This blog post describes in a clear and concise way how the (non-negative and low-rank) matrix factorization technique (Lee and Seung, 1999) can be used in a recommender system to discover the latent features. As a means, author describes an example for recommending films, in which five users rates four films and the unconstrained minimization of the SSE of the representation is accomplished by gradient descent (it could alternatively be computed with an ANN, e.g.). A section on how to regularize the above objective function is included, and a downloadable 23-lines Python implementation with the full source code and this small example along with four interesting references are given as well.

A.A. Yeung. Matrix Factorization: A Simple Tutorial and Implementation in Python. In quuxlabs.com, Blog Spots section, September 2010.

Pablo Guerrero-Garcia

University of Malaga