What is recall?

Recall is a simple score that tells you how many of the items you were looking for a system actually found. It’s the proportion of relevant results that were correctly identified out of all the relevant items that exist.

Let's break it down

Imagine you have three groups: true positives (the items the system got right), false negatives (the relevant items it missed), and false positives (the items it got wrong but counted as relevant). Recall only cares about the first two. The formula is: Recall = True Positives ÷ (True Positives + False Negatives). In other words, it’s “found relevant items” divided by “all relevant items”.

Why does it matter?

Recall matters when missing a relevant item is a big problem. If a medical test fails to detect a disease, the consequences are serious, so you want a high recall. It helps you understand how complete your system’s results are.

Where is it used?

  • Medical diagnosis tools (catching as many true cases as possible)
  • Spam filters (making sure most spam gets caught)
  • Search engines (returning all pages that match a query)
  • Fraud detection (identifying as many fraudulent transactions as possible)
  • Recommendation systems (showing all items a user might like)

Good things about it

  • Easy to understand: a single percentage tells you how many relevant items you captured.
  • Focuses on completeness, which is critical in safety‑critical or health‑related applications.
  • Works well with imbalanced data where the “positive” class is rare; it highlights how well the minority class is being detected.
  • Can be combined with precision to give a fuller picture (e.g., F1 score).

Not-so-good things

  • Ignores false positives, so a system could have perfect recall but also return a lot of junk.
  • Can be misleading if used alone; high recall doesn’t guarantee useful results.
  • Often requires a trade‑off with precision-improving recall may lower precision.
  • In some contexts (like search ranking), users care more about the relevance of the top results than about catching every possible match.