Towards efficient and robust correntropy-based anchor tensor learning for multi-view subspace clustering

Multi-view clustering aims to learn a latent shared representation by integrating complementary information from multiple views, thereby comprehensively capturing the underlying data structure. However, its practical applications are constrained by three key challenges: 1) High computational complexity, 2) Lack of high-order correlation information, 3) Compromised model robustness. To address these issues, this paper proposes a correntropy-based anchor tensor learning method for multi-view subspace clustering (CATMSC). Specifically, we first construct a tensor from the anchor graphs to capture the high-order cross-view correlation. Then, a globally consistent structure is promoted by enforcing low-rank constraint on this tensor, while simultaneously exploiting the local Gaussian kernel property of correntropy-induced metric (CIM) to adaptively suppress the impact of noise. Finally, an iterative optimization algorithm based on half-quadratic optimization is designed to solve the optimization problem induced by CATMSC. Experimental results on multiple real-world datasets demonstrate that the proposed CATMSC outperforms state-of-the-art clustering methods in both clustering accuracy and efficiency, while exhibiting superior robustness, especially in scenarios involving non-Gaussian noise or outliers.