Netflix awards $1M prize to improve movie picks

It was a close call, but BellKor's Pragmatic Chaos narrowly beat out a rival group called the Ensemble for the Netflix Prize, the Los Gatos, Calif.-based company said at an awards ceremony in New York.

For those more excited by algorithms than touchdowns, following the Netflix Prize has been like the Super Bowl. And the winning method could have implications well beyond Netflix recommendations; any business that uses people's preferences to sell products could learn from the exercise.

Tens of thousands of people have pored over the problem since the contest began in October 2006. At last count, there were more than 51,000 contestants from 186 countries.

In the end, one-time rivals joined forces to form the two remaining powerhouses, BellKor and Ensemble.

"We had a bona fide race right to the very end," Netflix CEO Reed Hastings said in a statement. "Teams that had previously battled it out independently joined forces to surpass the 10 percent barrier. New submissions arrived fast and furious in the closing hours."

The winning team consists of two researchers at AT&T Inc., two engineers from Montreal, a research scientist at Yahoo Inc. and two machine-learning researchers from Austria. Netflix said all seven met in person as a group for the first time Monday.

BellKor in June became the first team to cross the 10 percent threshold. That kicked off a 30-day period during which other contestants could try to beat them. The Ensemble submitted its solution in late July, with just a few minutes to spare before the deadline.

According to Netflix's prize leaderboard, the Ensemble improved predictions for what movies people will like by 10.1 percent, while BellKor had improved predictions by 10.09 percent.

But Netflix actually had another set of data, visible only to Netflix. That one showed a 10.6 percent improvement by both teams. BellKor won because it submitted its final entry about 10 minutes before Ensemble.

Netflix is now planning a second contest - a sequel, if you will. While the first contest required contestants to improve predictions for subscribers who regularly provide ratings on movies they've watched, the second will involve those who don't rate movies often or at all.

©2009 The Associated Press. All rights reserved. This material may not be published, broadcast, rewritten or redistributed.