Suppose a fraud detector catches 80 of 100 fraudulent transactions. Its recall is 80%: of the cases that really were positive, it identified 80. Recall is also called the true positive rate (TPR). Calculate it as true positives divided by true positives plus false negatives. The missed frauds matter because recall counts what the model failed to catch, not just what it flagged. section 5.2
Same recall, very different workload
Two detectors each catch 80 of 100 fraudulent transactions among 100,000 genuine ones, so both have 80% recall. One also flags 20 genuine transactions; the other flags 920. Recall alone cannot distinguish their false-alarm burden. Check precision and false positive rate alongside it, especially when positive cases are rare. Changing the number of fraud cases while keeping each detector's recall at 80% also changes precision. section 5.2
Choose recall in light of the decision it supports. A high value says that few actual positives were missed; it does not establish that positive predictions are trustworthy or that the system will meet its business goal. The metric used to assess a model may differ from its training loss, and choosing an unsuitable metric can steer development toward poor real-world results. section 5.2
Recall within a search shortlist
In a photo-search experiment, a retrospective check found that 97% of all clicked photos appeared in the new model's top 1,000 results. This Recall@1000 finding suggested a possible next step: a two-stage pipeline with a retrieval stage. It was an observation about the shortlist, not a substitute for the experiment's user and business metrics. section 12.4.2
Do not confuse classifier recall with the sensitivity of a statistical test. In A/B-test simulations, sensitivity means the estimated probability that a test detects a real difference; it is assessed by repeatedly simulating an uplift and checking how often the null hypothesis is rejected. section 12.2.5
