Matrix factorization
Approximating a large user-item rating matrix as the product of lower-dimensional latent factor matrices, a core Netflix Prize technique.
Matrix factorization represents a sparse ratings matrix as the product of two smaller matrices: one encoding user tastes, the other item traits, in a shared latent space. Predicted ratings are inner products of those vectors, which generalizes to unseen user-item pairs.
Yehuda Koren, Robert M. Bell, and Chris Volinsky described factorization variants that dominated the Netflix Prize leaderboard alongside ensemble methods. The technique remains a baseline in collaborative filtering before deep models took over some domains.