Research map / Label-efficient and robust learning / Semi-supervised learning
RankUp: Boosting Semi-Supervised Regression with an Auxiliary Ranking Classifier
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.
Adapts FixMatch-style semi-supervised classification to regression by training an auxiliary pairwise ranking classifier (ARC) alongside the regression head, plus a Regression Distribution Alignment (RDA) step that refines pseudo-labels.
Problem
Semi-supervised regression cannot directly use confidence-threshold and low-density techniques from semi-supervised classification because regression models lack confidence measures and decision boundaries.
Approach
Convert regression into a pairwise ranking problem solved by an auxiliary softmax ranking classifier (inspired by RankNet), so FixMatch can be applied to it; add RDA, which aligns sorted regression pseudo-labels to the sorted labeled-label distribution.
Key results
- On UTKFace with 50 labels, RankUp + RDA reaches MAE 9.33, R^2 0.552, SRCC 0.770, versus best baseline MixMatch at 11.44, 0.401, 0.674.
- On UTKFace with 250 labels, RankUp + RDA reaches MAE 6.57 and R^2 0.782 vs. supervised 9.42 and 0.540.
- On BVCC (250 labels), RankUp + RDA achieves MAE 0.463 and R^2 0.598, outperforming the best prior method UCVME (0.498, 0.553).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add ARC to the supervised regression model (ablation) | RankUp · UTKFace, 250 labels | MAE: 9.42 -> 7.06 | Helped |
| Add RDA to RankUp (ablation) | RankUp · UTKFace, 50 labels | MAE: 9.96 -> 9.33 | Helped |
| Add RDA to RankUp (ablation) | RankUp · UTKFace, 250 labels | MAE: 7.06 -> 6.57 | Helped |
| Add RDA to RankUp (ablation) | RankUp · BVCC, 250 labels | MAE: 0.470 -> 0.463 | Helped |
| Add RDA to RankUp (ablation) | RankUp · Yelp Review, 250 labels | SRCC: 0.829 -> 0.810 | Hurt |
| Use RDA only, without ARC (ablation) | RankUp · UTKFace, 50 labels | MAE: 14.13 -> 14.34 | Hurt |
| Train ARC with FixMatch rather than supervised ARC (ablation) | RankUp · UTKFace, 250 labels | MAE: 9.03 -> 7.06 | Helped |
| Apply weak augmentation to labeled data (ablation) | RankUp + RDA · UTKFace, 250 labels | MAE: 7.16 -> 6.57 | Helped |
Surprising findings
- RDA alone does not consistently beat the supervised baseline on UTKFace with 50 labels; it helps mainly when combined with ARC.
- On Yelp Review (only five distinct labels), adding RDA lowered SRCC relative to RankUp alone (0.829 to 0.810).
Limitations
- RDA relies on the assumption that labeled and unlabeled label distributions are similar and that pseudo-label rankings are accurate.
- RDA may underperform when the label set has few distinct values (e.g., Yelp Review SRCC drop).
- Experiments use only three datasets and a small set of base architectures.
Experimental setup
Datasets: UTKFace BVCC Yelp Review
Benchmarks: UTKFace BVCC Yelp Review
Metrics: MAE R^2 SRCC
Baselines: Supervised Π-Model Mean Teacher CLSS UCVME MixMatch
Builds on: FixMatch RankNet USB MixMatch distribution alignment
Closest papers by meaning
- Semi-Supervised Regression by Preserving Ranking Relationships Between Close Unlabeled Samples (AAAI 2026 · similarity 0.90)
- Nonparametric Regression with Comparisons: Escaping the Curse of Dimensionality with Ordinal Information (ICML 2018 · similarity 0.85)
- RankMatch: A Novel Approach to Semi-Supervised Label Distribution Learning Leveraging Rank Correlation between Labels (NeurIPS 2025 · similarity 0.84)
- Enhancing Semi-Supervised Learning via Representative and Diverse Sample Selection (NeurIPS 2024 · similarity 0.84)
- CGMatch: A Different Perspective of Semi-supervised Learning (CVPR 2025 · similarity 0.84)
- RegMixMatch: Optimizing Mixup Utilization in Semi-Supervised Learning (AAAI 2025 · similarity 0.83)
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.