Tensor double arc-tangent norm minimization for multi-view clustering

Abstract

Tensor structures are widely utilized in clustering tasks due to their ability to represent high-order relationships among multi-view data. The rank and decomposition of tensors are essential characteristics that have recently received considerable attention. Most approaches approximate tensor rank using the tensor nuclear norm (TNN) and non-convex rank functions. We introduce a multi-view clustering method based on an enhanced self-expression with a tensor double arc-tangent norm. Initially leveraging tensor singular value decomposition, this method impartially approximates the true rank of the tensor, avoiding rank overestimation. The enhanced self-expression further utilizes the latent connections among the original data. Then, we propose an improved alternating direction method of multipliers (ADMM) for rapid optimization of the model. Experiments conducted on six datasets confirm the effectiveness of the proposed model.