Multi-view subspace clustering using drop out technique on points.

Multi-view subspace clustering methods have been very popular among all the multi-view clustering approaches. Overall, in the subspace clustering methods, self-expressive representation of data points gives a coefficient matrix. In multi-view approaches coefficient matrices of all views are utilized to achieve one coefficient matrix that gives a clustering of points. In the previous proposed methods, to determine clusters, the focus was often on the sparsity among different clusters. However, the sparsity may cause that data points of the same clusters does not have strong connections, i.e. some points in the same cluster may not be connected that causes a cluster to be a union of two components and it causes the increase of over-segmentation. To address this problem, we apply drop out technique on columns for all views for several times. In each step, some columns are dropped out and just the remaining points have role in the self-representative combinations of points. Then we consider all self-expressive representations, simultaneously. This method improves the intra-clusters connectivity as well as some other criteria. Our method is evaluated on some real world multi-views data set for confirming its performance. The results depict that the method works well and the proposed algorithm has excellent connectivity. Moreover, some other criteria, which have been improved, are compared with some other contemporary approaches. [ABSTRACT FROM AUTHOR]