Automatic Playlist Generation by Similarity
What is the problem you are facing?
It's not really a problem, but rather an additional way of experiencing sound and music in funkwhale.
If you're listening "Random Radio" (or either radio specified by genre) there are sometimes really huge differences in regards of tempo, style, tone etc. There could be huge differences between song A and B even "within" the same genre. Imagine Funkwhale compiling playlist like a real DJ*ane. And those playlists are compiled using techniques which are based on actual research in psycho-acoustics.
In fact there had been Open Source development in the past and I can recall very good. Somebody from the field of psycho-acoustics wrote a plug-in "mirage" for the long abandoned banshee.fm music library (a former gnome application). I used to use it in these days and it was really good! This plugin, mirage, "simply" analysed each and every song in the library and attached a so called "timbre/similarity"-model to them.
What are the possible drawbacks or issues with the requested changes?
I do not know anything about both coding or psycho-acoustics. I guess if you want to develop something like "mirage" for funkwhale you would likely have to do some serious research into acoustic theory and so on. The first (and only) starting point I can possibly give you is the link to the plug-in page. There you'll find links furthering you to other knowledge resources: