The first track that played was a System of a Down tune. Which is about as pop as it gets. The ones after didn't get much better. If they really want to use this to push less played songs which have potential, they should actually better get some.
There's a way lesser-know bands can upload their songs into their database for free, this leads to lesser-known songs being tagged with adjectives, this leads to a music search engine that you can use to find those lesser-known songs you like.
Stop being arithmetic supergeeks wanting to put everything inside a box, and start figuring out how to get all these weird unpredictable people to input valuable data into your system.
Google figured this out more than a decade ago, so why are we still seeing stupid mathematical and "pattern-based" algorithms every year?
The neural network is trained on crowd-sourced data. TRANSLATION: These supergeeks actually DID figure out how to get all these weird unpredictable people to input valuable data into their system. The solution they designed is that people will play their Facebook game, herdit.org, and the statistically significant answers to the quiz game are tagged to the song clips. These tagged song clips are then used as a training data set for the neural network. The machine algorithm is a result of the collective intelligence of all the players of herdit.org
Anyone can make an omelet with eggs. The trick is to make one with none.