Follow Trump or Clinton on Twitter? These researchers might’ve analyzed your face.
Image: Flickr/Gage Skidmore
In the lead up to election night in 2012, there was rampant speculation that a top secret computer program could ultimately be responsible for the outcome of the race. Far from a conspiracy theory, the computer program, known as Project Narwhal, was the Obama campaign's big data solution to hyper-individualized voter outreach. It allowed the campaign to target voters based on a profile compiled from several different databases, emailing them about issues the program thought would matter most to them based on their profile characteristics.
Although Narwhal made use of highly complex algorithms to assemble voter profiles, a recent study published on arXiv demonstrated that voter analysis need not be so complex. In fact, a relatively sophisticated demographic analysis of a candidate's voter base requires little more than users' social media profile pictures.
The study, led by Yu Wang, a political science PhD candidate at the University of Rochester, analyzed the Twitter profiles of Trump and Clinton followers to determine their geographic location, social influence, gender, race, and age. If you were following either of these candidates between September 17 and December 22 last year, there's a decent chance you may have been included in Wang's analysis.
To begin their study, Wang and his colleagues analyzed the Twitter avatars of over 28,000 Clinton followers and 29,000 Trump followers using a machine learning algorithm that compared the avatars against a database of some 55,000 faces to determine the Twitter users' age, gender and race.
As far as gender goes, Wang found that women make up approximately 45 percent of Clinton's followers, which confirms previous observations of gender affinity voting (the likelihood of a voter to vote for someone of the same gender). Moreover, this finding stands in opposition to evidence that Clinton's support among female voters has dropped sharply in recent months.
Perhaps even more surprising, Trump has almost the exact same gender breakdown—about 45 percent of his analyzed followers were female, despite Trump coming under fire for his sexist comments about Fox News pundit Megyn Kelly.
The racial analysis of each candidate's Twitter followers was a different story. Wang found that Clinton supporters were more likely to be African American or Hispanic than Trump supporters and that the overwhelming majority of Trump supporters were white, a pattern that Wang says is "consistent with historical voting patterns" for the Democratic and Republican parties. As far as age goes, Trump's Twitter base seems to confirm the stereotype that Republicans tend to be old white people, although he also has a substantial following among the very young. Most of these users appeared to be too young to vote.
After using facial recognition software to analyze the race, gender and age of Trumpists and Clintonists, Wang and his colleagues examined the number of followers each of the Trump and Clinton supporters had as a measure of their social influence. They found that the social influence of Trumpists was more polarized than that of the Clintonists. Trump's followers tended to only have a few followers or several hundred followers, whereas Clinton's followers generally had somewhere between a few dozen and 200 followers.
According to Wang and his colleagues, such an analysis could provide important information for candidates, since race and gender have both been shown to be important factors in voting behavior. However, there are some obvious shortcomings to the study.
In the first place, not all Twitter users had avatars that can be analyzed by facial recognition software. Moreover, the number of followers analyzed in the study represents only a fraction of each candidate's total Twitter base—Clinton has over 5 million followers and Trump has over 7 million at the time of writing. There's also the fact that following someone on Twitter is not necessarily indicative of support (especially in the case of Trump, whose garbage fire of a Twitter account appeals to supporters and detractors alike for its entertainment value). And finally, the number of Twitter users pales in comparison to other social media platforms, so it can hardly be considered an accurate portrayal of the demographics of the American electorate.
Still, as facial recognition software becomes more accurate, its application to other social media platforms could be a useful tool for understanding the shifting allegiances of voters in real time and may very well prove to be a crucial campaign tool in future elections.