[vidarh]

The problem with "people who liked this song also liked that song" is that very often I don't want to listen to that song now even if it's something I really love.

If I'm listening to ambient music I don't suddenly want to be ambushed by something uptempo. If I'm listening to e.g. Debussy, you might be excused for suggesting something vaguely new age in a similar mood and tempo, but certainly not rock.

Another problem is that you can't just take raw overlap in tastes, because some people like "everything", and the fact their tastes overlap with mine does not mean I'll like everything else they like.

I've yet to hear a recommendation system that chooses music I want to listen to reliably enough that I can generally stand to listen to them for more than a few songs at a time without it turning into an endless annoying sequence of skipping.

Respecting genre (segues need to be gradual, if at all), respecting mood and tempo needs to come first. Then you can consider what others who likes the same songs within those constraints also likes within those constraints. Honestly if I have to choose between personalised recommendation and precise control of genre and mood/tempo, I'd take genre and mood/tempo over personalisation any day.

Another pet peeve of mine is lack of visibility into how to teach a system what I want. E.g. if I dislike or skip a song, will it get that it doesn't fit my current mood or what I want to listen to now, or will it wrongly infer I don't like the song at all?

Sometimes it feels as if the people designing these systems don't use them.

[crispyambulance]

> you can't just take raw overlap in tastes, because some people like "everything", and the fact their tastes overlap with mine does not mean I'll like everything else they like.

I think that these recommendation systems, especially youtube's, are much more nuanced than you're suggesting. And I say this as someone who finds these recommendation systems lacking and of limited utility compared to discovery from reading Pitchfork, for example. They're trying to balance quite a lot of inputs and actually somehow make money from it.

Also, "genre" and "mood" of a piece of music are not easy to define, let alone measure. You're asking for a lot and I don't think what you're asking for is realistic from a piece of software.

[vidarh]

I'm sure they are more nuanced in that of course they will take into account differences in breadth etc. to some extent, but they are clearly not nuanced enough, because they barely work.

> Also, "genre" and "mood" of a piece of music are not easy to define, let alone measure. You're asking for a lot and I don't think what you're asking for is realistic from a piece of software.

Yet I think if you asked a bunch of humans to attach labels to tracks they like, you'd find they overlap closely enough.

I base that on the fact that while many will struggle to identify niche sub-genres or discriminate between genres they don't listen to, people certainly have a relatively large shared idea about major classifications.

What is more: You'd be able to find the labels of people who usually label the same way as you.

And it doesn't need to be precise. It just needs to better.