Genre, Manifolds, and AI. | Ben Schmidt
▻http://benschmidt.org/post/2021-03-08-genre-and-manifolds/genres-and-manifolds
The theory is this. If you assume that music is distributed in a high-dimensional feature space (as they surely do) the distribution of pieces in that space is almost sure to be highly uneven. Some areas (recordings of the Beethoven string quartets) will be densely populated; others (suites for toy piano) will be quite sparse.
If you then use k-nearest neighbors approaches to serve up recommendations for music (Spotify built the best-known library, so we know that they use it), you’ll likely hit music on the periphery of its local clusters far more often than music at the center.