Recall

Recall answers one question: of all the positive cases that truly exist, how many did the model catch?

Imagine sweeping a room for lost coins. Ten coins are hidden in the room, and you find seven. Your recall is 70%, no matter how many bottle caps you also picked up by mistake. Recall only compares what you found against what was there to find.

The formula: recall = true positives ÷ (true positives + false negatives). A false negative is a real positive case the model missed.

When recall is the metric that matters#

Care about recall when missing a case is expensive. A cancer screening test with low recall sends sick patients home. A fire alarm with low recall stays silent during a real fire. Here a false alarm is annoying, but a miss is a disaster.

Recall is easy to cheat. A model that flags everything as positive scores 100% recall and is useless. That is why recall is always read next to precision, and why the f1-score folds the two into one number.

Where to go next#

• Full lesson: Model evaluation
• Related terms: precision, f1-score, confusion-matrix, class-imbalance