Research map / Label-efficient and robust learning / Noisy labels
Automatically Identify and Rectify: Robust Deep Contrastive Multi-view Clustering in Noisy Scenarios
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 AIRMVC, a deep contrastive multi-view clustering framework that identifies noisy samples with a GMM over soft predictions, rectifies them via hybrid interpolation, and uses a noise-robust contrastive loss, evaluated on six benchmarks under 10-70% synthetic noise.
Problem
Deep multi-view clustering degrades substantially when some views contain noisy data, and existing noise-robust methods do not explicitly identify and rectify noisy samples.
Approach
Reformulates noisy-sample identification as anomaly detection by fitting a two-component GMM on the clean-probability derived from cross-view soft predictions and representations; a hybrid rectification interpolates each view's soft prediction with the first (assumed clean) view's prediction, and a contrastive loss uses soft-prediction agreement to pick reliable positive pairs.
Key results
- On BBCSport at 10% noise, AIRMVC reaches 49.45 ACC / 28.63 NMI / 49.44 PUR, above the best baseline in each metric per the paper.
- On WebKB at 10% noise, AIRMVC reports 83.73 ACC versus 78.12 for the strongest baseline (RMCNC/MFLVC).
- Ablation shows removing the identification-and-rectification module hurts more than removing the contrastive mechanism.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AIRMVC instead of the strongest listed baseline | Multi-view clustering · BBCSport, 10% noise | ACC: 47.06 -> 49.45 | Helped |
| Use AIRMVC instead of the strongest listed baseline | Multi-view clustering · WebKB, 10% noise | ACC: 78.12 -> 83.73 | Helped |
| Use AIRMVC instead of the strongest listed baseline | Multi-view clustering · Reuters, 10% noise | ACC: 45.25 -> 48.25 | Helped |
| Use AIRMVC instead of the strongest listed baseline | Multi-view clustering · UCI-digit, 10% noise | ACC: 88.20 -> 93.95 | Helped |
| Use AIRMVC instead of the strongest listed baseline | Multi-view clustering · Caltech101, 10% noise | ACC: 20.53 -> 21.45 | Helped |
| Use AIRMVC instead of the strongest listed baseline | Multi-view clustering · STL10, 10% noise | ACC: 27.58 -> 28.81 | Helped |
| Add noisy identification and rectification (ablation) | AIRMVC · Six benchmark datasets, noisy scenarios | clustering performance | Helped |
| Add the noise-robust contrastive mechanism (ablation) | AIRMVC · Six benchmark datasets, noisy scenarios | clustering performance | Helped |
| Add noisy identification and rectification plus noise-robust contrastive learning (ablation) | Autoencoder backbone · Six benchmark datasets, noisy scenarios | clustering performance | Helped |
| Increase the contrastive confidence threshold (ablation) | AIRMVC · Six datasets, 10% noise | clustering performance | Helped |
| Set the rectification-loss weight alpha and contrastive-loss weight beta to extreme values (ablation) | AIRMVC · Six datasets, 10% noise | clustering performance | Hurt |
Surprising findings
- Several classical deep MVC methods perform worse than single-view clustering under noisy multi-view input, according to the toy experiments.
Limitations
- Assumes the first view is clean, which may not hold in real scenarios.
- Noise is synthetically injected rather than from real-world sensor corruption.
- Hyperparameters alpha, beta fixed at 1.0 and tau chosen by sensitivity analysis on test benchmarks.
Experimental setup
Datasets: BBCSport WebKB Reuters UCI-digit Caltech101 STL10
Benchmarks: BBCSport WebKB Reuters UCI-digit Caltech101 STL10
Metrics: ACC NMI PUR
Baselines: CoMVC SiMVC MFLVC DealMVC SURE CANDY DIVIDE TGM-MVC SCE-MVC RMCNC MVCAN
Builds on: contrastive learning GMM autoencoder
Closest papers by meaning
- EASEMVC:Efficient Dual Selection Mechanism for Deep Multi-View Clustering (CVPR 2025 · similarity 0.88)
- Investigating and Mitigating the Side Effects of Noisy Views for Self-Supervised Clustering Algorithms in Practical Multi-View Scenarios (CVPR 2024 · similarity 0.88)
- Decoupled Contrastive Multi-View Clustering with High-Order Random Walks (AAAI 2024 · similarity 0.87)
- ROLL: Robust Noisy Pseudo-label Learning for Multi-View Clustering with Noisy Correspondence (CVPR 2025 · similarity 0.87)
- Medusa: A Multi-Scale High-order Contrastive Dual-Diffusion Approach for Multi-View Clustering (CVPR 2025 · similarity 0.87)
- Incomplete Multi-view Clustering via Diffusion Contrastive Generation (AAAI 2025 · similarity 0.87)
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.