arXiv · 2611.16749
ORCA:Orthogonal Residual Consensus Alignment for Multi-View Clustering
Abstract
Deep multi-view clustering often suffers from redundant information accumulation during fusion and indiscriminate alignment across representation levels. We propose Orthogonal Residual Consensus Alignment (ORCA) to address these challenges. Its semantic incremental recurrent fusion module extracts cross-view orthogonal residuals as non-redundant increments and adaptively weights recurrent states through divergence-guided gating. Its consensus-calibrated hierarchical contrastive alignment module uses the global consensus to weight contrastive alignment of view-level, pairwise, and recurrent representations. Together, these modules balance view-specific structure with cross-view consistency. Experiments on five benchmarks against fifteen competing methods demonstrate ORCA's effectiveness, including a 5.25 percent accuracy improvement over the strongest baseline on 100leaves. Ablation studies validate both modules. Code is available at https://github.com/XiaojianDing/2026-NeurIPS-ORCA.
Explore related subjects
Keep this discovery
Explore connections, maps & timelines
Xiaojian Ding, Xian Li, Xiaoying Zhu, Lin Zhao, Kaixiang Wang. 2026-11-12. ORCA:Orthogonal Residual Consensus Alignment for Multi-View Clustering. https://arxiv.science/abs/2611.16749
Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite arXiv Science by Indiciae: BibTeX · RIS.