Sketched multi-view subspace clustering

In this paper, we focus on the multi-view subspace clustering (MvSC) problem, where the task is to cluster the data points given multi-view data. Even though the existing MvSC methods perform well, they incur high computation costs. In this work, we aim to reduce the computation cost involved in MvSC using the tools from randomized linear algebra. We propose three MvSC algorithms assuming that the available multi-view data admit a linear or non-linear subspace representation and propose efficient solvers based on a coordinate descent algorithm. The proposed methods are computationally efficient and incur a lower computation cost than the existing methods. We theoretically evaluate the proposed methods in terms of representation error as a function of the sketching dimension. Finally, we demonstrate the efficacy of the proposed method on various synthetic and real-world datasets.