Single-phase incomplete multi-view clustering via joint anchor refactoring and tensorial bipartite graph learning.

Abstract

The inherent incompleteness of multi-view data poses significant challenges for conventional clustering methods, particularly those based on bipartite graph construction. Although existing approaches have made notable progress by refining graph architecture, they often overlook the critical role of anchor quality in determining overall clustering performance. To address this limitation, we propose a Single-phase Incomplete Multi-view Clustering via Joint Anchor Refactoring and Tensorial Bipartite Graph (SIMC-ARTB), a novel unified framework that jointly integrates anchor refactoring, aligned graph fusion, and single-phase label learning. Specifically, SIMC-ARTB first learns high-quality anchors through sample projection matrices and introduces a novel anchor refactoring strategy to enforce distributional consistency between anchors and original data. To effectively fuse incomplete view-specific bipartite graphs, we further propose an alignment-based fusion method that incorporates orthogonal transformation matrices and adaptive weights, yielding a consistent and well-aligned graph representation. To capture high-order structural correlations across multiple views, we formulate a weighted tensor nuclear norm regularization term, which enhances the robustness and low-rank consistency of the fused graph. Moreover, departing from the widely adopted two-stage paradigm, SIMC-ARTB embeds discrete label learning directly into the optimization process, enabling mutual reinforcement between consistency graph learning and discrete label assignment. Extensive experimental results on benchmark datasets demonstrate that SIMC-ARTB consistently outperforms state-of-the-art incomplete multi-view clustering methods in terms of clustering accuracy, stability, and scalability.