The Netflix Prize

On 2 October 2006 Netflix published 100,480,507 ratings that 480,189 customers had given to 17,770 films, and offered a million dollars to the first team to predict held-out ratings ten percent more accurately than its own system. Thousands of teams entered. The prize was claimed on 21 September 2009 by BellKor's Pragmatic Chaos, a merger of three teams, twenty minutes ahead of a rival ensemble that had matched their score.

The competition changed how the field worked. It established that a public dataset, a fixed metric and a leaderboard would draw more effort than any laboratory could fund, a model that Kaggle turned into a business in 2010 and that ImageNet adopted for vision. And its winning methods, matrix factorisation of the user–film table into learned vectors, blended with hundreds of other models, taught a generation that embeddings and ensembles beat clever rules.

Netflix never deployed the winning system, judging the engineering cost too high, and cancelled a second prize in 2010 after researchers showed the anonymised ratings could be linked to named people. Both lessons, that benchmarks are not products and that data has privacy in it, recurred at larger scale in the 2020s.