Precision vs Recall: What's the Difference?

Precision and recall are two evaluation metrics for classification models. Precision answers 'of everything the model flagged as positive, how much was actually positive?' while recall answers 'of all the real positives, how many did the model catch?'. They trade off against each other, so which one matters more depends on whether false positives or false negatives are more costly in your problem.

The formulas

Both metrics are built from the counts in a confusion matrix. Precision equals true positives divided by (true positives plus false positives), so it drops when the model raises false alarms. Recall, also called sensitivity or the true positive rate, equals true positives divided by (true positives plus false negatives), so it drops when the model misses real positives. Each ranges from 0 to 1, where higher is better.

The precision-recall trade-off

Most classifiers output a probability, and you choose a threshold above which a prediction counts as positive. Raising the threshold makes the model more conservative, which usually increases precision but lowers recall; lowering it does the opposite. Because of this, you rarely maximize both at once, and the precision-recall curve shows the achievable combinations across all thresholds.

When to prioritize precision vs recall

Pick the metric that matches the cost of each error type.

• Favor recall when missing a positive is dangerous or expensive: disease screening, fraud detection, or safety alerts, where a false negative is the worst outcome.
• Favor precision when a false alarm is costly or erodes trust: spam filtering, or flagging content for human review, where too many false positives waste effort.
• Favor a balance (F1) when both errors matter and the classes are imbalanced.

F1 score and combining the two

The F1 score is the harmonic mean of precision and recall, equal to 2 times precision times recall divided by (precision plus recall). It rewards models that keep both reasonably high and punishes those that sacrifice one for the other, which makes it a common single-number summary for imbalanced problems. If one error type is genuinely more important, a weighted F-beta score lets you tilt toward precision or recall.

Why accuracy can mislead you

On imbalanced data, accuracy is deceptive. If 99 percent of transactions are legitimate, a model that labels everything legitimate is 99 percent accurate while catching zero fraud, giving it a recall of 0. Precision, recall, and F1 expose that failure where raw accuracy hides it, which is why they are the go-to metrics for rare-event problems.

How TensorTurn reports these for you

Every classification run in TensorTurn reports precision, recall, F1, and the full confusion matrix alongside accuracy, so you can judge the model on the metric that matches your risk rather than a single headline number. Because it also runs dataset health checks first, you can trust that those metrics are not being inflated by leakage or duplicate rows.