Precision answers how much a team can trust a flag. Recall answers what share of the actual positive cases the model found. In confusion matrix terms, precision is true positives divided by everything the model flagged, and recall is true positives divided by every actual positive case. Google’s Machine Learning Crash Course explains the two with the same formulas.
CUBIG’s UCI Iranian Churn experiment (October 2026) shows the two moving in opposite directions. For logistic regression on the same 630 evaluation customers, recall rose from 0.4242 to 0.7879 after data preparation, and precision fell from 0.8400 to 0.6555. The model found 78 of 99 churners instead of 42, and flagged 41 non-churners instead of 8.
Which one to prioritize:
- recall, when a missed case is expensive, such as an undetected failure or a churned customer nobody called
- precision, when every flag costs review time and the team can handle only a fixed number
- both, through a combined measure such as the F1 score, when neither error clearly dominates
How are precision and recall different from accuracy?
- Accuracy counts all correct predictions, including correct negatives, so it can look strong when the positive class is rare.
- Precision and recall focus on the positive class and separate false alarms from misses.
- The F1 score combines them into one number but loses the information about which one moved.
Where do precision and recall break in AI work?
Both depend on the decision threshold. Lowering the threshold usually raises recall and lowers precision, so a comparison between two models is fair only at the same threshold on the same evaluation records. A team that reports recall from one run and precision from another can make a model look better than it is.
Both also depend on the labels. If the actual outcomes in the evaluation data are wrong, as described under label noise, precision and recall measure agreement with those wrong labels. CUBIG’s Ready for Model experiments report recall, precision and F1 together for the before and after versions under the same evaluation conditions, so a gain on one metric is read alongside the cost on the other.