Research map / Learning theory and algorithms / Clustering
Decoupled Contrastive Multi-View Clustering with High-Order Random Walks
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 DIVIDE, a contrastive multi-view clustering method that uses multi-step random walks on an affinity graph to rectify false negatives/positives globally, combined with decoupled intra- and inter-view contrastive learning via cross-view decoders.
Problem
Contrastive multi-view clustering suffers from false negatives (intra-cluster samples treated as negatives) and, with neighborhood-based correction, false positives from under- and over-rectification; also needs robustness to missing views.
Approach
Replace first-order epsilon/kNN neighborhood pair construction with t-step random-walk transition matrices on a heat-kernel affinity graph to soft-rectify contrastive targets (identifying out-of-neighborhood positives and in-neighborhood negatives), and perform intra-view and inter-view contrastive learning in separate view-specific spaces linked by cross-view decoders.
Key results
- On Scene-15 complete setting DIVIDE reaches 49.1 ACC / 48.7 NMI / 31.6 ARI, above the best baseline ProImp (43.6 / 45.0 / 26.8).
- On Caltech-101 incomplete setting DIVIDE reaches 63.4 ACC vs 56.2 for the next best (DAIMC).
- Ablation on Scene-15 shows the random-walk rectification adds about 1 point ACC over no rectification (49.1 vs 48.0 ACC with decoder).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DIVIDE instead of the strongest listed baseline | MvC clustering · Scene15 incomplete | ACC: 41.6 -> 46.8 | Helped |
| Use DIVIDE instead of the strongest listed baseline | MvC clustering · Caltech101 incomplete | ACC: 56.2 -> 63.4 | Helped |
| Use DIVIDE instead of the strongest listed baseline | MvC clustering · Reuters incomplete | ACC: 51.9 -> 54.7 | Helped |
| Use DIVIDE instead of the strongest listed baseline | MvC clustering · LandUse21 incomplete | ACC: 23.1 -> 30.0 | Helped |
| Use DIVIDE instead of the strongest listed baseline | MvC clustering · Scene15 complete | ACC: 43.6 -> 49.1 | Helped |
| Use DIVIDE instead of the strongest listed baseline | MvC clustering · Caltech101 complete | ACC: 57.5 -> 62.2 | Helped |
| Use DIVIDE instead of the strongest listed baseline | MvC clustering · Reuters complete | ACC: 56.5 -> 59.3 | Helped |
| Use DIVIDE instead of the strongest listed baseline | MvC clustering · LandUse21 complete | ACC: 26.2 -> 32.3 | Helped |
| Replace no FN identification with random-walk FN identification (ablation) | DIVIDE · Scene15 | ACC: 48.0 -> 49.1 | Helped |
| Replace no FN identification with random-walk FN identification (ablation) | DIVIDE · Caltech101 | ACC: 54.3 -> 62.2 | Helped |
| Use random walk rather than k-nearest-neighbors for FN identification (ablation) | DIVIDE · Scene15 | KL divergence: 22.3 -> 8.2 | Helped |
| Use random walk rather than epsilon-neighborhood for FN identification (ablation) | DIVIDE · Caltech101 | KL divergence: 8.5 -> 5.1 | Helped |
Limitations
- Gains over the nearest robust baseline are modest on some settings (e.g., about 1 point ACC from the rectification module)
- Evaluated only on four small multi-view benchmarks with fully connected encoders
Experimental setup
Datasets: Scene-15 Caltech-101 Reuters LandUse-21
Benchmarks: Scene-15 Caltech-101 Reuters LandUse-21
Metrics: ACC NMI ARI KL divergence
Baselines: DCCAE BMVC PMVC DAIMC EERIMVC SURE DSIMVC DCP ProImp
Builds on: InfoNCE momentum contrast SURE DCP
Closest papers by meaning
- Medusa: A Multi-Scale High-order Contrastive Dual-Diffusion Approach for Multi-View Clustering (CVPR 2025 · similarity 0.90)
- Incomplete Multi-view Clustering via Diffusion Contrastive Generation (AAAI 2025 · similarity 0.89)
- Robust Contrastive Multi-view Clustering against Dual Noisy Correspondence (NeurIPS 2024 · similarity 0.89)
- EASEMVC:Efficient Dual Selection Mechanism for Deep Multi-View Clustering (CVPR 2025 · similarity 0.89)
- Self-supervised Trusted Contrastive Multi-view Clustering with Uncertainty Refined (AAAI 2025 · similarity 0.88)
- Incomplete Contrastive Multi-View Clustering with High-Confidence Guiding (AAAI 2024 · similarity 0.88)
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.