As part of my exploration of data science and my love for music, I decided I'd try to create some genre playlists on Spotify by using their API. The code for this can be found on my Github.
Unfortunately, there are a few issues to work out still. For one, Spotify doesn't actually store genre information for their songs (this is probably why they haven't done genre playlists themselves). Why that is the case is beyond me - maybe it's too much work to categorize songs into genres? Maybe they don't want to commit things like that and then have people complain when something comes up in a genre that is from a different genre? Who knows.
My next idea was to use some unsupervised learning and cluster songs based on their similarity. Again, Spotify kind of foiled my plans - all they provide you with regarding track features are somewhat abstract concepts that I assume they're using as a latent space to describe tracks, such as "danceability" or "energy". Naturally, these latent concepts aren't defined in any other way. I'm guessing they found these by running PCA/SVD on whatever original features they had describing each song.
This is still a big WIP, of course.