Story here. Something about predicting what kind of movies people will like is as far as I got.
see here for the original rules. What they wanted was an algorithm that - given a lot of data about a customer - would predict which films that customer would rate highly, and do it better than the system they had in place already. That’s it, basically.
The problem is this:
You give a 1-5 rating on every movie you rent. Then Netflix makes recommendations to you based on what ratings people with similar tastes have given to stuff you haven’t seen. When you subsequently rate a movie that was recommended, you are not just rating the movie, but the quality of the recommendation.
The Netflix prize was for finding a way to make recommendations that would result in higher ratings, thereby implicitly improving the recommendation system itself, in a way that could be measured statistically.
The previous posters have pretty well nailed it. The new contest is largely similar, except that the data set contains basic demographic information in addition to customers’ previous movie ratings. There’s a brief description at the end of this article.
I’m hoping they find a winner. My propensity to watch even lousy sci-fi movies has caused NetFlix to assume I hate all sci-fi. It’s very hard to explain to the current algorithm that I really like good sci-fi and merely tolerate the rubbish because nothing else is on.
The new algorithm is supposed to be 10% more effective than the current system.
The current system actually works fairly well in my experience. I have rated over 400 movies (the more movies you rate, the better the recommendations they’ll give you), and I have gotten recommendations for movies I’ve never heard of which after I watched them, became some of my favorite movies.
Yeah, their current algorithms are good at making very “safe” recommendations, IME. “Oh, you liked crime dramas X and Y? Here, let me recommend crime drama Z – it’s almost the same movie!” I wish it was better at coming up with recommendations for movies that are out of my comfort zone, because I think I’ve exhausted the genres I usually go to.
In my case, Netflix gave up on giving me anime recommendations, since I rated one series which I absolutely love, and another which I completely hated*. Both series were pretty similar in their gross features, involving lots of explosions and robots and overly-complicated scifi plots, so now I think the algorithm is completely confused. This is probably a case of my own idiosyncratic preferences making it hard to extrapolate based on what other people like, since there’s not a large set of people that absolutely love the one series but hate the other.
*Ghost in the Shell and Neon Genesis Evangelion, respectively
Wait until you add a few more. I’m up over 4,000 rated titles. (Not only have I been a member for a long time, but I rate anything I’ve ever seen, and I frequently use the Watch Instantly feature as a replacement for regular TV - yay for no commercials!)
If you want more info on the Netflix prize, Wired did an article about it back in Feb 2008 when BellKor was only 8.43% better than Netflix’s engine. You can read the article online here.