Research map / Label-efficient and robust learning / Semi-supervised learning
Correlation-Induced Label Prior for Semi-Supervised Multi-Label Learning
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 PCLP, a pseudo-labeling method for semi-supervised multi-label learning that uses a structural causal model as a correlation-induced label prior inside a variational label enhancement framework trained with a GAN-style discriminator.
Problem
Semi-supervised multi-label learning with very few labeled multi-label images; reliable label correlation is hard to estimate from minimal labeled data, so prior SSMLL methods fail to exploit inter-label correlation for pseudo-labeling.
Approach
Model the joint label prior with a nonlinear structural causal model over a DAG (edges from chain-rule ordering of labels) and integrate it into a variational label enhancement objective so that pseudo-labels for unlabeled images conform to the learned correlated prior, optimized jointly on labeled and unlabeled data.
Key results
- PCLP reaches 84.59 mAP on VOC at p=0.20 and 71.52 on COCO at p=0.20, the best among compared methods in Table 1.
- On NUS at p=0.05, PCLP achieves 46.39 mAP versus 44.99 for PLC and 44.98 for CAP.
- Ablation shows adding SCM prior over plain variational inference improves mAP by about 0.8 to 1.2 points on VOC (e.g., 82.21 vs 81.45 at p=0.10).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use PCLP instead of the strongest reported baseline | SSMLL with ResNet-50 · VOC, p=0.05, 0.10, 0.15, 0.20 | mAP: 75.90 -> 77.25; 81.83 -> 82.21; 83.10 -> 83.72; 84.32 -> 84.59 | Helped |
| use PCLP instead of the strongest reported baseline | SSMLL with ResNet-50 · COCO, p=0.05, 0.10, 0.15, 0.20 | mAP: 62.88 -> 64.43; 67.18 -> 69.02; 68.99 -> 70.86; 70.43 -> 71.52 | Helped |
| use PCLP instead of the strongest reported baseline | SSMLL with ResNet-50 · NUS, p=0.05, 0.10, 0.15, 0.20 | mAP: 44.99 -> 46.39; 48.70 -> 48.83; 49.99 -> 50.57; 51.37 -> 52.45 | Helped |
| add variational inference (VI) for pseudo-label generation (ablation) | PCLP · VOC, COCO, and NUS; p=0.05, 0.10, 0.15, 0.20 | mAP: VOC p=0.05: 75.18 -> 76.38; COCO p=0.05: 62.17 -> 63.38; NUS p=0.05: 44.61 -> 45.44 | Helped |
| add SCM-correlated label prior to VI pseudo-label generation (ablation) | PCLP · VOC, COCO, and NUS; p=0.05, 0.10, 0.15, 0.20 | mAP: VOC p=0.05: 76.38 -> 77.25; COCO p=0.05: 63.38 -> 64.43; NUS p=0.05: 45.44 -> 46.39 | Helped |
| increase the weight of the unlabeled-data regularization term | PCLP · VOC, COCO, and NUS | mAP | Helped |
Surprising findings
- Instance-based pseudo-labeling (IAT-1) underperforms the simple BCE baseline on NUS at larger labeled proportions.
Limitations
- Lambda fixed to 1 across datasets; sensitivity to the tradeoff weight alpha shows performance depends on its value.
- Theoretical guarantees rely on infinite-capacity assumptions and a Bayesian-informed generator definition.
Experimental setup
Datasets: Pascal VOC 2012 MS-COCO 2014 NUS-WIDE
Benchmarks: Pascal VOC 2012 MS-COCO 2014 NUS-WIDE
Metrics: mAP
Baselines: BCE ASL IAT-1 IAT-K IAT LL-R LL-CT LL-CP PLC ADSH FreeMatch DRML
Builds on: variational label enhancement structural causal model (Yu et al., 2019) FreeMatch CAP instance-aware threshold (IAT)
Closest papers by meaning
- RankMatch: A Novel Approach to Semi-Supervised Label Distribution Learning Leveraging Rank Correlation between Labels (NeurIPS 2025 · similarity 0.90)
- DiCaP: Distribution-Calibrated Pseudo-labeling for Semi-Supervised Multi-Label Learning (AAAI 2026 · similarity 0.89)
- Collaboration Based Multi-Label Learning (AAAI 2019 · similarity 0.89)
- Semi-supervised Multi-label Learning with Balanced Binary Angular Margin Loss (NeurIPS 2024 · similarity 0.88)
- Can Class-Priors Help Single-Positive Multi-Label Learning? (NeurIPS 2025 · similarity 0.87)
- Bi-Level Optimization for Semi-Supervised Learning with Pseudo-Labeling (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.