Khái niệm cốt lõi
Proposing PUUPL, an uncertainty-aware pseudo-labeling framework, improves performance in imbalanced Positive-Unlabeled Learning scenarios.
Thống kê
Train an ensemble
Predict with uncertainty
Select with uncertainty
Positive-unlabeled
Pseudo-labeled
FPR: 0.1, TPR: 0.2
Trích dẫn
"PUUPL yields substantial performance gains in highly imbalanced settings."
"Uncertainty quantification improves pseudo-label quality and predictive performance."