Source-linked AI summary
The role of bot squads in the political propaganda on Twitter
Guido Caldarelli, Rocco De Nicola, Fabio Del Vigna, Marinella Petrocchi, Fabio Saracco
TL;DR
This paper examines how social bots contribute to significant Twitter traffic about migration from Northern Africa to Italy. It combines bot detection with an entropy-based null model, finding that politically aligned hubs share bots and that bots are disproportionately present among followers of some political leaders.
Problem
Bot activity can amplify low-credibility content, and analyses of complex systems have rarely accounted for random noise in assessing message exchange.
Source: §1 · p. 2, §3 · p. 15Method
The paper combines bot-detection techniques with an entropy-based null model to analyze Twitter content exchange and uses verified-user interactions to infer political inclination.
Source: §3 · p. 15Results
Some hubs with similar political views share bots, and bots significantly retweeting Salvini and Meloni make up a greater fraction than bots’ incidence in the validated network.
Source: §1 · p. 3, §3 · p. 16Takeaways & Limitations
The findings indicate that some bots supporting multiple accounts are validated as non-random and that bots’ retweeting activity contributes to the visibility of hubs’ messages.
Source: §3 · pp. 16–17Almost half of unverified users did not interact with a verified user during the study period and therefore were excluded from the polarization procedure.
Source: §2.2.2 · p. 7
Abstract
from arXiv · showhide
Social Media are nowadays the privileged channel for information spreading and news checking. Unexpectedly for most of the users, automated accounts, also known as social bots, contribute more and more to this process of news spreading. Using Twitter as a benchmark, we consider the traffic exchanged, over one month of observation, on a specific topic, namely the migration flux from Northern Africa to Italy. We measure the significant traffic of tweets only, by implementing an entropy-based null model that discounts the activity of users and the virality of tweets. Results show that social bots play a central role in the exchange of significant content. Indeed, not only the strongest hubs have a number of bots among their followers higher than expected, but furthermore a group of them, that can be assigned to the same political tendency, share a common set of bots as followers. The retwitting activity of such automated accounts amplifies the presence on the platform of the hubs' messages.
1 Introduction
Social bots are automated accounts that can amplify messages, but malicious bots also spread low-credibility content and target influential users. This paper combines bot detection with an entropy-based null model to examine significant Twitter traffic about migration from Africa to Italy, finding that successful hubs have bot followers and some politically aligned hubs share bots.
1 Introduction
- Social media have become prominent channels for information and news, with traditional newspapers and agencies also moving to social networks.Source: §1 · p. 2
- Social bots can amplify messages, but malicious accounts have been linked to political tampering, low-credibility content amplification, hateful targeting, and interactions shaped by users’ political opinions.Source: §1 · p. 2
- Estimates put social bots at 9% to 15% of active Twitter users, while evolving bots have proved difficult for people and detection tools to distinguish from legitimate users.Source: §1 · p. 2
- Prior work has used profile, network, and posting characteristics to detect bots, while entropy-based null models provide an unbiased benchmark by filtering random noise.Source: §1 · pp. 2–3
- The study combines a lightweight bot classifier with an entropy-based null model to examine significant Twitter traffic about migration from Africa to Italy over one month.The analysis focuses on retweeting without assessing the truthfulness of the messages.Source: §1 · p. 3
- Some successful hubs with similar political views share bots, which appear to increase the visibility of their messages.Source: §1 · p. 3
2 Results
The results identify politically structured communities and a smaller validated network of significant content exchange. Within that network, leading hubs share bot followers, while polarization assignment remains harder for bots and many unverified users.
2.1 Data collection and processing
- The corpus comprised 1,082,029 tweets from 127,275 accounts collected over one month, with 9,396 accounts classified as bots.Collection covered January 23 to February 22, 2019, and focused on migration flows from Libya to Italy.Source: §2.1 · pp. 3–4
2.2.1 Political affiliation
- The verified–unverified interaction network involved nearly half of unverified users and retained strong communities, with three large communities identified by Louvain.For migration propaganda, M5S supporters were not separated from the right wing, and left-wing representatives inside and outside the PD were closer than in the election-campaign study.Source: Figure 1 · p. 5, §2.2.1 · p. 6
2.2.2 Polarization of unverified accounts
- Nearly half of unverified users did not interact with a verified user during the observation period, limiting direct polarization assignment.The authors conjecture this may reflect the topic-specific corpus, unlike election-campaign data where candidates may have engaged standard users.Source: §2.2.2 · p. 7
- After ten rounds of polarization contagion, the procedure assigned clear political membership to 27% more users; validated accounts saw an increase of almost 58%.The method iteratively assigns a user's prevailing interaction-partner polarization, stopping when neighboring accounts show no clear agreement.Source: §2.2.2 · pp. 7–8
- Bots were harder to assign polarization than users overall, and their community fractions were broadly similar except for a slight increase in purple bots.The authors suggest the relative purple-community excess may reflect automated news spreaders increasing newspaper visibility.Source: §2.2.2 · p. 8, Figure 3 · p. 8
2.3 The backbone of the content exchange on Twitter
- The entropy-based null model filters user activity and tweet virality, leaving a directed validated network of 14,883 users and 34,302 links.Its bipartite directed configuration model preserves users' original-tweet counts, each message's retweet count, and each user's retweet count.Source: §2.3 · pp. 8–9
2.3.1 Hubs and bots
- The top 20 hubs all belonged to the blue community; the first non-blue account, TgLa7, ranked 176th.A migrant-assistance NGO had the fifth-highest out-degree (kout = 1104) but a hub score of 4 × 10−4, ranking 452nd.Source: §2.3.1 · pp. 10–11
- 22 genuine accounts shared 22 bots, including 9 of the top 10 hubs; the largest shared-bot group belonged almost exclusively to the blue community.The subgraph contains 172 nodes; its largest node represents Matteo Salvini.Source: §2.3.1 · p. 12, Figure 7 · p. 13
3 Discussion
The discussion uses entropy-based filtering to distinguish significant Twitter activity from random user and tweet activity, then examines bots' role in political content circulation. It reports that validated bots are unevenly distributed across hubs and that some politically aligned hubs share bot followers.
3 Discussion
- The validated network contains 14,883 of 127,275 users, and its bot incidence is almost one third that of the full dataset.The authors interpret this as indicating that only a minority of bots have retweets incompatible with random activity.Source: §3 · p. 16
- The validated user set is more polarized: unpolarized accounts exceed 40% of all accounts and 50% of bots overall, but fall to around 10% and 5% in the validated network.The authors report that polarized bots pass validation more easily and contribute more significantly to message spreading.Source: §3 · p. 16
- The hubs most effective at delivering messages are in the blue community, which includes government representatives and the right wing; TgLa7 is the first account outside it, ranked 176th.Source: §3 · p. 16
- Bots significantly retweeting Salvini and Meloni make up a greater fraction of their followers than bots in the validated network, while other similarly ranked hubs have fewer bots.Source: §3 · p. 16
- Some top hubs share bots, and the largest such group is the right-wing group, with 172 nodes versus 58 in the other group.The second group’s strongest hub ranks 176th, and earlier analyses of the same problem observed only starlike subgraphs.Source: §3 · p. 16
- The authors present entropy-based filtering as revealing bots’ retweet amplification of messages and identify shared bot support as a contribution beyond earlier starlike findings.They distinguish this analysis from fake-news studies by not focusing specifically on fake news, and suggest analyzing exchanged messages, including their sentiment and mentions.Source: §3 · pp. 16–17
4 Methods
The methods combine a lightweight account classifier with entropy-based validated projections. They infer users’ political groupings from verified–unverified interactions, then extract significant user-to-user message flows while accounting for activity and tweet virality.
4.1 Bot detection classifier
- The reconstructed J48 classifier uses profile-based features and achieved the same classification performance as the original model on its training set, which classified more than 95% of training data correctly.The original training set combined purchased fake accounts with accounts certified genuine by analysts.Source: §4.1 · p. 17
4.2 Validated projection of the bipartite network and users polarization
- Political affiliation is inferred from an undirected verified–unverified retweet network by comparing shared contacts against a BiCM benchmark that accounts for both layers’ degree sequences.The authors restrict their check to verified users because Twitter’s verification procedure makes only their information fact-checkable.Source: §4.2 · pp. 18–19
4.2.1 The Bipartite Configuration Model
- The BiCM constrains average degree and factorizes graph probabilities into independent link probabilities, with likelihood matching the expected degree sequence to the observed one.Source: §4.2.1 · p. 19
4.2.2 Monopartite validated projection
- The validated projection links same-layer nodes when their observed overlap exceeds the BiCM expectation, testing overlap significance with a Poisson-binomial distribution and FDR correction.The resulting network is clustered with Louvain, repeated under shuffled node orders, retaining the maximum modularity.Source: §4.2.2 · pp. 19–20
4.3 Extraction of the backbone of the traffic activity on Twitter
- The traffic backbone uses a directed BiDCM projection over user–tweet layers to identify significant author-to-retweeter flows while accounting for tweet virality and user activity.Source: §4.3 · p. 20
Declaration of authors
The authors state that all authors approved the submitted version and accept personal accountability for their contributions and for investigating and documenting integrity or accuracy concerns.
Declaration of authors
- All authors approved the submitted version and agreed to accountability for their contributions and for resolving and documenting accuracy or integrity questions.Source: Declaration of authors · p. 21
A.1 The Bipartite Configuration Model
The Bipartite Configuration Model builds an ensemble of bipartite graphs with degree constraints by maximizing Shannon entropy. It fits the ensemble’s expected degrees to the observed network and uses a sparse-network approximation when appropriate.
A.1 The Bipartite Configuration Model
- The BiCM extends entropy-based null models to bipartite networks represented by a biadjacency matrix, with nodes’ degrees as constraints.The model assigns probabilities over an ensemble of graphs with the same layer sizes as the observed network.Source: §A.1 · pp. 27–28
- The model maximizes Shannon entropy, preserving specified average quantities while remaining as general as possible about other network features.Source: §A.1 · pp. 27–28
- The ensemble is tailored by matching its expected degree sequence to the real network, yielding a degree-matched comparison while leaving other features as general as possible.Source: §A.1 · p. 28
- For sparse networks with small fitnesses, the link probability is approximated as p_iα ≃ x_i y_α, the Chung-Lu Approximation.The authors use this approximation because the exact system has no analytic solution and the network is large.Source: §A.1 · p. 28
A.1.1 Validated monopartite projection
The BiCM projects shared links after discounting degree-sequence contributions. It tests whether observed common links are statistically significant against a null distribution, then applies FDR control.
A.1.1 Validated monopartite projection
- The BiCM discounts degree-sequence contributions when projecting information from the original network.The projection compares observed common links with their theoretical distribution.Source: §A.1.1 · pp. 28–29
- Common-link counts follow a Poisson-binomial distribution, approximated by a Poisson distribution when motif probabilities are low.The study uses the Poisson approximation because V-motif probabilities are particularly low.Source: §A.1.1 · p. 29
- Significant common-link counts are identified from p-values using α = 0.01 and FDR validation.FDR controls false positives when testing many p-values together.Source: §A.1.1 · p. 29
A.2 The Bipartite Directed Configuration Model
The BiDCM represents authorship and retweeting as separate directed links between users and posts. It constrains the directed degree sequence, with the single-author-per-post property simplifying the model.
A.2 The Bipartite Directed Configuration Model
- The model represents users and posts with separate biadjacency matrices for tweets and retweets.T records authorship links from users to posts; R records retweet links from posts to users.Source: §A.2 · pp. 29–30
- Each post has in-degree one because it can have only one author, simplifying subsequent calculations.The model's link probabilities distinguish a user's authorship of a post from retweeting it.Source: §A.2 · p. 30
- Maximizing entropy while constraining the directed degree sequence yields a graph probability that factorizes into link probabilities.The model is the Bipartite Directed Configuration Model (BiDCM).Source: §A.2 · p. 30
A.2.1 Directed validated projection
The directed validated projection links significant post sources to significant retwitters after accounting for user activity and post virality. It also examines non-retweeted posts and bot contributions to content flow.
A.2.1 Directed validated projection
- The projection produces directed user-layer links from significant post sources to significant retwitters, beyond activity and post virality.The validation assigns p-values using a Poisson approximation and applies FDR at α = 0.01.Source: §A.2.1 · pp. 30–31
- Unlike the earlier approach, this projection includes non-retweeted posts and examines bot contributions to content flow.Including non-retweeted posts changes the q probabilities and tests whether accounts retweet all of an author's posts or select among them.Source: §A.2.1 · p. 31
- The authors regard retweeting all posts from one user as probably more critical for message spreading than selecting only some.This is presented as a probable consideration linked to the activity analysis.Source: §A.2.1 · p. 31
B Iterative procedure for assigning a political inclination to unverified users
The iterative procedure assigns unverified users the inclination of the majority of their followers and followees when their polarization index is at least 0.5. It runs for at most 10 steps and assigns inclinations to nearly 58% of unpolarized users in the directed validated projection.
B Iterative procedure for assigning a political inclination to unverified users
- The procedure iteratively assigns unpolarized users the inclination of the majority of their followers and followees when ρ is at least 0.5.Users below 0.5 have mostly non-polarized neighbours and re-enter the polarization step; the procedure also includes nodes grey in the first bipartite validation.Source: B Iterative procedure for assigning a… · p. 31
- After 10 steps, the procedure assigned political affiliation to almost 27% of unpolarized users, or 15% of all accounts.The procedure stops after 10 steps.Source: B Iterative procedure for assigning a… · p. 31
- In the directed validated projection, polarization by contagion assigned inclination to almost 58% of unpolarized nodes.Unpolarized users made up 8.3% of validated accounts.Source: B Iterative procedure for assigning a… · p. 32
C Alternative approaches to assign a political inclination
The authors favor validated projection because it produces clearer communities and fewer evidently wrong verified-user assignments than the non-validated partition. The two approaches differ in how they group several political accounts, despite some similarities.
C Alternative approaches to assign a political inclination
- The left-panel partition mixes verified accounts from the right and purple groups, while the validated partition separates them and reveals an M5S politician-and-journalist group.The M5S group also includes newspapers, a gossip site, and TV newscasts described as far from M5S.Source: C Alternative approaches to assign a… · pp. 32–33, Figure 9 · p. 32
- The partitions also differ in the treatment of Salvini: his account appears in the left panel’s dark-blue group, while the official Lega accounts remain in its blue group.The dark-blue group also includes French and German politicians allied with Lega in the European Parliament, but also Italian Democratic Party parliamentarians.Source: C Alternative approaches to assign a… · p. 33
- The validated projection has far fewer evidently wrong verified-user assignments than the non-validated partition, motivating its adoption.The remaining errors are limited to a few center-left journalists assigned to the right-wing community.Source: C Alternative approaches to assign a… · p. 33
D Bot Squads
Bot squads are groups of automated accounts following multiple genuine users, and the paper identifies two such groups in the validated traffic backbone. Their bots retweet across hubs that rarely retweet one another significantly.
D Bot Squads
- One subgraph contains 22 hubs sharing 22 bots, including 12 of the top 20 hubs and 9 of the top 10.Each of the 22 bots retweets content from more than one hub in the subgraph.Source: D Bot Squads · p. 33
- The larger shared-bot group includes Salvini, Toninelli, Giorgia Meloni, official Lega accounts, and other politicians and media figures.The passage also identifies a Casa Pound-supported news website and journalists among its members.Source: D Bot Squads · pp. 33–34
- The violet subgraph has 58 accounts versus 178 in the blue subgraph, and its strongest hub, a popular newscast, ranks 176th overall.The authors describe its genuine accounts as less efficient at delivering their messages.Source: D Bot Squads · p. 34
- The violet subgraph includes NGOs and their representatives, La7 journalists, and small left-wing parties.Examples of NGO representatives come from Comunità di Sant’Egidio and MOAS.Source: D Bot Squads · p. 34
- 87% and 79%: bot incidence is similar in the two subgraphs, but shared-bot-to-genuine-user ratios are 1 and around 0.58, with only 3 significant links among hubs.The authors say the bots undertake the task of spreading partner hubs’ content.Source: D Bot Squads · p. 34