Oh man, I should write a blog post about that, as I built that feature myself. It was meant to be a stopgap until we could get some real matchine learning in th

Oh man, I should write a blog post about that, as I built that feature myself. It was meant to be a stopgap until we could get some real matchine learning in there, but nothing else we tried did as well. First, for efficiency all recommendations were artist to artist,nl not song to song. That works well for a lot of genres but is pretty bad for others. We started with a free DB of artist similarities, I don't remember where we got that from, maybe musicbrainz? We built a shitty internal interface for adding and removing links between artists and adjusting the weights of those links and then made it available to all employees to mess with. As you might imagine just about everyone there was passionate about music so it didn't take long to crowdsource a huge catalog of quality recommendations and then for really obscure stuff we would fall back to the open db. So the actual algorithm would look at your seeds - artists you put in the queue before turning on radio or artists with songs that you liked while radio was on, pull the top n linked artists for each of your seed artists, and do some weighted shuffling. It would also make sure to space out artists so you don't hear the same one too often etc. Then for genre radio we just secretly selected a bunch of artists we felt were representative of the genre and used those as the seeds. Oh yeah and if you disliked a song we'd prevent that artist from playing for the rest of your session. We also would look at anomalies like popular artists with not many recommendations, or artists that, when used as seeds, lead to shorter listening sessions (implying that the recommendations need to be cleaned up). Most attempts to replace this with something smarter ran into 2 problems: 1. Popular stuff is popular, so it looks like a good recommendation for anything, and 2. ML is hard and takes a lot of time, which we never had enough of