Research map / Label-efficient and robust learning / Recognition and long-tail
Batch Selection for Multi-Label Classification Guided by Uncertainty and Dynamic Label Correlations
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 an uncertainty-based mini-batch selection method for multi-label classification that combines per-label entropy with sliding-window prediction fluctuation and dynamic mutual-information label correlations, improving Macro-AUC and convergence over random, Balance, Active Bias, Recency Bias and Hard Imbalance on MPVAE, CLIF and DELA.
Problem
Mini-batch selection for training deep multi-label classifiers: existing uncertainty-based batch selection methods are designed for single-label data, ignore fluctuations within sliding windows and the current confidence, and do not exploit dynamic uncertainty-based label correlations.
Approach
Measure per-label uncertainty as a mix of the current entropy and the mean absolute difference between successive predictions in a sliding window, then weight each sample's uncertainty by a mutual-information label-correlation matrix computed over per-label uncertainties each epoch, and sample mini-batches with exponentially decaying selection pressure.
Key results
- On Macro-AUC across 13 datasets and three base models, the proposed method ranks first on average (average rank 1.23 for MPVAE, 1.46 for CLIF, 1.38 for DELA), with Wilcoxon signed-rank p-values below 0.05 in most comparisons.
- On yahoo-Arts with CLIF, the method reaches Random's best AUC at epoch 32 versus epoch 85 for Random, taking 1280 s versus 1560 s.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use uncertainty-based batch selection with current and sliding-window uncertainty plus dynamic label correlations | CLIF · Macro-AUC across datasets | Macro-AUC | Mixed |
| use uncertainty-based batch selection with current and sliding-window uncertainty plus dynamic label correlations | CLIF · yahoo-Arts with CLIF | time to reach the same AUC: 1560 seconds at epoch 85 -> 1280 seconds at epoch 32 | Helped |
| add current-prediction uncertainty and sliding-window fluctuations to the full method (ablation) | CLIF on bibtex · bibtex | Macro-AUC | Helped |
| add sliding-window uncertainty to current-prediction-only uncertainty (ablation) | CLIF on bibtex · bibtex | Macro-AUC | Helped |
| add current-prediction uncertainty to sliding-window-only uncertainty (ablation) | CLIF on bibtex · bibtex | Macro-AUC | Helped |
| add uncertainty-based label-correlation matrix C (ablation) | CLIF on bibtex · bibtex | Macro-AUC | Helped |
| add label-correlation matrix C to Recent (ablation) | CLIF on bibtex · bibtex | Macro-AUC | Helped |
| add label-correlation matrix C to Active (ablation) | CLIF on bibtex · bibtex | Macro-AUC | Helped |
| rely on current-prediction uncertainty alone (ablation) | CLIF on bibtex · bibtex | Macro-AUC | Hurt |
Surprising findings
- Recent batch selection underperforms on datasets with many labels, and Active/Recent summed-label uncertainty can even fall below the random baseline on some datasets.
Limitations
- Relies on entropy-based uncertainty, which the authors note may need temperature scaling for calibration.
- Hyperparameters (lambda_1, selection pressure s_t, window T) selected by grid search, with sensitivity only on two datasets.
- Batch selection methods are compared only with three base multi-label deep models and single-run results per fold.
Experimental setup
Datasets: scene yeast Corel5k rcv1subset1 rcv1subset2 rcv1subset3 yahoo-Arts yahoo-Business bibtex tmc2007 enron cal500
Benchmarks: scene yeast Corel5k rcv1subset1 rcv1subset2 rcv1subset3 yahoo-Arts yahoo-Business bibtex tmc2007 enron cal500
Metrics: Macro-AUC Ranking Loss Hamming Loss
Baselines: Random Balance Active Bias Recency Bias Hard Imbalance
Builds on: Active Bias Recency Bias Balance Hard Imbalance
Closest papers by meaning
- Noise-Robust Learning from Multiple Unsupervised Sources of Inferred Labels (AAAI 2022 · similarity 0.88)
- A no-regret generalization of hierarchical softmax to extreme multi-label classification (NeurIPS 2018 · similarity 0.87)
- Addressing Multi-Label Learning with Partial Labels: From Sample Selection to Label Selection (AAAI 2025 · similarity 0.87)
- Enhancing Multi-Label Classification via Dynamic Label-Order Learning (AAAI 2024 · similarity 0.86)
- Neural Collapse in Multi-label Learning with Pick-all-label Loss (ICML 2024 · similarity 0.85)
- Theory-Inspired Deep Multi-View Multi-Label Learning with Incomplete Views and Noisy Labels (CVPR 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.