Research map / Learning theory and algorithms / Clustering
Collaborative Similarity Fusion and Consistency Recovery for 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 SFCR, a graph-based incomplete multi-view clustering method that jointly learns a unified similarity graph via view-to-sample weighted fusion and recovers missing samples under that graph's consistent structure, yielding cluster labels directly from the graph's connected components.
Problem
Incomplete multi-view clustering, where some samples are absent in some views; existing methods either discard missing samples or recover them at the view level without accounting for per-sample differences or cross-view structural consistency, and usually need K-means post-processing.
Approach
Fuses view-wise similarity graphs with view-to-sample weights (per sample, per view) and recovers missing samples via self-representation guided by the shared graph, with a rank constraint on the Laplacian so the graph's connected components directly give cluster labels; graph learning and recovery are optimized alternately.
Key results
- On ORL at 0% missing ratio, SFCR reaches 90.25% ACC versus 78.25% for the best competitor SCSL.
- On Leaves at 0% missing ratio, SFCR reaches 91.19% ACC versus 81.71% for PIMVC.
- Average rank across missing ratios is best for SFCR (about 1.17 to 1.33 for ACC), with PIMVC second.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SFCR instead of PIMVC | Incomplete multi-view clustering · 3sources, 20% missing | ACC: 63.86±2.10 -> 74.67±1.79 | Helped |
| Use SFCR instead of PIMVC | Incomplete multi-view clustering · 3sources, 20% missing | NMI: 61.33±4.27 -> 66.83±3.35 | Helped |
| Use SFCR instead of UEAI | Incomplete multi-view clustering · Webkb, 20% missing | ACC: 72.90±3.84 -> 73.39±0.77 | Helped |
| Use SFCR instead of UEAI | Incomplete multi-view clustering · Webkb, 20% missing | NMI: 38.04±3.35 -> 40.43±2.04 | Helped |
| Use SFCR instead of IMSR | Incomplete multi-view clustering · MSRC-v1, 20% missing | ACC: 89.27±3.42 -> 91.80±1.23 | Helped |
| Use SFCR instead of PIMVC | Incomplete multi-view clustering · MSRC-v1, 20% missing | NMI: 85.52±1.10 -> 84.64±2.64 | Hurt |
| Use SFCR instead of PIMVC | Incomplete multi-view clustering · ORL, 20% missing | ACC: 74.47±0.71 -> 81.80±3.22 | Helped |
| Use SFCR instead of PIMVC | Incomplete multi-view clustering · ORL, 20% missing | NMI: 88.91±0.74 -> 90.95±1.45 | Helped |
| Use SFCR instead of PIMVC | Incomplete multi-view clustering · Leaves, 20% missing | ACC: 66.85±0.82 -> 79.51±1.53 | Helped |
| Use SFCR instead of PIMVC | Incomplete multi-view clustering · Leaves, 20% missing | NMI: 82.95±0.45 -> 87.14±0.80 | Helped |
| Use SFCR instead of PIMVC | Incomplete multi-view clustering · COIL100, 20% missing | ACC: 68.72±0.36 -> 73.63±0.93 | Helped |
| Use SFCR instead of PIMVC | Incomplete multi-view clustering · COIL100, 20% missing | NMI: 88.60±0.09 -> 91.25±0.26 | Helped |
Surprising findings
- Simple best-single-view baseline with zero-filled missing samples is competitive on some settings, and SCSL performs worse than BSV on several datasets at higher missing ratios.
Limitations
- Relies on a manually tuned trade-off parameter lambda, with performance sensitive to its value.
- Scalability to large datasets is noted as an open concern; complexity is cubic in the number of samples.
- Experiments are limited to six datasets with synthetically generated missingness at 0/20/30% ratios.
Experimental setup
Datasets: 3sources Webkb MSRC-v1 ORL Leaves COIL100 Three-Ring
Benchmarks: 3sources Webkb MSRC-v1 ORL Leaves COIL100
Metrics: ACC NMI average rank running time
Baselines: BSV UEAI AWGF IMSR PIMVC SCSL
Builds on: adaptive neighbor graph learning (Nie et al. 2014) UEAI IMSR AWGF
Closest papers by meaning
- Sample-Level Cross-View Similarity Learning for Incomplete Multi-View Clustering (AAAI 2024 · similarity 0.92)
- Incomplete Contrastive Multi-View Clustering with High-Confidence Guiding (AAAI 2024 · similarity 0.91)
- Incomplete Multi-view Clustering via Hierarchical Semantic Alignment and Cooperative Completion (NeurIPS 2025 · similarity 0.91)
- Fast Incomplete Multi-view Clustering with Adaptive Similarity Completion and Reconstruction (AAAI 2025 · similarity 0.90)
- Attribute-Missing Multi-view Graph Clustering (CVPR 2025 · similarity 0.90)
- Incomplete and Unpaired Multi-View Graph Clustering with Cross-View Feature Fusion (AAAI 2025 · similarity 0.90)
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.