Jiebo Luo and Yu Wang did not set out to predict who would win the 2016 U.S. presidential election. However, their exhaustive, 14-month study of each candidate鈥檚 Twitter followers鈥揺nabled by machine learning and other data science tools鈥搊ffers tantalizing clues as to why the race turned out the way it did.
Unlocking big data
A Newscenter series on how Rochester is using data science to change how we research, how we learn, and how we understand our world.
鈥淲e wanted to understand how each of the candidate鈥檚 campaigns evolved, and be able to explain why someone won or lost,鈥� says Luo, an associate professor of computer science.
Luo and Wang, a dual PhD candidate in political and computer science, summarized their findings in eight papers during the course of the campaign, including these observations:
- The more Donald Trump tweeted, the faster his following grew鈥揺ven after he performed poorly in debates against other Republican candidates, and even after he sparked controversies, such as proposing a ban on Muslim immigration.
- When Trump accused Hillary Clinton of playing the 鈥渨oman card,鈥� women were more likely to follow Clinton and less likely to 鈥渦n-follow鈥� her during the week that followed. But it did not affect the gender composition of Trump followers.
- Moreover, a 鈥済ender affinity effect鈥� seen in other elections鈥搘omen tending to vote for women鈥揹id not appear to be working for Clinton as the primaries drew to a close. The percentage of female Twitter followers in the Clinton camp was no larger than that in the Trump camp. Moreover, though 鈥渦n-followers鈥� were more likely to be female for both candidates, the phenomenon was 鈥減articularly pronounced鈥� for Clinton.
- At the same time, several polls, including ABC/Washington Post and CBS/New York Times, suggested that some Bernie Sanders supporters might 鈥渏ump ship鈥� from the Democratic column, and end up voting for Trump if Sanders dropped out. Luo and Wang found supporting evidence, reporting that the number of Bernie Sanders followers who were also following Trump was increasing鈥揵ut the number also following Clinton was declining. The dual Sanders/Trump followers were also disproportionately (up to 64 percent) male.
鈥淚n the end, even though we chose not to make any predictions, we were not surprised at all that Donald Trump won,鈥� says Luo.
Why Twitter?
Barack Obama鈥檚 use of social media in the 2008 presidential race helped establish Twitter and other social media platforms as powerful tools for candidates to quickly reach and receive feedback from large numbers of potential voters鈥揳nd to attack their opponents.
Since then, there鈥檚 been a burgeoning interest in scholarly research employing data science to analyze elections based on social media postings.
Twitter, in particular, is a rich source of data because the millions of tweets posted by its members each day are easily accessible using an application programming interface.
The key for Luo, Wang, and their colleagues was to collect as much of this data as possible, starting early in the campaign, and to then 鈥渕ine鈥� it in innovative ways.
鈥淭he very nature of this data is that it will disappear tomorrow, so we had to start capturing it from an early stage and design a research framework so we could continue to collect data all along,鈥� said Wang.
From September 2015 through October 2016, the team began accumulating a huge data set that included:
- The number of Twitter followers of each of the major candidates in the initially crowded field鈥搖pdated every 10 minutes.
- 8 million tweets sampled from the followers of Clinton and Trump.
- 1 million images of the candidates鈥� followers on Twitter.
- 5 million Twitter IDs that include all candidate followers in early April 2016.
Using advanced computer vision tools, the researchers trained an artificial neural network (what’s called a convolutional neural network) to determine鈥搘ith 90 percent accuracy or more鈥搕he age, gender, and race of the candidates鈥� followers using their Twitter photos. This helped the researchers analyze the role of each of those factors in the campaign, as they tracked the changes in each candidate鈥檚 followers before and after debates, for example, and how followers reacted to the candidates鈥� own tweets.
Twitter mining has its limits compared to the responses gleaned from traditional telephone polling. There鈥檚 no opportunity to ask follow-up questions, for example, and tweets are difficult to place geographically, limiting their application for studying trends in swing states. (Even geotagged tweets may be sent while the sender is on vacation or attending a rally in another state.)
But Twitter mining also has its advantages鈥揺nabling researchers to quickly, continually, and inexpensively sample data on a scale that far surpasses the 1,000 or so responses that pollsters increasingly struggle to gather using traditional techniques. In one study, for example Luo and Wang were able to characterize 322,116 Trump or Clinton followers who subsequently became 鈥渦n-followers.鈥�
鈥淭his is an approach that is broadly applicable,鈥� Luo says. 鈥淚f you want to test public reaction to the next generation of iPhones, or to a new model of car, you can use the same approach to see what consumers like or don鈥檛 like. It enables us to track millions of people and get reliable readings on their preferences.鈥�
Other Election 2016 papers by Luo, Wang, and their colleagues look at:
- 听
- 听
- 听
- 听
- .
