The Netflix Prize was an open competition for the best collaborative filtering algorithm to predict user ratings for films, based on previous ratings without any other information about the users or films, i.e. without the users being identified except by numbers assigned for the contest. The competition was held by Netflix, a video streaming service, and was open to anyone who was neither connected with Netflix (current and former employees, agents, close relatives of Netflix employees, etc.) nor a resident of certain blocked countries (such as Cuba or North Korea). On September 21, 2009, the grand prize of US$1,000,000 was given to the BellKor's Pragmatic Chaos team which bested Netflix's own algorithm for predicting ratings by 10.06%. == Problem and data sets == Netflix provided a training data set of 100,480,507 ratings that 480,189 users gave to 17,770 movies. Each training rating is a quadruplet of the form . The user and movie fields are integer IDs, while grades are from 1 to 5 (integer) stars. The qualifying data set contains over 2,817,131 triplets of the form , with grades known only to the jury. A participating team's algorithm must predict grades on the entire qualifying set, but they are informed of the score for only half of the data: a quiz set of 1,408,342 ratings. The other half is the test set of 1,408,789, and performance on this is used by the jury to determine potential prize winners. Only the judges know which ratings are in the quiz set, and which are in the test set—this arrangement is intended to make it difficult to hill climb on the test set. Submitted predictions are scored against the true grades in the form of root mean squared error (RMSE), and the goal is to reduce this error as much as possible. Note that, while the actual grades are integers in the range 1 to 5, submitted predictions need not be. Netflix also identified a probe subset of 1,408,395 ratings within the training data set. The probe, quiz, and test data sets were chosen to have similar statistical properties. In summary, the data used in the Netflix Prize looks as follows: Training set (99,072,112 ratings not including the probe set; 100,480,507 including the probe set) Probe set (1,408,395 ratings) Qualifying set (2,817,131 ratings) consisting of: Test set (1,408,789 ratings), used to determine winners Quiz set (1,408,342 ratings), used to calculate leaderboard scores For each movie, the title and year of release are provided in a separate dataset. No information at all is provided about users. In order to protect the privacy of the customers, "some of the rating data for some customers in the training and qualifying sets have been deliberately perturbed in one or more of the following ways: deleting ratings; inserting alternative ratings and dates; and modifying rating dates." The training set is constructed such that the average user rated over 200 movies, and the average movie was rated by over 5000 users. But there is wide variance in the data—some movies in the training set have as few as 3 ratings, while one user rated over 17,000 movies. There was some controversy as to the choice of RMSE as the defining metric. It has been claimed that even as small an improvement as 1% RMSE results in a significant difference in the ranking of the "top-10" most recommended movies for a user. == Prizes == Prizes were based on improvement over Netflix's own algorithm, called Cinematch, or the previous year's score if a team has made improvement beyond a certain threshold. A trivial algorithm that predicts for each movie in the quiz set its average grade from the training data produces an RMSE of 1.0540. Cinematch uses "straightforward statistical linear models with a lot of data conditioning." The performance of Cinematch had plateaued by 2006. Using only the training data, Cinematch scores an RMSE of 0.9514 on the quiz data, roughly a 10% improvement over the trivial algorithm. Cinematch has a similar performance on the test set, 0.9525. In order to win the grand prize of $1,000,000, a participating team had to improve this by another 10%, to achieve 0.8572 on the test set. Such an improvement on the quiz set corresponds to an RMSE of 0.8563. As long as no team won the grand prize, a progress prize of $50,000 was awarded every year for the best result thus far. However, in order to win this prize, an algorithm had to improve the RMSE on the quiz set by at least 1% over the previous progress prize winner (or over Cinematch, the first year). If no submission succeeded, the progress prize was not to be awarded for that year. To win a progress or grand prize a participant had to provide source code and a description of the algorithm to the jury within one week after being contacted by them. Following verification the winner also had to provide a non-exclusive license to Netflix. Netflix would publish only the description, not the source code, of the system. (To keep their algorithm and source code secret, a team could choose not to claim a prize.) The jury also kept their predictions secret from other participants. A team could send as many attempts to predict grades as they wish. Originally submissions were limited to once a week, but the interval was quickly modified to once a day. A team's best submission so far counted as their current submission. Once one of the teams succeeded in improving the RMSE by 10% or more, the jury would issue a last call, giving all teams 30 days to send their submissions. Only then, the team with the best submission was asked for the algorithm description, source code, and non-exclusive license, and, after successful verification; declared a grand prize winner. The contest would last until the grand prize winner was declared. Had no one received the grand prize, it would have lasted for at least five years (until October 2, 2011). After that date, the contest could have been terminated at any time at Netflix's sole discretion. == Progress over the years == The competition began on October 2, 2006. By October 8, a team called WXYZConsulting had already beaten Cinematch's results. By October 15, there were three teams who had beaten Cinematch, one of them by 1.06%, enough to qualify for the annual progress prize. By June 2007 over 20,000 teams had registered for the competition from over 150 countries. 2,000 teams had submitted over 13,000 prediction sets. Over the first year of the competition, a handful of front-runners traded first place. The more prominent ones were: WXYZConsulting, a team of Wei Xu and Yi Zhang. (A front runner during November–December 2006.) ML@UToronto A, a team from the University of Toronto led by Prof. Geoffrey Hinton. (A front runner during parts of October–December 2006.) Gravity, a team of four scientists from the Budapest University of Technology (A front runner during January–May 2007.) BellKor, a group of scientists from AT&T Labs. (A front runner since May 2007.) Dinosaur Planet, a team of three undergraduates from Princeton University. (A front runner on September 3, 2007 for one hour before BellKor snatched back the lead.) The algorithms used by the leading teams were usually an ensemble of singular value decomposition, k-nearest neighbor, neural networks, and so on. On August 12, 2007, many contestants gathered at the KDD Cup and Workshop 2007, held at San Jose, California. During the workshop all four of the top teams on the leaderboard at that time presented their techniques. The team from IBM Research—Yan Liu, Saharon Rosset, Claudia Perlich, and Zhenzhen Kou—won the third place in Task 1 and first place in Task 2. Over the second year of the competition, only three teams reached the leading position: BellKor, a group of scientists from AT&T Labs (front runner during May 2007 – September 2008) BigChaos, a team of Austrian scientists from Commendo Research & Consulting (single team front runner since October 2008) BellKor in BigChaos, a joint team of the two leading single teams (a front runner since September 2008) === 2007 Progress Prize === On September 2, 2007, the competition entered the "last call" period for the 2007 Progress Prize. Over 40,000 teams from 186 countries had entered the contest. They had thirty days to tender submissions for consideration. At the beginning of this period the leading team was BellKor, with an RMSE of 0.8728 (8.26% improvement), followed by Dinosaur Planet (RMSE = 0.8769; 7.83% improvement), and Gravity (RMSE = 0.8785; 7.66% improvement). In the last hour of the last call period, an entry by "KorBell" took first place. This turned out to be an alternate name for Team BellKor. On November 13, 2007, team KorBell (formerly BellKor) was declared the winner of the $50,000 Progress Prize with an RMSE of 0.8712 (8.43% improvement). The team consisted of three researchers from AT&T Labs, Yehuda Koren, Robert Bell, and Chris Volinsky. As required, they published a description of their a
Read more →
