Reproduce “Uncertainty-aware Pseudo-label Selection for Positive-Unlabeled Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Emilio Dorigatti, Jann Goschenhofer, Benjamin Schubert, Mina Rezaei, Bernd Bischl
PUUPL introduces an iterative uncertainty-aware pseudo-labeling framework for Positive-Unlabeled (PU) learning, improving binary classification performance, especially in imbalanced datasets. The method leverages deep ensembles for uncertainty quantification, enhancing model calibration and outperforming baselines on diverse datasets, including a proteasomal cleavage prediction task with an AUROC of 87.20%.
Key insights
- Explicitly quantifying and utilizing epistemic uncertainty is crucial for effectively selecting and refining pseudo-labels, mitigating confirmation bias, and improving model calibration in PUL.
- An iterative pseudo-labeling process, coupled with uncertainty-aware selection, can effectively expand the training signal, particularly for minority positive classes in imbalanced PU learning scenarios.
- The framework's modality-agnostic design and compatibility with various PU loss functions enable its broad applicability across diverse data types and enhancement of existing PUL methods.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.