Incomplete multiview subspace clustering based on multiple kernel low-redundant representation learning

Abstract

Subspace clustering is a widely used technique for clustering high-dimensional data. However, its effectiveness is limited in the context of incomplete multiview clustering, where intact subspaces cannot be obtained due to missing instances. To address this issue, we present a novel approach for incomplete multiview subspace clustering based on multiple kernel completion, low-redundant representation learning, and weighted tensor low-rank constraint. First, a carefully designed kernel completion scheme is employed to obtain intact kernels, from which the complete low-redundant representations are learned to obtain intact and compact subspaces. Second, unlike the traditional pairwise subspace fusion, we propose to fuse the multiview subspaces with a weighted tensor low-rank constraint, which not only explores higher-order relationships among views but also assigns appropriate weights to each view. Finally, we propose a unified model that jointly learns low-redundant representations, view-specific subspaces, and their low-rank tensor structure. Extensive experiments conducted on four publicly available datasets demonstrate the effectiveness of the proposed method.