We’ll Tell You What to Like
dan tynan on Jul 03 2008 at 9:21 am | Filed under: Da Web, US Airways Magazine, Web 2.0
Recommendation engines know what you want — even if you don’t
(A version of this post originally appeared in the July 2008 issue of US Airways Magazine.)
Some people remember their first kiss or the first time they drove a car. I remember the first time I recorded a show on TiVo. It was an old Rockford Files episode, nearly five years ago. The next time I turned on the TV I found TiVo had recorded an entire season of Diff’rent Strokes reruns.
This was the machine’s way of telling me that if I liked James Garner, I’d love Gary Coleman.
TiVo was relying on an internal recommendation engine: software that looks at your behavior, like the TV shows you’ve recorded, then makes an educated guess about what else you might like. (Some guesses are more educated than others.)
Since then, recommendation engines have grown more sophisticated and found a permanent home on the Web. Buy earrings from Amazon.com, and you’ll see jewelry displayed on the home page the next time you log on. Tell Netflix you loved Shaun of the Dead, and it suggests you rent Army of Darkness. Read a story on WashingtonPost.com, and it tells you what other readers of that story have looked at.
Recommendation engines are popular for a simple reason: We are drowning in choices. There’s too much content, not enough brain cells. Recommendation engines tell us what we ought to be looking at so we can safely ignore the rest.
They do it in essentially three ways: by creating profiles based on your behavior, analyzing content you’ve already consumed and finding stuff that’s similar, or pulling information from users who are more or less like you. Increasingly, engines are using a combination of these tactics.
For example, TargusInfo’s predictive analytics software helps retailers and web sites predict what you’ll want to buy in the future by looking at what you’ve bought in the past. They combine that with information from third-party databases, fit you into one of 232 predefined profiles, and then suggest products that match your profile, says Ken Inman, vice president of research and development.
Music streaming site Pandora takes the second tack. For its Music Genome Project, Pandora hired 50 professional musicians to analyze millions of songs, breaking them down into 400 attributes like “a subtle use of vocal harmony” or “major key tonality.”
The result of this “completely, completely insane” project, says Pandora founder Tim Westergren, is a musical fingerprint that lets each song be matched to tunes with similar fingerprints. So if you like Elvis Costello, Pandora assumes you’ll like Van Morrison or Ben Folds Five. Tell it you’re a Louis Armstrong fan, and you’ll hear Billie Holiday or the Dukes of Dixieland. Pandora also relies on the wisdom of the crowd: If enough users click thumbs down on a song, it plays less often.
It’s the social aspect of recommendation engines that has the most potential. Install Loomia’s SeenThis? application inside Facebook, for example, and it tells you what your friends are reading on sites like The Wall Street Journal. Visit the Journal’s site, and you’ll see recommendations from your Facebook pals. Loomia factors in behavioral and contextual data as well – sorting its recommendations by topic and by the stories most often read by others who fit your profile, says CEO Dave McMurtry.
By tapping into the opinions of people you trust, SeenThis? and other social recommendation engines like StumbleUpon (Web sites), Criticker (movies), or Foodio54 (restaurants) hope to avoid problems with software that can’t tell one 1970s TV show from another, or search algorithms that push products the site’s sponsors would like you to have, instead of things you might actually want.
Another advantage: If your friends recommend something and you absolutely hate it, at least you’ll know who to blame.
