The Netflix Prize: Alternating Least Squares in MPI

In 2006 Netflix announced a million dollar prize to the first team that could beat their Cinematch recommendation system by 10% on a particular test data set. Specifically, given over 100 million ratings, 1-5, from 480,189 different users and 17,770 different movies, the goal was to produce predictions for the test set that minimize the root mean square error. Cinematch scored an rmse of .9525, so the goal was to score less than or equal to .8572. My goal in this project was to simply beat Cinematch, in parallel. The prize was won in September, 2009 by a team using a blend of many different techniques. One such technique that played a large part in the winning blend was matrix factorization. The 480,000x17,000 (sparse) ratings matrix, R, is approximated by a product of two much smaller matrices. A singular value decomposition can produce the two matrices and conveniently minimizes the square norm. After choosing a number of “features”, f, you want an fx480,000 user matrix U and an fx17,000 movie matrix M whose product, UM comes as close as possible to the given training data in R. A number of algorithms exist to find these two matrices; I used an approach called alternating least squares with weighted-λ-regularization. [1]