The Taste of Others
How can all these websites presume to know what you might "also like?" Scott Kirsner pulls back the curtain to reveal the technology of taste that drives these consumer sites.
If pressed, Joe Konstan will admit that he watches "fewer than two movies a year." Yet Konstan has spent more than a decade creating software that attempts to figure out your taste in movies and suggest what you might want to watch next.
Konstan, a professor at the University of Minnesota, is just one of dozens of academic researchers trying to solve the problem of understanding our fickle predilections. In the commercial sector, last fall Netflix offered a $1 million prize to anyone who could help the DVD rental company improve its movie recommendations by at least 10 percent. Amazon has been a pioneer in examining the DVDs you've purchased (and others that you've searched for) and making suggestions about what you may want to buy next. Several start-up companies, including ChoiceStream and Matchmine, are also trying to design software that better intuits your likes and dislikes.
It's a tall challenge, and recommendation software can still be clueless at times: just because I once bought a how-to-knit DVD for my wife as a gift, Amazon constantly suggests that I might want to own more craft-oriented discs. Netflix doesn't have a way for users to let it know that they never watched a given DVD, even though it sat atop their TV for three weeks, which seems like it'd be an important piece of information.
The systems all need information about you to make a guess about what you like. On e-commerce sites like Amazon, that tends to come from DVDs you've bought or looked at; on movie rental sits like Netflix or Blockbuster Online, users rate movies they've seen. On Netflix, the typical user has rated about 200 movies (I'm at 316), and one especially passionate user has rated about 90 percent of the site's 85,000 movies, according to spokesman Steve Swasey. (No one at Netflix is convinced that this person has actually seen all those movies.)
Swasey explains that one way to make Netflix's recommendations more accurate is to create multiple profiles within an account. On his account, he says, because he and his daughter have different profiles, her movie ratings stay separate from his, and they each get different recommendations.
Konstain says that one thing that can skew recommendations is when users only rate movies they like. "That means you don't have enough negative examples to train the system on what kind of recommendations would be bad to make," he says. So letting the system know about movies you despise can produce a better set of suggestions.
One problem that the technology hasn't yet begun to deal with is our changing moods. "You might enjoy one movie when you're alone and feeling down, but things are different when you're going to be with a bunch of people," Konstan says. There's no way yet to tell Blockbuster that you're feeling burnt out on a Friday night and need some light entertainment, or that you owe your girlfriend a "date night" movie, or that you're in a documentary frame-of-mind.
Another issue is that none of the sites you visit share information about your tastes with anyone else. "Blockbuster thinks about the ratings you've given it as their information," explains Toffer Winslow, an executive at ChoiceStream, the software company that supplies Blockbuster with recommendation technology. He says that users may feel like their privacy has been compromised if information they've given to one site is used by another to serve up more personalized content.
But that can make it a hassle to ever switch from one service to another: if Netflix has an archive of your ratings, you'd have to start from scratch with Blockbuster.
A new start-up called Matchmine, which received funding recently from the owner of the New England Patriots football team, is trying to change that.
"The essence of our approach is to try and make personalization portable," says Mike Troiano, Matchmine's CEO. "It seems strange to us that the time you'd invest in rating movies at one Web site wouldn't benefit you on others." The company lets you store information about your taste on your own computer (rather than on a site's own servers), and share it with any site you want to. Right now, the technology works with just two movie sites: FilmCrave.com, a community for movie buffs, and PeerFlix, a service that lets users trade DVDs with one another.
Troiano says that making good recommendations is a tough nut to crack: "I've got three PhDs and 28 employees working on it full-time." About 2500 teams from around the world entered the Netflix contest, according to Swasey. In the end, however, no winner was announced.
But what provides the ultimate challenge for any piece of recommendation software are its rivals: human beings.
"The best technology in the world will never be as good as a guy who likes the same movies you do and tells you about them," admits Troiano.
And word-of-mouth recommendations from friends, family members, and colleagues have two important features that Web sites can't yet match: the fun of comparing notes on a recommended movie with them after you've watched it…and the guilt if you ignore their suggestion and they call you on it.
Scott Kirsner is the man behind the blog CinemaTech. He also writes about technology and innovation for such publications as the Boston Globe, The New York Times, Wired, Variety, Fast Company, the Hollywood Reporter, Salon.com, BusinessWeek, and Newsweek. Kirsner helped start (and continues to help run) three conferences: Future Forward, the Nantucket Conference on Entrepreneurship & Innovation, and Convergence: The Life Sciences Leaders Forum. He also often speaks and moderates at other people's conferences, and serves as a commentator on TV and radio.