A year ago, the online movie rental company Netflix offered $1 million to anyone who could come up with a way to improve its movie recommendations by 10 percent. Nobody has earned the money, but a team from AT&T Labs came close.
And other online retailers are watching the online video rental company closely, looking to improve their own shopping recommendations.
With 7 million subscribers, Netflix does provide computerized recommendations. Customers rate movies on a scale of one to five stars. And Vice President Jim Bennett says the company's track record is pretty good.
"We get basically three out of four correct, within half a star," Bennett said. "Which is pretty high, actually."
Netflix uses what computer scientists call a "neighborhood approach."
"So if you rated Gone with the Wind and Anna Karenina both positively, and we find many people who have done that, then we find a correlation between those two," Bennett said.
"So that if someone comes along and rates Gone with the Wind, but hasn't told us anything about Anna Karenina, we should adjust their prediction for Anna Karenina positively."
But the Netflix system only goes so far. Tens of thousands of statisticians and computer scientists have been working to improve it. No one scored the 10 percent improvement needed to claim the million-dollar grand prize. But the team from AT&T Labs came close.
By slicing and dicing millions of Netflix ratings, the team was able to find patterns that seem to suggest common factors that might lead certain customers to like certain films.
The factors behind the patterns could be dialogue or action or popular stars. But the team doesn't pretend to know. Statistician Bob Bell says he's not even much of a film buff. What interests him is the data.
"As far as we were concerned, the movies were just an ID number," Bell said. Netflix has awarded his team a "progress prize" for its work — and urged them to keep at it.
But if Bell's approach seems a little too mathematical, you might want to try another movie recommendation site, What To Rent (whattorent.com). One of the company's founders, Adam Geitgey, says the site asks would-be renters a series of personal questions, sort of like a matchmaker about to arrange a blind date.
"People relate to movies like they relate to people," Geitgey said. "If we can figure out what kind of person you are and what kind of person the movie would be, we can match you with a movie."
But that's not really practical on a large scale. Netflix Vice President Bennett says the beauty of his company's algorithm is that it can work automatically, and with any kind of product. That's why not only Netflix but all kinds of e-commerce sites are anxious to see it work better.
And recommendation algorithms aren't just good for companies that want to move more merchandise. Computer scientist John Riedl of the University of Minnesota says they are also good for customers who may be looking to find an out-of-the-way film that wasn't a blockbuster hit.
"What I'm excited about in recommenders is the possibility of letting the little hit matter," Riedl said, "for letting all of us who have our own quirky, special little movie or book taste be recommenders for people who like the kinds of things we recommend."
If Riedl is right, moviegoers and other online merchants could be giving Netflix's new and improved algorithm millions of thumbs up.