Evaluation Measures next up previous
Next: Collection of Pages Up: Experimental Environment Previous: TREC Implementation

Evaluation Measures

We compare the effectiveness of our implementation of TREC ad-hoc algorithm to four commercial search engines: Excite, Google, Lycos, and AltaVista Raging. For a given query, if a page is not found in the top ten ranks by a search engine, that engine gets no credit for that particular query. The assumption here is that if a user can't find a page in the top result page, the user will simply give up. This assumption is strongly supported by the fact that almost 85% of users don't request beyond just the first results screens for their query [18]. For every system we count the number of queries for which it retrieves the desired site at rank-1, up to rank-2, up to rank-3, and so on, and plot this on a graph (see Section 5). The higher the number of queries for which an engine retrieves the desired site at a certain rank, the better is the engine.

Using the top ten pages per query also allowed us to manually judge every run. Even though it is simple in principle to find if two URLs will get you the same page, in light of redirections (via the refresh HTML meta-tag), pages generated by javascripts, mirror sites, etc., this becomes a non-trivial exercise in the current web environment. Therefore, we check all the results by hand to find ranks of the relevant pages retrieved (as they may be retrieved under a completely different URL).


next up previous
Next: Collection of Pages Up: Experimental Environment Previous: TREC Implementation
Amit Singhal 2001-02-18