3.5.1: Measures

A more detailed description can be given in the following way:

ZUGANG3.5.1.1: Relevance

According to this definition relevance depends only on the query and the document. Other influences like the documents already seen or the knowledge of the user cannot be taken into account.

ZUGANG3.5.1.2: Precision and Recall

Precision gives the fraction of relevant documents within the retrieved documents, recall gives the fraction of relevant documents that were retrieved.

Best values for precision are obtained if all retrieved documents are relevant, for recall if all relevant documents were retrieved.

Extreme values: one single relevant document: precision = 1; All documents: recall = 1.

Precision and recall are antagonistic: A small result set from a specific query will result in high precision and low recall; a large result set from a very general query will result in low precision an high recall.

In case of ranked result sets this antagonism can be displayed:

ZUGANG3.5.1.3: Precision-Recall-Diagram

ZUGANGAbb. 21: A Precision Recall Diagram

ZUGANGAbb. 22: Precision - Recall - Diagram Displayed in the Plane

Comparing two systems: A system is "better" if its precision values are higher at all recall levels.

One dimensional comparison of systems with ranked results:


© 1998 / HTML-Version 17. 11. 1998: R. Ferber