Research map / Label-efficient and robust learning / Recognition and long-tail
AUC Optimization from Multiple Unlabeled Datasets
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 U^m-AUC, an AUC optimization method for learning from multiple unlabeled datasets that only needs the relative order of class priors, reducing the problem to m-1 multi-label sub-problems trained with a min-max square-loss objective.
Problem
Learning an AUC-optimal classifier from m unlabeled datasets with different class priors, where only the relative ordering (not exact values) of the priors is known.
Approach
Converts the pairwise U^m AUC risk into a multi-label AUC problem with m-1 labels, and rewrites the square-loss pairwise objective as a min-max problem solvable in O(n) per epoch via stochastic primal-dual updates, requiring only the ordering of class priors.
Key results
- On K-MNIST with m=10 under the uniform prior setting, U^m-AUC reaches 0.938 test AUC versus 0.911 for U^m-SSC (true-prior baseline U^m-SSC is the closest competitor).
- On CIFAR-10 with m=50 under the uniform prior setting, U^m-AUC reaches 0.889 test AUC versus 0.857 for LLP-VAT.
- Across all 24 dataset/m/prior configurations in the main table, U^m-AUC attains the best mean test AUC, and on imbalance tests it degrades only slowly as the size-reduction ratio decreases.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use U^m-AUC instead of the strongest reported baseline | U^m classification · K-MNIST, m=10 and m=50, four prior distributions | test AUC | Helped |
| Use U^m-AUC instead of the strongest reported baseline | U^m classification · CIFAR-10, m=10 and m=50, four prior distributions | test AUC | Helped |
| Use U^m-AUC instead of the strongest reported baseline | U^m classification · CIFAR-100, m=10 and m=50, four prior distributions | test AUC | Helped |
| Reduce the size of half the datasets using the stated reduction ratio | U^m-AUC · K-MNIST, CIFAR-10, and CIFAR-100; m=10 and m=50 | test AUC | Mixed |
Surprising findings
- The method outperforms baselines that are given exact class priors even though it only uses the relative order of priors.
Limitations
- Requires at least two unlabeled sets with different priors and knowledge of their relative ordering.
- Experiments use synthetic bags built from benchmark datasets with binarized labels and a small MLP/ResNet-32 model, not real-world U^m data.
Experimental setup
Datasets: Kuzushiji-MNIST CIFAR-10 CIFAR-100
Benchmarks: Kuzushiji-MNIST CIFAR-10 CIFAR-100
Metrics: test AUC
Baselines: LLP-VAT LLP-VAT* U^m-SSC U^m-SSC*
Builds on: LLP-VAT U^m-SSC deep AUC maximization
Closest papers by meaning
- Semi-Supervised AUC Optimization Without Guessing Labels of Unlabeled Data (AAAI 2018 · similarity 0.91)
- Consistent Multi-Class Classification from Multiple Unlabeled Datasets (ICLR 2024 · similarity 0.88)
- Towards Decision-Friendly AUC: Learning Multi-Classifier with AUCµ (AAAI 2023 · similarity 0.88)
- Does It Pay to Optimize AUC? (AAAI 2023 · similarity 0.87)
- Classification from Pairwise Similarity and Unlabeled Data (ICML 2018 · similarity 0.87)
- Positive-unlabeled AUC Maximization under Covariate Shift (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.