School of
Information Management & Systems
Previously School of Library & Information Studies
Michael Buckland,
Professor.
buckland@sims.berkeley.edu
The Trade-off between Recall and Precision
Abstract:
Michael K. Buckland and Fredric Gey.
The relationship between Recall and Precision.
Journal of the American Society for Information Science
45, no. 1 (Jan. 1994): 12-19.
Summary:
The traditional measures of retrieval performance are Recall
(completeness of retrieval) and Precision (purity of retrieval).
Empirical studies of retrieval performance have shown a
tendency for Precision to decline as Recall increases.
Analysis of the relationships between Recall, the number of
items retrieved, and Precision shows that there is a definable
region for all feasible retrieval results. For all cases of consistently
better-than-random retrieval, Recall curves tend to follow an
increasing curve rising from the origin, and a trade-off between
Precision and Recall is inherent, not just an inconvenient
empirical finding.
More generally, a trade-off between Precision
and Recall is entailed unless, as the total number of documents
retrieved increases, the marginal retrieval performance is equal to or better than
overall retrieval performance thus far.
There is a fundamental relationship between Precision and
Recall which, for a given model of Recall, constrains the
behavior of Precision. In particular, if Recall is modeled by a
polynomial function of proportion of documents found, then
Precision is modeled by a lower order polynomial function of
the same variable.
A simple geometric transformation can produce a quadratic
model of Recall that satisfies tangency to perfect retrieval at
the origin and yields reasonable looking Recall-Precision
trade-offs.
Two-stage, or, more generally, multistage retrieval procedures,
whereby a retrieved set is used for subsequent, more detailed
search, is likely to achieve the goal of improving both
Precision and Recall simultaneously even though
the trade-off between them cannot be avoided.
Go to
OASIS research program
or to Michael Buckland's
home-page.
Revised Sept 2, 2005.