I posted a few months ago about the potential problems with tailoring search results to a profile for each user based on a history of searches. My issue with this kind of personalization is not privacy, but that I want a fresh start each time I go looking for something. I don’t want to decrease my chance of finding something totally new.
After reading Greg Linden’s post about a 2007 Microsoft Research paper on personalization, I see that one kind of personalization he and others are pursuing is short-term, click-based tailoring of results (within a search session).
The authors conclude that click-based techniques “work well”, but profile-based “improve the search accuracy on some queries, but they also harm many queries” and are “not as stable as click-based”.
So I understand the attention being paid to this now. If you could make a decent guess at whether sequential queries are related, it seems like looking at a series could inform the relevance of results more than looking at queries one at a time. And I agree as the paper suggests that long-term, profile-based will help and hurt depending on the situation.
Looking over a section of my web history, I couldn’t find many strings of queries that were related at all. And I can’t find an example in the AOL search logs of a series of queries that would helpfully inform each other. Although, as with natural language search, no one really knows how new tools could change behavior.
More on personalized search from Linden:
There may be some disadvantages to the approach Google is using for personalized search. For example, using long-term, high-level profiles means that the search engine can shift results slightly toward general preferences, but it cannot make immediate changes based on what a searcher is doing right now. In particular, it cannot help much when searchers are on a mission, doing a series of related searches, but not finding what they want.
Linden’s site Findory is an interesting experiment in personalization divided into news, blogs, video, and podcasts. After reading some above-average articles in terms of interestingness, it recommended an article because I had just read an article on the same news item. In terms of short-term patterns, you’d want to suppress situations like this: