Recall Machine Learning
Recall answers the question precision doesn't: of everything that was actually positive, how many did the model actually catch? It's computed as true positives divided by true positives plus false negatives, and in medical statistics the same quantity goes by the name sensitivity — a mammography screening study enrolling 100 patients who truly have cancer, where the test correctly flags 92 of them and misses 8, reports a recall (sensitivity) of 92 divided by 100, or 0.92. Like precision, the term traces back to Cyril Cleverdon's 1960s Cranfield experiments in information retrieval, where it measured the fraction of all relevant documents a search actually returned, long before it became standard vocabulary for evaluating any classifier.
Recall and precision start from opposite ends of the confusion matrix, and conflating them is the single most common mistake in reading a classifier's reported performance. Precision starts from the set the model flagged and asks how many of those were right; recall starts from the set that was actually positive and asks how many got flagged — a model that flags every single case as positive scores a perfect 1.0 recall by definition, since it's impossible to miss a true positive if nothing is ever predicted negative, while its precision would likely collapse under the flood of false alarms that same strategy produces. The two numbers measure different failure modes, laid out together in confusion-matrix, and a single number from either one, reported alone, is close to meaningless without its counterpart.
Recall is the metric that governs which mistake a system is built to tolerate, and the domains that demand near-perfect recall share one trait: missing a true case is catastrophic, and a false alarm is cheap to absorb. Cancer screening programs are designed around high recall deliberately, accepting that some healthy patients will get flagged for follow-up testing they didn't need, because the alternative — a missed cancer case — is immeasurably worse than an unnecessary biopsy; airport security screening and structural-defect inspection in manufacturing follow the identical logic, favoring more false alarms over any single missed threat or flaw. This is the mirror image of the spam-filter example that drives precision's design instead, where a false positive (deleting a real email) is the costly mistake and recall can afford to suffer in comparison.
Lowering the decision threshold is the standard lever for buying recall, and it costs precision in direct, predictable proportion. Accepting a positive prediction at 30% confidence instead of 50% catches more of the genuinely positive cases that were sitting just below the old threshold, which raises recall, but it also lets through more borderline cases that turn out to be false positives, lowering precision — the exact same threshold-moving mechanism covered under classifier-precision, just read from the opposite direction. Neither metric can be pushed toward 1.0 for free; every gain in one, at a fixed model, is paid for somewhere in the other.
Recall's entire value is in forcing an answer to the question precision can't address at all — not "how trustworthy are the positive predictions" but "how much of the real problem did the model actually find" — and a system tuned only for one of these two numbers while ignoring the other is, almost by construction, optimized for exactly the kind of failure nobody checked for.