Research map / Label-efficient and robust learning / Noisy labels
Toward Robustness in Multi-Label Classification: A Data Augmentation Strategy against Imbalance and Noise
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 BalanceMix, a data augmentation method for multi-label image classification that handles imbalanced and noisy labels via confidence-based minority oversampling, Mixup, and label-wise GMM-based label management.
Problem
Multi-label classification suffers from imbalanced label distributions and noisy (mislabeled or partial) labels in training data, which degrade generalization.
Approach
Combines two minority samplers that generate diverse minority-augmented instances via Mixup with label-wise management that models per-label clean/noisy status with GMMs, so each label is treated as a candidate and corrected or kept individually rather than instance-wise.
Key results
- BalanceMix improves mAP over the same TResNet-L backbone by about 0.6 on a compared setting in Table 6 (rebuttal states the gain is 0.3 in one reading).
- Reports significant mAP improvements even at noise ratios up to 40%, including a label flipping attack scenario.
- Claims state-of-the-art performance over five compared methods, all run within the ML-Decoder framework.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use BalanceMix under label-flipping attack and exceptionally high label-noise ratios | Multi-label classification · Label-flipping attack; label-noise ratios up to 40% | mAP | Helped |
| Use BalanceMix instead of the compared methods | ML-Decoder framework · MS-COCO testing data | mAP | Helped |
Limitations
- Label correction assumes clean labels have lower loss than noisy ones, which systematic noise can violate.
- Method relies on BCE loss and is designed for image domain via image mixing.
- Only one real noisy multi-label dataset (DeepFashion) is used; systematic noise is not evaluated.
Experimental setup
Datasets: PASCAL-VOC MS-COCO DeepFashion
Benchmarks: PASCAL-VOC MS-COCO DeepFashion
Metrics: mAP
Baselines: Co-teaching T-estimator ML-Decoder
Builds on: Mixup ML-Decoder Balanced-mixup
Closest papers by meaning
- Learning Imbalanced Data with Beneficial Label Noise (ICML 2025 · similarity 0.89)
- Rebalancing Multi-Label Class-Incremental Learning (AAAI 2025 · similarity 0.88)
- Addressing Multi-Label Learning with Partial Labels: From Sample Selection to Label Selection (AAAI 2025 · similarity 0.86)
- ConR: Contrastive Regularizer for Deep Imbalanced Regression (ICLR 2024 · similarity 0.86)
- Taming the Sigmoid Bottleneck: Provably Argmaxable Sparse Multi-Label Classification (AAAI 2024 · similarity 0.86)
- Revisiting Interpolation for Noisy Label Correction (AAAI 2025 · similarity 0.85)
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.