In this day and age, social media has infiltrated every aspect of our life, expanding beyond the social realm and into the personal, financial, and even political sphere. Research by Manish Raghavan, a current computer science Ph.D. student, and Prof. Jon Kleinberg, computer and information science, delved into just how social media and politics interacted leading up to the 2016 presidential election between Hillary Clinton and Donald Trump.
On first glance, it may seem like social media platforms amplified the echo chamber effect in the months leading up to the election because, according to Raghavan, these platforms tend to “give you what they think you’re going to like.”
Raghavan said that the algorithms used to make active content recommendations to users “do play a role in making you feel that the world is a particular way, when that’s really the world that they’ve curated because they think it’s the world you want to see.”
However, Raghavan and Kleinberg’s research calls into question the political significance of echo chambers in the months before the election. Indeed, one of the most counter-intuitive conclusions drawn in their paper is that cross-spectrum political interaction actually spiked in the months leading up to the election.
While the researchers analyzed user behavior on both Twitter and Reddit to measure political interaction, Raghavan said Twitter is more helpful than Reddit in understanding large-scale sociopolitical patterns.
“Twitter is just bigger, which means it’s easier to find these patterns,” said Raghavan, as opposed to Reddit, where “the data just isn’t there… it’s far too noisy, there’s far too little data to draw such conclusions.”
According to their paper, user behavior was modeled by a type of mathematical network known as an invocation graph.
An invocation graph, as defined and studied by Kleinberg and Raghavan, is a directed, weighted graph between political domains like The New York Times or Breitbart, where a directed edge is formed when a social media user invokes domain X to reply to a user who posted domain Y on their social media, denoted X -> Y. The weight of this edge is the amount of times this invocation occurs.
To further explain, a directed graph is one where the edges are somehow asymmetrical; the network of all Twitter users forms a directed graph, where an edge from person A to person B, denoted A -> B, signifies that A follows B, but not necessarily the reverse. A directed graph is said to be weighted if the edges also carry some numeric component; for example, the network of all Paypal transactions forms a directed, weighted graph, where the weight is the amount in the transaction and the direction is from the sender to the recipient.
Raghavan discussed other types of graphs, such as other forms of directed graphs which can be used to study the influence of social media on politics. For example, he discussed how other researchers have examined how political article authors may reference one another in their writing. If author A were to reply to author B, then a directed edge would be formed from A to B.
The invocation graph is unique, though, in that it allows researchers to focus on “what insights you can actually get from the way that these [domains] are being used,” said Raghavan. The distinction lies in measuring the intended impact of the authors’ articles versus the real-world usage of the articles by readers.
When asked if the main cause behind the counter-intuitive conclusion lay in the structure of the networks or election-driven sociological changes, Raghavan responded, “[I] wouldn’t say it’s a network effect because that same network is present before and after the election … what has changed is how much people care, and how much people are trying to change each other’s minds.”
Courtesy of Prof. Jon Kleinberg and Manish Raghavan
Across domains | This graph indicates the correlation between how pro-Trump a domain is and how often it is rebutted by other domains.
Raghavan discussed some of the challenges faced during the research process, which included difficulties in measuring and collecting appropriate data.
“There’s some real-world level insight that you’re trying to pull out, how to translate that down into a mathematical formulation, into something that you can actually run on a dataset, that was a challenge for us,” Raghavan said.
In contrast, one of the benefits of this large-scale network analysis is that, in some ways, it can confirm our intuitions about the categorization of certain political domains.
“You can put a number to the politicization of how these news sources are being used and actually how that kind of reflects your prior belief of what you thought it would have been,” Raghavan said.