What's losing my interest more and more is that their music recommendations are terrible, rarely match my taste...
The really aggravating thing is that we know they have the tech to do better. There was that demo of a 2D embedding of genre-space that actually did a pretty decent job of grouping similar-sounding songs together, but their recommendations all seem to be based off of what was popular at the same time, or what came from the same label.
E.g. when I ask for recommendations for a song like "Mellon Collie and the Infinite Sadness," I should get "Avril 14th," not "Bullet with Butterfly Wings."
I don't see the connection between Smashing Pumpkins and Aphex Twin though, what kind of genre matching do you think would enable that on your specific example?
Not nit-picking, just piqued my curiosity about how you would connect those artists.
That's exactly the problem, those songs are not in either artist's genre. The artists don't sound similar, but the songs do. By "genre" I mean the effective genre (mood/vibe, instrumentation, pacing, chord progression, etc) of the song, not the nominal genre of the artist.
I would listen to those two songs in a row and be completely satisfied by the transition.
I was a huge Smashing Pumpkins fan when I was young, I'm a big Aphex Twin/Richard D. James fan and I don't know if I'd ever connect those two songs on a playlist or even play them together. That might be the issue, it's too much a subjective judgment.
I've worked and interviewed for companies that do music analysis to find moods, atmospheres, similar timbres, chords and structure to mix and match automated playlists for business needs (restaurants, cafés, retail stores and so on), would like to see what kind of similarity scores they could get between the songs you mentioned.
The really aggravating thing is that we know they have the tech to do better. There was that demo of a 2D embedding of genre-space that actually did a pretty decent job of grouping similar-sounding songs together, but their recommendations all seem to be based off of what was popular at the same time, or what came from the same label.
E.g. when I ask for recommendations for a song like "Mellon Collie and the Infinite Sadness," I should get "Avril 14th," not "Bullet with Butterfly Wings."