Research On Multi-view Subspace Clustering Algorithms

Multi-view subspace clustering algorithm is a machine-learning algorithm which combines multi-view learning and subspace clustering.It mainly uses the relationship between views to describe data more comprehensively to improve clustering performance.However,it is found that most of the multi-view subspace clustering algorithms only assign the same weight to the views when they are merged.If a view is damaged,the negative impact on the final unifed clustering result will be increased.Secondly,to block diagonalization of the subspace representation matrix,most of the multi-view subspace clustering algorithms are based on sparse or low-rank representation,but they both have drawbacks.Sparse representation mainly focus on local structural information and low-rank representation mainly focus on global structural information.Finally,many multi-view subspace clustering algorithms ignore the difference learning between views so that they may lose much useful information in the subspace representation.In response to the above-discovered problems,this thesis studies the multi-view subspace clustering algorithm.The specific research content is as follows:1.The self-weighted multi-view subspace clustering algorithm based on low-rank sparse constraint is studied.Considering that low-rank sparse constraint can be used to obtain global and local structural information at the same time,the low-rank sparse subspace clustering algorithm is extended to the field of multi-view clustering.In the process of learning the unified subspace representation,a self-weighted method is introduced to solve the problem that the unified representation matrix error increases because the views are assigned the same weight.The self-weighted method's main idea is calculating the distance between the subspace representation matrix of each view and the unified representation matrix and then using the inverse ratio of them to assign a suitable weight to each view.Experiments on the standard datasets show that the proposed algorithm is superior than other considered algorithms.2.The multi-view low-rank sparse subspace clustering algorithm based on diversity is studied.In view of the common problem of ignoring the mutual learning between views in many algorithms,the mutual learning between views is used to increase the complementarity between the views' subspace representation matrices,so that the subspace representation matrix of each view needs to be as different as possible.We introduce diversity concept into the framework of the multi-view low-rank sparse subspace clustering algorithm to obtain more comprehensive information.At the same time,in order to obtain a unified clustering result,the spectral clustering process is added to the framework for joint optimization to improve the performance of clustering.Experiments on three image datasets have proved the effectiveness of the algorithm.This thesis contains 3 figures,19 tables and 104 references.