Recall and precision


Recall is a measure of how well an information search and retrieval system finds ALL relevant documents on a searched for topic, even to the extent that it includes some irrelevant documents.

Precision is a measure of how well such a system finds ONLY relevant documents on a searched for topics, even to the extent that it skips irrelevant documents.

Recall is like throwing a big fishing net into the pond. You may be sure to get all the trout, but you've probably also pulled up a lot of grouper, bass, and salmon, too. Precision is like going spear fishing. You'll be pretty sure to ONLY get trout, but you'll no doubt miss a lot of them, too.

The Cranfields Tests in the 1960's claimed to find an inverse relationship between recall and precision. As your recall rate went up, your precision rate fell; as your precision rate rose, your recall rate went down.


See these citations for more info:

Cleverdon, C. W., and E. M. Keen. 1966. Factors determining the performance of indexing systems, volume 1: design, volume 2: test results. Cranfield, England: Aslib Cranfield Research Project.

Sparck Jones, K. 1981. The Cranfield tests. In: Information Retrieval Experiment, K. Sparck Jones (ed.). London: Butterworths: pp. 256-284.

Ottaviani, J. S. 1994. The fractal nature of relevance: a hypothesis. Journal of the American Society for Information Science, May 1994, v. 45, n. 4, pp. 263-272. This article proposes a new model for relevance, based on fractal geometry.


Return to the main menu.