recall
/ ree-KAWL /
Recall asks the mirror-image question to precision: of all the things that were truly yes, how many did the model actually catch? If 20 emails in your inbox are genuinely spam and the filter catches 16 of them, its recall is 16/20 = 80 percent. The four it missed slipped through into your inbox. Recall measures the misses, not the false alarms.
Formally, recall is true positives divided by all the actual positives — true positives plus false negatives. It looks at every case that should have been flagged and asks what fraction the model found. It is also called sensitivity or the true positive rate, especially in medicine.
Recall matters most when missing a true case is dangerous. A cancer screen with high recall rarely tells a sick patient they're fine; a fraud detector with high recall lets little fraud slip by. The catch is that you can hit perfect recall trivially — just flag everything as positive and you'll catch every real case, at the price of drowning in false alarms. That is exactly why recall is meaningless without precision beside it: the two together describe the whole trade-off.
A disease affects 50 patients in a study. The screening test flags 40 of them as positive and misses 10. Recall = 40/50 = 80%. One in five sick patients was sent home reassured but still ill — the price of imperfect recall.
Recall = of the truly positive cases, the fraction the model caught.
Watch the language: "sensitivity" (medicine) and "true positive rate" (signal detection, ROC curves) are the same thing as recall. Don't be fooled by a model boasting 99% recall — it may be flagging nearly everything, which destroys precision.