Research map / Label-efficient and robust learning / Recognition and long-tail
GBRIP: Granular Ball Representation for Imbalanced 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 GBRIP, which builds a coarse-grained granular-ball feature representation via unsupervised 2NN clustering and a GB graph with a multi-center loss to handle inter- and intra-class imbalance in long-tailed partial label learning.
Problem
Partial label learning under class imbalance, where label disambiguation relies on inter-class features and ignores intra-class imbalance, yielding unreliable pseudo-label confidence matrices.
Approach
Uses unsupervised 2NN clustering to split the feature space into roughly equal-sized granular balls acting as sub-classes, builds a GB-based weighted graph (within-ball and between-ball edges) to estimate label confidences, and adds a multi-center loss that pulls samples toward their ball centers alongside cross-entropy and a class-prior regularizer.
Key results
- On CIFAR10-LT (psi=0.3, gamma=200), GBRIP reaches 85.78% vs. 80.57% for the second-best HTC, a reported 5.21% improvement.
- On CIFAR100-LT (psi=0.1, gamma=50), GBRIP reaches 56.91% vs. 51.26% for HTC.
- On real-world Lost, GBRIP reaches 85.73% vs. 82.71% for HTC; on Yahoo! News it is slightly below HTC (72.38% vs. 72.61%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GBRIP instead of the strongest baseline | CIFAR10-LT · CIFAR10-LT (all six ψ and γ settings in Table 1) | accuracy: 91.54\textpm 0.12 vs 88.14\textpm 0.94; 88.97\textpm 1.55 vs 85.66\textpm 1.44; 85.78\textpm 1.55 vs 80.57\textpm 1.40; | Helped |
| Use GBRIP instead of the strongest baseline | CIFAR100-LT · CIFAR100-LT (all six ψ and γ settings in Table 1) | accuracy: 67.15\textpm 1.24 vs 65.83\textpm 0.43; 63.42\textpm 0.97 vs 61.13\textpm 1.15; 57.94\textpm 0.38 vs 53.26\textpm 1.90; | Helped |
| Use GBRIP instead of HTC | GBRIP · Lost, Bird Song, Soccer Player | accuracy: 85.73 vs 82.71; 78.93 vs 77.98; 62.78 vs 60.11 | Helped |
| Use GBRIP instead of HTC | GBRIP · Yahoo! News | accuracy: 72.38 vs 72.61 | Hurt |
| add CGR and MCL (ablation) | GBRIP · CIFAR10-LT (ψ=0.5, γ=100) | accuracy: 37.06 -> 86.71 | Helped |
| add MCL (ablation) | GBRIP · CIFAR10-LT (ψ=0.5, γ=100) | accuracy: 70.59 -> 86.71 | Helped |
| use CGR instead of KMeans (ablation) | GBRIP · CIFAR10-LT (ψ=0.5, γ=100) | accuracy: 68.40 -> 86.71 | Helped |
| add Mixup (ablation) | GBRIP · CIFAR10-LT (ψ=0.5, γ=100) | accuracy: 73.60 -> 86.71 | Helped |
| add logit adjustment (ablation) | GBRIP · CIFAR10-LT (ψ=0.5, γ=100) | accuracy: 75.07 -> 86.71 | Helped |
| add sample selection (ablation) | GBRIP · CIFAR10-LT (ψ=0.5, γ=100) | accuracy: 75.53 -> 86.71 | Helped |
Surprising findings
- Removing the CGR and MCL modules together drops CIFAR10-LT accuracy from 86.71% to 37.06%, a much larger drop than removing either the MCL alone (70.59%) or the logit adjustment (75.07%).
Limitations
- Gains over HTC are marginal or absent on Yahoo! News (72.38% vs. 72.61%).
- Reported SUN397 results are shown only in a figure and the setup is partly synthetic (flipped partial labels).
- Code is only promised 'in supplemental materials' and no release link is given.
Experimental setup
Datasets: CIFAR10-LT CIFAR100-LT Lost Bird Song Soccer Player Yahoo! News SUN397
Benchmarks: CIFAR10-LT CIFAR100-LT Lost Bird Song Soccer Player Yahoo! News SUN397
Metrics: accuracy many/medium/few-shot accuracy
Baselines: MSE VALEN LWS PRODEN PiCO RECORDS Solar PLRIPL HTC CAVL
Builds on: RECORDS Solar PLRIPL HTC PiCO
Closest papers by meaning
- Rebalancing Multi-Label Class-Incremental Learning (AAAI 2025 · similarity 0.84)
- Rethinking Self-Distillation: Label Averaging and Enhanced Soft Label Refinement with Partial Labels (ICLR 2025 · similarity 0.84)
- Toward Robustness in Multi-Label Classification: A Data Augmentation Strategy against Imbalance and Noise (AAAI 2024 · similarity 0.84)
- Semi-supervised Multi-label Learning with Balanced Binary Angular Margin Loss (NeurIPS 2024 · similarity 0.84)
- Taming the Sigmoid Bottleneck: Provably Argmaxable Sparse Multi-Label Classification (AAAI 2024 · similarity 0.83)
- Navigating Real-World Partial Label Learning: Unveiling Fine-Grained Images with Attributes (AAAI 2024 · 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.