Precision, Recall & the Confusion Matrix
A disease test that says healthy to everyone scores 90% accuracy if only 10% of patients are sick — and catches nobody. Whenever classes are imbalanced, accuracy lies, so classifiers are judged on the confusion matrix instead: the four-way count of caught cases, missed cases, false alarms, and correct all-clears. From it come the two numbers that matter: precision (of everything flagged, how much was real) and recall (of everything real, how much was caught). They pull against each other, and the threshold slider is the lever — where to set it is a question about costs, not math.
Drag the threshold. Watch the four confusion-matrix cells trade against each other, and find the settings where accuracy stays impressively high while recall — the number that actually matters here — collapses.
Screening test: 100 patients, only 10 actually sick.
The threshold trades precision against recall. Which way to lean is not a math question: missing a cancer costs more than a false alarm, so screening leans toward recall.
Check yourself
For a cancer screening test, would you rather have 90% precision with 60% recall, or 60% precision with 90% recall — and what does each error cost?
Go deeper (free): Google ML Crash Course — Classification metrics ↗