In August 2017, a “Unite the Right” rally took place in Charlottesville, Va., motivated by the removal of Confederate statues from a nearby park. Attendees included neo-Nazis, members of the “alt-right,” and other far-right extremist groups. Violent clashes ensued between rally attendees and counter-protesters, resulting in the death of paralegal Heather Heyer when an attendee drove his car into a crowd of counter-protesters. In the aftermath, President Trump declared there to be “very fine people on both sides” . White supremacists celebrated Trump’s response, with one former leader stating that his comments marked “the most important day in the white nationalist movement” .
Did support for Trump’s address of the events in Charlottesville extend beyond white supremacists and the alt-right? Did the response split along partisan lines, or was there a more nuanced reaction? Marisa Eisenberg, Sarah Cherng, Mason Porter, and I investigate these questions by using network science and data science to examine the structure of the online Twitter conversation about Charlottesville .
Social media platforms are important mechanisms for shaping public discourse, and big data in the context of social media is a large and rapidly-growing area of research . While important differences exist between Twitter users and the overall domestic population, Twitter data are nonetheless important; Russian influence campaigns targeting Twitter in the 2016 presidential election are well-documented , and the platform has been key for white supremacist efforts to shape public discourse on race and immigration . These features, together with ease of data access, motivated our study of Twitter data .
We examined the network structure of nearly half a million tweets containing the hashtag #Charlottesville that were posted shortly after the Unite the Right rally. Twitter data have a natural network structure, as Twitter accounts can be linked to one another in several different ways, i.e., through follows, mentions, and retweets. We focused on the “retweet network,” where nodes signify Twitter accounts participating in the online conversation about Charlottesville, with a directed edge pointing from account \(j\) to account \(i\) corresponding to the number of times that \(j\) retweeted \(i\). In short, we treated the retweet network as a weighted, directed network.
Dimension Reduction: Node Characterization and Community Structure
The Charlottesville retweet network consists of about half a million edges and approximately 300,000 nodes — too much data to directly visualize. To deduce structure, we apply two dimension-reduction techniques: principal component analysis (PCA) and community detection. We use PCA to characterize the nodes in the network and community detection to simplify the network by identifying tightly-knit groupings of nodes. Together these tools allow us to study the structure of the online Twitter conversation and specifically examine divisions in the online response concerning the Charlottesville rally.
Our idea for node characterization is a simple one based on online media preferences. Studies indicate that preferred media sources vary with readers’ political affiliations . We make use of this knowledge by analyzing followership for different media accounts on Twitter. Specifically, we accessed complete lists of Twitter users who follow 13 media accounts spanning the conservative/liberal spectrum (e.g., from conservative outlets like FoxNews and BreitbartNews to liberal outlets like thenation and dailykos), and applied PCA to these “media choice” data. The resulting first principal component has a natural interpretation as left/right media orientation. PCA also provides a left/right media score, which allows us to extract a binary classification of the nodes as “left” or “right” based upon their media choice.
To reduce the network dimension, we next look for tightly-knit groups of nodes — known as “communities” in the retweet network. Many algorithms can detect communities [3, 5]. Deploying two widely-used methods reveals roughly 200 communities in the Charlottesville retweet network: a number small enough to allow for characterization and visualization when combined with left/right media orientation.
Divided We Tweet
Figure 1. Diversity of community composition (measured by the Shannon diversity index, which is commonly used to assess diversity in ecological studies) versus left/right media principal component analysis (PCA) score. Left-leaning communities correspond with negative values, and right-leaning communities correspond with positive values. Bubble size corresponds to community size.
The data show that the #Charlottesville conversation split markedly according to left/right media preference. More than 90 percent of retweets occurred between nodes of the same left/right media orientation, and communities are largely segregated in terms of media preference (see Figure 1). The largest communities on the left and right are nearly homogeneous. While there are exceptions—the Twitter accounts of Republican Senator Tim Scott and Republican Representative Carlos Curbelo are in left-leaning communities, for example—the Twitter data are consistent with the response to Charlottesville splitting largely along partisan lines.
Examination of tweet content shows that “Trump” is the most common word from both the left and right. However, the accompanying language is very different. For example, “Nazis” and “impeachment” often occur on the left, but “MAGA” and “POTUS” frequently occur on the right. The Twitter data show a very divided reaction in the aftermath of Charlottesville, with fault lines centered on the president.
Finally, we consider which groups on the left and right participated in the Twitter discussion following Charlottesville. We do this by characterizing the detected communities in terms of each community’s influential (central) nodes . Doing so shows right communities associated with white supremacist symbols and influential alt-right personas. These groups are consistent with the organizers and core participants of the Charlottesville rally. An additional community on the right, and in fact the largest right community in these data, includes well-known right-wing personas and FoxNews. The Twitter data thus show a highly-polarized response to Charlottesville that split largely according to media preference, with online support for Trump extending beyond neo-Nazis and the alt-right to include right-wing accounts with large online followings.
Joseph Tien presented this work during a minisymposium on the “Dynamics of Democracy” at the 2019 SIAM Conference on Applications of Dynamical Systems, which took place in May in Snowbird, Utah. A recording of the presentation is available here.
Acknowledgements: I am grateful to Mason Porter (University of California, Los Angeles) for providing detailed feedback on an earlier version of this piece.
 Boyd, D., & Crawford, K. (2012). Critical questions for big data: provocations for a cultural, technological, and scholarly phenomenon. Inform., Comm., and Soc., 15(5), 662-679.
 Daniels, J. (2018). The algorithmic rise of the “alt-right”. Contexts, 17(1), 60-65.
 Fortunato, S., & Hric, D. (2016). Community detection in networks: A user guide. Phys. Rep., 659, 1-44.
 Mitchell, A., Gottfried, J., Kiley, J., & Masa, K.E. (2014). Political polarization & media habits (Pew Research Center technical report). Retrieved from http://www.journalism.org/2014/10/21/political-polarization-media-habits/.
 Newman, M.E.J. (2018). Networks (2nd ed.). Oxford, U.K.: Oxford University Press.
 Shear, M.D., & Haberman, M. (2017, August 15). Trump defends initial remarks on Charlottesville: again blames ‘both sides’. The New York Times, p. A1. Retrieved from https://www.nytimes.com/2017/08/15/us/politics/trump-press-conference-charlottesville.html.
 Tien, J.H., Eisenberg, M.C., Cherng, S.T., & Porter, M.A. (2019). Online reactions to the 2017 ‘Unite the Right’ rally in Charlottesville: measuring polarization in Twitter networks using media followership. Preprint, arXiv:1905.07755.
 United States of America v. Internet Research Agency LLC. (2018). Case 1:18-cr-00032-DLF. 18 U.S.C. §§2, 371, 1349, 1028A. Retrieved from https://www.justice.gov/file/1035477/.