Research map / Label-efficient and robust learning / Noisy labels
Sample Selection via Contrastive Fragmentation for Noisy Label Regression
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 ConFrag, which discretizes regression labels into fragments, trains MoE experts on maximally contrasting fragment pairs, and selects clean samples via neighborhood agreement, evaluated on six curated noisy-label regression benchmarks.
Problem
Learning regression models under noisy continuous labels, where existing noisy-label methods are designed mainly for classification and noise rates are often unknown.
Approach
Partition the label range into fragments, pair the most distant fragments into contrastive pairs to train binary expert feature extractors (turning closed-set noise into open-set noise), and select clean samples by a fragment-prior-weighted mixture of neighborhood agreements, with neighborhood jittering as a regularizer.
Key results
- ConFrag achieves the lowest MRAE in most symmetric and Gaussian noise settings across six benchmarks, e.g., 5.08% vs 16.18 vanilla at 20% symmetric noise on IMDB-Clean-B; 2.74% on AFAD-B at 20%.
- Co-ConFrag achieves the best MRAE in most settings, e.g., 0.54% on AFAD-B at 20% symmetric noise and -8.86% on IMDB-WIKI-B real noise.
- ConFrag-R (regression-based experts) underperforms classification-based ConFrag in all reported datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ConFrag rather than the noise-free-trained model | AFAD-B · AFAD-B symmetric 20% | MRAE (%): 9.37 -> 2.74 | Helped |
| Use ConFrag rather than the noise-free-trained model | IMDB-Clean-B · IMDB-Clean-B symmetric 40% | MRAE (%): 32.05 -> 12.64 | Helped |
| Use ConFrag rather than the noise-free-trained model | SHIFT15M-B · SHIFT15M-B symmetric 40% | MRAE (%): 17.96 -> 6.18 | Helped |
| Use ConFrag rather than the noise-free-trained model | MSD-B · MSD-B symmetric 40% | MRAE (%): 18.43 -> 4.94 | Helped |
| Use classification rather than regression for feature-extractor learning (ablation) | ConFrag · IMDB-Clean-B symmetric 40% | MRAE (%): 22.73 -> 12.64 | Helped |
| Add self agreement to neighborhood agreement (ablation) | ConFrag sample selection · IMDB-Clean-B symmetric 40% | MRAE (%): 22.22 -> 12.64 | Helped |
| Add neighborhood agreement to self agreement (ablation) | ConFrag sample selection · IMDB-Clean-B symmetric 40% | MRAE (%): 16.97 -> 12.64 | Helped |
| Use the union rather than only predictive selected samples (ablation) | ConFrag sample selection · IMDB-Clean-B symmetric 40% | MRAE (%): 14.18 -> 12.64 | Helped |
| Use the union rather than only representational selected samples (ablation) | ConFrag sample selection · IMDB-Clean-B symmetric 40% | MRAE (%): 14.06 -> 12.64 | Helped |
| Use the union rather than the intersection of predictive and representational selected samples (ablation) | ConFrag sample selection · IMDB-Clean-B symmetric 40% | MRAE (%): 13.08 -> 12.64 | Helped |
| Use contrastive fragment pairing rather than all fragments together (ablation) | ConFrag · IMDB-Clean-B symmetric 40% | MRAE (%) | Helped |
| Increase fragment count (ablation) | ConFrag · IMDB-Clean-B | MRAE (%) | Hurt |
Surprising findings
- Co-ConFrag and ConFrag achieve MRAE below the noise-free model in several settings, e.g., -8.86% on IMDB-WIKI-B real noise.
- Injecting open-set label noise was found to be less harmful than closed-set noise in the authors' experiments.
Limitations
- Mixture-of-experts parameter count grows linearly with the number of fragments F.
- Introduces a new hyperparameter (number of fragments/experts) and experts partition the data, possibly limiting per-expert data use.
- Scalability concerns for MoE in computation and memory, mitigated only with compact extractors.
Experimental setup
Datasets: AFAD IMDB-Clean IMDB-WIKI UTKFace SHIFT15M MSD
Benchmarks: AFAD-B IMDB-Clean-B IMDB-WIKI-B SHIFT15M-B MSD-B UTKFace-B
Metrics: MRAE MAE selection rate ERR (Error Residual Ratio) MSE
Baselines: CNLCU-S CNLCU-H Sigua SPR BMM DY-S SuperLoss C-Mixup RDI CDR D2L AUX
Builds on: mixup co-teaching superloss selfie
Closest papers by meaning
- Enhanced Sample Selection with Confidence Tracking: Identifying Correctly Labeled Yet Hard-to-Learn Samples in Noisy Data (AAAI 2025 · similarity 0.85)
- Learning from Noisy Labels via Conditional Distributionally Robust Optimization (NeurIPS 2024 · similarity 0.85)
- ConR: Contrastive Regularizer for Deep Imbalanced Regression (ICLR 2024 · similarity 0.85)
- Leveraging Data Symmetries to Select an Optimal Subset of Training Data under Label Noise (NeurIPS 2026 · similarity 0.85)
- Unlocking the Power of Open Set: A New Perspective for Open-Set Noisy Label Learning (AAAI 2024 · similarity 0.84)
- Adaptive Integration of Partial Label Learning and Negative Learning for Enhanced Noisy Label Learning (AAAI 2024 · similarity 0.84)
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.