[vintermann]

Spotify, hire musicologists

Strongly disagree with that. For one thing, Spotify already has an army of tastemakers spending all day assembling curated playlists. For another, they have the Echonest data which relies heavily on manual labeling.

But the major innovation in Discover Weekly was to use machine learning directly on mel spectrograms to figure out meaningful features for human taste. They still want to rely on their experts as much as possible (and hey, I don't blame them for not wanting to fire people), so they try to combine their expert's features with the algorithm's. But this introduces human biases again.

The problem is that when it comes to music, everyone's a missionary. Everyone wants the world to listen to the music they are excited about. Professional opinion-haver about music is the dream job for many adults, much like chocolate factory taste QA expert is for 6-year olds. And they just can't separate their own opinions from objective truths very well. It's hard to be objective about something you love.

The real great thing about AI in recommendations isn't really the intelligence part. It's the "AO" - artificial objectivity. The algorithm is probably inferior to humans in some aspects (it can't interpret the themes in lyrics very well, for instance), but the advantage is that you have full confidence about

1. What information it actually might use, and

2. What it tries to optimize.

From point one, you can be sure that it's opinion on Smashing Pumpkins isn't affected by that annoying kid in 8th grade that used to listen to them. For point two, you can be sure it's really trying to find the music you will love, not what it thinks you should love.

To get it slightly back on track: I can't wait until an AI can do music history, or etymology, or economics, or history. Or matchmaking in dating! It will be useful long before it can match humans on intelligence. How great wouldn't it be to get results in those fields which you could trust were from a disinterested party.

[mbrock]

I don't imagine the musicologists charged with personally recommending music.

A more important task would be to restructure the "information architecture," for example to improve the experience of looking for classical music or jazz. There's a lot to do that isn't just based on opinions.

The dream of AI doing music history seems kind of bizarre to me... as well as the whole idea that human knowledge is bad because it's biased...

[vintermann]

Information architecture of music is what Echonest was all about, and I'm pretty sure they keep doing what they were doing when Spotify bought them.

To the degree that judging good jazz or classical is different from judging other types of music, I think that yes, it's based a lot on opinion. In particular the opinion of authorities - critics and other performers.

It's not that this is entirely unreasonable. With music as a social phenomenon, you might prefer to not be "into" the wrong kind of music, even if you would like it for the music itself. Spotify and Echonest have actually talked a bit about how listening patterns can reveal "shameful" tastes, different from the tastes we would like to project.

The job of a recommendation system then, if we should look cynically at it, is to show you the "right" kinds of music that you would like to like, but actually like too - and to not tempt you with "wrong" kinds of music that you would like despite yourself.

And yes, I'd bet you'd need musicologists (or human analysts) for that. It would by definition be very hard to figure out from listening patterns or acoustic features. But isn't this a bit cynical as I said? Shouldn't we try to not be ashamed of what we actually like?

[mbrock]

I'm basically not interested in automated music recommendations, and it's simply not what I'm talking about.

Categorizing, labelling, organizing, displaying music information is not about judging quality.

That aside, I also don't believe it's possible to separate a "pure taste" from a "cultural taste," philosophically.