Multiple Kernel Low-Redundant Representation Learning Based Incomplete Multiview Subspace Clustering

Abstract

multiview data because self-representation technique relies on complete data for subspace construction. In this paper, we propose an incomplete multiview subspace clustering method based on multiple kernel low-redundant representation learning. Firstly, in order to learn the intact subspace representation, we implement incomplete kernel completion, and fully consider both of view-specific local information preservation and global information fusion. And then, by revealing the redundancy in kernel space, we also propose to infer multiview low-redundant representation to seek the intact subspace representation instead of original incomplete data. Secondly, different from the pairwise learning in conventional approaches, a weighted tensor low-rank constraint is introduced for multiview subspaces fusion to learn intrinsic low-rank tensor subspace structure, which can not only explore the higher-order relationship among views but also unequally treat their significance. Finally, we put these terms into a unified model to jointly learn low-redundant representation, view-specific subspace, and low-rank tensor subspace structure. Extensive experimental results onfour publicly available datasets demonstrate that our proposed method outperforms the existing cutting-edge incomplete multiview clustering approaches in terms of several evaluation metrics. The demo code for our method is uploaded as attachment.

Keywords: incomplete multiview clustering, representation learning, subspace learning, tensor analysis