Multi-view Subspace Clustering Based On Unified Measure Standard

In recent years,multi-view subspace clustering methods have attracted extensive attention because they can access comprehensive information from multiple data features.With respect to multi-view clustering,to obtain better performance,the information from multi-view data must be reasonably fused to well extract the consistency and specificity from different views.Therefore,how to explore the consistency and specificity of multi-view data is the focus of our research.To improve the clustering performance,previous methods try to explore the consistency and specificity between different views by making the common representation coefficient matrix as close to the representation coefficient matrices learned in each view as possible.However,the value of the element corresponding to the similar degree of the strong or weak relationship often has different magnitude levels in the representation coefficient matrices learned in each view.In this situation,the above strategy will make the information of some view ignored or magnified,and further lead to some issue,i.e.it is difficult for us to find a representation coefficient matrix expressing the specificity and consistency information of all views well according to the previous algorithms,and thus the clustering performance will be deteriorated.To overcome this limitation,we propose the multi-view subspace clustering based on unified measure standard in this thesis,our proposed method can normalize the degree of the strong or weak relationship in each view to the unified measure standard by scaling the representation coefficient matrices learned in each view,the consistency and specificity between different views will be mined more effectively.Therefore,better clustering results can be obtained.In addition,we provide the theoretical analysis about the convergence and computation complexity of our algorithm.The experimental results on several multi-view data sets indicate that our proposed method is effective and efficient for multi-view subspace clustering both in theory and in practice,which also demonstrates the effectiveness of multi-view subspace clustering based on unified measure standard in dealing with multi-view data sets.