Twitter connections between white nationalists and Attorney General Sessions

On February 8, 2017, the United States Senate confirmed Jefferson Sessions as U.S. Attorney General. This appointment has been widely criticized over Sessions’s civil rights record and hard-line stance on immigration. The Trump administration defends the appointment, citing Sessions’s vote to confirm Attorney General Eric Holder and efforts to award Rosa Parks the Congressional Gold Medal. It has also been reported that Sessions is well-liked by his colleagues in the Senate.

Twitter data shows that someone else likes Sessions, too: white nationalists.

Our findings have been reported by Mother Jones. Here, we describe our methods and results in more detail.

Senators in Twitter space

We have seen previously that useful information can be gleaned from examining co-followership of Twitter accounts. In our previous blog post, we used publicly available data from Twitter to look at network relationships between nodes of interest; these focal nodes included white nationalist groups, media sources, and politicians. Here we turn our attention to Twitter relationships between white nationalist groups and U.S. Senators, again using tools from network science. We included Twitter accounts for 9 white nationalist groups and all U.S. Senators, giving a total of 109 focal nodes. We obtained a list of followers for each focal node  using the Twitter API1Twitter data pulled in a two week time span from November 22 to December 6, 2016, giving more than 11.4 million unique followers in total. We used these data to construct a co-follower network between these focal nodes using the methods described previously.

To illustrate that the co-follower network conveys meaningful information, we used community detection. Specifically, we performed modularity maximization using a Louvain algorithm to look for structure in the sub-network consisting of only the 100 Senators (see Figure 1). Community detection yielded three groups: one group consisting entirely of Democrats together with the independent Angus King, and the other two consisting almost entirely of Republicans.2The following are exceptions: Senator Tester is a Democrat placed in the first Republican community, and Senator Donnelly and Senator Heitkamp were placed with the second Republican community. This partisan homogeneity within the detected communities indicates that the co-follower network is indeed capturing important features of the focal nodes and the political interests of their followers.

Figure 1. Communities in the network of senator co-followership.

Senators and white nationalists

To examine the relationship between Senators and the white nationalist community (the set of 9 white nationalist focal nodes), we compute the distance between these two groups on the co-follower network. Specifically, we calculate the geodesic (shortest path) distance from Senators to the white nationalist community (see Figure 2), with distance across a directed edge equal to the reciprocal of the edge weight between nodes.3We computed shortest paths between focal nodes using Dijkstra’s algorithm. To compute the distance to the white nationalist community, we calculated a weighted average (with weights according to PageRank) of the distance to each white nationalist focal node.

Figure 2. Distance between focal nodes in the co-follower network.

Figure 3 shows the resulting distance from each Senator to the white nationalist community, with color indicating a senator’s ideology score (as computed by GovTrack.us).4An analysis of distance in the other direction, from the white nationalist community to Senators, gives similar but less dramatic conclusions.. There is a strong correlation between follower size and distance, with larger numbers of followers associated with greater distances from the white nationalist community.5This reflects the way the co-follower network is constructed, as the edge weights in the co-follower network correspond to the proportion of shared followers. Thus, if node A has a far larger number of followers than node B, the maximum for the edge weight from A to B is the number of followers of B divided by the followers of A (giving a small number). Senator Sessions had nearly 86,000 followers at the time of data retrieval, ranking 22/100 among the Senators. Despite this reasonably large following, the distance from Sessions to the white nationalist community was the closest of all Senators. Indeed, Sessions appears as an outlier in Figure 3. This closeness of Sessions to the white nationalist community reflects an unexpectedly high degree of co-followership. For example, more than 7% of followers of white nationalist nodes also follow Sessions, including 10% of followers of Matthew Heimbach6 Heimbach was featured recently in a New York Times article on the `alt-right’ and 9% of followers of ‘The Right Stuff’. This is the 6th highest proportion of co-followership with the white nationalist community among all Senators, and the five Senators with higher proportions all have far larger followings on Twitter (on the order of 1 million followers or more).7 The six Senators with highest white nationalist co-followership are, in order, Rand Paul, Ted Cruz, Marco Rubio, Bernie Sanders, John McCain, and Jeff Sessions.

Figure 3. Co-followership distance from senators to the white nationalist community. Color indicates each senator’s ideology score, with blue indicating more liberal and red more conservative.
Mouse-over any data point to see the Senator’s name.

 

What other factors influence the distance from Senators to white nationalists? Party and ideology are obvious candidates, and there appear to be some relationships between these factors and distance. For example, ideology scores closer to 1, indicating more conservative legislative behavior, may correlate with lower white nationalist distances for a given number of followers (i.e. the bottom edge of the dots appears to be redder). A similar pattern can be seen if we color the data by party instead of ideology score. (An interesting exception is Mitch McConnell, who has a very large distance to white nationalist communities given his number of followers. This may indicate a broader range of network ties, perhaps due to his position as Senate Majority Leader). However, in both cases, there is quite a bit of variation in placement, and more analysis would need to be done to confirm these observations.8For example, one could do a regression analysis of ideology score with white nationalist distance, controlling for number of followers.

Geography is another potential factor, but it shows no obvious pattern. Figure 4 shows a map of the average distance from the Senators in each state to the white nationalist community. Any influence of geography, ideology, or party appears secondary to the effect of follower size, where Sessions is a clear outlier.

Figure 4. Co-followership distance from Senators to the white nationalist community by state (averaged over both Senators from each state).

Why is Sessions an outlier? Our data cannot answer that question. Nor can our analysis indicate Sessions’s personal views or the stances that he will take as U.S. Attorney General. But the data do show that white nationalist groups have a strong interest in our new Attorney General.

References   [ + ]

1. Twitter data pulled in a two week time span from November 22 to December 6, 2016
2. The following are exceptions: Senator Tester is a Democrat placed in the first Republican community, and Senator Donnelly and Senator Heitkamp were placed with the second Republican community.
3. We computed shortest paths between focal nodes using Dijkstra’s algorithm. To compute the distance to the white nationalist community, we calculated a weighted average (with weights according to PageRank) of the distance to each white nationalist focal node.
4. An analysis of distance in the other direction, from the white nationalist community to Senators, gives similar but less dramatic conclusions.
5. This reflects the way the co-follower network is constructed, as the edge weights in the co-follower network correspond to the proportion of shared followers. Thus, if node A has a far larger number of followers than node B, the maximum for the edge weight from A to B is the number of followers of B divided by the followers of A (giving a small number).
6. Heimbach was featured recently in a New York Times article on the `alt-right’
7. The six Senators with highest white nationalist co-followership are, in order, Rand Paul, Ted Cruz, Marco Rubio, Bernie Sanders, John McCain, and Jeff Sessions.
8. For example, one could do a regression analysis of ideology score with white nationalist distance, controlling for number of followers.

Twitter without the tweets: white nationalists, Breitbart, and Trump

Following the election, there has been a surge of interest in white nationalists and the ‘alt-right’. Critics claim that President-elect Donald Trump courted and provided a platform for white nationalists. Spokesperson Kellyanne Conway has dismissed these claims as sour grapes and political correctness run amok. Trump’s choice of Steve Bannon as chief strategist in particular has been a lightning rod for criticism. Bannon, of course, was former chief executive of Breitbart, which has attracted attention for controversial columns and as a home for the ‘alt-right’. Discussions of Bannon’s personal views on race and whether Bannon is anti-semitic have yielded conflicting accounts, perhaps distorted through a political lens.

Given the venomous rhetoric and misinformation that are pervasive in today’s polarized political landscape, it can be hard to resolve disparate accounts such as these. Here we take a scientific approach, looking directly at what the data say concerning associations between Trump, Steve Bannon, and white nationalists. Specifically, we analyze publicly available data from Twitter, using tools from network science.

The Twitter network

Donald Trump is notably fond of Twitter.

Twitter has a natural network structure: the nodes of the network correspond to Twitter users, and an edge or link joins two nodes together if one user follows the other. For example, if user A follows user B, an edge points from A to B in the network. Lots of edges point to Donald Trump, as @realDonaldTrump has more than 15 million followers — but many other individuals and organizations are active on Twitter. For example, Breitbart News has about 500,000 followers, and David Duke has more than 24,000.

Data on who follows whom on Twitter is publicly available. These are the data we use, to look for patterns and structure in the network.1Twitter data pulled in a two week time span from November 22 to December 6, 2016. The Twitter graph changes over time as followers are added or deleted, for example. Here, we look at followers of specific Twitter accounts of interest, called focal nodes. These focal nodes are made up of Twitter accounts of three different types:

  • news organizations
  • politicians
  • white nationalist groups

For news organizations, we included mainstream organizations with lots of followers (e.g. Washington Post, Fox News, NPR), as well as organizations on the left (e.g. dailykos, the Nation) and the right (e.g. the Blaze, Drudge Report, Breitbart News). A total of 13 news groups spanning the political spectrum were included.

Politicians included Paul Ryan and the 6 U. S. Senators with more than 1 million Twitter followers: Ted Cruz, Marco Rubio, John McCain, Cory Booker, Elizabeth Warren, and Bernie Sanders.

To identify white nationalist groups, we used the Southern Poverty Law Center’s listing of extremist groups, searching under ‘white nationalist’ and ‘racist skinhead’. Google searches for these groups yielded 7 twitter accounts. To this list we added David Duke, as well as the Twitter account associated with the white nationalist site ‘The Right Stuff’.White nationalist propaganda from this site was posted on at least two college campuses following the election. Both David Duke and The Right Stuff have over 20,000 Twitter followers apiece.2Richard Spencer is a prominent white nationalist whose twitter account was suspended at the time of this analysis.

To top things off, we included Donald Trump. In total, we have 30 focal nodes, with more than  42.7 million total unique followers and nearly 74 million edges — quite a bit of information.

To look at things in a simple manner, we construct the mutual follower network. The nodes in this network consist only of the focal nodes. The strength of the tie between two nodes is determined by the number of co-followers they share, i.e. people who follow both focal nodes. However, the same number of, say, 100 shared followers might mean something very different for a node with a large number of followers vs. one with only a few. There are thus several ways to represent this data, shown in the figure below. For the analysis here we use a directed graph, similar to the bottom example in the figure. To get the edge weight from one focal node to another, we look at the proportion of the followers of one focal node that also follow the other node. For example, if 20% of the followers of A also follow B, we would have an edge in the mutual follower network pointing from A to B with weight 0.2. So this gives a weighted, directed graph, where the edge weights indicate the connection strength between nodes.

Different ways to view the twitter network of shared followers.

Different ways to view the twitter network of shared followers.

So what we’ve done is extract information from Twitter based purely upon who follows who — Twitter, without the tweets. This is perhaps the simplest way to look at Twitter relationships between focal nodes from a network standpoint.

Communities in the network: News polarization and white nationalists

To understand the network structure, we use tools from network science to detect communities within the Twitter network.3Specifically, we use modularity maximization via a Louvain algorithm. First, to give a sense that the community detection algorithm works for the mutual follower network, we consider just the 13 news organizations. Running the algorithm results in 3 groups: Left (Mother Jones, dailykos, the Nation, Christian Science Monitor, 538), Right (the National Review, the Blaze, Drudge Report, Breitbart News), and Mainstream (Wall Street Journal, Washington Post, Fox News, NPR).4Mainstream in part reflects large twitter followings. There is certainly variation in left / right leanings within this community. So community detection successfully recovers polarization of the news. This is hardly a conclusion that requires tools from network science to reach, but it serves as a nice check that the network as we’ve constructed it and the methods we are using to study it yield sensible results.

Twitter network: Polarization in the news. Three communities emerge, shown in blue (left side of figure), grey (center), and red (right).
Double-click any node to highlight just its connections (double-click again to turn off highlighting).
Drag nodes to clarify connections and make it easier to see things.

 

What happens if we run community detection on the entire mutual follower network? In this case, the white nationalist nodes all group together. This is expected, at least for the seven groups identified as extremist by the Southern Poverty Law Center — but the fact that the community detection algorithm correctly places these known extremists in the same group gives us confidence in the algorithm. David Duke and The Right Stuff are also placed with these white nationalist groups, again consistent with what we knew beforehand.

Communities among focal nodes. Three communities emerge, shown on the left (purple), middle (red) and right (orange).
Double-click any node to highlight just its connections (double-click again to turn off highlighting).
Drag nodes to clarify connections and make it easier to see things.
Mouse-over smaller nodes to see node names.

In addition to the white nationalist community, the algorithm yields two more communities, the purple and red communities shown above. The red community corresponds to Breitbart News, the Drudge Report, the Blaze, the National Review, and the Republican members of congress Ted Cruz, Paul Ryan, and John McCain. The purple community corresponds to the Nation, dailykos, Mother Jones, 538, the Christian Science Monitor, the Washington Post, Wall Street Journal, Fox News, NPR, Democratic senators Cory Booker, Elizabeth Warren, and Bernie Sanders, as well as Republican senator Marco Rubio. The purple community is not strictly partisan -- instead, it is a combination of the ‘Left’ and ‘Mainstream’ news communities detected earlier, together with politicians from both parties.

But the real surprise comes from the president-elect: Donald Trump gets placed in the same community with the white nationalist groups. It’s important to note that this does not mean that a lot of Trump's followers also follow these extremist groups (almost all Trump followers do not). Rather, it means that followers of the extremist groups are very likely to also follow Trump. Indeed, 73% of all white nationalist Twitter followers also follow Donald Trump. If we compute this percentage using a weighted average of the arc weights from the white nationalist nodes to Trump (with weights according to PageRank), we find a similar number—on average, 75% of followers of the white nationalist twitter accounts also follow Trump.

Co-followership between the alt-right/white nationalists and Donald Trump. 73% of followers of white nationalist Twitter accounts also follow Trump, although these followers make up only a small fraction (<1%) of Trump's overall followers.

 

Whether or not Trump has provided a platform for white nationalists, we can reasonably conclude that

White nationalist groups really like Donald Trump.

This is a conclusion drawn purely from the Twitter data on who follows whom.

Breitbart news and white nationalists

What about the connection between white nationalists and Breitbart? We can again use the mutual follower network to examine this question. For each node within the white nationalist community (as detected above), we look at the proportion of followers of this node that also follow a particular news source. The figure below shows news preferences among the different extremist groups.


Co-followership between the alt-right/white nationalists and news organizations.

 

If we rank each news source in terms of its popularity among followers of an extremist node, the results are striking: 5 out of 9 extremist focal nodes prefer Breitbart, 3 focal node groups prefer Fox News, and 1 prefers the Drudge Report.5Meaning that out of all news sources considered, the highest proportion of followers of this extremist node also follow Breitbart. This preference for Breitbart is in spite of the fact that Fox News has a far larger following on Twitter (approximately 12 million) than Breitbart (about 500,000). There is overlap between followers of different extremist nodes, but even if we examine the overall group of all 21,302 people who follow both a white nationalist node and one of the three top news sources, we find that Breitbart is still the most popular — 66% follow Breitbart, 61% follow the Drudge Report, and 53% follow Fox News.6Many people follow more than one news source, which is why the percentages don’t add up to 100%. If we look only at the 3 most central nodes within the extremist community, these all have Breitbart ranked number one.7One can see the same results using several different centrality measures, such as PageRank and degree centrality.

Breitbart is the preferred news source for white nationalists.

Whatever Steve Bannon's personal views, the Twitter data directly show that he has shaped a company that white nationalist groups prefer.

Summary

There is quite a lot of information in Twitter, even without considering the tweets. And these data show that Donald Trump has a significant following among white nationalists, as does Breitbart News as shaped by Steve Bannon.

Updated Jan. 27, 2017 to correct an error regarding the community assignment of Fox News.

References   [ + ]

1. Twitter data pulled in a two week time span from November 22 to December 6, 2016. The Twitter graph changes over time as followers are added or deleted, for example.
2. Richard Spencer is a prominent white nationalist whose twitter account was suspended at the time of this analysis.
3. Specifically, we use modularity maximization via a Louvain algorithm.
4. Mainstream in part reflects large twitter followings. There is certainly variation in left / right leanings within this community.
5. Meaning that out of all news sources considered, the highest proportion of followers of this extremist node also follow Breitbart.
6. Many people follow more than one news source, which is why the percentages don’t add up to 100%.
7. One can see the same results using several different centrality measures, such as PageRank and degree centrality.