Research map / Learning theory and algorithms / Clustering
Towards Resource-friendly, Extensible and Stable Incomplete Multi-view Clustering
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 ToRES, a hyper-parameter-free, linear-complexity incomplete multi-view clustering method that builds prototype-sample affinity and directly optimizes discrete cluster labels, with zero variance across runs.
Problem
Incomplete multi-view clustering (IMVC) suffers from high time and memory costs from self-expression affinities, hyper-parameter sensitivity that hurts extensibility, and non-zero variance in clustering results.
Approach
Replace self-expression affinity with a small prototype-sample affinity built from view-wise and cross-view prototypes aggregated through one matrix, jointly learn representations and discrete labels by alternating optimization (with two equivalent solvers for the orthogonal-constrained subproblem), so no fusion stage, hyper-parameters, or spectral embedding are needed.
Key results
- On Webkb at 20% incomplete, ToRES (Ours-1) reaches 86.20 ACC and 32.61 NMI, with zero variance across runs.
- On EMNIST (280,000 samples), ToRES runs in 635.10 seconds, and it reports the lowest memory in most datasets (e.g., 2.36 GB on AWA10 vs. 13.86 GB for IMSC-AGL).
- ToRES is compared against 20 IMVC baselines and outperforms most in ACC/NMI/Purity, though it is slightly worse than some baselines at 60% incomplete data and on some NMI entries (e.g., AWP on SUNRGB-D at 20%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add view-wise prototypes to cross-view-prototype-only ToRES (ablation) | ToRES · Webkb, PID 20% | ACC: 64.99 -> 86.20 | Helped |
| add cross-view prototypes to view-wise-prototype-only ToRES (ablation) | ToRES · Webkb, PID 20% | ACC: 51.19 -> 86.20 | Helped |
| add view-wise prototypes to cross-view-prototype-only ToRES (ablation) | ToRES · Wikifea, PID 20% | ACC: 54.47 -> 56.28 | Helped |
| add cross-view prototypes to view-wise-prototype-only ToRES (ablation) | ToRES · Wikifea, PID 20% | ACC: 52.51 -> 56.28 | Helped |
| add view-wise prototypes to cross-view-prototype-only ToRES (ablation) | ToRES · AWA10, PID 20% | ACC: 16.51 -> 28.88 | Helped |
| add cross-view prototypes to view-wise-prototype-only ToRES (ablation) | ToRES · AWA10, PID 20% | ACC: 16.70 -> 28.88 | Helped |
| add view-wise prototypes to cross-view-prototype-only ToRES (ablation) | ToRES · SUNRGB-D, PID 20% | ACC: 15.52 -> 20.93 | Helped |
| add cross-view prototypes to view-wise-prototype-only ToRES (ablation) | ToRES · SUNRGB-D, PID 20% | ACC: 15.40 -> 20.93 | Helped |
| add view-wise prototypes to cross-view-prototype-only ToRES (ablation) | ToRES · AwAfea, PID 20% | ACC: 6.68 -> 8.96 | Helped |
| add cross-view prototypes to view-wise-prototype-only ToRES (ablation) | ToRES · AwAfea, PID 20% | ACC: 4.23 -> 8.96 | Helped |
| add view-wise prototypes to cross-view-prototype-only ToRES (ablation) | ToRES · EMNIST, PID 20% | ACC: 37.15 -> 47.18 | Helped |
| add cross-view prototypes to view-wise-prototype-only ToRES (ablation) | ToRES · EMNIST, PID 20% | ACC: 13.31 -> 47.18 | Helped |
Surprising findings
- ToRES is slightly worse than some baselines at 60% incomplete ratio, which the authors attribute to misaligned prototypes and cluster imbalance.
- Two algorithmic solutions (feasible-region QP and objective-transformation SVD iteration) yield identical clustering results.
Limitations
- Equal weighting of view-wise and cross-view prototypes; the authors suggest a weighting scheme could be more reasonable.
- Gm update dominates runtime due to SVD on dm x dm matrices.
- Sub-optimal results at 60% incomplete data on some datasets (e.g., Webkb).
Experimental setup
Datasets: Webkb Wikifea AWA10 SUNRGB-D AwAfea EMNIST
Benchmarks: Webkb Wikifea AWA10 SUNRGB-D AwAfea EMNIST
Metrics: ACC NMI Purity running time memory overhead
Baselines: IMSC-AGL AWP APMC IMG TMBSD IKMKC IMVTSC-MVI CPM-Nets LSIMVC GSRIMC COMPLETER TCIMC
Builds on: self-expression affinity IMVC spectral clustering
Closest papers by meaning
- Deep Variational Incomplete Multi-View Clustering: Exploring Shared Clustering Structures (AAAI 2024 · similarity 0.92)
- DVSAI: Diverse View-Shared Anchors Based Incomplete Multi-View Clustering (AAAI 2024 · similarity 0.92)
- Bifurcate then Alienate: Incomplete Multi-view Clustering via Coupled Distribution Learning with Linear Overhead (ICML 2025 · similarity 0.92)
- Incomplete Contrastive Multi-View Clustering with High-Confidence Guiding (AAAI 2024 · similarity 0.91)
- Incomplete Multi-view Clustering via Diffusion Contrastive Generation (AAAI 2025 · similarity 0.91)
- Fast Incomplete Multi-view Clustering by Flexible Anchor Learning (ICML 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.