Does interpolation change the returned ranking?

No. It transforms the evaluation summary while leaving the ranked results untouched.

Can interpolated precision be lower than an observed precision at the same recall?

Under the stated maximum rule, no. That observed point is itself among the eligible candidates.

Is every average-precision metric based on this curve?

No. Definitions differ, and common retrieval average precision uses precision at relevant-document ranks without this interpolation. State the exact convention.

AI Glossary / Causal inference and experimentation / Intermediate

Interpolated precision

Definition

Interpolated precision is a summary of a precision - recall curve that assigns each recall level the highest observed precision obtainable at that recall or any higher recall. It forms a non-increasing upper envelope over the observed values. In ranked retrieval, this makes a jagged sequence of measurements easier to compare across queries. The word interpolated here refers to this maximum-based rule, rather than ordinary straight-line interpolation between neighboring points.

What it means

A search engine returns documents in order. After each additional result, precision measures the fraction examined that are relevant, while recall measures the fraction of all relevant documents that have been found. These quantities do not change smoothly: irrelevant results reduce precision, and later relevant results may increase it again.

Interpolated precision asks what the best observed precision is if at least a specified recall must be achieved. It therefore summarizes an observed trade-off rather than reporting the precision at one fixed rank. The calculation does not reorder documents, remove bad results, or improve the search system. It changes how the evaluation curve is represented.

How it works

Begin with a ranked list and relevance judgments, including the number of relevant documents used as the recall denominator. Calculate precision and recall at the evaluation cutoffs. For a requested recall r, retain points whose recall is at least r and take their maximum precision.

Repeating that operation yields the interpolated curve. As the required recall rises, fewer candidate points remain, so the maximum cannot rise. Evaluators can then report values at common recall levels, such as the traditional eleven points from zero through one.

Document the evaluation convention, especially truncated rankings, unattained recall levels, and aggregation across queries. A named average-precision score may use a different definition, so its implementation should never be inferred from the word precision alone.

Three practical examples

1. Search benchmark: researchers compare retrieval systems at shared recall levels even when relevant documents occur at different ranks for each query.

2. Document discovery study: an analyst examines how much precision can be maintained while retrieving a larger proportion of a judged relevant collection.

3. Metric teaching: students plot raw and interpolated curves for a short ranking to see which fluctuations the maximum-based envelope hides and which trade-offs remain visible.

Think of it this way

Think of choosing the best observed deal among all offers that satisfy a minimum quantity requirement. Raising the minimum quantity removes options, so the best remaining deal cannot improve under a fixed comparison score. The analogy is limited: retrieval precision and recall are linked by a particular ranked list, and an interpolated value does not promise the same experience at the user's chosen stopping point.

Worked example

Assume four documents in a collection are relevant. A ranking places them at positions 1, 3, 4, and 8. At those positions, the recall - precision pairs are (0.25, 1), (0.50, 2/3), (0.75, 3/4), and (1.00, 1/2).

At recall 0.50, interpolated precision is the maximum of 2/3, 3/4, and 1/2: it is 0.75. This differs from precision at rank 3, which is approximately 0.667. Looking one document further reaches a higher recall and a higher precision.

At recall 0.80, only the point reaching full recall qualifies among these relevant-document cutoffs, so interpolated precision is 0.50. At recall 0.25 it is 1.00. The ranking has not changed; the calculations summarize its observed operating points.

Illustrative case study

Fictional case: Archive Lane compares two document-ranking prototypes. A report shows an improved interpolated score, and a manager assumes the first three results must now be more useful. An analyst checks the actual rankings and finds that the improvement comes from relevant documents appearing later in the list.

The team retains the interpolated curve for understanding broader retrieval coverage but adds precision at the short cutoff used in its interface. It also reviews queries whose judgments are incomplete. The revised report distinguishes the metric's trade-off summary from the experience of users who rarely examine further results. This fictional example claims no real product performance.

Common misconceptions and mistakes

Do not connect two curve points with a straight line and assume that produces this definition of interpolated precision. The rule takes a maximum over points meeting the recall requirement.

Do not equate interpolated precision at a recall level with precision at a fixed rank. Those are different questions.

Avoid treating smoothing as new evidence or silently mixing interpolated and uninterpolated average precision. Changes in relevance judgments, recall denominators, or ranking depth can alter the result. A polished curve cannot compensate for a poorly constructed test collection.

Business use

Interpolated precision is useful when a business needs to understand retrieval quality over different coverage requirements. An internal research search tool and a short product-search page may favor different parts of the curve.

Decide which recall levels or cutoffs correspond to actual work, and report metrics aligned with those needs. Keep evaluation data and aggregation rules consistent when comparing versions. Where exhaustive relevance judgments are unavailable, explain the limitation in the recall estimate. Use the interpolated curve as one view of performance alongside representative queries and errors that reveal why a result is useful or distracting.

Check your understanding

At recall 0.50, eligible observed precisions are 0.60, 0.72, and 0.45. What is interpolated precision there?

Answer: 0.72, the highest eligible precision. It is not their average and does not require that the winning point occur at exactly 0.50 recall.

Related terms

Precision
Recall
Mean average precision
Precision at k (precision@k)
Ranked retrieval
precision–recall curve
average precision
precision at k
relevance judgment

Learn next

Mark relevant documents in a short ranked list, calculate precision and recall after each result, and construct the maximum-based envelope. Compare interpolated precision at a chosen recall with precision at a chosen rank. Write down which of those questions better matches the intended user's task.

Sources

Primary teaching source: Manning, Raghavan, and Schütze, Introduction to Information Retrieval, Evaluation of ranked retrieval results, https://nlp.stanford.edu/IR-book/html/htmledition/evaluation-of-ranked-retrieval-results-1.html . The authors' definition and distinction from uninterpolated average precision were consulted. The ranking arithmetic and fictional product discussion are original examples. This entry uses the standard retrieval definition; benchmark-specific average-precision conventions should be checked separately.

Previous: Interoperability Next: Interpolation search