Key points are not available for this paper at this time.
We describe a framework of bootstrapped hypothesis testing for estimating the confidence in one web search engine outperforming another over any randomly sampled query set of a given size. To validate this framework, we have constructed and made available a precision-oriented test collection consisting of manual binary relevance judgments for each of the top ten results of ten web search engines across 896 queries and the single best result for each of those queries. Results from this bootstrapping approach over typical query set sizes indicate that examining repeated statistical tests is imperative, as a single test is quite likely to find significant differences that do not necessarily generalize. We also find that the number of queries needed for a repeatable evaluation in a dynamic environment such as the web is much higher than previously studied.
Jensen et al. (Sat,) studied this question.