A statistical model of query log generation

Abstract

Query logs record past query sessions across a time span. A statistical model is proposed to explain the log generation process. Within a search engine list of results, the model explains the document selection - a user's click - by taking into account both a document position and its popularity. We show that it is possible to quantify this influence and consequently estimate document "un-biased" popularities. Among other applications, this allows to re-order the result list to match more closely user preferences and to use the logs as a feedback to improve search engines.