AI Glossary / Causal inference and experimentation / Beginner
Precision
Definition
Precision is the fraction of a model's positive predictions that are actually positive according to the reference labels. It is calculated as true positives divided by true positives plus false positives. If a system flags records for review, precision answers: among the flagged records, how many really meet the target condition? In information retrieval, the corresponding question is how many retrieved items are relevant. This entry concerns evaluation precision, not numerical storage precision.
What it means
Precision focuses on the reliability of selected or flagged results. A high value means relatively few of those positive predictions are false alarms on the evaluated dataset. It does not say how many actual positives the system failed to find.
For example, an assistant may identify a small set of clear refund requests very accurately while overlooking many less explicit requests. Its precision can be high even though its coverage is poor. Recall is the companion measure that asks about that coverage.
The positive class must be defined before the number is meaningful. “Urgent,” “duplicate,” and “relevant” require labeling rules, and disagreements or incomplete labels can affect the measured result.
How it works
Collect predictions and trusted reference labels for an appropriate evaluation set. Count positive predictions that match actual positives as true positives, or TP. Count positive predictions that are actually negative as false positives, or FP.
Compute precision = TP/(TP + FP). True negatives and false negatives do not appear directly in this denominator because the question concerns the selected positives. If there are no positive predictions, the denominator is zero and the mathematical ratio is undefined; software may apply a documented convention.
For score-producing classifiers, precision depends on the chosen decision threshold. It also depends on the evaluation population, including how common the positive condition is. In multiclass or multilabel evaluation, specify whether precision refers to one class or an averaging procedure across classes. Those summaries can answer different questions.
Three practical examples
1. Duplicate-record review: precision measures what fraction of flagged pairs are genuine duplicates, helping estimate how much review effort goes to false matches.
2. Search results: precision measures what fraction of retrieved documents meet the query's relevance criteria, reflecting how much irrelevant material users must inspect.
3. Product-quality alerts: precision measures the proportion of flagged items that actually have the defined defect, alongside recall for defects that were missed.
Think of it this way
Think of collecting shells from a beach with a scoop and then checking the scoop's contents. Precision asks what proportion of the collected objects are the shells you wanted. It does not count the desired shells still on the beach. The analogy is limited because real labels can be uncertain, and a model's selection threshold and operating population affect the result.
Worked example
Suppose a fictional model reviews 200 support messages for refund requests. Reference labeling identifies 50 actual requests. The model flags 40 messages, of which 32 are genuine requests and 8 are not.
Precision is 32/40 = 80%. Recall is 32/50 = 64%, because 18 actual requests were missed. These two percentages describe different denominators: flagged messages versus actual requests.
A stricter operating threshold flags only 20 messages, with 19 genuine requests and 1 false alarm. Precision becomes 19/20 = 95%, while recall becomes 19/50 = 38%. In this invented example, cleaner alerts come with lower coverage. Raising a threshold does not guarantee a perfectly monotonic precision increase on every finite dataset, so measure the actual trade-off instead of assuming it. The appropriate operating point depends on the workflow's cost of false alerts and missed requests.
Illustrative case study
Fictional case: Pinecrest Archives evaluates a document finder that returns only the most obvious matches. Reviewers praise the relevance of the returned documents, producing a high precision score. A second review discovers many useful documents that were never returned.
The team adds recall-oriented evaluation and asks users how much material they need to find for their task. For a quick lookup, a small clean result set may be useful; for a comprehensive review, missing relevant records matters much more. The case illustrates why precision must be interpreted in context, not a real archive's measured results.
Common misconceptions and mistakes
Precision is not overall accuracy. Accuracy counts all correct classifications, including true negatives, while precision concentrates on positive predictions. A system can score highly on one and poorly on the other.
Do not call 100% precision a complete success without examining the number of positive predictions and recall. One correct alert with many missed positives can produce perfect observed precision. A score based on very few alerts also has substantial uncertainty.
Another mistake is assuming a precision measured on an artificially balanced dataset will transfer unchanged to operations. Different prevalence and case mix can change it. Also distinguish precision from the false-positive rate: false-positive rate divides false positives by actual negatives, not by all positive predictions.
Business use
Precision helps teams estimate the usefulness of review queues and the credibility of automated alerts. It can connect model evaluation with workloads: if an illustrative queue contains 100 alerts at 80% precision, about 20 are false positives in that assessed set.
Pair the metric with alert volume, recall, and the consequences of errors. Document the threshold and sample composition, and inspect important groups or input types separately. Business decisions need both the cleanliness of selected cases and an understanding of what the system leaves behind.
People also ask
Can precision exceed 100%?
No, under the standard unweighted count formula. True positives cannot exceed the total number of positive predictions.
What if the model predicts no positives?
The ratio is undefined. Report the situation and any software convention rather than interpreting an assigned value as evidence of useful detection.
Is precision the same as positive predictive value?
In a binary classification confusion-matrix setting, those names refer to the same ratio, subject to the same labeling and population context.
Check your understanding
A model makes 60 positive predictions. Forty-five are correct, and 15 are false alarms. What is precision, and can recall be determined from these counts alone?
Answer: Precision is 45/60 = 75%. Recall cannot be determined without the number of actual positives, including those the model missed.
Related terms
- Recall
- Positive class
- True positive
- False positive
- Confusion matrix
- Classification threshold
- Precision-recall curve
- Accuracy
Learn next
Build two confusion matrices with the same precision but different recall. Explain how both can have equally clean alert lists while finding different proportions of all target cases. Then identify which additional counts a business reviewer should request.
More in Causal inference and experimentation
- Plackett-Luce model - The Plackett - Luce model is a probability model for rankings that assigns each item a positive worth and represents an ordering as successive choices from...
- Policy evaluation - Policy evaluation is the task of estimating how much future reward an agent can expect when it follows a specified policy.
- Post-sample evaluation - Post-sample evaluation assesses a forecasting model using observations from after the period on which it was estimated.
- Precision at k (precision@k) - Precision at k, written precision@k or P@k, is the fraction of the first k items in a ranked result list that are relevant.
- Preferential Bayesian optimization - Preferential Bayesian optimization is a sequential search method that uses comparisons between alternatives to seek a preferred solution.
- Probabilistic Programming - Probabilistic programming is an approach to expressing statistical models as programs that represent uncertainty through probability distributions.
Sources
Google, Classification: Accuracy, recall, precision, and related metrics: https://developers.google.com/machine-learning/crash-course/classification/accuracy-precision-recall
Google, Machine Learning Glossary, precision: https://developers.google.com/machine-learning/glossary#precision
The primary teaching references ground the classification definition. Numerical comparisons, fictional case, and exercises are original; no universal precision target is prescribed.
Previous: Precedence constraints Next: Precision at k (precision@k)