News for hackers. Stuff that matters.
Re:Last.fm Turns 20 by Rodeoclash on Wednesday November 23, 2022 @05:11PM (#33725164)
Two things I like to see trailed in music recommendation algorithms would be:

1. Associations happen at a song level. At the moment when artists are recommended you get what I call the "Red Hot Chilli Peppers" effect. That is, the artist that basically everyone seems to have in common in their playlists is RHCP. Instead, I think the recommendations should be more granular, at a song level. Artists can diverge quite wildly in style between individual songs (Ween being an extreme example of this). Take for example Planet Caravan by Black Sabbath (https://www.youtube.com/watch?v=SvrOzYtnLMA) this is great song but wildly different from say "War Pigs". In fact, it's a dubby, groovy slow jam.

2. Weight follow on plays more highly. So taking War Pigs in the point above. I listen to Planet Caravan and it makes me think of this great remix of Wicked Game by Trentemoller (https://www.youtube.com/watch?v=ph4z3u7IwrM). A totally unrelated artist and genre yet somehow linked in my mind. My assumption is that people that listen to Planet Caravan would also have a high likelihood of enjoying the Wicked Game remix because it has a similar... vibe I guess?

Anybody want to collaborate on building a recommendation engine around these two points? I'd be more than happy to record what I listen to as training data for it.

Re:Last.fm Turns 20 by sph on Thursday November 24, 2022 @03:30AM (#33728978)
> Weight follow on plays more highly

What if I listen to Black Sabbath and then my niece wants to listen to the Peppa Pig soundtrack after that? What if it's a party and everyone is queueing their favourite song?

The issue with recommendation engines is that they're dumb as rocks. They cannot tell that the Trentemoller remix might be related to War Pigs, but not Peppa Pig, without ears and a brain.

So, to be safe, they go for simple statistical aggregation which creates safe and boring recommendations that are unable to surface interesting connections.

Re:Last.fm Turns 20 by CrypticShift on Thursday November 24, 2022 @07:09AM (#33730396)
What you are talking about (simple statistical aggregation) may be true 10 years ago. Deep learning is truly different.

I'm not saying it is "smart" or listen the way we are. But is not "dumb" either in that it does not just follow prescribed rules.

I personally prefer human curation. But I've heard many many stories of how spotify's AI sometimes surprises people with uncanny magical recommendations (and dumb ones too). And it is just a start.

Re:Last.fm Turns 20 by sph on Thursday November 24, 2022 @07:38AM (#33730625)
Deep learning is not statistical aggregation, but still wouldn't solve the very example OP and I are talking about.

Deep learning is not smart nor is able to replace actual ears and understanding of music. Even a random recommendation system would be able to surprise once in a while.

Personally, I think Spotify's system is terrible and gives me a good suggestion every 100 wrong ones.

Re:Last.fm Turns 20 by CrypticShift on Thursday November 24, 2022 @08:58AM (#33731320)
> Even a random recommendation system would be able to surprise once in a while.

True. And because there is no way to "measure" the quality, if you say it is now horrible your you, I'll take your word for it.

I don't believe in "General AI", But this is not a binary outcome. It WILL get better. How much? We'll see.