A Million Dollars, Baby

“Well, I’m not really interested in the money,” says Geoffrey Hinton, a U of T computer science professor, standing at a scribbled-upon whiteboard in his office. “But they are,” he adds, jabbing a friendly thumb toward two of his graduate students, Andriy Mnih and Ruslan Salakhutdinov, who both grin sheepishly. For Mnih and Salakhutdinov, the prospect of advancing the frontiers of machine learning (a branch of artificial intelligence) would be great. But a $1-million prize? Now we’re talking. The students are working with Hinton to win a $1-million competition sponsored by Netflix, an online DVD-rental service based in California. Teams worldwide are trying to devise a set of algorithms that improve the company’s movie recommendation software by 10 per cent.“

Collaborative filtering is what it’s called, and it has many applications besides movies,” says Hinton. At the moment, a Netflix user can rank movies out of five stars on the company’s website. The system can estimate (usually within one star) what ranking a user will give a new movie – based on the user’s past rankings, and the rankings of those with similar tastes. If, for example, he or she has given five stars to classic science-fiction movies and low marks to musicals, Netflix will recommend, say, Forbidden Planet instead of My Fair Lady.

That may seem obvious to a human being, but for a computer, prediction – the ability to examine a large tangle of data, find meaningful patterns and extrapolate what might come next – is a huge and complex problem. Even small steps toward solving it could yield huge improvements in fields such as data compression, speech recognition and image correction.

More than 20,000 teams have entered the competition.The U of T team started in first place, but has since slipped to third, having achieved a 7.07 per-cent improvement in ranking accuracy. At press time, a mysterious competitor called “BellKor” leads with 7.8 per cent. But Mnih says the U of T group can still catch up. (Teams can refine their algorithms and resubmit results as often as they like.) The competition, which began last October, will run until October 2011 or until the 10 per-cent mark is reached. “It gets more difficult the closer you get to 10 per cent,” says Mnih. “That last two per cent is going to take a while.”