Evaluating Search Engines by Clickthrough Data
- Cite this paper as:
- He J., Li X. (2010) Evaluating Search Engines by Clickthrough Data. In: Patel-Schneider P.F. et al. (eds) The Semantic Web – ISWC 2010. ISWC 2010. Lecture Notes in Computer Science, vol 6497. Springer, Berlin, Heidelberg
It is no doubt that search is critical to the web. And it will be of similar importance to the semantic web. Once searching from billions of objects, it will be impossible to always give a single right result, no matter how intelligent the search engine is. Instead, a set of possible results will be provided for the user to choose from. Moreover, if we consider the trade-off between the system costs of generating a single right result and a set of possible results, we may choose the latter. This will naturally lead to the question of how to decide on and present the set to the user and how to evaluate the outcome.
In this paper, we introduce some new methodology in evaluation of web search technologies and systems. Historically, the dominant method for evaluating search engines is the Cranfield paradigm, which employs a test collection to qualify the systems’ performance. However, the modern search engines are much different from the IR systems when the Cranfield paradigm was proposed: 1) Most modern search engines have much more features, such as snippets and query suggestions, and the quality of such features can affect the users’ utility; 2) The document collections used in search engines are much larger than ever, so the complete test collection that contains all query-document judgments is not available. As response to the above differences and difficulties, the evaluation based on implicit feedback is a promising alternative employed in IR evaluation. With this approach, no extra human effort is required to judge the query-document relevance. Instead, such judgment information can be automatically predicted from real users’ implicit feedback data. There are three key issues in this methodology: 1) How to estimate the query-document relevance and other useful features that useful to qualify the search engine performance; 2) If the complete ”judgments” are not available, how can we efficiently collect the most critical information from which the system performance can be derived; 3) Because query-document relevance is not only feature that can affect the performance, how can we integrate others to be a good metric to predict the system performance. We will show a set of technologies dealing with these issues.