I worked full-time on algorithmic music recommendation for about 5 months last year (started working on it as a side project in 2016). I've always thought that, ultimately, an algorithm should be the best--after all, you can have it use human-curated data as an input. While my algorithm isn't amazing, it works far better for me than Spotify or Pandora ever did, so I've been using it myself regularly. I pivoted to a general-purpose recommender system idea last november, but I'd love to go back and work more on music later on.
(Music was just too hard to get started in--I spent far more time trying to find a way to integrate my algorithm into playback sources than actually working on the algorithm)