Learning from Uncertain Similarity and Unlabeled Data

2025/09/15 by Wei, Meng, Zhongnian Li, Peng Ying +4
Computer Science · #Benchmark (surveying) #Convergence (economics) #Data Mining Algorithms and Applications #Estimator #FOS: Computer and information sciences #Machine Learning (cs.LG) #Machine Learning and Data Classification #Parametric statistics #Pattern recognition (psychology) #Semantic Web and Ontologies #Semi-supervised learning #Similarity (geometry) #Supervised learning

paper · pdf · doi:10.48550/arxiv.2509.11984

published in arXiv (Cornell University) (Cornell University)

openalex publication_date 2025/09/15 · openalex created_date 2025/10/12 · openalex updated_date 2026/08/05

Abstract

Existing similarity-based weakly supervised learning approaches often rely on precise similarity annotations between data pairs, which may inadvertently expose sensitive label information and raise privacy risks. To mitigate this issue, we propose Uncertain Similarity and Unlabeled Learning (USimUL), a novel framework where each similarity pair is embedded with an uncertainty component to reduce label leakage. In this paper, we propose an unbiased risk estimator that learns from uncertain similarity and unlabeled data. Additionally, we theoretically prove that the estimator achieves statistically optimal parametric convergence rates. Extensive experiments on both benchmark and real-world datasets show that our method achieves superior classification performance compared to conventional similarity-based approaches.

Citations

• A Universal Unbiased Method for Classification from Aggregate Observations
• Unlocking the Power of Open Set : A New Perspective for Open-Set Noisy Label Learning
• Barely-Supervised Learning: Semi-Supervised Learning with very few labeled images
• Learning from Similarity-Confidence Data
• Pointwise Binary Classification with Pairwise Confidence Comparisons
• Unbiased Risk Estimators Can Mislead: A Case Study of Learning with Complementary Labels
• Progressive Identification of True Labels for Partial-Label Learning
• Mitigating Overfitting in Supervised Classification from Two Unlabeled Datasets: A Consistent Risk Correction Approach
• Symmetric Cross Entropy for Robust Learning with Noisy Labels
• Generative-Discriminative Complementary Learning
• On Symmetric Losses for Learning from Corrupted Labels
• Complementary-Label Learning for Arbitrary Losses and Models
• On the Minimal Supervision for Training Any Binary Classifier from Only Unlabeled Data
• Classification from Pairwise Similarity and Unlabeled Data
• Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms
• Virtual Adversarial Training: A Regularization Method for Supervised and Semi-Supervised Learning
• Positive-Unlabeled Learning with Non-Negative Risk Estimator
• Learning with Multiple Complementary Labels
• Gradient-based learning applied to document recognition

Related

• Learning from Similarity-Confidence Data
• Classification from Pairwise Similarity and Unlabeled Data
• Complementing Semi-Supervised Learning with Uncertainty Quantification
• Classification from Pairwise Similarities/Dissimilarities and Unlabeled Data via Empirical Risk Minimization
• Learning from Similarity-Confidence and Confidence-Difference