Efficient multi-view subspace clustering with coupled sparse gradient and anchor-driven Laplacian regularization

Multi-view clustering (MVC) leverages the inherent diversity of multiple views to identify clustering structures that encompass distinct feature representations within datasets. Existing methods recognize the importance of exploring inter-view consistency, but the in-depth exploration of latent manifold structures and the reduction of computational costs remain understudied. To this end, we propose efficient multi-view subspace clustering with coupled sparse gradient (SG) and anchor-driven Laplacian regularization (EMCSA), a novel multi-view subspace clustering (MSC) approach. Specifically, SG and Laplacian regularization enhances inter-cluster discriminability, while the ℓ_2,log-norm induces column-wise sparsity to prune redundant features. The tensor γ^∗-norm serves as a tighter convex surrogate for rank approximation. We design an algorithm to fuse across-view information and rigorously prove convergence of the iterates to a Karush–Kuhn–Tucker (KKT) critical point. Extensive experiments highlight the effectiveness and robustness of EMCSA across diverse scenarios.