The blog post should clear up your questions:
tl;dr - it’s a machine learning model that runs on your server and places each of your tracks, albums, and artists at a point in n-th dimensional space, and calculates the distance between each item to figure out similarity.
