How Cambridge Analytica’s Facebook targeting model really worked • Hacker News Board

Re: How Cambridge Analytica’s Facebook targeting model really worked

#11

Post by acdha » Fri, Mar 30, 2018, 9:27 PM UTC

rjurney wrote: post #5

godelski wrote: post #4

> has revealed that his method worked much like the one Netflix uses to recommend movies. I'm not sure this is the model you want to emulate. The suggestions are terrible and continually getting worse.

It's a hard problem but Netflix's model represents the state of the art in machine learning for recommendations. Still scared of the singularity? :)

There’s a bug confound with the catalog: Netflix’s streaming catalog is much smaller than their DVD selection was and it changes as licensing deals expire. No model can make up for that entirely.

Re: How Cambridge Analytica’s Facebook targeting model really worked

#12

Post by tomrod » Fri, Mar 30, 2018, 9:28 PM UTC

lordnacho wrote: post #10

Finally, I was waiting for someone to talk about the model itself. It makes sense that SVD or something like it (PCA, co-occurrence, etc) would be used. But I also wonder what exactly you are going to do with the predictions. What exactly do you show to someone to make them more likely to go and vote if they are inclined to vote your way, or make them stay at home otherwise? Is there evidence that whatever you're sho…

> What exactly do you show to someone to make them more likely to go and vote if they are inclined to vote your way, or make them stay at home otherwise?

Qualitatively: show things that get them angry.

Quantitatively: test and control pop splits.

Re: How Cambridge Analytica’s Facebook targeting model really worked

#13

Post by daenz » Fri, Mar 30, 2018, 9:28 PM UTC

CiPHPerCoder wrote: post #8

daenz wrote: post #3

Am I understanding this correctly? Facebook user data (likes/profile info) was scraped to produce low-dimension feature vectors for users (similar to word2vec). These feature vectors were then run through some ML model to predict...what exactly? Targetability for effective political ads?

[ Deleted. Nothing I say on HN ever matters. Move along ]

"What sort of lies" is pretty hand-wavy when it comes to labeling training data for a model. Are you summarizing from a source? I'm interested in the technical details of what happened.

Re: How Cambridge Analytica’s Facebook targeting model really worked

#14

Post by mattmanser » Fri, Mar 30, 2018, 9:30 PM UTC

tomrod wrote: post #12

lordnacho wrote: post #10

Finally, I was waiting for someone to talk about the model itself. It makes sense that SVD or something like it (PCA, co-occurrence, etc) would be used. But I also wonder what exactly you are going to do with the predictions. What exactly do you show to someone to make them more likely to go and vote if they are inclined to vote your way, or make them stay at home otherwise? Is there evidence that whatever you're sho…

> What exactly do you show to someone to make them more likely to go and vote if they are inclined to vote your way, or make them stay at home otherwise? Qualitatively: show things that get them angry. Quantitatively: test and control pop splits.

How do you test anything? There's only one vote, you can't iterate.

Re: How Cambridge Analytica’s Facebook targeting model really worked

#15

Post by sebleon » Fri, Mar 30, 2018, 9:30 PM UTC

tomrod wrote: post #12

lordnacho wrote: post #10

Finally, I was waiting for someone to talk about the model itself. It makes sense that SVD or something like it (PCA, co-occurrence, etc) would be used. But I also wonder what exactly you are going to do with the predictions. What exactly do you show to someone to make them more likely to go and vote if they are inclined to vote your way, or make them stay at home otherwise? Is there evidence that whatever you're sho…

> What exactly do you show to someone to make them more likely to go and vote if they are inclined to vote your way, or make them stay at home otherwise? Qualitatively: show things that get them angry. Quantitatively: test and control pop splits.

> Quantitatively: test and control pop splits.

How do you actually do this? Presidential elections come once every 4 years.

Re: How Cambridge Analytica’s Facebook targeting model really worked

#16

Post by danso » Fri, Mar 30, 2018, 9:33 PM UTC

dazc wrote: post #9

rjurney wrote: post #5

Earlier quoted context omitted.

It's a hard problem but Netflix's model represents the state of the art in machine learning for recommendations. Still scared of the singularity? :)

How hard can it be? Any film that I didn't watch all the way to end ought to be a strong signal that I didn't enjoy it. Not that I want to watch other movies just like it.

OTOH, that you even chose to watch a movie is a sign that you are interested in the genre. Just because I abruptly stopped watching “Star Trek IX”, doesn’t mean that I have abandoned the sci-fi genre or even Star Trek

Re: How Cambridge Analytica’s Facebook targeting model really worked

#17

Post by metaobject » Fri, Mar 30, 2018, 9:35 PM UTC

mattmanser wrote: post #14

tomrod wrote: post #12

Earlier quoted context omitted.

> What exactly do you show to someone to make them more likely to go and vote if they are inclined to vote your way, or make them stay at home otherwise? Qualitatively: show things that get them angry. Quantitatively: test and control pop splits.

How do you test anything? There's only one vote, you can't iterate.

Maybe with polling?

Re: How Cambridge Analytica’s Facebook targeting model really worked

#18

Post by godelski » Fri, Mar 30, 2018, 9:36 PM UTC

rjurney wrote: post #5

godelski wrote: post #4

> has revealed that his method worked much like the one Netflix uses to recommend movies. I'm not sure this is the model you want to emulate. The suggestions are terrible and continually getting worse.

It's a hard problem but Netflix's model represents the state of the art in machine learning for recommendations. Still scared of the singularity? :)

Personally, I used to get substantially better results. It would suggest to me a movie I'd never seen before with a score above 90, I'd watch, and enjoy. Now I'm getting things like kids shows suggested to me. I double checked by history to make sure no one watched one on my account, but they keep popping up with high percentage. I also really like horror and get weird suggestions [1] for similar movies. After I watched Alien it suggested the Great British Baking show. I literally have 0 trust in the new ranking system. They push certain shows way too hard and TV shows I'm watching show up on page 3 of "continue watching".

This is far from my previous stance of "Huh, movie I've never seen before and in a different language? It's 95% so yeah, I'll give it a go."

And before someone says anything, I do vote up and down quite frequently. But can't help but notice that my suggestions were better when I had the more nuanced star system.

[1] https://i.imgur.com/MOt3XlL.png

[1.1] How I'd rate these. Cars 3: not really interested. Liked the first though. Stitches: idk, doesn't look appealing. Teeth: Classic cult film but yeah... Big Mouth: I have ZERO interest in watching this, please stop suggesting. I have downvoted this! Waterboy: I like it, but far from 95%. I'll give it like an 80.

Re: How Cambridge Analytica’s Facebook targeting model really worked

#19

Post by emodendroket » Fri, Mar 30, 2018, 9:38 PM UTC

I am really puzzled by the Cambridge Analytica scandal. It's not particularly savory, but is there something happening here that it wasn't basically already known about how Facebook worked? By the protests of their own executive, the system was working as designed, and at worst Cambridge Analytica misled them about how they intended to use the data, right? There was no actual security breach here, as far as I can understand it.

Re: How Cambridge Analytica’s Facebook targeting model really worked

#20

Post by emodendroket » Fri, Mar 30, 2018, 9:39 PM UTC

daenz wrote: post #3

Am I understanding this correctly? Facebook user data (likes/profile info) was scraped to produce low-dimension feature vectors for users (similar to word2vec). These feature vectors were then run through some ML model to predict...what exactly? Targetability for effective political ads?

It seems like the purpose was narrowly tailoring messages, which is something political campaigns are really keen to do now (Obama's campaign was kind of a trailblazer here, right?).