Netflix would like to improve their movie recommendation system, and they’re willing to pay $1,000,000 to anyone that can do it. Even you can participate! Here’s the details:
To encourage users to keep their subscriptions active, Netflix uses a movie recommendation system called Cinematch. It predicts whether someone will enjoy a movie based on how much they liked or disliked other movies. The results are displayed to users, titled “Movies You’ll Love.” Cinematch can make some impressive connections. People who enjoy The Patriot also tend to like Pearl Harbor; both are history-war-action movies. Cinematch’s suggestions drive 60% of Netflix’s rentals.
Can Cinematch work better?
Cinematch does an incredible job predicting user preferences. Netflix has improved upon the algorithms since 2000, identifying success by comparing actual user ratings to Cinematch’s predicted user ratings. But Netflix wondered if they could achieve better results. Currently, Cinematch can predict whether a user will like a movie, on a scale 1 to 5 stars, by about 1 star. But Netflix is curious if someone can think of algorithms that could beat Cinematch’s prediction accuracy–perhaps there is a new mathematical or statistical method that Netflix has not tried.
Netflix will pay $1,000,000 to anyone who can improve Cinematch’s prediction accuracy by 10%. Essentially, Cinematch can predict a user’s movie rating within 1 star. Netflix will award $1,000,000 to anyone who can create a method that predicts ratings within 0.9 stars. The competition was announced in October 2006–so far thousands of programmers have made attempts, but no one reached 10% and claimed the prize.
Why this is brilliant
Netflix derives much of their business from the movie recommendation system. Consider their business model: users pay a fixed price ($5-$20/month) to rent movies. At first, New customers will likely rent any hit movies they have not seen, but over time their initial interest dwindles and customers may stop renting. Every month, Netflix must justify the monthly cost and direct users to new movies that they never considered watching. This is the value of Cinematch; users consume more content with the movie recommendation system. Therefore, an accurate system that correctly judges a user’s preferences will provide infinite value to Netflix. Netflix tried for years to improve Cinematch, with only incremental results. A marginal performance increase of 10% could generate revenue well in excess of the $1,000,000 prize.
How it works
Teams register on the Netflix Prize site. Contestants can then download an enormous data set of 480,000 customers that rated 18,000 movies. Teams write algorithms to predict the customer ratings, and then Netflix tests the code against a different ratings data set, which has been kept secret.
Marketing Note: high-minded marketers will likely classify the prize as “web 2.0,” “mass collaboration,” “wikinomics,” or “crowd surfing.” In reality, Netflix based the contest off of the British Longitude Prize of 1714. Eight years ago, Goldcorp also paid half a million dollars to anyone who could find gold deposits using its mining data.