Multi-view clustering is an interesting and much talked about direction in the field of clustering.Due to the diversity of data acquisition and the development of deep learning,the presentation of data features has become more complex and diverse.In the face of complex multi-view data,researchers have used different methods to achieve cluster classification for multi-view data.Although there are many different multi-view clustering algorithms,the main problems can be summarized as follows: 1.how to explore the structure of multi-view data so that it can be appropriately clustered;2.how to obtain a suitable matrix representation that can be used for data clustering.Since subspace learning shows good performance on multiple views,its learned subspaces can be used not only for multi-view clustering alone,but also as the result of upstream tasks for downstream clustering implementation.At the same time,subspace learning can also be used to explore the data structure and perform information fusion of multiple views to obtain better clustering results.However,both existing multi-view subspace algorithms and other types of related algorithms ignore another simple and straightforward way of fusing view information of multiple views,and the algorithms that consider this part of information ignore the information complementarity between multiple views,which will directly lead to partial loss of data information.Therefore,thesis explores the possibility and effectiveness of subspace-based fusion information clustering from different view fusion directions.thesis we obtain the consensus matrix used for clustering by different fusion methods of multi-view information,the main elements of which are as follows.1.A multi-view union and a multi-view clustering algorithm under multi-view learning are proposed.Specifically,the multi-view data needs to be united into the overall data and the potential representation matrix is learned from it through parametric constraints.At the same time,in order to exploit the consensus and complementarity of multiple views,the representation matrix of each view is learned and then linked to the representation matrix of the union data by a regularization formula so that the learned representation matrix not only contains the overall data information,but also can be recognized by all views.Finally,experiments are then performed to prove the superiority of the algorithm.2.A multi-view clustering algorithm based on feature direct linkage and structured constraints is proposed.The algorithm integrates the subspace information of the original multiple views and the feature-directly connected views to obtain a better subspace representation.A more suitable consensus representation is also obtained by error reconstruction and structured constrained subspaces.At the same time,the weight relationship between multiple views and feature-directly connected views is also considered.Finally,the effectiveness of view fusion information on clustering is further verified by experiments. |