{"id":43049,"date":"2018-03-30T13:57:07","date_gmt":"2018-03-30T20:57:07","guid":{"rendered":"http:\/\/world.edu\/?p=43049"},"modified":"2018-03-30T13:57:07","modified_gmt":"2018-03-30T20:57:07","slug":"how-cambridge-analyticas-facebook-targeting-model-really-worked-according-to-the-person-who-built-it","status":"publish","type":"post","link":"https:\/\/world.edu\/how-cambridge-analyticas-facebook-targeting-model-really-worked-according-to-the-person-who-built-it\/","title":{"rendered":"How Cambridge Analytica\u2019s Facebook targeting model really worked \u2013 according to the person who built it"},"content":{"rendered":"

\"\"
\nThe researcher whose work is at the center of the\u00a0Facebook-Cambridge Analytica data analysis and political advertising uproar<\/a>\u00a0has revealed that his method worked much like the one\u00a0Netflix uses to recommend movies<\/a>.<\/p>\n

In an email to me, Cambridge University scholar Aleksandr Kogan explained how his statistical model processed Facebook data for Cambridge Analytica. The accuracy he claims suggests it works about as well as\u00a0established voter-targeting methods<\/a>\u00a0based on demographics like race, age and gender.<\/p>\n

If confirmed, Kogan\u2019s account would mean the digital modeling Cambridge Analytica used was\u00a0hardly the virtual crystal ball<\/a>\u00a0a few have claimed<\/a>. Yet the numbers Kogan provides\u00a0also show<\/a>\u00a0what is \u2013 and isn\u2019t \u2013\u00a0actually possible<\/a>\u00a0by\u00a0combining personal data<\/a>\u00a0with machine learning<\/a>\u00a0for political ends.<\/p>\n

Regarding one key public concern, though, Kogan\u2019s numbers suggest that information on users\u2019 personalities or \u201cpsychographics<\/a>\u201d was just a modest part of how the model targeted citizens. It was not a personality model strictly speaking, but rather one that boiled down demographics, social influences, personality and everything else into a big correlated lump. This soak-up-all-the-correlation-and-call-it-personality approach seems to have created a valuable campaign tool, even if the product being sold wasn\u2019t quite as it was billed.<\/p>\n

The promise of personality targeting<\/strong><\/h4>\n

In the wake of the revelations that Trump campaign consultants Cambridge Analytica used\u00a0data from 50 million Facebook users<\/a>\u00a0to target digital political advertising during the 2016 U.S. presidential election, Facebook has\u00a0lost billions in stock market value<\/a>, governments on\u00a0both sides of the Atlantic<\/a>\u00a0have\u00a0opened investigations<\/a>, and a nascent\u00a0social movement<\/a>\u00a0is calling on users to\u00a0#DeleteFacebook<\/a>.<\/p>\n

But a key question has remained unanswered: Was Cambridge Analytica really able to effectively target campaign messages to citizens based on their personality characteristics \u2013 or even their \u201cinner demons<\/a>,\u201d as a company whistleblower alleged?<\/p>\n

If anyone would know what Cambridge Analytica did with its massive trove of Facebook data, it would be Aleksandr Kogan and Joseph Chancellor. It was\u00a0their startup Global Science Research<\/a>\u00a0that collected profile information from\u00a0270,000 Facebook users and tens of millions of their friends<\/a>\u00a0using a personality test app called \u201cthisisyourdigitallife.\u201d<\/p>\n

Part of\u00a0my own research<\/a>\u00a0focuses on understanding\u00a0machine learning<\/a>methods, and\u00a0my forthcoming book<\/a>\u00a0discusses how digital firms use recommendation models to build audiences. I had a hunch about how Kogan and Chancellor\u2019s model worked.<\/p>\n

So I emailed Kogan to ask. Kogan is still a\u00a0researcher at Cambridge University<\/a>; his collaborator\u00a0Chancellor now works at Facebook<\/a>. In a remarkable display of academic courtesy, Kogan answered.<\/p>\n

His response requires some unpacking, and some background.<\/p>\n

From the Netflix Prize to \u201cpsychometrics\u201d<\/strong><\/h4>\n

Back in 2006, when it was still a DVD-by-mail company, Netflix offered a\u00a0reward of $1 million<\/a>\u00a0to anyone who developed a better way to make predictions about users\u2019 movie rankings than the company already had. A surprise top competitor was an\u00a0independent software developer using the pseudonym Simon Funk<\/a>, whose basic approach was ultimately incorporated into all the top teams\u2019 entries. Funk adapted a technique called \u201csingular value decomposition<\/a>,\u201d condensing users\u2019 ratings of movies into a\u00a0series of factors or components<\/a>\u00a0\u2013 essentially a set of inferred categories, ranked by importance. As Funk\u00a0explained in a blog post<\/a>,<\/p>\n

\u201cSo, for instance, a category might represent action movies, with movies with a lot of action at the top, and slow movies at the bottom, and correspondingly users who like action movies at the top, and those who prefer slow movies at the bottom.\u201d<\/em><\/p>\n

Factors are artificial categories, which are not always like the kind of categories humans would come up with. The\u00a0most important factor in Funk\u2019s early Netflix model<\/a>\u00a0was defined by users who loved films like \u201cPearl Harbor\u201d and \u201cThe Wedding Planner\u201d while also hating movies like \u201cLost in Translation\u201d or \u201cEternal Sunshine of the Spotless Mind.\u201d His model showed how machine learning can find correlations among groups of people, and groups of movies, that humans themselves would never spot.<\/p>\n

Funk\u2019s general approach used the 50 or 100 most important factors for both users and movies to make a decent guess at how every user would rate every movie. This method, often called\u00a0dimensionality reduction<\/a>\u00a0or matrix factorization, was not new. Political science researchers had shown that\u00a0similar techniques using roll-call vote data<\/a>\u00a0could predict the votes of members of Congress with 90 percent accuracy. In psychology the \u201cBig Five<\/a>\u201d model had also been used to predict behavior by clustering together personality questions that tended to be answered similarly.<\/p>\n

Still, Funk\u2019s model was a big advance: It allowed the technique to work well with huge data sets, even those with lots of missing data \u2013 like the Netflix dataset, where a typical user rated only few dozen films out of the thousands in the company\u2019s library. More than a decade after the Netflix Prize contest ended,\u00a0SVD-based methods<\/a>, or\u00a0related models for implicit data<\/a>, are still the tool of choice for many websites to predict what users will read, watch, or buy.<\/p>\n

These models can predict other things, too.<\/p>\n

Facebook knows if you are a Republican<\/strong><\/h4>\n

In 2013, Cambridge University researchers Michal Kosinski, David Stillwell and Thore Graepel published an article on the\u00a0predictive power of Facebook data<\/a>, using information gathered through an online personality test. Their initial analysis was nearly identical to that used on the Netflix Prize, using SVD to categorize both users and things they \u201cliked\u201d into the top 100 factors.<\/p>\n

The paper showed that a factor model made with users\u2019 Facebook \u201clikes\u201d alone was\u00a095 percent accurate<\/a>\u00a0at distinguishing between black and white respondents, 93 percent accurate at distinguishing men from women, and 88 percent accurate at distinguishing people who identified as gay men from men who identified as straight. It could even correctly distinguish Republicans from Democrats 85 percent of the time. It was also useful, though not as accurate, for\u00a0predicting users\u2019 scores<\/a>\u00a0on the \u201cBig Five\u201d personality test.<\/p>\n

There was\u00a0public outcry<\/a>\u00a0in response<\/a>; within weeks Facebook had\u00a0made users\u2019 likes private<\/a>\u00a0by default.<\/p>\n

Kogan and Chancellor, also Cambridge University researchers at the time, were starting to use Facebook data for election targeting as part of a collaboration with Cambridge Analytica\u2019s parent firm SCL. Kogan invited Kosinski and Stillwell to join his project, but it\u00a0didn\u2019t work out<\/a>. Kosinski reportedly suspected Kogan and Chancellor might have\u00a0reverse-engineered the Facebook \u201clikes\u201d model<\/a>\u00a0for Cambridge Analytica. Kogan denied this, saying his project \u201cbuilt all our models<\/a>\u00a0using our own data, collected using our own software.\u201d<\/p>\n

What did Kogan and Chancellor actually do?<\/strong><\/h4>\n

As I followed the developments in the story, it became clear Kogan and Chancellor had indeed collected plenty of their own data through the thisisyourdigitallife app. They certainly could have built a predictive SVD model like that featured in Kosinski and Stillwell\u2019s published research.<\/p>\n

So I emailed Kogan to ask if that was what he had done. Somewhat to my surprise, he wrote back.<\/p>\n

\u201cWe didn\u2019t exactly use SVD,\u201d he wrote, noting that SVD can struggle when some users have many more \u201clikes\u201d than others. Instead, Kogan explained, \u201cThe technique was something we actually developed ourselves \u2026 It\u2019s not something that is in the public domain.\u201d Without going into details, Kogan described their method as \u201ca multi-step\u00a0co-occurrence<\/a>\u00a0approach.\u201d<\/p>\n

However, his message went on to confirm that his approach was indeed similar to SVD or other matrix factorization methods, like in the Netflix Prize competition, and the Kosinki-Stillwell-Graepel Facebook model. Dimensionality reduction of Facebook data was the core of his model.<\/p>\n

How accurate was it?<\/strong><\/h4>\n

Kogan suggested the exact model used doesn\u2019t matter much, though \u2013 what matters is the accuracy of its predictions. According to Kogan, the \u201ccorrelation between predicted and actual scores \u2026 was around [30 percent] for all the personality dimensions.\u201d By comparison, a person\u2019s previous Big Five scores are about\u00a070 to 80 percent accurate<\/a>\u00a0in predicting their scores when they retake the test.<\/p>\n

Kogan\u2019s accuracy claims cannot be independently verified, of course. And anyone in the midst of such a high-profile scandal might have incentive to understate his or her contribution. In his\u00a0appearance on CNN<\/a>, Kogan explained to a increasingly incredulous Anderson Cooper that, in fact, the models had actually not worked very well.<\/p>\n