Research map / Label-efficient and robust learning / Recognition and long-tail
Positive-Unlabeled Learning by Latent Group-Aware Meta Disambiguation
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 LaGAM, a positive-unlabeled learning framework that combines hierarchical contrastive learning over latent clusters with meta-learning to directly distill true labels of unlabeled data.
Problem
PU learning lags far behind supervised learning because unreliable binary labels with poor semantics hinder learning discriminative representations and label disambiguation.
Approach
Use k-means cluster-based and kNN-based contrastive positives, cut off by classifier predictions (dichotomized cutoff), to learn group-aware representations; then use an online bi-level meta-learning procedure with projected gradients and EMA to directly update pseudo-labels of unlabeled data rather than reweighting them.
Key results
- LaGAM reaches 96.6% on CIFAR-10 and 89.8% on CIFAR-100 in ablation, with the abstract claiming an average 6.8% accuracy gain over state of the art.
- On Alzheimer, LaGAM achieves 77.8% accuracy, 84.5 F1 and 83.2 AUC versus 71.7% accuracy for the best baseline Dist-PU.
- With a ResNet backbone, LaGAM reaches 96.2% on CIFAR-10-1 versus 88.8% for Dist-PU.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LaGAM instead of Dist-PU | PU classification with CNN · CIFAR-100-2 | accuracy: 72.9±0.6 -> 85.9±0.1 | Helped |
| Use LaGAM instead of P3MIX-C | PU classification with CNN · STL-10-2 | accuracy: 84.1±0.3 -> 81.9±1.5 | Hurt |
| Use LaGAM instead of Dist-PU | Alzheimer PU diagnosis · Alzheimer | AUC: 77.1±0.7 -> 83.2±1.7 | Helped |
| Add meta-disambiguation to nnPU with Lcont (ablation) | PU classification · CIFAR-10 | accuracy: 91.2 -> 96.6 | Helped |
| Add Lcont to meta-disambiguation (ablation) | LaGAM · CIFAR-100 | accuracy: 73.5 -> 89.8 | Helped |
| Add meta-disambiguation and Lcont instead of naive BCE (ablation) | PU classification · CIFAR-100 | accuracy: 50.0 -> 89.8 | Helped |
| Use normalized-gradient soft label updates instead of projected-gradient EMA updates (ablation) | LaGAM without Lcont · CIFAR-10 | accuracy: 87.6 -> 88.4 | Helped |
| Remove EMA from projected-gradient label updates (ablation) | LaGAM without Lcont · CIFAR-10 | accuracy: 87.6 -> 91.2 | Helped |
| Add neighbor smoothing to the positive-set construction (ablation) | LaGAM contrastive learning · CIFAR-10 | accuracy: 96.3 -> 96.6 | Helped |
| Add dichotomized cutoff and neighbor smoothing (ablation) | LaGAM contrastive learning · CIFAR-10 | accuracy: 91.8 -> 96.6 | Helped |
| Set the k-means cluster number to 10 instead of 5 (ablation) | LaGAM · CIFAR-10 | accuracy: 96.4 -> 97.0 | Helped |
| Set the k-means cluster number to 20 instead of 5 (ablation) | LaGAM · CIFAR-100 | accuracy: 88.4 -> 90.2 | Helped |
Surprising findings
- Projected-gradient EMA label updating performs worst without contrastive loss but best with it, reversing the ranking of update strategies.
- Freezing all classifier layers except the head during meta-learning speeds training and, per the authors, yields higher converged accuracy.
Limitations
- Requires a small labeled support set of clean samples for meta-learning evaluation.
- Relies on k-means clustering with a chosen cluster number; accuracy degrades slightly at small cluster counts.
Experimental setup
Datasets: CIFAR-10 STL-10 CIFAR-100 Alzheimer MRI dataset
Benchmarks: CIFAR-10 STL-10 CIFAR-100 Alzheimer
Metrics: classification accuracy F1 score AUC
Baselines: uPU nnPU Self-PU PAN VPU Dist-PU P3MIX-E P3MIX-C Supervised
Builds on: nnPU Self-PU P3MIX Dist-PU MixUp
Closest papers by meaning
- A Closer Look to Positive-Unlabeled Learning from Fine-grained Perspectives: An Empirical Study (NeurIPS 2025 · similarity 0.88)
- Adaptive Integration of Partial Label Learning and Negative Learning for Enhanced Noisy Label Learning (AAAI 2024 · similarity 0.85)
- Positive and Unlabeled Learning with Controlled Probability Boundary Fence (ICML 2024 · similarity 0.84)
- Learning from positive and unlabeled examples -Finite size sample bounds (NeurIPS 2025 · similarity 0.84)
- Leveraging Group Classification with Descending Soft Labeling for Deep Imbalanced Regression (AAAI 2025 · similarity 0.84)
- Prior-Constrained Association Learning for Fine-Grained Generalized Category Discovery (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.