In 2009 Yehuda Koren, Robert Bell, and Chris Volinsky, members of the team that won the Netflix Prize, published an article in IEEE Computer (volume 42, issue 8, pages 30-37) explaining the technique at the heart of their winning system. The core idea of matrix factorization is to represent both users and items as vectors of hidden factors learned from the data. A movie might score high on a “comedy” factor and low on a “serious drama” factor; a user gets a matching vector describing how much they like each factor. The predicted rating is just the dot product of the user vector and the item vector, so the model can guess how a person will rate a film they have never seen.
The paper’s contribution was less the basic algorithm than a clear, practitioner-focused account of how to make it work on real data. The authors describe adding bias terms (some users rate everything high, some movies are universally liked), folding in implicit feedback such as which titles a user browsed or rated at all, modeling how tastes drift over time, and weighting observations by confidence. They favored learning the factors by stochastic gradient descent or alternating least squares rather than classic singular value decomposition, because the rating matrix is mostly empty and naive SVD overfits.
Matrix factorization became the default workhorse of recommendation for the better part of a decade and still underpins many production systems. The article is one of the most cited works in the field and is the standard reference for anyone learning how recommenders actually predict preferences.
For a business reader, this is the paper that turned a contest-winning trick into an industry standard: it is why the “because you watched” and “customers also bought” features across streaming and retail are built the way they are.