- Block 15 — The classification metrics
- Objective
- define recall rigorously as the fraction of actual positives a rule
recovers, and derive from the one fact that its denominator is supplied by the world its
invariance to prevalence, its monotonicity in the threshold, and its standing as the stable
half of the precision-recall pair; reconcile the four names it carries across machine
learning, medicine, signal detection and information retrieval; exhibit the trivial
construction reaching recall 1.0; define specificity, per-class recall and balanced
accuracy; work a costed example whose optimum has precision 0.27 and is correct; define
recall@k and coverage, and measure a two-stage cascade end to end; compute macro, micro and
weighted recall on multiclass and multilabel data; measure recall under label noise and
establish that it is an upper bound when the negative class hides positives; compute
confidence intervals governed by the positive count; and derive LR+ and LR- and use them to
update pre-test odds.
- Estimated duration
- 85 minutes
- Prerequisites
- chapters 011, 029, 030, 034, 050, 051, 052, 053 and 054
- Associated quizzes
055.1-quiz-definition-and-denominator.md to
055.10-quiz-reporting-recall.md