Fig 1.
Weekly volume of Slovenian Twitter data, collected over the three year period.
The retweet network observation window is 24 weeks (blue and red lines), with exponential weight decay (half-time of 4 weeks, green curve), and one week sliding window (difference between the red and blue line). Note a large increase of Twitter activities at the emergence of the Covid-19 pandemic, which also coincided with the change of the left-wing to the right-wing government in Slovenia.
Fig 2.
Differences between the adjacent network partitions measured by the F1 score.
The red line at the top shows weekly differences F1(Pt|Pt−1) at timepoints t = 1, 2, …, 132. The five selected partitions are denoted by P0, P22, …, P132. The middle blue line shows the theoretical maximum max-F1 differences between distant partitions at the selected timepoints t = 0, 22, 68, 91, 132. The bottom black line shows the standard F1 differences.
Fig 3.
A Sankey diagram showing major transitions between the five selected timepoints P0, P22, …, P132.
The numbers indicate core nodes (black), new nodes (brown, at top), and lost nodes (yellow, at bottom) between two adjacent network partitions. The differences between the adjacent partitions are quantified by max-F1(Pi|Pi−1), shown with blue line in Fig 2. Note a relatively large in- and out-flow of new and lost nodes between the partitions.
Fig 4.
A Sankey diagram showing transitions between the five largest communities C1, …, C5 at the selected timepoints P0, P22, …, P132.
The remaining smaller communities are labeled as Small, and new and lost nodes are not shown here. The differences between the adjacent partitions are quantified by F1(Pi|Pi−1), shown with black line in Fig 2. The left-leaning communities are in shades of red, the right-leaning communities are in shades of blue, and the Sports community is green.
Fig 5.
Identification of super-communities from the meta-networks.
Nodes are detected communities C1, …, C7 at different timepoints, and edges denote average external influences. A node diameter is proportional (cube-root) to the number of community members, darker area corresponds to the internal influence, and lighter area to the external influence. Red communities are part of the Left-wing super-community, blue communities are part of the Right-wing super-community, and the remaining Sports community is not shown. Dashed edges show rare and relatively weak links between the Left- and Right-wing communities.
Fig 6.
Total weighted out-degree influence for both super-communities.
A super-community size is proportional to the number of its members. Total influence is the sum of weighted out-degree influences of all super-community members. Note that the influence of the Right-wing super-community is at least twice as large as the influence of the Left-wing super-community and increasing with time, despite the fact that it is considerably smaller.
Table 1.
Top ten influential users from each super-community, ranked by the overall h-index.
Individual Twitter users are anonymized and their handles start with $. Left-to-Right denotes users which moved from the Left-wing to the Right-wing super-community (@vladaRS: the official Slovenian government account, and @ukclj: University Medical Centre Ljubljana). The top users in each super-community are $PM-mar20-now (current prime minister), and $PM-sep18-mar20 (former prime minister), respectively. Each user is assigned the h-index rank (h-rank), the h-index (h-ind) for the overall three year period and the five selected timepoints (P0, …, P132), and the overall unweighted out-degree (out-deg). Note that the top Left-wing influential users barely reach the h-index rank of top 100.
Fig 7.
A comparison of the BCubed F1 measure with Adjusted Rand Index (ARI) and Normalized Mutual Information (NMI).
The comparison is run on the initial G0 Slovenian retweet network. On the x-axis is the number of standard Louvain trials, N = 10, 20, …, 100. For each N, all the resulting partitions are pairwise compared by the three measures, ARI, F1, and NMI (y-axis). Solid lines show the mean values and shaded areas the 95% confidence intervals.