Table 1.
Weekly tweets and unique users in our filtered dataset.
Week 4, in which the Paris Agreement announcement was made, includes many more tweets than any other week, accounting for 35% of all the tweets studied.
Fig 1.
Schematic diagram of the bipartite network construction and unipartite projection.
Each user is connected to the URLs they share in the study week. The unipartite projection creates edges between two URLs whenever they are shared by the same person. Multiple edges in the projection indicate that multiple users have shared the pair of URLs, and this information is tracked by edge weights. A user’s edge contribution to the projection increases quadratically with the number of URLs they share, potentially leading to a dominance of highly active users in the unipartite projection, for example User 3 (red) in the projection; this is handled by a hyperbolic weighting scheme (see Section Network construction).
Fig 2.
Average domain ideological positions assigned by each coder and the overall average across all coders.
Values should be thought of as standard deviations from the mean. Vertical bars indicate ± one standard deviation across the coders.
Fig 3.
Information-sharing networks are polarised.
Plot shows the URL co-sharing network for the week in which the US withdrawal from the Paris Agreement was announced (Week 4 of the study period). The five largest communities by total share count are displayed (67.69% of 7, 496 nodes). Communities 1 − 5 are coloured blue, yellow, green, red and purple respectively and node size is proportional to the square root of total share count. Node placement uses a force-directed algorithm [50] which groups densely connected nodes together; this layout highlights two large clusters, with four communities on the left and a single community on the right.
Fig 4.
Climate media content is politicised.
Mean political ideology (left-to-right) and climate opinion (environmentalist-to-sceptic) expressed in content from the 62 coded web domains over the six coders (see Ideological coding of source domains). Point size is proportional to the square root of total share count and lines indicate ± one standard deviation. Labels are shown for 10 most frequently shared domains in the coded list.
Fig 5.
Network clusters are ideologically biased.
The two large clusters within the URL co-sharing network for Week 4 shown with URLs coloured by: (a) the average political bias of their source domain; and (b) the average climate change bias of their source domain. Red denotes left-wing domains, blue denotes right-wing domains, green denotes environmentalist domains, orange denotes sceptic domains. White denotes any domain coded as neutral and domains not coded are in gray.
Fig 6.
Source domains and content of media articles shared within the five largest communities in the information-sharing network.
Tokens in each plot are weighted (using TF-IDF) to make distinctive tokens prominent. Circle size is scaled to indicate the total number of URL shares within the community. Terms coloured black are the highest weighted terms required to reach 15% of the total weight in a community. For visual clarity, each stemmed token is represented by the most common token that maps to it (or pair of tokens for bigrams).
Table 2.
Characterisation of the five largest communities in Week 4.
Fig 7.
Consistent network structure over time.
Network diagrams of the top five communities across the six remaining weeks. Each figure is oriented such that the left-wing cluster is on the left and the right-wing cluster is on the right. In each case node colour signifies community membership and size is proportional to the square root of total share count. Communities are labelled 1 − 5 in decreasing order of size, and colored blue, yellow, green, red and purple respectively. Node placement is determined by the Python implementation of the ForceAtlas2 algorithm [50].
Table 3.
Similarity scores for users, URLs and domains between each of the seven weeks calculated using Eq (3).
Similarity is directional. The similarity given in cell (4, 1) is the proportion of the users/URLs/domains in Week 4 also in seen Week 1. Cool shades indicate values smaller than the mean while warm shades indicate values greater than the mean.
Fig 8.
Levels of political and climate change bias over the course of the seven week study period.
These are measured as the mean coded bias of domains weighted by total shares (see Ideological coding of source domains). Bias in each of the five largest communities are represented by the scatter points in each week, and the bias across the whole network is given by the lines. The colour of the community points is consistent with other figures. In most weeks, the average network bias is left of centre and more environmentalist than sceptic. This trend continues to the individual communities, with the majority being left-wing and environmentalist.