Research map / Label-efficient and robust learning / Recognition and long-tail
Learning Only When It Matters: Cost-Aware Long-Tailed Classification
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 AugARP, a cost-aware long-tailed classification method that learns an implicit region partition via attention and uses it to guide mixup-based augmentation toward high-cost tail instances.
Problem
Long-tailed classification where training and test class distributions match but misclassification costs vary, with only some tail instances being high-cost; standard cost-agnostic long-tail methods assume balanced test sets and uniform costs.
Approach
Builds online feature-space region partitions via an attention module trained to match feature similarity and cost similarity, then increases mixup ratios for sample pairs with high attention scores; theory shows low within-region risk variance identifies high-cost tail regions.
Key results
- On CIFAR100 long-tail setting, AugARP-MFW reaches 0.5714 weighted accuracy vs 0.4769 for MFW and 0.4443 for ERM.
- On Amazon Products Sales 2023, AugARP-Remix reaches 0.7073 weighted accuracy vs 0.7054 for Remix.
- On iNaturalist, AugARP-Remix reaches 0.5865 weighted accuracy vs 0.4453 for Remix.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add adaptive region partition to MFW | MFW · CIFAR10 long-tail | weighted accuracy: 0.8667 -> 0.8930 | Helped |
| add adaptive region partition to Remix | Remix · CIFAR10 long-tail | weighted accuracy: 0.9013 -> 0.8886 | Hurt |
| add adaptive region partition to MFW | MFW · CIFAR10 step | weighted accuracy: 0.9255 -> 0.9282 | Helped |
| add adaptive region partition to Remix | Remix · CIFAR10 step | weighted accuracy: 0.9488 -> 0.9438 | Hurt |
| add adaptive region partition to MFW | MFW · CIFAR100 long-tail | weighted accuracy: 0.4769 -> 0.5714 | Helped |
| add adaptive region partition to Remix | Remix · CIFAR100 long-tail | weighted accuracy: 0.5682 -> 0.5631 | Hurt |
| add adaptive region partition to MFW | MFW · CIFAR100 step | weighted accuracy: 0.5916 -> 0.6064 | Helped |
| add adaptive region partition to Remix | Remix · CIFAR100 step | weighted accuracy: 0.5047 -> 0.5544 | Helped |
| add adaptive region partition to MFW | MFW · Amazon Products Sales 2023 | weighted accuracy: 0.6168 -> 0.6999 | Helped |
| add adaptive region partition to Remix | Remix · Amazon Products Sales 2023 | weighted accuracy: 0.7054 -> 0.7073 | Helped |
| add adaptive region partition to MFW | MFW · iNaturalist 2018 | weighted accuracy: 0.2707 -> 0.4905 | Helped |
| add adaptive region partition to Remix | Remix · iNaturalist 2018 | weighted accuracy: 0.4453 -> 0.5865 | Helped |
Surprising findings
- Traditional long-tailed methods such as LDAM and Focal loss perform poorly in the cost-aware setting, e.g., LDAM reaches 0.0757 on CIFAR100 long-tail.
- On CIFAR10, AugARP is slightly behind the best baseline (Remix), so the gains are not uniform across datasets.
Limitations
- The attention module used for implicit region partitioning may not be the most effective structure.
- Mixup-based augmentation may not be the best way to exploit region partition information.
- Cost assignment is assumed to be correlated with features; incorrect cost estimates could harm minority classes.
Experimental setup
Datasets: CIFAR10 CIFAR100 iNaturalist 2018 Amazon Products Sales 2023
Benchmarks: CIFAR10 CIFAR100 iNaturalist 2018 Amazon Products Sales 2023
Metrics: weighted accuracy (cost-weighted)
Baselines: ERM LDAM Focal loss LDAM-DRW MFW Remix
Builds on: MFW Remix LDAM-DRW ORIWAL ARBAL
Closest papers by meaning
- Long-Tailed Learning as Multi-Objective Optimization (AAAI 2024 · similarity 0.87)
- Towards Heterogeneous Long-tailed Learning: Benchmarking, Metrics, and Toolbox (NeurIPS 2024 · similarity 0.85)
- Breaking Long-Tailed Learning Bottlenecks: A Controllable Paradigm with Hypernetwork-Generated Diverse Experts (NeurIPS 2024 · similarity 0.85)
- Long-Tailed Partial Label Learning by Head Classifier and Tail Classifier Cooperation (AAAI 2024 · similarity 0.85)
- Feature Fusion from Head to Tail for Long-Tailed Visual Recognition (AAAI 2024 · similarity 0.85)
- A Square Peg in a Square Hole: Meta-Expert for Long-Tailed Semi-Supervised Learning (ICML 2025 · similarity 0.84)
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.