Building a non-negative matrix factorization model

A general improvement on the basic cross-wise nearest-neighbor similarity scoring of collaborative filtering is a matrix factorization method, which is also known as Singular Value Decomposition (SVD). Matrix factorization methods attempt to explain the ratings through the discovery of latent features that are not easily identifiable by analysts. For instance, this technique can expose possible features such as the amount of action, family friendliness, or fine-tuned genre discovery in our movies dataset.

What's especially interesting about these features is that they are continuous and not discrete values and they can represent an individual's preference along a continuum. In this sense, the model can explore shades of characteristics, for example, perhaps a critic in the movie reviews dataset, such as action flicks with a strong female lead that are set in European countries. A James Bond movie might represent a shade of that type of movie...