« Pre-order Gretchen Rubin's new book | Main | Interview with Kevin Murphy »
The Netflix prize has been won
Details of the competition are here. The winning team is here. Background on the prize is here. Previous MR coverage is here.
Bravo! I should note that current "recommender" systems don't much help me make further selections.
I thank Jüri Saar for the pointer.
Posted by Tyler Cowen on June 26, 2009 at 04:42 PM in Current Affairs | Permalink
Comments
Wait, the winning team's webpage says other competitors have the next month to enter in their best model.
Posted by: Mike Gibson at Jun 26, 2009 5:08:06 PM
Yeah, it seems like this entry only triggered the lightning round. Someone has reached the clearing condition, which means that now there's a month left.
Posted by: John Thacker at Jun 26, 2009 5:18:09 PM
I really hope they follow this up with another similarly themed contest.
Posted by: Massimo at Jun 26, 2009 5:25:14 PM
The contest itself should win some sort of brilliancy prize. Netflix has gotten far more than $1,000,000 worth of good PR out of it, along with an improved algorithm for cheap.
The Bellkor guys get bragging rights worth easily $1,000,000 over their collective lifetimes, even if somebody else comes during the next 30 days.
I'm guessing every team in the top 20 will do well economically out of this.
Here's one of the guys in the top 20 (#16) using similar techniques to predict wine prices:
http://justaguyinagarage.blogspot.com/
Posted by: ZBicyclist at Jun 26, 2009 6:10:34 PM
How do they avoid curve fitting their models?
Posted by: John at Jun 26, 2009 7:20:59 PM
Perhaps unusually, I get a lot of value out of current recommender systems. They point me toward categories that I would otherwise ignore, like nature documentaries or network comedy shows, and end up liking. Also, low recommender ratings on movies by great directors (Bergman, Antonioni) are a reminder not to be too swayed by prestige and snob appeal.
Posted by: BenjaminL at Jun 26, 2009 9:34:53 PM
@John: I think contestants got a training data set, and predictions were then tested against subsequent, real ratings from customers (see http://www.netflixprize.com//rules). So overfitting to the training set would not have been an advantage.
Posted by: CFox at Jun 27, 2009 4:05:17 AM
The recommender system is my least favorite part of Netflix as it seems to take the place of more normal search capabilities. I use the 'watch instantly' streaming quite a bit and, quite annoyingly, there's no way to search watch instantly selections by, say, director or actors. Other people have been frustrated enough to create their own netflix searchers, for example:
http://netflix.softworksdevelopment.com/
Posted by: Slocum at Jun 27, 2009 6:47:17 AM
"The recommender system is my least favorite part of Netflix"
Worse than the bill?
Posted by: josh at Jun 27, 2009 12:02:28 PM
