One of my colleagues is the author of a remarkable way of building
hybrid recommender systems. The general idea is that when one
calculates an SVD for a matrix $A$, the SVD of the Gramian matrix
$AA^T$ is received for free. But the Gramian matrix itself is a
similarity matrix. So if we have not only collaborative similarity
but some content-based one, we still can find an SVD of this
similarity and use it to improve collaborative filtering.
