Predicting query performance using query, result, and user interaction features

Abstract

The high cost of search engine evaluation makes techniques for accurately predicting engine effectiveness valuable. In this paper we present a study in which we use features of the query, search results, and user interaction with the search results to predict query performance. We establish which features are most useful, study the effect of different classes of features, and examine the effect of query frequency on our predictions. Our findings show that performance predictions using result and interaction features are substantially better than those obtained using only query features. Such results can support automated search engine evaluation methods and new query processing capabilities.