AT&T researchers Chris Volinsky and Robert Bell, scientists who work on visualizing and analyzing large networks with AT&T Labs-Research, were part of the team presented with the Netflix Prize, a multi-year contest to improve upon the advanced Netflix movie recommendation system. More than 40,000 teams from 186 countries participated in the competition.
The $1 million prize was awarded to Volinsky, Bell and other members of the winning coalition in a news conference hosted by Netflix. Volinsky and Bell are twice winners of “Progress Prizes” in the competition – first in 2007 with former AT&T Labs colleague Yehuda Koren, now a Senior Research Scientist with Yahoo! Labs, and again in 2008 after combining with Big Chaos, an Austrian research team. Earlier this year, Bell and Volinsky reached out to Canadian team Pragmatic Theory to create the Grand Prize winning coalition.
The Netflix Prize was launched in 2006 to reward the team that could achieve a 10 percent improvement over the accuracy of the Netflix movie recommendation system at that time. The largest such data set ever released – 100 million anonymous movie ratings ranging from one to five stars – was made available to contestants. All personal information identifying individual Netflix customers was removed from the prize data, which contained only movie titles, star ratings and dates but no text reviews.
Jon Sanders, engineer in charge of the Netflix ratings database and one of the people overseeing the Netflix Prize says "[this] has been a worldwide competition launched in October 2006 to determine who can improve upon the Netflix recommendations system by ten percent. The contest made available to contestants 100 million anonymous movie ratings ranging from one to five stars, the largest such data set ever released. All personal information identifying individual Netflix members was removed from the prize data, which contained only movie titles, star ratings and dates but no text reviews."
The AT&T Labs researchers advanced two key methodologies central to recommendation systems: neighborhood models and latent factors models.
Neighborhood models predict a user’s interest in a new item, such as a movie or product, based on the user’s ratings or purchase history of a set of similar items. For example, a neighborhood for the latest action movie remake may include its sequels, other movies featuring the same stars or director, or other remakes of older films. The AT&T Labs researchers developed new methodology to better combine information across many neighbors, resulting in more accurate and meaningful recommendations.
Latent factors models use aggregate information on past user choices to establish a set of unobserved features for each item (for example, latent factors for a specific movie might include amount of violent content, level of comedic content, or level of interest to young children). These factors then enable an automated system to estimate and predict a user’s interest in a product based on these features. Along with using customer ratings themselves, the AT&T Labs researchers incorporated information about the types of items a user rated in order to improve estimates of a user’s interest in different products.
Additionally, the AT&T Labs researchers developed innovative ways to integrate neighborhood and latent factor models in ways that utilized the complementary strengths of each technology.