TracksStudyAbout
Sign in
Start studying
Module 1 · Machine Learning Module contents
Module 1  ·  Card 45 / 55beginner

Recall

Recall asks: of all the real positives, how many did the model actually catch?

It is true positives divided by all actual positives.

High recall means few misses.

It is the metric you care about when missing a case is dangerous.

Failing to detect a disease is exactly that kind of costly mistake.

Back Next
Learn in depth ↓
Module 01 · Machine Learning · 55 concepts
In depth  /  15 lines

Recall, in more detail

Recall answers the opposite question: of all the real cases, how many did we catch?

It ignores false alarms entirely.

So it measures the cost of missing something.

It is the metric you want when a miss is dangerous.

An undetected tumour, a fraud that goes through, an alarm that stays silent.

You can push recall to a perfect score by flagging absolutely everything.

The model then cries wolf constantly, and recall still looks perfect.

That is why recall on its own is never enough either.

It has to be read next to precision, always.

Lowering the threshold lifts recall and lowers precision, every time.

In screening problems high recall usually wins, with a human checking afterwards.

The first stage catches nearly everything and the second stage removes the noise.

Recall is sometimes called sensitivity, or the true positive rate.

Those are the same number wearing different names in different fields.

Report the threshold you used, or the number cannot be compared with anything.

Want the full technical article? Wikipedia ↗
Deep dive

What is underneath Recall

The formula

Recall=TPTP+FN\mathrm{Recall} = \frac{TP}{TP + FN}

The denominator is everything that actually was positive, so recall asks: of the things that were there, how many did it find?

  • TPTP: present and found.
  • FNFN: present and missed.

Note what is missing: FPFP does not appear. Recall knows nothing about how many false alarms it raised getting there.

It goes by other names in other fields: sensitivity in medicine, true positive rate in signal detection. Same formula.

What it costs to get it wrong

Recall is the metric that matters when missing a case is expensive.

A cancer screen with low recall sends sick people home. A retrieval system with low recall never shows the document that had the answer.

In both, the cost is invisible at the time. Nobody complains about the result they were never shown, which is what makes low recall so much easier to ship than low precision.

Trivially maximised, in the opposite direction

Recall=PP+0=1.0\mathrm{Recall} = \frac{P}{P + 0} = 1.0

Predict positive for everything. Nothing is missed, so recall is perfect and precision is the base rate.

That is the mirror image of the precision trap, and it is why the two are always reported together.

The trade, made explicit

  1. Lower the threshold so more things are flagged.
  2. More real positives make it through, so recall rises.
  3. More negatives make it through too, so precision falls.
  4. Somewhere along that curve is the operating point you actually want.
  5. Which point that is depends on the cost of a miss against the cost of a false alarm, and the data cannot answer that.
Test yourself

One question on this card. Every option explains itself once you pick, right or wrong.

Question 1
You must never miss a case of a serious disease. Which metric matters most?
Concepts these five lines lean on
Confusion MatrixPrecision
Machine Learning

55 concepts in this module, then a graded test with every answer explained.

Study this module →
↑↓ move↵ openesc close