1. 5
  1.  

  2. 1

    Well-written and interesting! It satisfied some of my curiosity about how these kind of media-recommendation systems work.

    It’s odd that your system works best with only 8-dimensional vectors, since I’d expect musical styles to have higher dimensionality than that. Is it possible to figure out if there’s some kind of human-comprehensible meaning of those 8 dimensions, like “heaviness” or “dissonance” or “Beatles vs Stones”?

    I want to try out the Music Galaxy site, but while it renders fine on my iPad, I can’t navigate at all because it only responds to keyboard commands. Adding touch gesture support would be awesome! (I’ll definitely try it out next time I’m at a ‘real computer’.)

    1. 2

      Hey - thanks for reading it!

      It’s odd that your system works best with only 8-dimensional vectors

      As far as I can tell, this is mainly due to the limited detail that the source data I use to produce the embedding provides. With a max of only 20 related artist connections per artist, there really isn’t a ton of detail available to represent those kinds of complex relationships. I believe that having access to a larger number of related artists as well as the degree of similarity (how much their audiences are similar) would help with that.

      I can’t navigate at all because it only responds to keyboard commands

      Ah I think that’s an issue with my mobile detection code; there are touch controls that work on mobile phones but that must not be getting activated on iPad. I’ll fix that right now, thanks for letting me know!