I'm astounded that they only got a "data scientist" 3 years ago. Wonder who/what controlled their recommendation engine before that?
All-in-all that was a pretty well balanced article. I like the bit about PhDs - don't need 'em, but useful since their whole schtick is solving hard problems independently.
Recommendations were likely produced by applying collaborative filtering to their expert-extracted features (the whole "Music Genome" thing). Collaborative filtering is easy to implement, and the massive feature extraction investment likely provided a rich dataset from which they could recommend songs.