Russian invasion of Ukraine in 2022
On February 24, 2022 Russia launched the attack against Ukraine. Since the beginning of this Russian invasion in Ukraine, the number of false information has increased considerably throughout the world. This false information aims to distort people's perception of the causes of the conflict, thus creating hostilities against the Ukrainian cause.
Mapping Coordinated Networks That Circulate Problematic Information on the War in Ukraine
The project “Mapping Coordinated Networks That Circulate Problematic Information on the War in Ukraine” is a project proposed by scientists from the University of Urbino. It aims to map the networks of actors with Coordinated Link Sharing Behaviour (CLSB) and who share false information about the war in Ukraine. “CSLB refers to a specific coordinated activity performed by a network of Facebook pages, groups and verified public profiles (Facebook public entities) that repeatedly shared the same news articles in a very short time from each other”. The work is based on the use of CooRnet. This is a library based on CrowdTangle to get a list of Facebook/Instagram accounts (pages, public groups or profiles) with CLSB.
The initial data contains 25870 original URLs classified as problematic. After recovering these problematic URLs, the authors then used CooRnet to collect all public shares on Facebook of these URLs 7 days after their publication. This method collected 300,836 shares for 14,187 news articles. Using a 10-second coordination interval, the authors identified 818 coordinated accounts (115 pages and 703 groups organised into 95 networks) that quickly shared at least four different posts deemed problematic. Using CrowdTangle (CT), they created lists made up of these accounts (one list comprising 115 pages and another comprising 682 groups), collecting messages posted on the Ukrainian war.
Using CooRnet, they extracted all the links shared in these posts and tracked the networks of accounts that shared these links in a coordinated fashion within a short period of time. In total, they identified 1,509 entities that performed CLSBs, organised into 117 components and 122 clusters. Facebook groups sharing content do so on different themes around a fairly dense informational ecosystem. They share content on different themes (politics, security, health, etc.). The content they share is most often produced by several media sources (social media, web media, alternative media, etc.). An analysis of the shared content makes it possible to identify communities that have similar profiles, their media sources and the topics shared.After analysing the initial database (containing the 122 clusters), cluster 5 was selected as the focus of this work.
The Informational Ecosystem of French Anti-Establishment Facebook Groups
This cluster 5 contains a French-speaking audience that is characterised by anti-establishment attitudes. The first mapping of cluster 5 (on Gephi), as the following graph underlines, identified 7 different categories of audience or communities in this cluster who share the same content on Facebook. From this categorization into seven (7) communities, a database was built. It contains the posts that have been shared to reconstruct communities in order to identify which audiences belong to the French-speaking protest movement on Facebook. From this collection, the analysis of the network of actors revealed groups and pages with a great variety of political positioning. Some refer to traditional political parties (far left or far right) while the majority of them seem to reject politics altogether (“Gilets jaunes”/“Yellow Jackets”). It appears that a common point was defiance towards government officials (President Macron) and, to a lesser degree, mainstream media.
Therefore the cluster is labelled “Anti-Establishment”. In addition, a particular attention was attached to the media ecosystem. The analysis of links and media sources revealed a diversified choice of websites, from more mainstream media to more alternative media sources.
Network 1 : The cluster 5’s sub communities from the coordinated networks that circulate problematic information on the war in Ukraine
The facebook groups as there are a vast majority of groups in our dataset against only one facebook page, each row in the dataset is described by a list of attributes (most of are qualitative) that gives a sufficient description, one of the most useful attributes that helps clarifying the image of the datasets are :
Page Admin Top Country
Post created date
After gathering all the posts from the french-speaking facebook groups from cluster 5 over a period of 12 months, it seemed interesting to question the relation between the new fund sub-clusters, their topics and the use of different types of media.
To what extent the French anti-establishment Facebook groups regrouped by topic of interest are sharing a specific media landscape?
In order to analyse its implication, three different reading grids have been identified ; the actors themselves, their behaviour regarding the type of links they share and the content of their narrative. For each level of analysis, a specific question is asked :
Actors : Who belongs to the French anti-establishment movement on Facebook?
Behaviour : What kind of media sources are shared during the year 2022?
Content : What are the main stories shared by the anti-establishment movement on Facebook?
The main hypothesis is to have one cluster of actors associated with a main type of media. However, concerning the contents, the topic modelling would show main tendencies for each sub-clusters but some stories would be shared across different groups. Indeed, because they are french-speaking groups, the traditional polarisation of political entities as it can be observed in the U.S between left and right should not be illustrated here. On the contrary, the line between those parties are more blurred. They even seem to be associated together in opposition to the current political power in office. Therefore, they are forming the global group of anti-establishment making a point of being against all societal subjects supported by the government. In the end, the use of media in the anti-establishment on facebook should reflect their behaviours regarding the stories they share.
In order to answer the questions, the work is based on the technique of social network analysis. Several mappings were carried out in an iterative way with the help of Gephi software and the online tool Cortext for the textual analysis of the contents. In parallel, the database of shared posts was enriched with the online tool Crowdtangle and exploited through data science techniques such as Natural Language Processing.
Preliminary network mapping to guide the analysisA first mapping of the actors' network has been done from the dataset provided. This mapping was carried out using Gephi software. The nodes are the actors and there is a link between the actors if they share the same content. The size of the nodes is a function of the number of people who subscribe to the page. The force atlas 2 algorithm was used for the spatialization of the map using the following parameters :
Then a detection of the communities was realised thanks to the statistical module of Gephi software and the metric "modularity". From this metric, the nodes and links were coloured when they belong to the same community. After a first analysis of the created map, a further study on the cluster that concentrated a majority of French-speaking actors has been done. A new mapping (graph 1) was created on Gephi software by selecting only the actors of this French-speaking cluster and by using the same parameters. The "modularity" metric made it possible to highlight sub-communities.
Furthermore, the Facebook pages of these sub-communities were explored and categorised according to their title. Based on this first analysis the Crowdtangle tool was used. That is a tool developed by Meta that allows us to extract all the publications of a Facebook page or group over a given period. From a python script, a csv file containing the url of the facebook pages of the whole cluster was created. From this file and thanks to Crowdtangle all the posts of the facebook pages over the last 12 months were downloaded. That is a total of 1 128 826 posts. These posts are grouped in a CSV file containing as many rows as there are posts and 34 columns. This csv file made it possible to analyse more precisely the actors, their behaviour and the content of shared posts.
ActorsMany tools have been used along the data journey, starting from Python on jupyter notebook, different python libraries have been helpful during the data processing journey such as Pandas, Numpy… For Data Visualization, Microsoft Power Bi was the main tool used. It is a unified, scalable platform for self-service and enterprise business intelligence (BI). Connect to and visualise any data, and seamlessly infuse the visuals into the apps you use every day, so PowerBI was used to draw 3 first graphs in the Finding Part (Area Chart - Stacked Area Chart). The last graph is a streamgraph that has been plotted using RawGraphs which is quite helpful when using time series. Some python libraries such as ‘Matplotlib’ and ‘Seaborn’ were also helpful in the Data exploratory step.
*Behaviour*From the Crowdtangle csv file a new mapping has been made. Only the following csv file’s columns were kept : cluster identifier and the post URL. The domain name of each url was extracted thanks to a python script, which allowed it to categorise the 50 most represented domain names.
This categorization is based on a media typologie created for this analisis with following categories :
ContentsRegarding the content, two different complementary approaches were used with the platform CorText : first a network mapping with the list of terms extracted with the highest chi2, and then a topic modelling to get all the most used terms per topics, all of it can be associated with the results from the actors.
First, from the data, 100 words were extracted from the textual field “title” in the dataset, using the term extraction tool with a maximal length of 3 and ranking principle of chi2. From this extraction, the network mapping can be done with the parameters ; First and second field: Terms, Number of nodes: 150 max, and Ranking principle: chi2.
Second, the script called “topic modelling” from CorText was used on this dataset. The topic modelling was done also based on the chi2 score applied to the column “title”. According to the past results, the parameters were set to choose between 3 to 7 topics with a maximum of 10 steps, and to associate one topic per post. The final number of found topics by the algorithm was 3. In the end to a better reading, the topic modelling has been illustrated by word clouds as the designer of this project recommended through the python library wordcloud.
The first part of this analysis was to gather statistical information about the dataset in order to better contextualise the posts.
Top 5 domains linked in the Last 12 months amongst the posts of the French anti-establishment groups on Facebook
From the graph, the top 5 most appearing domains in the last year are all Social Media domains:
The line plot of each social media domain seems similar and normal, except for TikTok that had a peak from february to april and a huge degradation after that.
Number of posts in the Last 12 months amongst the French anti-establishment groups on Facebook
From the graph, it seems that the number of the posts is decreasing regularly with time.
The Top 5 media domains linked in the Last 12 months amongst the posts of the French anti-establishment groups on Facebook
This graph, which was plotted by RawGraph, is a stream graph that clarify the repartition of the top 5 media domains :
BehavioursThe cluster is mapped through network analysis. Links between Facebook groups (nodes in the graph) are created whenever two groups have shared the same URL. The node size depends on the number of links.
Network 2 : Cluster 5’s sub communities and media type shared amongst the posts of the French anti-establishment groups on Facebook
Thanks to the modularity metric from Gephi, 3 main subclusters were identified. Those identification has been confirmed by the topic modelling explained latter on :
Blue nodes are affiliated to pro-Russia groups, such as the group “Putin’s great Russia”. They mainly share posts from alternative media.
Green nodes are affiliated to the traditional left-wing, like “The voice of the France insoumise” group. They mainly share posts from mainstream media.
Purple nodes form the populist nebula opposing COVID-related sanitary measures such as “Stop the sanitary dictature” or “Say no to the vaccinal pass”. They mainly share posts from alternative media.
Network 3 : Mapping with CorText and Gephi of the most used terms in the title of the posts amongst the French anti-establishment groups on Facebook
Two predominant clusters can be distinguished, the first is in green and represents a political cluster with the name of the president in the centre: “Emmanuel Macron”. The other political figures can also be distinguished, such as “Marine Le Pen” or “Jean-Luc Mélenchon” and “Éric Zemmour”. It is also interesting to note that on the right a small cluster linked to this one on the war in Ukraine is emerging. The second major cluster is purple, it is a cluster on the vaccine with as main words: “Vaccinal Pass”, “Vaccin anti-Covid”, “Sanitary Pass”. As the main word is 'Courrier des Stratèges' which is a news site that has talked a lot about the coronavirus. This cluster is linked to a smaller one, in blue, which has the coronavirus as its main theme: “Big Pharma”, “Bill Gates” (who donated a lot of money for the development of the vaccine), and 'Reiner Fuellmich' (covidosceptic lawyer). Finally, there are some very small clusters like the one in blue on the left about life from a religious point of view with the words: “Is life”, “Why life”.
This part puts in relation the two network mapping as it links the narratives and the type of media that is shared by those particular groups. Indeed, three main topics have been identified through a topic modelling with a LDA model (by the platform CorText:). For each topic, a representative post (“best-prototype”) is selected. From this list a “best-prototyp” post on facebook and the link shared for each topic has been screenshotted to better illustrate.Topic Modeling Topic Modelling of the narratives in the title of the posts amongst the French anti-establishment groups on Facebook with associated Worclouds
The first topic is pro-Russia and the war in Ukraine. However, the stories are in favour of Putin and Russia, but mainly in opposition to France and Macron. The narrative places Putin as a strong leader, a good representative of power to follow as Macron in France is painted as a bad leader. Nevertheless, there is quite a big engagement regarding news in Dumbas that can be explained by a very active newspaper on the subject. It is interesting to note that the social media “telegram” is associated with pro-Russia people, meaning a lot of news about this is shared on this online platform.
Paragons post and shared article from topic 1 Pro-Russia
The second topic is around the anti-vaccine movement. The main stories are to denounce the health restriction and obligation to get the vaccine. A lot of articles have pointed out the danger of the vaccine, saying that both it is a weapon to kill the citizens and a way for big pharma to get more money. The government with their health policy and big pharma are seen as the enemy to eradicate. For this topic, it seems that tiktok is the most selected platform to talk about this.
Paragons post and shared article from Topic 2 Anti-Vaccine
The third topic is around a large critic of the government. The main stories are strongly against all political decisions. From the french yellow jacket to the convoy or freedom, they are criticising all the public policy by states. The public debate such as “retirement”, “health” or election has been at the centre of complotist theories. For this topic, twitter is the main social media platform to express themselves.Paragon post and shared article from Topic 3 Anti-Government
In conclusion, 3 clusters can be distinguished : Pro Russia, Group affiliated to the traditional left-wing, Populist group. It is interesting to note that each group is affiliated to specific media. Indeed, for the pro-Russians, the media used are of alternative type such as Business Bourse and using Telegram. For left-wigs, they are associated with mainstream media like Mediapart. Finally, for the populist group, they are associated in majority with TikTok and with alternative media like The Exposé. The final poster shows that the anti-vaccine and anti-government groups are distinct and demonstrate distinct trends, but are also quite strongly linked.
For future research, our media have been classified into 3 groups: Alternative media, mainstream media and social media. However, limited by time, this classification was done manually. A more rigorous classification could be done.
Giglietto, F., Righetti, N., Rossi, L., & Marino, G. (2020). Coordinated Link Sharing Behavior as a Signal to Surface Sources of Problematic Information on Facebook. International Conference on Social Media and Society, 85–91. https://doi.org/10.1145/3400806.3400817
Giglietto, F., Righetti, N., Rossi, L., & Marino, G. (2020). It takes a village to manipulate the media: coordinated link sharing behavior during 2018 and 2019 Italian elections. Information, Communication and Society, 1–25. https://doi.org/10.1080/1369118X.2020.1739732
Giglietto, F., Righetti, N., & Marino, G. (2019). Understanding Coordinated and Inauthentic Link Sharing Behavior on Facebook in the Run-up to 2018 General Election and 2019 European Election in Italy.