Recommender systems in general try to solve an "explore / exploit" problem; I sometimes wonder when I get a load of odd recommendations from one if it is actually trying to explore my preferences (since a non-click is informative) more than it thinks I'd like the content.
Interesting, thanks - I hadn't seen that. It's certainly the case that click-through rates are extremely low in general, and I think his explanation makes sense.
Although it is obviously a bit dated now, I think it is still a very persuasive essay.
The 'light buyer' and the way that digital advertising is almost optimised not to reach them are quite counter-intuitive ideas, but have real explanatory power when you grasp them. People used to joke about how Amazon would see you had purchased a dehumidifier and conclude this was merely the start of a dehumidifier buying spree, but that is a very real online experience.
Vaguely related, this is a great Dan Davies essay (also nearly a decade old) - https://medium.com/bull-market/a-cynic-s-guide-to-fintech-3cd0995e0da3. Amongst other things, it sense checks the value of 'big data' by looking at the value of pure data analytics companies (fairly high, but not stratospheric).
I must say you're a far less cynical person than I am. I don't think the algorithm is making a 'mistake' recommending you anti-vaxxers - I think it is deliberately designed to recommend us all nutcases because that drives outrage and therefore engagement.
I agree that can be the effect, but I don't know if it's a deliberate thing. In fact I'm more worried about it sorting people into silos - when I first started using my alt account it was very keen for me to follow only COVID-cautious people for some reason for example, rather than mixing it up with provocations.
"... we called these techniques ‘machine learning’ ... before people realised they could get bigger grants ... by calling them ‘artificial intelligence’" 😆
I regularly gently amuse myself by asking all 'AI Fanz' to explain how what they're proselytising differs from ML
It's not a clear line in my mind, but I think there's maybe some kind of distinction just to do with the amount of computing power that's getting thrown at things these days. But I just wanted to remind people that deep down there's maths under the bonnet!
Recommender systems in general try to solve an "explore / exploit" problem; I sometimes wonder when I get a load of odd recommendations from one if it is actually trying to explore my preferences (since a non-click is informative) more than it thinks I'd like the content.
That's a cheerful thought, isn't it? (And yes, I realise that it's all a bit more complicated than I gave myself space to get into here!)
For some reason, this reminded me of a very sharp Ian Leslie essay from nearly a decade ago about the problems with targeted digital ads:
https://www.ft.com/content/cd1722ba-8333-11e5-8e80-1574112844fd
Interesting, thanks - I hadn't seen that. It's certainly the case that click-through rates are extremely low in general, and I think his explanation makes sense.
Although it is obviously a bit dated now, I think it is still a very persuasive essay.
The 'light buyer' and the way that digital advertising is almost optimised not to reach them are quite counter-intuitive ideas, but have real explanatory power when you grasp them. People used to joke about how Amazon would see you had purchased a dehumidifier and conclude this was merely the start of a dehumidifier buying spree, but that is a very real online experience.
Vaguely related, this is a great Dan Davies essay (also nearly a decade old) - https://medium.com/bull-market/a-cynic-s-guide-to-fintech-3cd0995e0da3. Amongst other things, it sense checks the value of 'big data' by looking at the value of pure data analytics companies (fairly high, but not stratospheric).
That's very interesting. Thanks for sharing. I think it still holds up pretty well.
I must say you're a far less cynical person than I am. I don't think the algorithm is making a 'mistake' recommending you anti-vaxxers - I think it is deliberately designed to recommend us all nutcases because that drives outrage and therefore engagement.
I agree that can be the effect, but I don't know if it's a deliberate thing. In fact I'm more worried about it sorting people into silos - when I first started using my alt account it was very keen for me to follow only COVID-cautious people for some reason for example, rather than mixing it up with provocations.
"... we called these techniques ‘machine learning’ ... before people realised they could get bigger grants ... by calling them ‘artificial intelligence’" 😆
I regularly gently amuse myself by asking all 'AI Fanz' to explain how what they're proselytising differs from ML
It's not a clear line in my mind, but I think there's maybe some kind of distinction just to do with the amount of computing power that's getting thrown at things these days. But I just wanted to remind people that deep down there's maths under the bonnet!