The original title is "Generalized Distribution-Free Semi-Supervised Learning with Risk Rewrite"

Short summary

This paper generalizes PNU learning — a distribution-free semi-supervised method — from binary to multiclass classification by constructing unbiased risk estimators from linear combinations of component risks. The authors derive minimum achievable variance, showing their estimator can outperform PNU in asymmetric loss scenarios, and prove a generalization bound linking variance reduction to better learning. Two practical SSL methods are introduced that empirically match or beat existing approaches on benchmarks.

• •Generalizes distribution-free PNU learning from binary to multiclass classification
• •Derives minimum variance estimators with provable generalization bounds
• •Two practical methods match or outperform existing SSL approaches on benchmarks

Generated with AI, which can make mistakes.