Tune YES / MAYBE thresholds from labelled data

Chooses cutoffs for np_tier() on a validation set. yes is the lowest score whose auto-accept precision still meets target_precision (so YES stays trustworthy while auto-deciding as many records as possible); maybe is the lowest score that still captures min_recall of true matches (anything below is safe to drop to NO). yes is never below maybe.

Usage

np_tune_thresholds(score, label, target_precision = 0.98, min_recall = 0.95)

Arguments

score, label

Scores and truth on the validation set.

target_precision

Minimum precision required of the YES tier.

min_recall

Fraction of true matches the MAYBE floor must retain.

Value

Named numeric c(yes=, maybe=), with achieved metrics attached as attribute metrics.