Hirose et al. Propose Generalized Distribution-Free SSL via Risk Rewrite

Typical semi-supervised learning (SSL) methods rely on distributional assumptions that often fail in real-world data, degrading performance. The PNU learning approach offered a distribution-free alternative but was limited to binary classification and had unclear variance optimality. In a new paper accepted to UAI 2026, Yushi Hirose, Hiroo Irobe, and Takafumi Kanamori propose a generalized framework that constructs unbiased risk estimators via linear combinations of component risks, subsuming PNU learning and extending it naturally to multiclass problems. The authors derive the minimum achievable variance for these estimators, proving they can attain lower variance than PNU in asymmetric loss scenarios, and establish a generalization bound that directly ties this variance reduction to improved learning performance.

Based on these theoretical insights, the team introduces two practical SSL methods. Empirical results show that these methods match or outperform existing SSL approaches on both binary and multiclass benchmarks, demonstrating the real-world applicability of the variance improvements. This work removes a key limitation of distribution-free SSL, offering a theoretically grounded, scalable solution that does not hinge on distributional assumptions—a significant step for robust machine learning in diverse, unlabeled data environments.