A Consensus Anchor-Guided Hypergraph Framework for Incomplete Multi-View Clustering

ICLR 2026 Conference Submission16897 Authors

Keywords: incomplete multi-view clustering; multi-view clustering; clustering;

Abstract: As a significant task within the field of unsupervised learning, Incomplete Multi-View Clustering (IMVC) faces considerable challenges in scenarios involving large-scale datasets, heterogeneous data, and missing views. Existing anchor-based clustering approaches primarily reduce computational and storage overhead by introducing anchors, yet they often focus on binary sample-anchor relationships. These methods lack robust learning of consensus anchors under missing conditions and fail to effectively model high-order relationships among samples. Furthermore, systematic discussions regarding implementation details and robustness mechanisms remain insufficient. To address this, this paper proposes a Missing-aware Consensus Anchor-guided Hypergraph Clustering (MCAHC) framework. This method constructs hypergraph through sample-anchor connections and anchor guidance to capture high-order relationships among samples, effectively mitigating view-missing and noise interference. Concurrently, it designs sample-level and view-level reweighting mechanisms to suppress inter-view imbalance and promote cross-view consistency, while explicitly down-weighting severely incomplete samples to prevent them from biasing anchor selection. Experimental results demonstrate that MCAHC provides an efficient and robust solution for multi-view clustering in large-scale and high-missing-value scenarios.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 16897