Curtis G. Northcutt, Tailin Wu, Isaac L. Chuang. Learning with Confident Examples: Rank Pruning for Robust Classification with Noisy Labels. 2017.
- Examines binary classification with false positive rate p1 and false negative rate p0
- Model combines previous methods including noise estimation, PN (pos/neg labeled, maybe wrongly), PU (some pos labelled, rest unlabled), any prob classifier, prob estim robustness, and added noise
- Rank pruning estimates noise rates and then uncovers true classifications
- Let subset of X and potentially false labels s have thresholds to predict chich pos and neg entries in s are incorrect.
- Use those to estimate noise rates.
- Using noise rates, prune set for unlikely pos and neg entries
- Retrain on more confident subset.
- How does this extend to multi-dimensional?
- Correlated noise?