Research map / Graph learning / Graph self-supervised learning
Disentangled Partial 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 TERIAL, a disentangled partial label learning method that builds an instance-label bipartite graph and uses a neighborhood routing mechanism with labeling confidences and distance-correlation independence loss to learn disentangled representations for disambiguation.
Problem
Partial label learning (PLL) where each instance has a candidate label set with only one valid label; existing PLL methods learn entangled representations that struggle to disambiguate candidate labels.
Approach
Formulates PL examples as an undirected instance-label bipartite graph and applies a tailored neighborhood routing mechanism (clustering in K factor subspaces, with labeling confidences gating edges) to produce factor-aware disentangled node representations, regularized by distance correlation for independence.
Key results
- TERIAL outperforms state-of-the-art PLL methods in 86 of 90 pairwise comparisons across 5 datasets and 3 values of r.
- On MNIST with r=3, TERIAL reaches 98.10% accuracy versus 97.85% for PRODEN.
- Ablation: TERIAL-0 (no disentangling layers) underperforms; L=2 layers chosen; K=1 gives 86.14% vs 87.53% at K=10 on KMNIST (r=5).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TERIAL instead of the strongest listed baseline | PLL on MNIST · MNIST, r=3/5/7 | accuracy: 97.97±0.16% -> 98.10±0.12%; 97.49±0.17% -> 97.71±0.09%; 96.19±0.19% -> 96.57±0.06% | Helped |
| Use TERIAL instead of the strongest listed baseline | PLL on KMNIST · KMNIST, r=3/5/7 | accuracy: 88.78±0.07% -> 89.24±0.11%; 86.98±0.11% -> 87.53±0.19%; 80.55±0.13% -> 82.97±0.16% | Helped |
| Use TERIAL instead of the strongest listed baseline | PLL on FMNIST · FMNIST, r=3/5/7 | accuracy: 88.17±0.10% -> 88.68±0.08%; 87.42±0.16% -> 87.95±0.09%; 86.02±0.23% -> 86.69±0.05% | Helped |
| Use TERIAL instead of the strongest listed baseline | PLL on SVHN · SVHN, r=3/5/7 | accuracy: 94.76±0.12% -> 95.24±0.09%; 94.25±0.11% -> 94.77±0.12%; 92.92±0.16% -> 93.92±0.07% | Helped |
| Use TERIAL instead of the strongest listed baseline | PLL on CIFAR-10 · CIFAR-10, r=3/5/7 | accuracy: 78.87±0.29% -> 79.01±0.23%; 75.59±0.34% -> 76.61±0.25%; 72.34±0.46% -> 65.07±0.37% | Mixed |
| Add disentangling layers (0 to 1) (ablation) | TERIAL on MNIST · MNIST, r=3 | accuracy: 97.73% -> 97.91% | Helped |
| Add a second disentangling layer (1 to 2) (ablation) | TERIAL on MNIST · MNIST, r=3 | accuracy: 97.91% -> 98.10% | Helped |
| Add a third disentangling layer (2 to 3) (ablation) | TERIAL on MNIST · MNIST, r=3 | accuracy: 98.10% -> 98.13% | Helped |
| Add disentangling layers (0 to 1) (ablation) | TERIAL on SVHN · SVHN, r=7 | accuracy: 92.67% -> 93.33% | Helped |
| Add a second disentangling layer (1 to 2) (ablation) | TERIAL on SVHN · SVHN, r=7 | accuracy: 93.33% -> 93.92% | Helped |
| Add a third disentangling layer (2 to 3) (ablation) | TERIAL on SVHN · SVHN, r=7 | accuracy: 93.92% -> 93.96% | Helped |
| Increase latent factor count from K=1 to K=5 (ablation) | TERIAL on KMNIST · KMNIST, r=5 | accuracy: 86.14% -> 86.62% | Helped |
Surprising findings
- Performance drops when the number of latent factors K gets too large, because chunk dimensionality shrinks.
- CAVL collapses on SVHN and CIFAR-10 at larger r (e.g., 35.65% on SVHN r=7), and TERIAL loses to PRODEN and RC on CIFAR-10.
Limitations
- Clustering-based routing is described as time-consuming due to iterative convergence requirements.
- Underperforms PRODEN and RC on CIFAR-10 in several settings.
- Performance is sensitive to the assumed number of latent factors K.
Experimental setup
Datasets: MNIST KMNIST FMNIST SVHN CIFAR-10
Benchmarks: MNIST KMNIST FMNIST SVHN CIFAR-10
Metrics: classification accuracy
Baselines: PRODEN RC CC LW VALEN CAVL
Builds on: capsule routing (dynamic routing) PRODEN distance correlation
Closest papers by meaning
- Partial Label Learning with Dissimilarity Propagation guided Candidate Label Shrinkage (NeurIPS 2023 · similarity 0.89)
- Distilling Reliable Knowledge for Instance-Dependent Partial Label Learning (AAAI 2024 · similarity 0.89)
- Navigating Real-World Partial Label Learning: Unveiling Fine-Grained Images with Attributes (AAAI 2024 · similarity 0.87)
- Graph-based Unsupervised Disentangled Representation Learning via Multimodal Large Language Models (NeurIPS 2024 · similarity 0.87)
- Realistic Evaluation of Deep Partial-Label Learning Algorithms (ICLR 2025 · similarity 0.86)
- Partial Label Causal Representation Learning for Instance-Dependent Supervision and Domain Generalization (AAAI 2025 · similarity 0.86)
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.