The Spatial Structures in the Austrian COVID-19 Protest Movement: A Virtual and Geospatial User Network Analysis

Abstract

The emergence of the COVID-19 pandemic, followed by policy measures to combat the virus, evoked public protest movements world-wide. These movements were formed not only in the virtual world but also through local protest gatherings. In contrast to previous research that studied movements in the virtual world through digital network analysis, this study recognizes the importance of the spatial dimension of social movements through local interaction. We therefore introduce a large-scale spatial–social network analysis of a georeferenced Twitter user network to understand the regional connections and transnational influences of the local movement through the virtual network. Our findings indicate that the virtual social network is distinctly structured along geographic and linguistic boundaries. Furthermore, our analysis of transnational influences reveals that the connections within Austria itself hold greater significance compared to their impact on external regions.

1. Introduction

The spread of COVID-19 in 2019 affected the lives of people around the world. National governments in numerous countries tried to prevent a global pandemic by introducing new policy measures, e.g., lockdowns and curfews, oftentimes leading to local protests and demonstrations against them (Kriesi and Oana 2023). Social media platforms gained popularity for the protest discourse during that time (Jarynowski et al. 2020), and their global reach enabled the spread of (mis-)information across protest participants (Chen et al. 2022; Muric et al. 2021). As a result, emerging “communities of fate” (Held 1997) developed a transnational understanding of their political claims, giving local COVID-19 protest movements a transnational dimension.

Some researchers proclaimed the “death of geography” (Bates 1999) in the early days of broad accessibility to the Internet, a theory that would render geospatial distance irrelevant due to the closeness in virtual space. Yet during the COVID-19 pandemic, activists still connected digitally and across borders with local, regional, national and international movement partners, organizing protest and negotiating political claims (Fominaya 2022). This created a virtual network structure in social media as well as a spatial structure according to the users’ physical locations.

Research around location-based social networks (LBSNs) in recent years has already established how methods of spatial network analysis can help gain insight into, for example, cultural differences between regional and inter-regional communication (Arthur and Williams 2019). However, most research studying COVID-19 protest movements has focused largely on the social network analysis, disregarding the spatial component (Ahmed et al. 2020; Hung et al. 2020). The combination of both components into a spatial–social network analysis helps in understanding which geographic regions are connected and if connections tend to extend to a transnational degree. This understanding, in turn, brings insight into the geopolitical relationships of digital protest mobilization.

The Austrian COVID-19 protest movement stood out due to its broad and heterogeneous mobilization in the Austrian population (Brunner et al. 2021) and also showed pronounced similarities to the COVID-19 protest movements in the two German-speaking neighboring countries of Switzerland and Germany (Nachtwey et al. 2020). On the other hand, research suggests that US protest movements originating in right wing groups against the “plandemic” spread via social media communities to Europe (Fominaya 2022). In order to understand these regional, national and transnational digital embeddings of the Austrian COVID-19 protest movement, our study contributes an extensive spatial–social network analysis for a Twitter user network of protest actors. We structure this analysis along the following research questions:

• How does the online social network manifest in geographic space?
• How are different geographical regions connected through the social network?
• How does the framework of socio-spatial network analysis help in understanding the transnationality of the underlying movement?

To answer these research questions, we construct a Twitter user–follower network using a snowball crawling system. Followership between users in Twitter can be modeled as a network graph: nodes represent Twitter users and the directed edge between them represents the followership. This information about a user’s followers, as well as the user’s specified location, is retrieved via Twitter’s API. By integrating the location information into the user network, we employ established techniques of spatial–social network analysis to gain novel insights into the geographic manifestation of the Austrian COVID-19 movement. We expect, for instance, a spatially non-uniform distribution of the network nodes and edges. The node distribution is likely to follow the population distribution, resulting in certain regions exhibiting a higher concentration of users and connections. We further assume that regions sharing linguistic similarities will show stronger connections through the social network. This phenomenon might be particularly evident among the German-speaking nations of Germany, Austria, Switzerland and Liechtenstein. In terms of transnational movement, we presume a significant influence from Austria to its neighboring German-speaking countries. The extent of Austria’s impact on the United States represents a particular research aspect.

2. Background and Theory

The potential global reach and connectivity of digital media, combined with the emergence of increasingly global issues, are linked to the development of “transnational” political online communication (Pfetsch et al. 2019), which refers to public communications “that go across borders and transform these borders by establishing structures and cultures of communication that exist beyond the interaction of national states” (Brüggemann and Wessler 2014). Research on transnationality has primarily focused on social movements and important global issues, such as globalization-critical protest movements (Conover et al. 2013), the environmental movement (Hellqvist 2022), and terrorist and far-right movements (Doerr 2017). Early approaches have tended to emphasize the potential benefits of social media and a shift from spatial organization to digitally networked activism (Sorce and Dumitrica 2022). Although digital media play a significant role in the development of a transnational protest, evidence increasingly suggests that political online communication also remains heavily localized in geographic space. Shared language is an important prerequisite for interactions between social media users, including for political debates and mobilization (Bastos et al. 2013; Takhteyev et al. 2012). Online networks also occur significantly along shared national identities and cultural backgrounds (Arthur and Williams 2019; Casero-Ripollés et al. 2020), and national states often remain the main target for social movements, as they usually represent the immediate responsibility for global problems (Brunnengräber 2012, p. 44). At the local level, (Casero-Ripollés et al. 2020) have shown that users in densely populated cities and in geographic proximity to political centers of power are more actively engaged in political debates. The digital communication networks of social movements are therefore fundamentally localized in spatial contexts but at the same time contain a transnational potential.

3. Related Work

Social network analysis to understand a social phenomenon has grown to an established approach to social and geospatial research. For example, Ahmed et al. (2020) study the spread of a specific conspiracy theory using social network analysis on keyword-filtered Twitter data. Their focus lies on determining central users by calculating their betweenness centrality scores, revealing users that have actively engaged in sharing conspiracy theories. Ref. Rauchfleisch and Kaiser (2020) present a study that analyzes a YouTube network of conspiracy-related channels that focused on a far-right political spectrum. They use a snowball-style crawling of the platform’s recommender system, where first a list of YouTube channels with politically far-right content is collected. In a second step, they retrieve recommended channels from this initial start list. Finally, the channels are connected with a directed edge if a channel is recommended by another. The methods applied here include a filtering of the network with a backbone extraction method (Serrano et al. 2009), leading to statistically significant edges, running a clustering algorithm to identify communities in the network and analyzing relevant comment topics. They are able to identify distinguishable communities with this approach. A methodologically similar study has been performed for the German Querdenken movement deriving network data from Telegram (Zehring and Domahidi 2023), identifying an influence of QAnon and far-right communities in the movement. Haupt et al. (2021) examine the COVID-19 Liberate protest movement by building up a network with Twitter users represented as nodes and network edges with retweets. They classify nodes as protest supporters and non-supporters using the Biterm Topic Model. Based on that classification, the authors compare the network of supporters and non-supporters separately, finding that the protest supporter network shows stronger influence of specific user nodes, thereby being more centralized.

All of the mentioned studies above focus on the analysis of the virtual online social network, disregarding the network’s spatial component. Scellato et al. (2010) analyze a Twitter user network with methods of spatial network analysis and collecting data with a snowball system. In order to describe each network’s manifestation in geographic space, they introduce the concept of node locality and a geographic clustering coefficient. The former is a way to quantify a node’s local embedding in contrast to global interactions, and the latter introduces a weighing based on node distance to the known clustering coefficient (Boccaletti et al. 2006). Using these metrics, the authors find that Twitter users are more likely to interact with users that are spread globally instead of users that are at closer distances. In a follow-up work, Scellato et al. (2011) analyze further spatial properties of LBSNs. They investigate the average geographic distance between users and investigated the probability of a link between network nodes as a function of their geographic distance. In a second step, the results are compared to two different statistical null models: one keeping the geographical location and shuffling the links, and the other one vice versa, keeping the links but shuffling locations. A notable finding here is that the occurring triangles, i.e., a direct connection between all three nodes, are spread on a wide geographical scale and that the length of a link does not affect the probability of belonging to a social triangle. Another study looking into the spatial properties of a Twitter network focuses on comparing the social and spatial distance of users within the US (Stephens and Poorthuis 2015). For each user in the network, an ego-alter subnetwork is created, and its transitivity and density are calculated. The results show that with increasing distance between the ego and its alters, the network density and transitivity decreases. This indicates that the closer the ego is to its alters, the higher the social clustering is. Another observation is that those networks with a standard distance of less than 500 km have high transitivity, which decreases with physical distance to a constant level. The authors draw the conclusion that the 500 km mark shows where information in the network flows most efficiently and hypothesize that this is most probably due to a common factor in the ego–alter relationship, like an offline relationship. Although much research has been performed on social networks structures, to the best of our knowledge, no publication has studied the Austrian COVID-19 protest with the methods of a socio-spatial network analysis, comparing the social and spatial network structures.

4. Methods of Spatial–Social Network Analysis

4.1. Overall Workflow

This study aims to understand the Austrian COVID-19 protest movement by means of spatial–social network analysis of a Twitter user network using the workflow displayed in Figure 1. The network data of user–follower relationships as well as each user’s location was retrieved using Twitter’s API. After geocoding and preprocessing the data, we aggregated the user network spatially and calculated network centralities, applied a community detection algorithm and performed a spatial hot spot analysis. The following sections will describe the single steps in more detail.

4.2. Data Retrieval

We retrieved the network data of user–follower relationships using Twitter’s API with a snowball crawling procedure, which started crawling the follower network from a selected set of start or seed users. A meaningful selection of seed users is important especially with regard to the transnationality analysis (Sorce and Dumitrica 2022, p. 167), as those users determine the network’s overall structure and ensure its relevance to the protest context. For this analysis, potential seed users were identified through extensive qualitative research of Austrian Corona protest media coverage. The selected protest actors had to have a Twitter account and had to be located in Austria based their profile information. Their active engagement with the protest was ensured through manual inspection of their Twitter accounts, specifically active posting and content re-tweeting behavior as well as the amount of followers. Furthermore, seed users had to reflect the heterogeneity of the protest actors with regard to their ideological background and functional role (Borbáth 2023). We emphasized selecting seed users that were opinion leaders on Twitter in terms of followers and in terms of the social media engagement they retrieved. The selected five seed users were characterized by the following:

• Right-wing and anti-system ideological backgrounds;
• Functional roles of politicians, experts and citizens;
• Different numbers of followers based on their overall follower distribution (few <500, medium 500–5000, many >5000).

We retrieved the network data of user–follower relationships using Twitter’s API with a snowball crawling. First, the followers of the seed users and then their followers were retrieved. More than one iteration of followers needed to be crawled to be able to observe a connected network structure instead of several unconnected trees. Crawling more iterations, however, might weaken the effects of interests in the network that are due to the selected seed users. For example, the current literature reports a value for the degree of separation as 4.59 in Twitter networks (Watanabe and Suzumura 2013). After retrieving the followers of these initial actors, the network was constructed. Users were modeled as network nodes, and directed edges were added from a user to their followers.

4.3. Location Information and Geocoding

The data retrieved in the previous step contain the user–follower relationships of the social network. As the goal of this study is to understand the spatial properties of the user network, information about each user’s location was needed. In this study, we assume that a user’s location is the one specified in their location profile (Serere et al. 2023). This information can be retrieved using Twitter’s API and then be passed to a geocoding service, which translates the textual information into geographic coordinates. OpenStreetMaps’s Nominatim was used for this purpose. All user nodes that did not specify a location were removed from the network as were all users with locations that could not be geocoded. The resulting georeferenced network consisted of approximately 800,000 nodes distributed over 160,000 unique locations and with 1.2 million connecting edges.

Table 1. Number of users, edges, and locations in the network before and after filtering for location and geocoding.

Characteristics	Absolute Number (%)
Users
Unique Users	1,609,811 (100%)
With location information in profile	884,227 (54.93%)
Geocodable location information	786,541 (48.86%)
Edges
Unique directed edges	4,325,470 (100%)
Edges between geocodable nodes	1,182,474 (27.34%)
Locations
Unique locations	244,239 (100%)
Locations with geocodable reference	159,506 (65.31%)

summarizes the network structure. The amount left after filtering for location was approximately 49% of all user nodes and 28% of all edges.

Less than one percent of the user nodes were geocoded into areas that have a population count of zero according to the world population dataset provided by (Center For International Earth Science Information Network-CIESIN-Columbia University, 2018). These user locations, e.g., “Antarctica” and “Bir Tawil”, appeared to be fake information upon manual inspection of the user profiles, and nodes with such locations were dropped. Furthermore, 13% of all users provided country or continent names as location information, which were geocoded as the centroid of the countries or continent’s bounding box. As this location information was too imprecise for analyzing regional connectivity, those nodes were removed as well.

4.4. Spatial Aggregation of User Network Data

One goal within this study is to understand the connectivity of geographical regions through the virtual user network. This can be achieved by defining geographical regions in the form of a regular grid, where each cell represents a region, and then aggregating the point network onto the grid layer by counting the connections between users per grid cell. This results in a network of connected grid cells, which allows to study the connections of regions through the underlying user network. Aggregation of the network was performed such that the original information of the user network was kept in the resulting grid cell network by applying the following procedure:

• The original number of users falling into each grid cell was counted and assigned as node weight to the grid cell in the resulting aggregated network.
• Two cells in the grid network were connected with a directed edge between them if there was at least one user–follower relationship with a user and follower located in the respective cells.
• The total number of directed connections between two grid cells was counted. This number was assigned as an edge weight in the resulting grid network.

Depending on the research question, we performed aggregations of the original user network into three distinct spatially aggregated networks. For instance, understanding the transnationality of the movement introduced the concept of nations into the analysis. In order to look into connections between nations, we aggregated the user network into the administrative borders of the world’s countries.

The question about the connectivity of regions through the user network was approached by aggregating the network nodes on a global hexagonal grid with a cell size of 80,000 km². Choosing an evenly sized grid as opposed to clustering points into variable cluster sizes was important in this analysis, as we aimed to calculate several network centralities and compare their magnitude and their spatial distribution. Using hexagons was beneficial for visualizing the results. The choice of cell size was of importance, as it impacted the results of the evaluation based on the spatial aggregation (Wong 2004). The size of a region for the world-wide analysis was chosen by visual analysis considering that a single cell should not cover whole or several European countries, as we studied an Austrian protest movement and wanted to identify regional connections. Analyzing the properties of this network indicated areas of interest (AOIs). Those areas could then be further examined by filtering the user network geographically for user nodes and edges that fall into the AOI and performing the spatial network analysis for Europe on an even finer grid with a cell size of 100 square kilometers.

4.5. Spatial Network Analysis

The aggregation in the previous steps resulted in a network of connected regions. We calculated four different centrality metrics to indicate regions of importance in the grid cell network.

The number of edges connecting a node is called the degree centrality (Freeman 1978) and reveals its importance in terms of the number of connections it has. As we investigate a Twitter user network, it is important to note that the user–follower relationships are not symmetric. When a user follows another, they retrieve information from the other’s feed but not vice versa. This means that the edge direction has to be regarded, and in- and outgoing edges have to be counted and evaluated separately. A node with an out-degree of N represents a user who is followed by N other users, whereas an in-degree of N means that this node follows N other users and retrieves their feed. In other words, the direction of the edge represents the direction in which information travels in the network.

Determining how close a node is to other nodes in the network can be performed with the closeness centrality (Freeman 1978; Wasserman and Faust 1994). Network nodes with a high closeness centrality value have the shortest path length to other nodes in the network and are therefore “closer”. For the grid network, this means that cells or regions that have a high closeness centrality are able to spread information through the network more efficiently. The betweenness centrality (Brandes 2001; Freeman 1977) indicates the importance of a network node with regard to information flow: nodes with a high betweenness centrality lie more frequently on the shortest path that connects two other nodes in the network. In the aggregated network, cells with a high betweenness centrality can be interpreted as regions with a high amount of control over the information flow.

Community detection algorithms detect clusters of nodes in the network such that nodes within the group are more densely connected with each other than with the rest of the network. Communities in the grid cell network can be regarded as geographical regions that are more strongly connected to each other than to other regions. There are many possible methods for detecting communities, and we used the Louvain algorithm (Blondel et al. 2008). This method optimizes on the network’s modularity and is particularly known to be a fast algorithm even on large networks (Mothe et al. 2017).

For the calculated centralities in the grid cell network, we performed a hot spot analysis using the Getis-Ord Gi* method (Getis and Ord 1992) in order to detect spatial clusters of central regions. These clusters are important, as they mark regions that are central for the underlying user network of COVID-19 protest actors.

5. Results

The results of the analysis are presented in this section in a descriptive fashion, whereas a thorough interpretation of the results will follow in the Discussion section. We first present the spatial properties of the retrieved user network and then walk through results of centrality calculation, community detection and hot spot analysis on the differently sized aggregated networks.

5.1. Spatial Properties of the Twitter User Network

Figure 2 shows a visualization of the network using a so-called edge bundling algorithm (Ersoy et al. 2011), where edges aligned based on similar positional information are bundled into “highways” of connections. The line width indicates the number of edges that follow along a connecting route: the more edges there are, the thicker the line. We can see that the users in the network are distributed over most populated parts of the world, and many connections lie within Europe, within the US, and also between those two continents.

In Figure 3, we can see the normalized geodesic edge length distribution using a bin size of 115 km, calculated using the Freedman–Diaconis rule. The distribution is right-skewed, peaks at around 350 km, and shows a second, significantly smaller increase in samples between 7000 and 8000 km. The dip between 3000 km and 6000 km delineates a “ring” around Europe defined by the Atlantic and Indian Oceans.

5.2. Spatial Aggregation on Country Level

A chord diagram visualizes the connections between countries through the network in Figure 4. For reasons of clarity, only the 15 most frequently occurring edges which make up 67% of all edges are visualized, finding that 44% of all edges occur within the same nations, namely, Germany (22.98%), the US (11.71%), and Austria (9.41%). In total, 4.73% of edges lead from Austria to Germany, 3.21% from Germany to the US, and 1.71% from the UK to the US.

In the next step, we examine the edges that originate in Austria, which indicates the countries with followers of Austrian users and, in terms of transnationality, which nations Austria potentially influences.

Table 2. Destination countries of edges out of Austria ordered by relative amount in % (left) and normalized by country population (right).

Absolute Number (%)		Normalized by Population (%)
Country	(%)	Country	(%)
Austria	46.89	Austria	60.96
Germany	23.63	Other	14.66
Other	11.73	Liechtenstein	11.21
US	7.27	Germany	3.28
UK	2.92	Switzerland	2.99
Switzerland	2.22	Luxembourg	1.83
France	1.52	Gambia	1.26
Canada	1.38	Monaco	1.14
Netherlands	0.82	Andorra	0.96
Spain	0.81	Ireland	0.93
Italy	0.81	Iceland	0.78

summarizes the total number of edges in percent that originate in Austria. It can be seen that the majority of edges originating in Austria (46.9%) are inner-Austrian edges, followed by Germany with 23.65% and the US with 7.27%. Other countries, like the UK, Switzerland, and France, each have less than a 3% share of the edges that start in Austria. When looking into the number of edges normalized by the country’s population, we find that the inner-Austrian edges dominate even more strongly and make up 60.96% of edges. Smaller countries like Liechtenstein, Switzerland and Luxembourg become more apparent compared to the absolute count of edges. Notably, the amount of edges going out to Germany is significantly smaller when looking into the normalized count (3.28% normalized compared to 23.63% absolute).

5.3. Spatial Aggregation on a Global Grid

We calculate the in- and out-degree, closeness and betweenness centralities of the grid cell network next. Cells with a high centrality indicate regions that have a high potential of influence over the information flow in the network. The distributions of the calculated centralities are shown in Figure 5. Betweenness and out-degree have heavily right-skewed distributions, where 87.72% of cells have a value of zero.

As the value distributions are not the same and not normally distributed, we use Spearman’s rank coefficient to test for a monotonic relationship between those four centralities. The resulting correlation matrix is visualized in Figure 6. It shows that betweenness and out-degree, as well as closeness and in-degree, are strongly related with correlations of 99.9 and 98.9%, respectively.

Knowing that the mentioned centrality pairs form a monotonic relationship in virtual space, the question arises if high or low values of the centralities also coincide in the same grid cell. We therefore cluster the centrality distributions into three value ranges: all cells with a centrality value of zero are assigned to the lowest bin. This is especially important for out-degree and betweenness centrality, as the majority of cells have a value of 0. For the remaining values, we calculate the median and assign the lower half of the distribution to the middle cluster and the upper half to the high cluster. Those low, middle, and high clusters of the two centralities are then visualized using a bivariate map as shown in Figure 7, where regions of high–high and low–low clusters become visible. For betweenness and out-degree centralities, we found a spatial cluster of cells falling into the high–high bins in central Europe. Single cells in the high–high value range could also be found in the US, especially on the East Coast. Most cells overall were low in both betweenness and out-degree. The closeness and in-degree centralities showed far more cells that fall into the high–high value bins, distributed along the more populated parts of the world: covering almost the entirety of the US and Europe, as well as parts of India and the coasts of South America and Africa.

Figure 8 illustrates the 10 largest communities detected within the grid cell network, with each grid cell color-coded to represent its assigned community as determined by the algorithm. The most extensive community, consisting of 833 cells and shaded in dark blue, forms a global presence encompassing English-speaking regions, like the US, the UK, India and large parts of Africa. The second-largest community, colored in orange, consists of 223 cells and has a noticeable presence in Russia, the South and Southeast of Brazil, and also Eastern Europe. Community number three is rooted in Central Europe, with single cells scattered across the globe. It is also noticeable that some communities redraw the geographical boundaries of countries: Mexico and Turkey are clearly distinguishable in brown and light green, respectively. There are also several communities that lack a clear regional concentration, e.g., community numbers four, five, and seven. Community numbers nine and ten are the smallest communities, consisting of 11 cells, and partly cover Spain and Italy, respectively.

5.4. Analyzing the European Subnetwork

The presented previous results indicate that with regard to the betweenness and out-degree centralities, the European continent and the US are the main AOIs. However, the results of community detection suggest that the US and the DACH region do not form a dedicated community; instead, the US seems to be embedded in a large community structure spanning all over the world. In order to identify important regions for the Austrian COVID-19 protest movement in more detail, we analyzed the European subnetwork. Therefore, we filtered for nodes and edges in the user network that were within a bounding box surrounding Europe. We then aggregated the subnetwork to a hexagonal grid of 100 square kilometer area size and re-calculated communities, as well as closeness and betweenness centralities on the finer grid. Figure 9 presents the outcomes of community detection within the European subnetwork, and Figure 10 presents the according distribution of the community cells across the countries. Some communities are spread over several countries, whereas the others show a clearly dominant presence in a distinct country. Specifically, the largest community, colored in blue, covers the area of the United Kingdom and is scattered across France, Italy and Spain. Community number two forms a distinct cluster centered in Germany with more than 50% of cells located there, and less than 10% being dispersed throughout other European countries Communities number three is strongly rooted in Austria, colored in orange in the map in Figure 9, with 20% percent of all cells being located there, as well as France, Italy and Spain. The fourth-largest community is dominant in Switzerland, and reaches into Germany and France. The remaining six communities are notably smaller in terms of the number of cells belonging to them, making up less than 10% of cells compared to the largest cluster. Those communities exhibit clear associations with country borders: Netherlands, Turkey, Switzerland, the Balkan countries, Denmark, and Spain are each associated with significantly smaller, but very distinct communities. Remarkably, the Catalan region around Barcelona stands out as a dedicated community, distinguishing itself from the broader Spanish context.

In Section 5.3, we already identified areas of interest with regard to high-centrality values on a coarse grid. This analysis can be extended further with spatial hot spot analysis of the European subnetwork’s centralities. These hot spots indicate regional clusters of users who were important for the information exchange in the virtual protest movement network. A hot spot analysis using the Getis-Ord GI* algorithm with 30 nearest neighbors as a conceptualization of spatial relationship was performed on the aggregated European subnetwork for betweenness and closeness centrality, and the results are shown in Figure 11.

Closeness centrality hot spots could be found in Germany, the German-speaking part of Switzerland and parts of Austria, especially Vienna, Graz, and the area of upper Austria. The cells in the UK are mostly non-significant, with the exceptions of the greater London area, which stands out as a hot spot with strong significance as well as weaker hot spots in the area of Manchester. Less significant hot spots can be found in Paris and Amsterdam. Ireland, Denmark and Serbia, however, are dominated by cold spots.

The hot spot map of the betweenness centrality is clearly dominated by cells classified as statistically not significant. There are clusters around urban areas, however, that are categorized as hot spots. Strong hot spots with 99% confidence can be found in the urban areas of Vienna, Berlin and London and in Munich. Less significant hot spots with 90 to 95% confidence are detected in the south of Stuttgart, as well as in the area around Essen.

6. Discussion

6.1. Interpretation of Results

This section discusses the methods and interprets the results obtained by the spatial–social network analysis. In answering the research questions in Section 1, we demonstrate how analyzing the the social network of protest actors helps further understanding the underlying protest movement.

6.2. How Does the Online Social Network Manifest in Geographic Space?

When examining the overall network plot, it is evident that the node distribution partly resembles the world population distribution. A high concentration of nodes can be found in the coastal areas of the United States, Europe, India and Brazil with fewer nodes in e.g., Russia, and China. The distribution of points, as well as the geodesic distribution of edge lengths closely mirrors the spatial nature of the network. Almost 50% of all edges have a length up to 900 km, which suggests that most edges in the network appear within continents.

The discovered communities on the globally aggregated network indicate a strong role of languages in structuring the network. Notably, the US and several countries, mostly English-speaking, are seemingly integrated into one large community. This suggests that the numerous connections observed in and to the US may be attributed to the overall large network, the shared language, and the strong use of social media in the US, rather than being an effect that is due to the initial network context of the Austrian COVID-19 protest movement.

The results of community detection within the European subnetwork further underline these findings, where the English-speaking UK is the center of the largest community. The high number of community cells in other European countries and urban centers could suggest a language-based followership. The significance of a shared language can also be observed in a community covering the Slavic-speaking region in Serbia and Bosnia.

In addition to this apparent linguistic structure in the network, the detected communities are clearly structured along country boundaries, as can be seen in the main communities in the Netherlands and Denmark. For the DACH region, we observe that Austria, Germany, and Switzerland each predominantly belong to one network community that also follows the respective country boundaries. However, a scattering of community cells across the other German speaking countries can also be overserved.

These results suggest that the online political movement is structured in primarily national protest communities through the Austrian protest actor network. This conclusion is supported by the fact that during the COVID-19 pandemic, nation-states increasingly initiated independent crisis management measures, which then became the target of protest movements with a strong national orientation (Kriesi and Oana 2023).

6.3. How Strongly Are Different Geographical Regions Connected through the Social Network?

In terms of closeness centrality hot spots, again, the DACH region stands out, with the highest number of hot spots. Interestingly, while we were able to detect several distinct communities, it is only within the DACH region that we see a well-connected region that plays a crucial role in the dissemination of information as measured by network centrality. Furthermore, the findings of both betweenness as well as closeness centrality hot spots analysis emphasize the importance of urban areas. The Austrian hot spots around Vienna, Graz and Linz, for example, not only represent the country’s largest cities, but are also regions of increased protest activities (Ann 2023). These results indicate that, in particular, urban regions are more strongly interconnected through the Austrian protest actor network, indicating a structure along an addition urban–rural dimension. This finding is consistent with existing research, identifying more active participation in online political debates, especially by users in densely populated cities (Casero-Ripollés et al. 2020).

6.4. How Does the Framework of Socio-Spatial Network Analysis Help in Understanding the Transnationality of the Underlying Movement?

Considering that the study subject is an Austrian protest movement, the high absolute number of inner-Austrian connections is expected as is the dominance of Germany as a highly populated German-speaking country, together with the US, which has the highest rate of social media use world-wide. However, the number of edges normalized by population of the countries emphasizes inner-Austrian edges even more and lessens the effect of German edges significantly. In line with this, community detection within the European subnetwork finds a distinct community centered on Austria that at the same time extends considerably to the DACH region and in particular Germany.

In summary, we observe a potential for transnationality for the Austrian protest network in the DACH region which appears to be supported by the shared language. With regard to online social networks, the shared language is a known prerequisite for interactions between social media users, especially with regard to political debates and mobilization (Bastos et al. 2013; Brunnengräber 2012, p. 44). Our findings are further supported firstly by the literature that observes pronounced similarities between the Corona protest movements of Switzerland and Germany, where actors network and mobilize across borders (Nachtwey et al. 2020), secondly by the prevailing national structure of the network, and lastly by the absence of a clear cross-country user community on Austrian territory. This interpretation is in line with existing literature finding that the main addressees for social movements often remain nation states (Brunnengräber 2012, p. 44).

6.5. Discussion of the Methodology

The network data were retrieved when starting the analysis in late July 2022, which is not the time when the political protest in Austria peaked. It was not possible to retrieve follower relationships for that time due to Twitter’s API limitations. In order to ensure that the retrieved network data were still relevant for the protest, we inspected the selected set of actors for starting the crawling. Manual inspection of the user profiles revealed that those users were still actively posting and engaging in the discourse on Twitter at the time of data retrieval.

The user locations within the network were approximated as the information provided in their profiles. It is important to acknowledge the inherent limitations of this approach, as the provided locations are textual references without means of verification for actual presence. However, the approximation is sufficient for our research aims, as we are not interested in the precise position of users but rather in understanding broader geospatial connectivity through the virtual network.

Missing location information from the profile and unsuccessful or imprecise geocoding resulted in the removal of almost 52% percent of nodes and 73% of all network edges. Despite this extensive filtering, we were still able to detect significant structures in both the virtual and spatial network as demonstrated through the results of community detection and hot spot analysis. Further analysis of user locations by, for example, the usage of location extraction techniques could enhance this aspect, but this was not the focus of this study.

The methods used for network analysis in this study are well established and considered state of the art within the field. While there are numerous centrality measures and community detection algorithms available, the chosen ones were considered appropriate for the goals of this analysis—providing considerable analytical insight rather than methodological advancement. The presented analysis assumes that the Austrian COVID protest plays a significant role for the networking and information exchange among the interconnected users. This assumption can be evaluated in future research by investigating the connection between geographical localization, virtual networks, and exchanged content. Our study did not research the dynamic changes in the network and instead focused on the spatial manifestation of the static protest user network. Further research can extend this work with a temporal analysis of changes in the virtual network and, from that, changes in the network’s spatial manifestation. This will allow to understand if and how the importance of geographic regions to the COVID protest changed over the course of the pandemic. Analyzing the communication network in the form of a tweet–retweet network would further allow to study strong and weak ties (Granovetter 1973) and exchanged content in the Austrian COVID protest network.

7. Conclusions

This study employed spatial–social network analysis to the Twitter network of Austrian COVID-19 protest actors. By examining the interconnections of geographic regions through the virtual social network, we gained further understanding into the underlying movement and sought to address a gap in the existing literature. Our findings revealed that the network is dominated by strong national clusters that also form along linguistic boundaries. On a finer geographic scale, hot spot analysis revealed the importance of urban centers in the virtual network. We further find that Austria’s potential for influence is weak among other nations and the inner-Austrian connections appear dominant, outweighing even the potential of influence into Germany, a country that speaks the same language and is ten times the size of Austria in terms of population. This underlines the importance of the local and regional factors in the Austrian protest movement over the potential significance of shared language. Although the presented approach does not allow to draw conclusions about the protest discourse and does not consider changes in the network over time, our work yielded valuable insights into the spatial structures in the Austrian protest network. Further research efforts are needed to understand the spatio-temporal evolution of the protest network, as well as the content spread in the network.