I'm interested in this too. I have unreleased music that I've made and it somehow generates reasonable similarities to other music in my library. It can't be simply pulling the info from the net since the artist name I'm using isn't out there anywhere. Some kind of spectral analysis maybe?
I could be completely wrong, but I know they have a tagging system in place. The tags presumably come from metadata providers (they give basic information about the song, like who wrote it and when it was made, as well as pictures for album covers and such). After that they can pretty easily look for similarities, eg. two items both tagged funny. If you were concerned that they upload your listening history somewhere to come up with recommendations, they do not.
Taking a quick look at the source code (of which I am not familiar so I could be barking up the wrong tree), it seems that Jellyfin gives you a list of candidates in the following order: similar to recently played, similar to liked, directed by recently played, actor from recently played, has liked director, has liked actor. Similarity is calculated by rating and production year, as well as shared number of connected people, genres, studios, and tags. For music, it also adds specific criteria for other albums the artist has worked on.
