Precision

Precision Evaluation Metric Retrieval Needs-Review

What it is

Precision is the fraction of retrieved documents that are relevant to a query. It measures the purity of retrieval results: precision = (# relevant retrieved) / (# total retrieved). High precision indicates few false positives (irrelevant results). Precision is particularly important in user-facing IR systems where showing irrelevant results frustrates users.

[illustrate: Retrieved set with relevant (green) and irrelevant (red) results; precision calculation]

How it works

Formula:

Precision = (# relevant retrieved) / (# retrieved)

At rank k:

P@k = (# relevant in top-k) / k

Properties:

  • Range: 0–1 (or 0–100%)
  • High precision: few irrelevant results
  • Trade-off: optimizing for precision often reduces recall
  • Emphasis: quality over quantity of results

Example

Query: "best machine learning libraries"
Retrieved:
1. Scikit-learn tutorial (relevant)
2. Restaurant menu (irrelevant)
3. TensorFlow guide (relevant)
4. Sports news (irrelevant)
5. PyTorch documentation (relevant)

P = 3/5 = 0.6
P@3 = 2/3 = 0.67
P@1 = 1/1 = 1.0

Variants and history

Precision and recall are classical IR metrics (1950s–60s). Precision@k focuses on top results (practical for users who see only top-k). Mean Average Precision (MAP) averages precision across recall levels. Average Precision integrates precision-recall curve. Modern metrics (NDCG, MRR) incorporate ranking quality. Precision alone incomplete; combined with recall for holistic evaluation.

When to use it

Use precision when:

  • User frustration with irrelevant results is costly
  • Showing wrong answer worse than showing no answer
  • High-precision retrieval is priority (Q&A, fact-checking)
  • Limited result viewing (users see only top-k)
  • Perfect recall unattainable or unnecessary

Precision is necessary but insufficient; balance with recall using F1 or MAP.

See also