“Spotify’s genre system places all songs as a point in 13 dimensions, grouping together those close-by points as genres. The dimensions include objective musical properties such as ‘loudness’ and ‘beats per minute’, as well as more subjective emotional properties, such as ‘energy’, ‘valance’ (sadness) and ‘danceability’. These latter, subjective measurements are established through listening sessions, where human subjects listen to pairs of songs and state which of them they think is saddest or more danceable. The algorithm learns the difference and classifies other songs appropriately.”
David Sumpter, Outnumbered (2018), pp. 74-75.
(This passage on Spotify’s Discover algorithm led me to Glenn McDonald’s everynoise.com, a fascinating graphical map of hundreds of musical styles. It’s beautiful. McDonald writes:
“This is an ongoing attempt at an algorithmically-generated, readability-adjusted scatter-plot of the musical genre-space, based on data tracked and analyzed for 1905 genres by Spotify. The calibration is fuzzy, but in general down is more organic, up is more mechanical and electric; left is denser and more atmospheric, right is spikier and bouncier.”)