Netflix Million Dollar Prize

I find this contest remarkably interesting. Netflix has a database of their customers' ratings of various movies. They use this database to make recommendations to their customers on which movies they might like to see next. The recommendation is not so much 'Goodfellas is a great movie' - but more 'Goodfellas was rated highly by people who like the same kind of movies that you do'. The competition is to improve the algorithm making the recommendations - prize $1 Million USD - open to anybody, anywhere (so long as you're not hated by the USA, including Quebec, harshly enough).

The target doesn't sound too ambitious. The ratings are between 1 and 5 stars. Their current system 'Cinematch' doesn't do too well. On average it's out by just under 1 star for each rating - so typically Cinemax will predict that a viewer will watch a movie and score it 4 stars - the viewer will actually score it 3 or 5 stars. You could probably get close to that level of competence by guessing that the viewer will rate every movie at 3 stars. Many times, you'd be right, and many times, you'd only be one star away. The vast majority of movies that I've seen are 2, 3 or 4 stars.

The contest is looking for just a 10% improvement on Cinematch to be eligible to win the prize. If I understand the small print of the contest correctly - it'll run for a minimum of 4 months, and provided at least one entry meets the 10% improvement, the algorithm that does best will win the prize.

This kind of contest is right up my street - I specialized in artificial intelligence subjects in University, but I've found it difficult to use those skills outside of personal projects. I'd have to regard this as a bit of fun too - but it's great to have this dataset to work with, and the potential prize makes it a lot more exciting.

Posted by Alexander at October 2, 2006 02:54 PM

Make Flash tests and assessments with the Question Writer, Quiz Software. Question Writer 4 is now available. Click here to download.

Trackback Pings

TrackBack URL for this entry: