It must be the same algorithm they use to propose new tracks, eg when we play “similar artists”.

“What is the best to listen after this track?” - just with a super small subset of tracks.