I logged in this week to find that Spotify had started recommending audiobooks. I guess why not? It’s a streaming service, why not put some books on there too? But man, these recommendations were weird.
Huh, Das Kapital, you say? “I've never read Marx before, but now that it's a Spotify Audiobook, sign me up!” Certainly, the best way to learn about Marx’s critique of capitalism is to listen to it on the Peloton.
Does Spotify think I’m a whiny, 19-year-old DSA member? You don’t know me at all, Spotify!
This book list is goofy, but it points to the brick wall we’ve hit with recommender systems. I don’t know what’s in Spotify’s book collection. The top of the Audiobook page is mostly a bunch of airport-ready pablum. But the four books it recommended to me are nowhere to be found on the Audiobooks page. So why did it recommend these four books? Do Recommender Systems mostly work by dumb luck?
I’m sure some sort of machine learning was used to generate these book recommendations. People who use Spotify in a similar way to me were blurred into archetypes, and then these blurs were matched against some blurry representations of books, and then… profit? Are we stuck in a rut with recommender systems because we focus so much on machine learning and so little on search and exploration?
The two extremes I have in mind here are Spotify and Bandcamp. Spotify shows why machine learning has its limits. Bandcamp shows why designing a search engine for exponentially expanding content is daunting.
Spotify floods you with recommended listening. You get a webpage of machine-generated suggestions and playlists based on your listening history. Some of these playlists are of artists you like, prompting you to listen to what you already know. Some try to push you to hear unfound artists and genres.
Bandcamp offers high-level curation on their main page, giving lists by microgenres and featuring editorial newsletters on up-and-coming bands. But once you’re on a band page, there are no external links for more engagement.
Neither of these are particularly good. As a listener, I always end up on Bandcamp because I heard of an artist in some other way. But I find their search, lists, or editorial content to quickly lead to dead ends. On Spotify, I used to love the things I’d find in recommended lists, but it’s definitely grown stale. The “mixes” it makes are of songs I’ve already listened to. The recommender engine can’t extrapolate at all. My Discover Weeklies used to be pretty engaging, but now it’s music I’ve mostly heard before (Portrait of Tracy by Jaco Pastorius this week?) or random noise bands I don’t want to listen to.
What’s missing on both of these websites is the record store. The shop I wrote about a few weeks ago. Run by some eccentric person who loves to think about what music they’ll stock, how they’ll arrange it, what they’ll play on the store’s loudspeakers today. The internet would be better if it had more independent record stores. Who’s going to open the first one?
I love this line:
> Are we stuck in a rut with recommender systems because we focus so much on machine learning and so little on search and exploration?
There are definitely limits to ML recommendation algorithms which must necessarily optimize for a single one dimensional metric. But our tastes and interests are multidimensional, and have structure beyond a feature vector. Why can’t we have recommender systems that are similarly more sophisticated, that we have more control over, that we can tune ourselves? I wrote about some related thoughts here: https://yeetgenstein.substack.com/p/on-the-music-snob