Research On Incomplete Multi-view Clustering

Real data are often coming from multiple incomplete sources,and incomplete multi-view clustering methods provide some natural ways to cluster these incomplete data.Previous studies assume that all views miss the same proportion of data and no noises appear in all views.However,in most real-world applications,incomplete views often have different incomplete rates(i.e.,unbalanced incomplete views),which results in strong views(views with low incomplete rate)and weak views(views with high incomplete rate).The unbalanced incompleteness maks the previous methods fail and noises lead to unreliable clustering results.In this paper,I propose Unbalanced Incomplete Multi-view Clustering(UIMC)and Auto-weighted Noisy and Incomplete Multi-view Clustering framework(ANIMC)to solve the above two problems separately.As far as I know,UIMC is the first study to handle the unbalanced-incompleteness problem by designing the rule of view evolution.It performs view evolution by proposing weighted multi-view subspace clustering to learn the optimal representations for all views with different incomplete rates and minimizing the difference between the cluster indicator matrix of each view and the consensus matrix.Compared with existing state-of-the-art methods,UIMC has following advantages:1)it proposes weighted multi-view subspace clustering to reduce the negative influence of unbalanced incompleteness;2)it designs low-rank and robust representation to diminishes the impact of noises for robustness.Unbalanced incomplete multi-view clustering experiments on four real datasets prove the effectiveness of UIMC.For the proposed ANIMC,a common latent feature matrix shared by all views and a proper weight of each view are learned automatically and update collaboratively based on adaptive multiple semi-nonnegative matrix factorization,which reduces the impact of noises.Moreover,the common latent feature matrix and same instances in different views are respectively aligned with the help of L_F-norm regularized regression and L_2,1-norm regularized regression,which desease the influence of missing instances.Noisy and incomplete multi-view clustering experiments on four real datasets prove the effectiveness of ANIMC.