BACVC: Bi-adaptive and cross-view consistency for incomplete multi-view subspace clustering

Abstract

Multi-view subspace clustering leverages complementary information from different views to uncover latent subspaces. However, incomplete multi-view data is prevalent, particularly in fields such as communication systems. Incomplete Multi-View Subspace Clustering (IMSC) addresses this challenge but faces two main challenges: (1) neglecting dissimilarities between views and samples, and (2) insufficient handling of cross-view consistency. To tackle these issues, we propose a novel IMSC framework, referred to as Bi-Adaptive and Cross-View Consistency (BACVC). BACVC improves incomplete data recovery and subspace structure discovery through view-adaptive tensor rank constraints, data-adaptive high-order correlations, and view-level contrastive learning. Specifically, we first apply tensor factorization with view-adaptive tensor rank approximation to enforce low-rank constraints on a stacked affinity tensor, capturing the view-specific subspace block-diagonal structure. We then introduce a data-adaptive non-uniform hypergraph-induced hyper-Laplacian regularization to model high-order correlations and guide the recovery of incomplete data. Finally, contrastive learning is applied to the soft clustering assignment of each view, ensuring cross-view structural consistency. Extensive experiments on four benchmark datasets show that BACVC outperforms eleven state-of-the-art methods, with improvements of up to 4.39%, 5.43%, and 3.95% in ACC, NMI, and purity, respectively. Experimental results demonstrate the robustness of BACVC in handling incomplete data and its effectiveness in practical applications.