The technical reason for this is just that the algorithm fails to distinguish between these two groups. Probably its just a hard problem to solve from a sample of what people listen to. I suspect a significant fraction of people that like black swing also like white swing and that results in the algorithm being unable to resolve that there are two features there and not one.
You really need a level of manual curation that a big data statistical model just can't provided at scale.
What I find curious about it though is that it's obviously recognizing the style - it plays 1-2 musicians from the "black" group, then circles into only playing Miller/Goodman/Rat pack and never comes back around to playing the music that I originally was trying to play. If it was behaving the way you are thinking, I'd expect it to mix the two styles.
I think the problem is that the algorithms are based on statistical probabilities from other users. I.e users who listen to X also like to listen to Y. So we’ll add Y to the queue. Then Y becomes the new reference point. I mean that is a gross simplification but essentially if your musical taste is outside of 2 standard deviations of the norm all the algorithms are gonna suck. For me they do.
I’ve definitely trained tidal that I prefer some pretty whacky sub genres (this Northern European country, but only metal with strong brass sections, or contemporary accordion, hurdy gurdy, and a dozen other clusters like that).
I’d guess if you created a profile and loaded it up with just black swing bands from the 30-50’s, it’d do OK.
If not, and I understand their algorithm correctly, it would not only be because no current listeners make that distinction (as discussed up thread).
It would also be because the metadata doesn’t give any signal for it. They seem to use information such as record labels, song writers, producers, guest musicians, etc.
If that metadata has no signal, then my guess is that you’re trying to get it to racially segregate music that was produced before the big interracial marriage scare.
People were worried that if their kids listened to the same musicians, then whites and blacks (or worse!) might marry, so they created white radio stations and black radio stations.
Before that, I imagine there was a lot more interracial collaboration, and the metadata wouldn’t find clean clusters along race boundaries.
It could also be that the old metadata was never digitized.
I wonder if this is the case, or if the model just has year as a heavily weighted factor, because bluntly, Ella and Louis were vastly superior musicians to Frank and Tony. I honestly can’t hear the similarity. It’s like thinking “Oh. You liked The Killers, here’s some One Direction”
You really need a level of manual curation that a big data statistical model just can't provided at scale.