[rightbyte]

It seems to be some sort of common principal component analysis.

E.g.

A likes a. B likes b. B likes c. A maybe likes c.

But with many more nodes. I think they use it for recommendations and subgenre sorting etc.