I’m quite behind on a lot of my readings, so I only got around to reading the IEEE Computer’s (August) special issue on search this weekend. (Abstracts are free but actual PDF’s require an expensive subscription or an expensive purchase.) It includes the following articles:
- Search Engines that Learn from Implicit Feedback
- A Community-Based Approach to Personalizing Web Search
- Sponsored Search: Is Money a Motivator for Providing Relevant Results?
- Deciphering Trends in Mobile Search
- Toward a PeopleWeb
The articles were written by a mix of university academics and researchers from Google and Yahoo. They seem targeted at giving the general practitioner a sampling of some of current research, rather than being comprehensive in any specific domain or deep in a particular research area.
For me, the most interesting article is “Search Engines that Learn from Implicit Feedback” by Thorsten Joachims and Filip Radlinski of Cornell University. It’s a very accessible summary of the research those two have been doing in the last few years. To start off their research, they used eye-tracking experiments to characterize how people react to search engine rankings. They found that the ranking order strongly biases what people view and therefore click on. A result in the top ranking will often be clicked on more often than a better result in the second or third ranking, as some users may not even have looked at the results beyond the first ranking. A straightforward assumption that a click is the equivalent of a positive vote is therefore naive. Instead, they examine results that were not clicked on but should have. For example, if results at ranking 3 and 4 are clicked on, but not the result at ranking 2, then one can be sure that the result at ranking 2 is worse than the ones at ranking 3 and 4 and can use that knowledge to improve the search engine. Note that if the result at ranking 1 was clicked on, nothing new is learned. People are so biased towards clicking the first result that only if it was not clicked on would that be considered informative.
Under that model, they can interleave the results from two different search engines (or algorithms) and evaluate which one is better based on users’ clickthroughs. This insight led them to develop a ranking SVM model to learn search engine rankings. The new algorithm was shown to create a better meta-search engine as well as a better domain-specific search engine.