Strategic polarization? Filter bubbles, bridges and the role of politicians on Twitter during 2019 European Election in Slovenia

1 Introduction

Social media has become an important platform for political communication and discussion about public affairs, however, some key questions regarding the nature of increasingly personalized communication networks remain unanswered. Contemporary networked public spheres consist of a multitude of dispersed publics which are constantly forming, connecting with each other and dissolving according to the shared interests of their members. As Bruns puts it, the “public sphere is now irretrievably fractured into a multiplicity of online and offline, larger and smaller, more or less public spaces that frequently (and often serendipitously) overlap and intersect with one another” (Bruns, 2023, p. 70). Political communication takes place in a dynamic social media environment that can, paradoxically, at the same time foster connections to more stakeholders than ever before, while also being increasingly fragmented and polarized. This has led to increased warnings that citizens online tend to cluster into communities of similar and increasingly extreme views, sometimes referred to as filter bubbles (Pariser, 2011). Despite the popularity of the topic, there is no consensus about the prevalence of filter bubbles online. Several studies using various methods confirm the existence of closed communication spaces (Hong and Kim, 2016), while other influential scholars find that social media are in fact much more open environments, and claim that the idea of a filter bubble is nothing more than a buzzword (Bruns, 2019). In this study, we use quantitative analysis to study information diffusion on Twitter^[1] in the context of pre-election political discussions, focusing on an under-researched area of Central and Eastern Europe, namely Slovenia. We build on existing research on political polarization on social media and contribute to the field by (1) providing empirical data on the extent of polarization; (2) examining the role of bridges which enable communication flows between different ideological communities; and (3) examining the role of politicians in polarized debates.

2 Literature review

3 Methodology

We rely on social network analysis (SNA) to examine the structure of online political discussion on Twitter. Our dataset consists of a total of 40,670 tweets, posted by 2845 unique users. Data was collected in the period of 30 days before the election via Twitter Streaming API, using a set of 15 election-related hashtags and keywords: these were the 10 most commonly used hashtags marking the European Election campaign and 5 different grammatical variations of the word “election” in Slovenian language. Data analysis was performed with Python programming language (version 3.5.6.) using TSM module (Freelon, 2018); with network visualizations done in Gephi (version 0.9.2.) using ForceAtlas2 algorithm (Jacomy et al., 2014).

We observe the flow of communication with the retweet network, as a retweet is the primary affordance through which the information is shared between users on Twitter (Firdaus et al., 2018). To examine the degree of fragmentation, we use the Louvain community detection method (Dugué and Perez, 2022). In addition to quantitative SNA measures, we also examine the five biggest communities by looking at their top users and hashtags, which helps us determine the main characteristics of each community and label them according to the prevailing political sentiment. Several different topologies have been used for categorizing social media users. Following the example of Bracciale et. al (2018), we extracted from these topologies 7 categories that represent users’ roles in the political discussion: ‘Media/Journalists’, ‘Politicians/Political parties’, ‘Institutions’, ‘Political supporters’, ‘Citizens’, ‘Commentators’ and ‘Organizations’.

To answer RQ1, we use a number of SNA measures that highlight the level of fragmentation between groups of users. Modularity and clustering coefficient are basic measures to find groups in the network. High modularity indicates dense connections between nodes within communities and sparse connections between nodes in different communities. Similarly, clustering coefficient tells us the degree to which nodes in the network tend to cluster together. Next, we analyze information flows within and among communities by using the E-I index to look at the amount of homophilic and heterophilic interactions (Krackhardt and Stern, 1988). The measure provides a normalized comparison of the number of interactions within a community (internal or homophilic) with the number of interactions between one’s own community and outside communities (external or heterophilic). Finally, we observe how different communities relate to one another by examining ties between each possible pair of a multitude of communities and directions. We use proximity matrix (Freelon, 2018), where proximity between communities A and B is defined as the proportion of shared ties between A and B of all the ties involving at least one member of A. Resulting matrix tells us how close each community is to each other. The closer the two communities are, the more content is able to flow between them and the less they resemble a filter bubble. It is a “relational measure that captures the specific relationships between different clusters rather than only measuring the coherence of a single cluster” (Lynch et al., 2017, p. 10).

To answer RQ2, we need to identify users that are able to facilitate communication between different clusters and bridge the ideological divide. Identification of such users is not straightforward, as one needs to consider both the intermediary capacity as well as influence of the node. Most studies use betweenness centrality as the measure for identifying broker positions, however, this measure does not take into account the polarized community structure of the network. As pointed out by Freelon, in polarized online networks, “what matters is often less the absolute number of nodes a given node connects and more the specific subgraphs a node connects” (Freelon, 2018, p. 7). The criteria for the identification of influential bridges are two-fold: (1) such users should have a high in-degree (meaning that they are often retweeted by other users), and (2) a notable proportion of their incoming ties should come from one or more communities other than their own (meaning that they are being retweeted by members of other communities). Nodes meeting both criteria lie near the borders of different communities, serving as a point of connection between them (Freelon, 2018).

To answer RQ3, we use SNA measures to define influential users, namely indegree, outdegree, eigenvector centrality and betweenness centrality. In the context of the retweet network, users with highest indegree are users that are most often retweeted – which is seen as an indicator of relevancy (Dubois and Gaffney, 2014). Outdegree reveals users that are retweeting the most and is an indicator of activity. The measure is connected to the concept of superparticipants (Graham and Wright, 2014), users that are the most active in disseminating content and may have the power to shape online discussions. We use eigenvector centrality as a measure that quantifies importance of a node in the network. Its score is higher when node’s connections are in turn highly connected, thus the measure is suitable for finding opinion leaders (Dubois and Gaffney, 2014). Lastly, betweenness centrality tells us how often a certain node is positioned on a path between two other nodes (Freeman, 1977). Users with high betweenness centrality can help establish connections among nodes that are not otherwise connected, thus the measure can help identify users that play a gatekeeper role (Ghajar-Khosravi and Chignell, 2017).

4 Results

To examine communication flows, a total of 27,561 retweets were mapped into a retweet network consisting of nodes (users) and 14,483 unique edges (retweets). Observed network has an average path length of 5.515 and an average clustering coefficient of 0.103. A random network of the same size following Erdős–Rényi model (Erdős and Rényi, 1960) would consist of 13,927 edges, have an average path length of 4.313 and an average clustering coefficient of 0.003. Compared to a random graph, the Slovenian retweet network has similar shortest path length, but a higher clustering coefficient. Average path length is within the expected range for social media networks, while a higher clustering coefficient suggests the presence of clusters, each of which has dense connections within. Modularity (Q = 0.426) is relatively high (Newman and Girvan, 2003), which further confirms fragmentation of the network. We identified 49 distinct communities in the observed network, out of which we closely examined the five biggest ones (together representing 92 % of all users and 98 % of all retweets). Network visualization can be seen in Figure 1.

Observed communities differ in size, intensity of communication, shared content and dominant ideological views.

• Community 1 consists of 39 % of all users and is the biggest community in the network. Its most prominent members are mainstream media accounts, state institutions and liberal or left/centre-left parties and politicians, including parties that formed a government coalition at the time; namely Lista Marjana Šarca (LMŠ, Renew), Social Democrats (SD, S & D) and the Left (Levica, European Left). We label this cluster “Left-wing, Mainstream”.

• Community 2 consists of 18 % of all users. Among the most prominent members are politicians and official accounts of the right-wing Slovenian Democratic Party (SDS, EPP), the biggest opposition party at the time of data collection. We label this cluster “Right-wing”.

• Community 3 consists of 26 % of all users. Among the most prominent members are right-wing politicians, opinion leaders and partisan media outlets, as well as content sympathizing with the opposition Slovenian Democratic Party (SDS, EPP). We mark this cluster “Right-wing supporters”.

• Community 4 consists of 6 % of all users. The most prominent members include politicians from the center-right Christian Democrats (NSi, EPP) as well as journalists. We label this cluster “Center-right”.

• Community 5 consists of 3 % of all users and therefore is the smallest community analyzed. It features far-right content and politicians (Alt-right, identitarian and populist accounts), as well as a group of foreign, likely astroturf profiles. We mark the label cluster as “Far right, Populist”.

Figure 1:

Retweet network. Dots represent users, lines represent retweets, colors represent communities.

We then calculated E-I index for each community to see whether retweets stay within communities or spread outside their borders. The three biggest communities (representing 84 % of all users) have a negative E-I index, which means that most of the retweets stay within their own community. This shows that communication happens largely within like-minded circles as users tend to predominantly share content posted by members of the same community. Nevertheless, all five of the biggest communities maintain at least some ties with other clusters, meaning that users are not entirely isolated from receiving information originating from other parts of the network. Observed communication patterns hint at fragmentation and polarization, but not isolation.

A look at the proximity matrix (Table 2) reveals which communities share information. Each cell depicts a share of ties involving the community mentioned in the row that are shared with the community mentioned in the column, divided by all ties involving the community mentioned in the row. Diagonal cells represent a share of internal connections. Reciprocal off-diagonal cells (e.g., 0–1 and 1–0) share the same numerators, however, their values vary significantly, because denominators are almost always distinct. Since we are observing retweets, the proximity matrix shows us that some communities tend to rebroadcast content from certain communities rather than others. The matrix reveals that Community 1 (mainstream, left-wing) has a low number of connections with other parts of the network, while Communities 2 and 3 (both right-wing) share a relatively large number of retweets among them. Numerous ties between these two communities are also seen in the network visualization, where both communities appear close together on the right side of the network.

In RQ2 we look at bridges, that is nodes that receive large numbers of ties from multiple communities. Table 3 presents the ten most prominent bridges in the network. The second column lists the categorization of the user, the third column lists their in-degree, the fourth column lists communities to which each user is connected (descending by number of ties), the fifth column lists the number of these connections, and the last column depicts the ratio of the tie count of the user’s second most-connected community to that of his own community. The extended table including user names can be found in Appendix B.

Bridges detected by this method are a diverse mix of commentators, partisan supporters and active citizens. Nearly all of them are members of the two big right-wing communities, and they (with the exemption of number 10^[2]) did not help to bridge content across the ideological divide. Instead, identified bridges facilitate information flow between ideologically similar communities (2 and 3); thereby linking different parts of a broader right-wing area of the network. This is clearly seen in the network visualization, where the bridges identified and their ties are highlighted (Figure 2): except one, all prominent bridges lie at the borders of ideologically similar Communities 2 and 3, facilitating exchange of information between them.

Figure 2:

Prominent bridges and their connections (highlighted) in the retweet network. Rectangles mark communities.

In RQ3 we examine what roles politicians play in polarized discussions on social media. We identify central users in the network, map them with community structure, and examine their communication patterns. The results show that political accounts play various roles in online discussions – most often as opinion leaders and gatekeepers, but rarely as most active users. When we consider communities to which these influential accounts belong, notable differences arise. The vast majority of the most active and most influential users in the network belong to right-wing Communities 2 and 3. This shows that members of the right-wing communities are more active, and their central accounts in turn receive considerably more attention and influence than central accounts from other communities (see Appendix A for the list of most influential users according to different centrality measures across five observed communities). The visibility of politicians within their communities varies greatly across the network. In the right-wing clusters, politicians dominate the discussion – while in the center-left part of the network, politicians do not hold central roles. The difference is most notable when comparing Communities 1 (left-wing, mainstream) and 2 (right-wing). As shown in Table 3, influential users in Community 1 are mainly media, commentators and institutions, while political accounts rarely make the list (despite the fact that all major political parties from the center-left and liberal spectre belong to this community) – hinting that politicians in this cluster struggle to set the agenda of the discussion. On the other hand, politicians and party accounts belonging to right-wing SDS party dominate discussion in Community 2, as they consistently score among the most influential accounts (Table 5). Extended tables with included usernames can be found in Appendix A.

5 Discussion

Communication patterns of Twitter users involved in a political discussion reveal communities of users with different political views emerging from the topological structure of the networks. The ideologically left-leaning part of the network consists of one big but sparsely connected community, while the tightly connected right-leaning part consists of ideologically similar conservative communities. Both parts of the network have roughly equal numbers of users, however, members of the right-wing clusters are disproportionally more active. On average, users belonging to right-wing communities are also able to transmit information to a larger audience than users belonging to left-leaning part of the network.

Results show that both right and left-leaning parts of the network tend to share information coming from ideologically congruent sources and that information hardly spreads between communities with different views. However, all communities observed maintain at least some ties with other clusters, which suggests that they are not completely isolated from encountering opposing views. Observed communication patterns confirm high communicative fragmentation and ideological polarization, but not isolation, as proposed by the filter bubble hypothesis (Pariser, 2011). Interestingly, this relates to empirical findings by Williams (2015), who analyzed Twitter discourse on climate change and found both open forums as well as entrenched homogenous communities of users engaging in negative discourse with the opposing side. Results are also in line with Törnberg et al. (2021), who see social media as polarizing not by enabling users to completely avoid opposing views, but by providing spaces for conflictual interaction. Even if social media users do encounter opposing views, worries about the consequences of like-minded communities remain, as pro-attitudinal exposure has been shown to foster polarization and increase hostility towards users with different views (Kubin and von Sikorski, 2021). Following these arguments, we want to avoid simplistic and techno-determinist understanding of filter bubbles as completely sealed-off communicative environments (e.g. Pariser, 2011; Sunstein, 2017). Nevertheless, empirical results show that while not completely isolated, homogeneous communities of like-minded users do form on Twitter, and provide further evidence for fragmentation of what was once perceived as the public sphere into a multitude of highly polarized, more or less overlapping public spaces, as predicted by Bruns (2023).

The results are partially in line with some authors who see the online environment as a sphere of interest driven communities, linked with intermediaries – users who act as bridges and facilitate communication flows between different groups (Bruns and Highfield, 2016). Bridges hold a key role in fostering exchange of opinions and enable what some see as a “complex system of distinct and diverse, yet inter-connected and overlapping publics” (Bruns and Highfield, 2016, p. 127). The results obtained by analyzing Slovenian Twitter users diverge from this thesis. We show that bridges between communities exist, however, prominent intermediaries are only bridging ideologically similar communities. From a theoretical perspective identifying bridges could well reveal users who are able to bring ideologically opposed communities together by promoting shared experiences and presenting common ground for cross-cutting debate. However, these results indicate the opposite – bridges, while present, do not foster a cross-cutting exchange of views, but help to establish communication flows among already connected and ideologically similar communities. In this case, prominent bridges simultaneously promote cohesion between similar communities as well as polarization of different ones, thereby contributing to deepening, not reducing, fragmentation of the network. These results encourage us to rethink not only what influence certain users have on the spread of information in the network, but also to reconsider some of the methodological approaches used for identifying and examining influential bridges. While basic measures from the SNA toolkit (such as betweenness centrality) could be used to easily identify nodes that act as links between different groups of users, more nuanced analysis is needed to interpret the results and confirm whether such users actually perform the roles of information brokers between ideologically different communities, thereby enabling cross-cutting information flows and reducing the effect of filter bubbles. The empirical approach used in this paper presents a potentially more robust measure for identifying influential brokers in polarized online networks.

Looking at central users across different communities, we find that the left-leaning part of the network shares content coming from a diverse mix of institutions and mainstream media outlets, while right-wing clusters seem to predominantly amplify content posted by politicians. Right-wing politicians are among the most prominent users in the network, and are able to dominate the discussion within their communities – they receive more attention and are more influential compared to politicians coming from the left-leaning part of the network. These politicians generally hold a considerable influence over content diffusion, albeit their influence remains limited to ideologically similar communities, while they do not facilitate communication across different clusters or attract sympathies of the wider network. This shows that some politicians, especially those belonging to right-wing communities, are using the fragmented social media environment in order to effectively spread their partisan messages among selected target groups and to mobilize their supporters. Politicians can gain visibility and rally support by sharing information that is ideologically aligned with the positions of their community and lands well with community members (Marwick and Lewis, 2017). While this study does not engage in qualitative analysis of messages posted, it shows that right-wing politicians are able to attract attention and visibility among their followers significantly better than left-wing politicians. The observations from this study are in line with global communication trends among conservative parties, which are moving away from the center and strategically using partisan discourse to attract attention, mobilize their supporters and engage with communities who hold more fringe and controversial views (Ceia, 2020; Kubin and von Sikorski, 2021), further increasing political polarization (Tucker et al., 2018).

The analysis revealed another interesting observation, namely that the mainstream media accounts are clustered within the same community as the government parties. The finding might suggest that the mainstream media are ideologically aligned with the left-wing government. Media alignment might be explained by the Slovenian political context, in which ruling coalitions most of the time have been formed by left-wing parties, giving them the opportunity to develop closer ties with the mainstream media. On the other hand, a network of alternative media outlets has emerged around right-wing political parties. While these partisan outlets are influential in the right-wing communities, their content does not appear to be shared across the ideological divide. Similar findings have been observed in other political contexts. Based on the use of social media by Spanish political actors, Ceia (2020) noted that parties frequently share links to content coming from a homophilic set of sources, which allows them to interpret events through their own frames. This finding empirically confirms the link between social media polarization and the polarization of the media system. Social media contributes to political polarization by creating communities of like-minded users, limiting exposure to diverse viewpoints, and reinforcing ideologically congruent media consumption. This polarization is further amplified by the role of political elites and partisan media, creating a complex interplay between social media dynamics and the broader political and media system (Kubin and von Sikorski, 2021). This is in line with the observations made within the American context by Wilson et al. (2020), who find that social media, political elites and partisan media collectively contribute to rising polarization and can create a self-perpetuating cycle of animosity and ideological division.

The study focuses on an underreached context of political communication in Central and Eastern Europe, namely Slovenia. There are structural reasons to believe that patterns of polarization on social media, observed mainly in the context of the U.S. (Kubin and von Sikorski, 2021), might differ in a different political (multiparty democracy, proportional voting system, coalition governments) and media system (significant media fragmentation along political lines). Nevertheless, empirical observations from Slovenia confirm similar patterns of ideological fragmentation compared to other Western democracies, thus providing further evidence that polarization in political discourse on social media is a prominent feature of the contemporary communication environment, present across various national contexts as well as different political and media systems.

Some limitations should be considered when interpreting the results. First, the study focuses on one social media platform in a specific context (pre-election period). We cannot generalize such studies to be representative of the national or global context – where social media is only one pillar in an increasingly connected hybrid media system in which political debate unfolds. Nevertheless, similar limitations apply to any case study of communication on social media. If we are to understand the characteristics of the contemporary communication environment, we need to first understand what is happening on individual platforms in a variety of specific contexts, as these platforms are becoming increasingly important arenas for political communication. Second, tools offered by SNA can uncover various communication patterns, however, many point out that the quantitative approach cannot sufficiently describe complex social phenomena. A more detailed study of interactions involving qualitative analysis of shared messages, images, videos and URLs would help to further illuminate the nature of online polarization. Lasty, the state of polarization on social media is far from fixed – new platforms that arise and become adapted for public communication can bring new discursive practices and interaction patterns.

6 Conclusion

Political communication on social media is often described as increasingly fragmented and polarized, with users forming closed communities of likeminded peers (so-called filter bubbles). The phenomenon has received broad academic attention, however, open questions about its extent in different contexts remain. We contribute to this debate by studying the structure of political discussion on social media in Slovenia, drawing on computational methods that go beyond the usual social network analysis toolkit. In line with similar studies, we find that political discussion among Slovenian Twitter users is highly fragmented and divided into distinct, ideologically homogeneous communities. Right-wing and left-wing parts of the network are roughly equal in size, however, right-wing clusters are considerably more active. The analysis of communication patterns and central users suggests that users tend to share messages from like-minded peers, and that communication between ideologically different communities is rare. At the same time, links between different parts of the network exist, thus rejecting the idea of isolated information spaces known as filter bubbles and empirically confirming the thesis about online environments increasingly fragmenting into a multitude of overlapping publics with rising polarization and disagreements between ideologically opposing groups. Second, we find that bridges – prominent users who foster connections between different communities – are present in the network, however, they predominantly enable information flows only among ideologically similar communities. The intermediaries who would facilitate an exchange of views among ideologically opposing communities are almost non-existent, which further confirms warnings about the dissolution of the networked public sphere into separate, unconnected groups of users with similar worldviews. Influential bridges identified in this study are thus contributing to deepening, not reducing, fragmentation of the network, which prompts us to rethink how we identify and examine users holding broker roles in the network. Third, we find that right-wing politicians are among the most prominent users in the network and are able to dominate the discourse within their communities – which cannot be said for left-wing politicians. This shows that right-wing politicians effectively use Twitter as a strategic communication tool to engage with their supporters, spread their partisan messages and make use of the polarized social media environment. The study provides quantitative evidence of the extent of polarization in a networked public sphere, and adds a comparative perspective by evaluating global communication trends in the context of a small, European multi-party democracy, where such topics have not been extensively studied before. The results provide further evidence that polarization in political discourse on social media is a prominent feature of the contemporary communication environment, present across various political and media systems.