generated by different views have potentially consistent information,this thesis uses subspace fusion regularity constraints to obtain shared subspace representation for multi-view data.In order to integrate self-representation learning,graph learning,subspace fusion and clustering into one optimization model,this thesis proposes a multi-view subspace clustering based on local and global information subspace fusion method to achieve multi-view subspace clustering in an end-to-end manner.Extensive experimental results verify the effectiveness of the proposed method.(3)Spectral Embedding Fusion Multi-view Subspace ClusteringMost multi-view subspace clustering methods ignore the influence of local structural information in the view data on the subspace representation when constructing the corresponding subspace representation for each view.As well as separating the self-representation learning,spectral embedding and clustering processes into three independent processes,ignoring the potential association information between them.This thesis proposes a multi-view subspace clustering based on local and global information spectral embedding fusion method,which integrates self-representation learning,graph learning,spectral embedding fusion and spectral rotation process into one optimization model to achieve a one-stage multi-view subspace clustering method.The effectiveness of the proposed method is validated on multiple public multi-view datasets.In this thesis,we propose a total of three different multi-view subspace clustering methods to consider the impact of multi-view information fusion on the clustering performance from different perspectives,respectively.