Big data and Brexit/Trump

Just a more precise and invasive method of psychological modeling. It relies on the existence of personality types (which is itself a flawed concept) in order to place individuals on a map via "like" association (the more data, the more methods). It interprets rather than predicts, though.
 
I posted the vid with the Cambridge Analytics CEO explaining the methods a couple of weeks ago. People thought that it wasn't that influential, Cruz even supposedly fired them.
 
Mciahel Goodman said:
Just a more precise and invasive method of psychological modeling. It relies on the existence of personality types (which is itself a flawed concept) in order to place individuals on a map via "like" association (the more data, the more methods). It interprets rather than predicts, though.
Click to expand...

It predicts in a statistical sense. If you establish that 77% of people in your sample who like Wu Tang Clan are heterosexual etc, you can predict a distribution from it.

You can however argue the validity of the prediction. As with all time-series, the representativeness can be dubious. In particular in anything but the short term.
 
I'll give this article a read -- second time today I've been pointed in its direction.

Analytical driven insight tends to sound a lot more scarier than it is in reality. I work for an online lender in the UK and my job completely revolves around using customer behavioural insights for predicting outcomes that maximize profit -- how much time do you spend on the site, are you single, how fast do you fill out the application.. etc..

It all boils down to fairly simple stuff but I've read articles that make it seem like the devil's work.
 
I very, very quickly skimmed the article because I was curious. What is described in the article is rudimentary stuff. There are companies all around us that are doing at least stuff that is as advanced as what is described in the article. And there are a number that are doing stuff that is far more advnaced.

There are profiles of all of you -- especially in countries where data is well organized -- that can be bought and profiled for targeting. In countries like India, it is not easy to do that.
 
zing said:
I very, very quickly skimmed the article because I was curious. What is described in the article is rudimentary stuff. There are companies all around us that are doing at least stuff that is as advanced as what is described in the article. And there are a number that are doing stuff that is far more advnaced.

There are profiles of all of you -- especially in countries where data is well organized -- that can be bought and profiled for targeting. In countries like India, it is not easy to do that.
Click to expand...

Could you elaborate on the more advanced concepts?
 
Counterargument here

 
Javi said:
Could you elaborate on the more advanced concepts?
Click to expand...

I know you asked about advanced concepts but let me start with simple examples for anyone else reading this who may be interested. You're sitting on redcafe and there must be ads on this page(I'm not sure, I have adblock enabled). Niall needs revenue for monetization of people, who are clicking through on his ads. A simple random experiment would involve Niall sticking the ads by the side instead of the bottom to see if that drives more clicks. This is likely to yield different results and one of this is going to be a better outcome for Niall. Next, he could segment it.. maybe placing it at the bottom works better for mobile users? Iteratively, you can get to more advanced concepts. You could move on to uplift modeling. As opposed to determining what treatment maximizes revenue on average for a customer segment, can I identify specific variables that allow me to say given x, y, z characteristics, what ad position works better for a segment?

That's sort of what the article is talking about but I am a little skeptical about what it purports to sell(how are they measuring success? I can think of ways in which they can do that but.. I have also seen agencies trying to sell us 'solutions' like the persona crap the article talks about which is garbage). For some other examples you can easily relate to https://blog.kissmetrics.com/how-netflix-uses-analytics/ -- is a good read because this is a product company that generates enormous amount of usable data. The word 'usable' being very important because I have seen so examples of companies crawling on Twitter and other social media data which turned up useless.

On a slight tangent, while not leveraging data in it self the way the above examples are, something like IBM's Watson is a very powerful use of machine learning. Imagine a human brain reading the words "hot dog" and you know it's about food because of your knowledge/experience that allows you to contextualize it. How does a computer processing the word know it's not about a dog that's hot? Now scale it up and IBM's watson is literally processing freeform text to generate answers to these questions -- the kind of advancement required to do that is leagues above what data backed insights like what the article is talking about.
 
donkeyfish said:
It predicts in a statistical sense. If you establish that 77% of people in your sample who like Wu Tang Clan are heterosexual etc, you can predict a distribution from it.
Click to expand...
That's my point. The inference comes after the fact instead of before.
 
I can think of some ways data science would be trialed in an election but I'm not sure how many of them usable. I am sure some of them are employed already. Measuring outcome on a population is easy. Tailoring it and getting to a stage where you can do uplift modeling and STILL measure outcomes. that is where the problem is.

Here's an approach Trump could take:
Run a poll. 20 pounds if you give your name, date of birth, which candidate you are going going to vote for.
You can use the personally identifiable information to back it up into some credit bureau's data files.
Now slam a region with marketing activity -- maybe talk about how Obamacare sucks in one region, talk about how guns are awesome in another region.
Do a poll again, try to get results from same people. Now you know what kind of people are likely to change their opinion, given the marketing message is a) about Obamacare b) guns being awesome.
Now, for each person, you know associated characteristics(single, likely to visit gambling sites, blah blah) from the credit bureau files and some other third party data. Some uplift modeling here to understand what kind of segment is more likely to change their opinion if you're saying Obamacare sucks.
Now, instead of telling everybody Obamacare sucks, purchase data from your credit buerau or whomever and based on the segments you've defined, just run different messages. You have their e-mail IDs -- put it out on Facebook. You have their phone numbers, put it out on Twitter.

Your marketing investment is likely to give you better outcomes than a blanketed one-size fits all targeting. The problem is.. how do you know it has actually influenced election results? You'll never really know as you can't A/B test an electoral outcome.

Like I said, this is just about being a bit creative. It's not advanced math.

I do think the forces that influence elections are a lot more complicated than this.
 
Last edited:
You must log in or register to reply here.