Research map / Label-efficient and robust learning / Recognition and long-tail
Evidential Mixture Machines: Deciphering Multi-Label Correlations for Active Learning Sensitivity
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 Evidential Mixture Machines (EMM), which predicts mixture-of-Bernoulli label-cluster weights with an evidential regression head and combines evidential and label-covariance uncertainty for multi-label active learning.
Problem
Multi-label active learning suffers from large, sparse label spaces and rare labels, and existing mixture-of-Bernoulli approaches (CBM, GP-B2M) either separate label correlations from features or give poor uncertainty estimates for sample selection.
Approach
Combine a mixture-of-Bernoulli label-cluster model with a deep evidential regression head that predicts component weights, plus a decoder that predicts proxy pseudo-count corrections to the Beta parameters of global clusters, and a multi-source uncertainty (MSU) score mixing evidential epistemic uncertainty, cluster-difference, and label-covariance terms for active sampling.
Key results
- EMM outperforms GP-B2M, MMC, Adaptive, CVIRS in micro-AUC active learning curves across Delicious, Corel5k, BibTex and NUS-WIDE.
- Ablation: replacing the evidential weight predictor with ridge regression (EMM-rr) or dropping proxy pseudo-count updates (EMM-fixed) reduces AUC; single-source uncertainty underperforms the MSU combination.
- Average precision improvement over GP-B2M is largest on the rarest labels (shown for the 50 rarest labels per dataset).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use EMM with MSU active sampling instead of the compared multi-label active-learning methods | EMM · Delicious, Corel 5k, BibTex, NUS-WIDE | micro-AUC | Helped |
| Use MSU sampling instead of entropy-based sampling | EMM · Real-world datasets | micro-AUC | Helped |
| Use evidential weight-coefficient regression instead of Ridge Regression (ablation) | EMM · Real-world datasets | micro-AUC | Helped |
| Add proxy pseudo-count updates and joint training instead of using fixed Bernoulli mixtures (ablation) | EMM · Real-world datasets | micro-AUC | Helped |
| Combine multiple uncertainty sources rather than use only coefficient epistemic uncertainty (ablation) | EMM MSU sampling · Real-world datasets | micro-AUC | Helped |
| Increase the balancing parameters for proxy-count and label-prediction uncertainty too much (ablation) | EMM MSU sampling · Real-world datasets | micro-AUC | Hurt |
| Increase the number of clusters from K = 3 to K = 10 (ablation) | EMM · Corel 5k and BibTex | micro-AUC | Hurt |
| Use an extremely low number of mixture clusters (ablation) | EMM · Corel 5k and BibTex | micro-AUC | Hurt |
Surprising findings
- Increasing the balancing parameters for the auxiliary uncertainty terms too much degrades sampling, so epistemic evidential uncertainty remains the dominant source.
- Larger numbers of clusters (K=10) performed worse than K=3 in one AL setting, which the authors attribute to overfitting.
Limitations
- Model complexity and compute may limit scaling to very large datasets or real-time use.
- Number of mixture components K must be chosen manually; no Bayesian selection of K.
- Macro-AUC is weaker on some datasets (e.g., NUS-WIDE 0.64) and comparisons to AUC-optimization methods are limited.
Experimental setup
Datasets: Delicious Bibtex Corel5k NUS-WIDE synthetic Gaussian-cluster dataset
Benchmarks: Delicious Bibtex Corel5k NUS-WIDE
Metrics: micro-AUC macro-AUC average precision improvement
Baselines: GP-B2M MMC Adaptive CVIRS EMM-entropy AUDI CS
Builds on: deep evidential regression mixture of Bernoulli GP-B2M conditional Bernoulli mixtures
Closest papers by meaning
- Uncertainty Herding: One Active Learning Method for All Label Budgets (ICLR 2025 · similarity 0.85)
- Batch Selection for Multi-Label Classification Guided by Uncertainty and Dynamic Label Correlations (AAAI 2025 · similarity 0.84)
- Learning from Label Proportions: Bootstrapping Supervised Learners via Belief Propagation (ICLR 2024 · similarity 0.84)
- Active Statistical Inference (ICML 2024 · similarity 0.83)
- Rethinking Learning from Label Proportions via Moment Matching (NeurIPS 2026 · similarity 0.83)
- Weakly-Supervised Residual Evidential Learning for Multi-Instance Uncertainty Estimation (ICML 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.