Research map / Learning theory and algorithms / Clustering
Scalable One-Pass Incomplete Multi-View Clustering by Aligning Anchors
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SOME-AS, a scalable one-pass anchor-learning method for incomplete multi-view clustering that aligns anchors across views while jointly learning an integral subspace with orthogonal cluster assignment and centroid factors.
Problem
Anchor-based multi-view clustering on large-scale incomplete multi-view data ignores learning an integral subspace with anchors from different views aligned, and does not jointly optimize anchor alignment and subspace learning, limiting clustering quality.
Approach
Builds a per-view anchor graph on incomplete data, learns orthogonal anchors and per-view alignment mappings, and jointly optimizes a shared integral subspace (centroid F and assignment Y) so that anchor alignment and partition learning improve each other within one unified objective solved by alternating updates.
Key results
- On SUNRGBD, SOME-AS improves over MKKM-IK by 1.20%, 3.48%, 2.12% and 3.25% across the four metrics.
- SOME-AS achieves the best ACC on most datasets, e.g., 98.35% on Cifar10 and 99.20% on MNIST, versus 96.20% and 98.10% for the next-best method (AUP-ID).
- Objective value decreases monotonically over iterations, and clustering is not significantly affected by the number of anchors in [1k, 3k].
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SOME-AS instead of MKKM-IK | incomplete multi-view clustering · SUNRGBD | ACC: +1.20% | Helped |
| Add anchor alignment and integral-subspace learning (ablation) | SOME-AS · ORL | clustering performance | Helped |
Surprising findings
- SOME-AS requires more running time than FIMVC-VIA due to the alignment mapping, which the authors acknowledge.
Limitations
- Requires more running time than FIMVC-VIA because of the alignment mapping.
- Optimization is non-convex and only converges to a local optimum.
- Hyperparameters alpha and beta are tuned on a single dataset (ORL).
Experimental setup
Datasets: ORL BDGP ProteinFold SUNRGBD NUSWIDEOBJ Cifar10 MNIST
Benchmarks: ORL BDGP ProteinFold SUNRGBD NUSWIDEOBJ Cifar10 MNIST
Metrics: ACC NMI F1-score Purity running time
Baselines: BSV MIC MKKM-IK AWP DAIMC APMC MKKM-IK-MKC EEIMVC V3H FIMVC-VIA AUP-ID
Builds on: anchor graph multi-view clustering Scalable Incomplete Multi-View Clustering with Structure Alignment (AUP-ID) FIMVC-VIA
Closest papers by meaning
- Fast Incomplete Multi-view Clustering by Flexible Anchor Learning (ICML 2025 · similarity 0.94)
- Alleviate Anchor-Shift: Explore Blind Spots with Cross-View Reconstruction for Incomplete Multi-View Clustering (NeurIPS 2024 · similarity 0.92)
- DVSAI: Diverse View-Shared Anchors Based Incomplete Multi-View Clustering (AAAI 2024 · similarity 0.91)
- Robust Consensus Anchor Learning for Efficient Multi-view Subspace Clustering (ICML 2025 · similarity 0.91)
- Learning Cluster-Wise Anchors for Multi-View Clustering (AAAI 2024 · similarity 0.91)
- Fast Incomplete Multi-view Clustering with Adaptive Similarity Completion and Reconstruction (AAAI 2025 · similarity 0.91)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.