Personalized Search

A personalized guitar

See this battered guitar, worn from years of playing rehearsals and concert gigs. I bet you it could tell us a story about the person who played it. It could tell us about rock-solid riffs, fiery solos and sweet melodies. It could, because this guitar has been personalized. The guitar has taken on some of the particulars of the musician, like the worn patches where the right arm and the pick rubs against the varnished wood. The wear and tear is a result of friction, and the guitar has given in and adapted itself.

Personalized search adapts itself, showing different results to different people based on their past behavior. Personalized search tries to disambiguate intent by using information not only about what you are doing now, but also what you did in the past. The purpose of personalization is to remove friction from the user experience, by removing what is irrelevant and highlighting what is more suited to your personal preferences. Precision is particularly important in web search, where user’s only view the first result page.

Some time ago, Google started personalizing web search results using your web history. Google will personalize the search result ranking whenever possible.They are very caution about this feature, though. According to an interview with Googler Marissa Mayer, personalized ranking will only lift two results into the top 10, never replacing the number one organic result. Facebook does another form of personalized search, letting your social network influence the on-site search ranking. Search anything; and your friends, groups and items will show up at the top.

A good personalization experience is an unobtrusive one. Wear a new pair of shoes for a week or so, and notice how the sole slowly shapes itself to your foot, and how that slight pressure on your ankles fades away. Would you wear those shoes if the material refused to give away and your ankles kept hurting? Probably not. I like the shoe metaphor, because it emphasizes that personalization is more about small adjustments rather than revolutionary changes. If you buy a pair of sneakers, you don’t want them to turn into sandals or boots, just slightly more comfortable sneakers.

What great improvements does personalized search bring with it? For one, relevancy is an individual measure. What to considered relevant is ultimately determined in relation to yourself. If I tell you something you already know, it’s not really relevant. Novelty is an issue, but familiarity is just as important. As many as 40% of all web queries are for the purpose of find back to something. Surfacing already visited links seems like a good idea. We may associate different things with the same query, and personalized search holds the promise to disambiguate these cases. An average query length of two is in many cases too short to predicts the user’s particular information need.

What are the success criteria for personalized search? Personalization can be based on both long-term and short-term history, and different techniques for profile-based and click-based personalization is used. Studies suggest that click-entropy (the amount of variation in the search results searchers click on) can be used as a simple measure to determine if a search result should be personalized. More variation means higher query ambiguity.

Profiling preferences with long-term history shares limitations with social recommenders. Systems like these have inertia built into the core, and they react slowly to changes in interests. I’m not so keen in books about Ankor Wat anymore, now that I’m back from my holiday in Thailand.

There are many good reasons to start working with personalized search results. Simpler categorical recommendation schemes, implicit feedback, and small steps are good startup parameters. Jakob Nielsen’s old post on personalization is still relevant and worth a read.