Netflix
In October of 2006, Netflix announced that it would give $1 million to anyone who could improve their Cinematch recommendation engine by at least 10%. This is the Netflix Prize competition. The Netflix dataset consists of over 100 million training examples, with over 500,000 users and 17,770 movies.
Along with two of my friends, David Lin and Lester Mackey, we started a team, Team Dinosaur Planet, named after the first movie in the database. Our approach, like that of the other top teams, is to find algorithms that capture different aspects of the data and blend them: over the past 12 months, we have implemented pretty much everything we have come across, as well as some new ideas of our own.
The final month of the 2006-2007 competition (culminating in a $50,000 Progress Prize) was pretty intense, and in the last week of the competition, we teamed up with Team Gravity (from Hungary) to form When Gravity and Dinosaurs Unite. For the last 23 hours, we reigned at the top of the leaderboard:
Unforunately, team BellKor from AT&T Research beat us out in the end, but it was an excellent competition and an awesome finish to the first year.
We are currently discussing writing up our work, so more details will come later...in the meantime, track our progress on the leaderboard.