Finding unobserved ratings using matrix factorization

Suppose that the utility matrix contains ratings of user on movies. How does matrix factorization find ratings for unobserved (user, movie) pair. Even though the purpose of the algorithm is to find the closest approximation of the original utility matrix.