Research On Multi-View Fuzzy Clustering Methods Based On Representation Learning

In recent years,multi-view clustering has received great attention,which aims to improve the clustering performance by using cooperative learning of different views.Although the existing methods have made considerable progress,they still face some issues that need to further research: First,multi-view data have specific information and common information,and most multi-view clustering methods only exploit common information.Second,most multiview clustering methods learn the hidden view representation and then perform clustering.This two-step learning framework ignores the correlation between the two process.Meanwhile,there is potential hierarchical information in multi-view data,and most multi-view clustering methods based on hidden views only exploit the single-level common information in multiview data,while ignoring its potential hierarchical information.In addition,real multi-view data contains redundant features and noises,leading to unsatisfying clustering performance.To address the above issues,two multi-view clustering methods are proposed in this thesis,and further did the application research based on the new methods.The main work and contributions in this thesis can be summarized as follows:(1)To address the issues that multi-view clustering methods based on hidden views split the representation learning and clustering and ignore the mining of hierarchical information,a new pseudo-label enhanced multi-view deep concept factorization fuzzy clustering is proposed.First,in order to mine hierarchical information in multi-view data,a deep concept factorization learning mechanism based on concept factorization technique is constructed to learn the deep representation of multi-view data and it is incorporated into the clustering process.Second,to further mine the common information among views,we learn the common representation of the deep representation by multi-view non-negative matrix factorization and incorporate it into the clustering process.Then,to further improve the discriminability of the common representation,a pseudo-label learning is introduced to improve the quality of representation learning.Finally,we integrate deep concept factorization learning,pseudo-label enhanced common representation learning,and clustering partition into an joint framework to negotiates with each other.(2)To address the issues that most multi-view clustering methods only mine specific information or common information and redundant features or noise in multi-view data affect the clustering performance,an end-to-end multi-view fuzzy clustering is proposed.First,we construct a multi-view fuzzy clustering framework to mine specific information of visible views.Second,to reduce the impact of redundant features and noises on the clustering performance,the orthogonal projection matrix is introduced into clustering framework to learn the lowdimensional representation of visible views.Meanwhile,this procedure is integrated into the clustering framework.Then,we explore the shared hidden view representation information between visible views by multi-view non-negative matrix factorization and integrate it into the clustering framework to realize visible-hidden view cooperation learning.Finally,shared hidden view representation learning between visible views,low-dimensional representation learning of visible views,and clustering partition of multi-view data can negotiate with each other in the end-to-end learning framework.(3)In order to further evaluate the effectiveness of the above new methods in this thesis,we further discuss the clustering analysis of enzyme data based on multi-view clustering.Specifically,first,the enzyme data are collected from the Protein Data Bank.second,we preprocess the enzyme data and extract multi-view enzyme data by feature extraction technique to generate a multi-view enzyme dataset.Then we perform the clustering analysis by the proposed methods in this thesis.Finally,we analyze the important role of enzyme data clustering analysis in downstream tasks such as enzyme function prediction.