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.
