Research map / Label-efficient and robust learning / Noisy labels
ROLL: Robust Noisy Pseudo-label Learning for Multi-View Clustering with Noisy Correspondence
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 ROLL, a multi-view clustering framework that handles both noisy pseudo-labels and noisy cross-view correspondence via a noise-tolerant pseudo-label loss and a robust multi-view contrastive loss.
Problem
Unsupervised multi-view clustering where self-generated pseudo-labels are often wrong (noisy pseudo-label problem) and cross-view sample alignment is corrupted (noisy correspondence problem), causing overfitting.
Approach
A two-stage pipeline: contrastive/autoencoder warm-up to produce per-view pseudo-labels, then training with a cross-view-weighted noise-tolerant pseudo-label cross-entropy loss plus a robust contrastive loss that interpolates between InfoNCE (r→0) and MAE-style loss (r→1) to limit the effect of misaligned pairs.
Key results
- On Scene-15 at 0% noise, ROLL reaches 47.14 ACC, 48.76 NMI, 30.54 ARI, the best among compared methods.
- On CUB at 20% noise, ROLL reaches 81.03 ACC, 76.81 NMI, 68.11 ARI, versus 71.77 ACC for RMCNC.
- Ablation on Scene-15 and CUB at 20% noise shows removing the RCL module or the NPL module degrades clustering performance substantially.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ROLL rather than the strongest listed baseline | Multi-view clustering · Scene15, 0% noise | ACC: 43.70 -> 47.14 | Helped |
| Use ROLL rather than the strongest listed baseline | Multi-view clustering · CUB, 0% noise | ACC: 81.20 -> 82.90 | Helped |
| Use ROLL rather than the strongest listed baseline | Multi-view clustering · Reuters, 0% noise | ACC: 50.07 -> 50.76 | Helped |
| Use ROLL rather than the strongest listed baseline | Multi-view clustering · NUS-WIDE, 0% noise | ACC: 64.29 -> 71.06 | Helped |
| Use ROLL rather than the strongest listed baseline | Multi-view clustering · XMediaNet, 0% noise | ACC: 80.33 -> 80.79 | Helped |
| Add reconstruction loss Lrec (ablation) | ROLL · Scene15, 20% noise | ACC: 18.29 -> 45.65 | Helped |
| Add reconstruction loss Lrec (ablation) | ROLL · CUB, 20% noise | ACC: 49.87 -> 81.03 | Helped |
| Add noise-tolerance pseudo-label loss Lnpl (ablation) | ROLL with Lrec and Lrcl · Scene15, 20% noise | ACC: 35.20 -> 45.65 | Helped |
| Add noise-tolerance pseudo-label loss Lnpl (ablation) | ROLL with Lrec and Lrcl · CUB, 20% noise | ACC: 43.93 -> 81.03 | Helped |
| Add robust multi-view contrastive loss Lrcl (ablation) | ROLL with Lrec and Lnpl · Scene15, 20% noise | ACC: 43.52 -> 45.65 | Helped |
| Add robust multi-view contrastive loss Lrcl (ablation) | ROLL with Lrec and Lnpl · CUB, 20% noise | ACC: 67.03 -> 81.03 | Helped |
| Replace cross-entropy CE with noise-tolerance loss Lnpl (ablation) | ROLL · Scene15, 20% noise | ACC: 43.31 -> 45.65 | Helped |
Limitations
- Performance at high noise rates (e.g., 80%) on some datasets remains low (e.g., Scene-15 39.84 ACC at 80%).
- Hyperparameters (alpha, beta, r) are fixed per dataset and sensitivity is shown only on one dataset and noise setting.
- Pseudo-label noise and correspondence noise are simulated by shuffling, which may not reflect real-world noise.
Experimental setup
Datasets: Scene-15 CUB Reuters NUS-WIDE XMediaNet
Benchmarks: Scene-15 CUB Reuters NUS-WIDE XMediaNet
Metrics: ACC NMI ARI
Baselines: MvCLN MFLVC SURE DCP DealMVC CVCL MSESC RMCNC ICMVC SCM-RE MVCAN
Builds on: InfoNCE RMCNC MFLVC PGCL
Closest papers by meaning
- Investigating and Mitigating the Side Effects of Noisy Views for Self-Supervised Clustering Algorithms in Practical Multi-View Scenarios (CVPR 2024 · similarity 0.89)
- Noisy Label Calibration for Multi-View Classification (AAAI 2025 · similarity 0.88)
- RPSC: Robust Pseudo-Labeling for Semantic Clustering (AAAI 2024 · similarity 0.88)
- Robust Multi-view Clustering against Imperfect Information (NeurIPS 2026 · similarity 0.87)
- Automatically Identify and Rectify: Robust Deep Contrastive Multi-view Clustering in Noisy Scenarios (ICML 2025 · similarity 0.87)
- Robust Contrastive Multi-view Clustering against Dual Noisy Correspondence (NeurIPS 2024 · similarity 0.86)
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.