Netflix 百万美元推荐大赛

2006 年 10 月，Netflix 宣布了一件在当时看来有点疯狂的事：悬赏 100 万美元，奖励任何能把其推荐算法准确率提升 10% 的团队。它公开了约 1 亿条匿名评分数据，全球任何研究者都能下载、建模、提交结果。一场关于「猜你喜欢」的全球竞赛就此拉开。

大赛持续了将近三年，吸引了来自 186 个国家的四万多支队伍。从研究生到业余爱好者，无数人在这 1 亿条数据上较劲。2009 年，一支融合了多个团队方法的「超级团队」以提升 10.06% 的成绩捧走大奖。夺冠方法不是某个天才想法，而是把几十种模型叠在一起——这个细节本身，就是对当时推荐系统研究的一次真实写照。

Netflix Prize 的意义远超一场比赛。它把「推荐算法」从公司的内部工程问题变成公开的学术热点，矩阵分解等一批技术由此进入主流。更重要的是，它第一次让「机器学习」这个词大规模走进公众视野：原来「猜你喜欢」背后的数字，值得一百万美金。

它也留下了争议。竞赛数据虽是匿名的，仍有研究者通过交叉比对识别出具体用户，Netflix 随后因隐私问题被起诉并支付和解金。数据开放与隐私保护的边界，在这场热闹的竞赛里被提前暴露。

回看 Netflix Prize，它示范了「数据+竞赛+开放协作」能怎样推动 AI 进步，也让推荐系统从边缘话题变成互联网产品的核心战场。今天每个人手机里的「猜你喜欢」，都多少带着 2006 年那场百万美元竞赛的影子。

In October 2006 Netflix announced something that seemed a little crazy at the time: a $1 million reward to any team that could improve its recommendation algorithm's accuracy by 10%. It released about 100 million anonymized ratings, and any researcher in the world could download, model, and submit. A global contest over "what you'll like" had begun.

The contest ran for nearly three years and drew more than 40,000 teams from 186 countries. Students, hobbyists, and professionals all wrestled with those 100 million ratings. In 2009 a "super-team" that merged many teams' methods took the prize with a 10.06% improvement. The winning approach was not a single genius idea but dozens of models stacked together—a detail that itself was an honest portrait of recommender research at the time.

The Netflix Prize's significance far exceeded one competition. It turned recommendation algorithms from an internal engineering problem into a public research hotspot, and brought techniques like matrix factorization into the mainstream. More than that, it put the phrase "machine learning" into public conversation for the first time at scale: the math behind "guess what you like" was worth a million dollars.

It also left controversy. Though the data was anonymized, researchers were able to de-anonymize specific users by cross-referencing, and Netflix was later sued over privacy and paid a settlement. The boundary between open data and privacy was exposed early by this very public contest.

Looking back, the Netflix Prize demonstrated how "data plus contest plus open collaboration" can drive AI progress, and turned recommendation systems from a niche topic into the core battlefield of internet products. Every "recommended for you" in your phone today carries a trace of that 2006 million-dollar contest.