PreprintPDF Available

Investigating Italian disinformation spreading on Twitter in the context of 2019 European elections


Abstract and Figures

We investigate the presence (and the influence) of disinformation spreading on online social networks in Italy, in the5-month period preceding the 2019 European Parliament elections. To this aim we collected a large-scale dataset oftweets associated to thousands of news articles published on Italian disinformation websites. In the observation period,a few outlets accounted for most of the deceptive information circulating on Twitter, which focused on controversialand polarizing topics of debate such as immigration, national safety and (Italian) nationalism. We found evidence ofconnections between different disinformation outlets across Europe, U.S. and Russia, which often linked to each otherand featured similar, even translated, articles in the period before the elections. Overall, the spread of disinformation onTwitter was confined in a limited community, strongly (and explicitly) related to the Italian conservative and far-rightpolitical environment, who had a limited impact on online discussions on the up-coming elections.
Content may be subject to copyright.
Alessandro Artoni
Dept. of Electronics, Information and Bioengineering
Politecnico di Milano
20133 Milano, Italy
Stefano Ceri
Dept. of Electronics, Information and Bioengineering
Politecnico di Milano
20133 Milano, Italy
Francesco Pierri
Dept. of Electronics, Information and Bioengineering
Politecnico di Milano
20133 Milano, Italy
October 28, 2019
We investigate the presence (and the influence) of disinformation spreading on online social networks in Italy, in the
5-month period preceding the 2019 European Parliament elections. To this aim we collected a large-scale dataset of
tweets associated to thousands of news articles published on Italian disinformation websites. In the observation period,
a few outlets accounted for most of the deceptive information circulating on Twitter, which focused on controversial
and polarizing topics of debate such as immigration, national safety and (Italian) nationalism. We found evidence of
connections between different disinformation outlets across Europe, U.S. and Russia, which often linked to each other
and featured similar, even translated, articles in the period before the elections. Overall, the spread of disinformation on
Twitter was confined in a limited community, strongly (and explicitly) related to the Italian conservative and far-right
political environment, who had a limited impact on online discussions on the up-coming elections.
In recent times, growing concern has risen over the presence and the influence of deceptive information spreading on
social media [
]. The research community has employed a variety of different terms to indicate the same issue, namely
disinformation, misinformation, propaganda, junk news and false (or "fake") news.
As people are more and more suspicious towards traditional media coverage [
], news consumption has considerably
shifted towards online social media; these exhibit unique characteristics which favored, among other things, the
proliferation of low-credibility content and malicious information [
]. Consequently, it has been questioned in
many circumstances whether and to what extent disinformation news circulating on social platforms impacted on the
outcomes of political votes [2, 3, 4, 5].
Focusing on 2016 US Presidential elections, recent research has shown that false news spread deeper, faster and broader
than reliable news [
], with social bots and echo chambers playing an important role in the diffusion of deceptive
information [
]. However, it has also been highlighted that disinformation only amounted to a negligible fraction
of online news [
], the majority of which were exposed to and shared by a restricted community of old and
conservative leaning people, highly engaged with political news [
]. In spite of such small volumes, a study
suggested that false news (and the alleged interference of Russian trolls) played an important role in the election of
Donald Trump [2].
As the European Union (EU) struggled to counter the financial crisis which took place since the end of 2009 (following
2008 financial crisis in the US), populist and anti-establishment movements emerged as new electoral forces in Europe
]. After the 2016 Brexit Referendum, anti-Europeans parties spread across the continent defining national identities
in terms of ethnicity and religion and supporting tighter immigration controls [
]. As Europeans were called to
elect their new representatives at the European Parliament–between the 23
and the 26
of May 2019–populist and
nationalist parties contrasted more traditional ones, such as European People’s Party (EPP), Socialists and Democrats
(S&D) and Alliance of Liberals and Democrats for Europe (ALDE), generally engaged in the defense of fundamental
values associated with the EU project. Eventually, the pro-European side prevailed on aforementioned disruptive forces
in most countries, but not in Italy where “Lega” amplified its electoral consensus (33%) and instead “Movimento 5
Stelle” declined (18%). Outside of our scope, a change of the Italian government occurred during the Summer of 2019.
For what concerns misbehavior on social platforms in European countries, recent research has highlighted the impact
and the influence of social bots and online disinformation in different circumstances, including 2016 Brexit [
], 2017
French Presidential Elections [
] and 2017 Catalan referendum [
]. A significant presence of disinformation
in online conversations concerning 2019 European elections has been recently reported across several countries
]. The European Commission has itself raised concerns–since 2015 [
]–about the large exposure of
citizens to disinformation, promoting an action plan to build capabilities and enforce cooperation between different
member states. In anticipation of 2019 European Parliament elections, they sponsored an ad-hoc fact-checking portal
) to debunk false claims relative to political topics, aggregating reports from several agencies
across different countries.
For what concerns Italy, according to Reuters [
], trust in news is today particularly low (40% of people trust overall
news most of the time, 23% trust news in social media most of the time), as result of a long-standing trend which is
mainly due to the political polarization of mainstream news organizations and of the resulting partisan nature of Italian
journalism. Previous research on online news consumption highlighted the existence of segregated communities [
and explored the characteristics of polarizing and controversial topics which are traditionally prone to misinformation
]. Remarkable exposure to online disinformation was highlighted by authors of [
], who exhaustively investigated
online media coverage in the run-up to 2018 Italian General elections; in particular, the study observed a rising trend
in the spread of malicious information, with a peak of interactions in correspondence with the Italian elections. This
result was later substantiated in a report of the Italian Authority for Communications Guarantees (AGCOM) [
]. A
very recent work [
] has collected electoral and socio-demographic data, relative to Trentino and South Tyrol regions,
as to directly estimate the impact of false news on the 2018 electoral outcomes, with a focus on the populist vote;
this study argues that malicious information had a negligible and non-significant effect on the vote. Furthermore, a
recent investigation by Avaaz [
] revealed the existence of a network of Facebook pages and fake accounts which
spread low-credibility and inflammatory content–reaching over a million interactions–in explicit support of "Lega",
"Movimento 5 Stelle" about controversial themes such as immigration, national safety and anti-establishment. Those
pages were eventually shut down by Facebook as violating the platform’s terms of use.
In this work we focus on the 5-month period preceding 2019 European elections; we carry out our research on a
consolidated setting, described in [
], for investigating the presence (and the impact) of disinformation in the
Italian Twittersphere. We recognize that our analysis has a few inherent limitations: first, according to Reuters [
Twitter is overtaken by far by other social platforms, accounting for only 8% of total users (with a decreasing trend)
when it comes to consume news online compared to Instagram (13%), YouTube (25%), WhatsApp (27%) and Facebook
(54%), which exhibit instead a rising trend. Second, these differences are even more accentuated when comparing with
the U.S. scenario [
], the focus of most of recent research. However, other aforementioned social media offer today
little opportunities to researchers to conveniently analyze the spread of online information, given the limitations they
impose on the acquisition of data and the different user experiences they offer. Our study sheds light on the Italian
mechanisms of disinformation spreading, and thus the outcomes of the analysis indicate directions for future research
in the field.
To collect relevant data, we manually curated a list of websites which have been flagged by fact-checking agencies
for fabricating and spreading a variety of malicious information, namely inaccurate and misleading news reports,
hyper-partisan and propaganda stories, hoaxes and conspiracy theories. Differently from [
], satire was excluded from
the analysis. Following literature on the subject [
], we used a "source-based" approach, and assumed that
all articles published on aforementioned outlets indeed carried deceptive information; nonetheless, we are aware that
this might not be always true and reported cases of misinformation on mainstream outlets are not rare [
]. Our analysis
was driven by the following research questions:
Figure 1: Time series for the number of tweets, containing links to disinformation articles, collected in the period from
07/01/2019 to 27/05/2019. We annotated it with some events of interest; network failures indicate when the collection
tool went down.
What was the reach of disinformation which circulated on Twitter in the run-up to European Parliament
elections? How active and strong was the community of users sharing disinformation?
What were the most debated themes of disinformation? How much were they influenced by national vs
European-scale topics?
Who are the most influential spreaders of disinformation? Do they exhibit precise political affiliations? How
could we dismantle the disinformation network?
Did disinformation outlets organize their deceptive strategies in a coordinated manner? Can we identify
inter-connections across different countries?
We first describe the data collection and the methodology employed to perform our analysis, then we discuss each of
the aforementioned research questions, and finally we summarize our findings.
Data Collection
Following a consolidated strategy [
], we leveraged Twitter Streaming API in order to collect tweets containing
an explicit Uniform Resource Locator (URL) associated to news articles shared on a set of Italian disinformation
websites. As a matter of fact, using the standard streaming endpoint allows to gather 100% of shared tweets matching
the defined query [27, 28, 8].
To this aim we manually compiled a list of 63 disinformation websites that were still active in January 2019. We relied
on blacklists curated by local fact-checking organizations (such as "", "" and "");
these include websites and blogs which share hyper-partisan and conspiratorial news, hoaxes, pseudo-science and satire.
We initially started with only a dozen of websites, and we successively added other sources; this did not alter the overall
collection procedure.
For sake of comparison, we also included four Italian fact-checking and debunking agencies, namely "",
"", "", "".
Figure 2: Time series for the number of shares on both Twitter (red) and Facebook (blue) for two disinformation outlets,
respectively "" (left) and "" (right), in the period from 07/01/2019 to 27/05/2019.
In accordance with current literature [
] we use a "source-based" approach: we do not verify each
news article manually but we assign the disinformation label to all items published on websites labeled as such (the
same holds for fact-checking articles).
In order to filter relevant tweets, we used all domains as query
parameters (dropping "www", "https", etc) in the
form "
byoblu com OR voxnews info OR ...
" as suggested by Twitter Developers guide (
). We built a crawler to visit these websites and parse URLs as to extract article text and other metadata
(published date, author, hyperlinks, etc). We handled URL duplicates by directly visiting hyperlinks and comparing the
associated HTML content. We also extracted profile information and Twitter timelines for all users using Twitter API.
The collection of tweets containing disinformation (see Fig 1) and fact-checking articles was carried out continuously
from January 1st (2019) to May 27th, the day after EU elections in Italy. We collected 16,867 disinformation articles
shared over 354,746 tweets by 23,243 unique users, and 1,743 fact-checking posts shared over 23,215 tweets by 9814
unique users.
We can observe that, in general, articles devoted to debunk false claims were barely engaged, accounting only for 6%
of the total volume of tweets spreading disinformation in the same period; such findings are comparable with the US
scenario [
], and they are in accordance with the very low effectiveness of debunking strategies which is documented in
]. We leave for future research an in-depth comparative analysis of diffusion networks pertaining to the two news
The entire data is available at:
Comparison with Facebook
In order to perform a rough estimate of the different reach of disinformation on Twitter compared to Facebook,
we collected data relative to the latter platform regarding two disinformation outlets, namely "" and
"", which have an associated Facebook page and are among Top-3 prolific and engaged sources of
malicious information (see Results).
We used
] to collect statistics on the number of daily shares of Facebook posts published by aforementioned
outlets, and we compared with the traffic observed on Twitter. As we can see in Fig 2, disinformation has a stronger
reach on Facebook than Twitter, for both sources, throughout the observation period; this is also shown in other works
], coherently with the Italian consumption of social news. An in-depth analysis of the Italian disinformation
on Facebook would be required, but it needs suitable assistance from Facebook for what concerns the disinformation
diffusion network.
Network analysis
Building Twitter diffusion network
We built Twitter global diffusion network–corresponding to the union of all sharing cascades associated to articles
gathered in our dataset–following a consolidated strategy [
]. We considered different Twitter social interactions
altogether and for each tweet we add nodes and edges differently according to the action(s) performed by users:
Tweet: a basic tweet corresponds to originally authored content, and it thus identifies a single node (author).
: whenever a tweet of user
contains a mention to user
, we build an edge from the author
of the
tweet to the mentioned account b.
Reply: when user areplies to user bwe build an edge from ato b.
Retweet: when user aretweets another account b, we build an edge from bto a.
Quote: when user aquotes user bthe edges goes from bto a.
When processing tweets, we add a new node for users involved in aforementioned interactions whenever they are not
present in the network. As a remark, a single tweet can contain simultaneously several actions and thus it can generate
multiple nodes and edges. Finally, we consider edges to be weighted, where the weight corresponds to the number of
times two users interacted via actions mentioned beforehand.
Building the network of websites
In order to investigate existing inter-connections among different disinformation websites, and to understand the nature
of external sources which are usually mentioned by deceptive outlets, we searched for URLs in all articles present in
our dataset, i.e. which were shared at least once on Twitter. We accordingly built a graph where each node is a distinct
Top-Level Domain–the highest level in the hierarchical Domain Name System (DNS) of the Internet–and an edge is
built between two nodes
has published at least an article containing an URL belonging to
the weight of an edge corresponds to the number of shared tweets carrying an URL with an hyperlink from
The final result is a directed weighted network of approximately 5k nodes and 8k edges. We used
package [31] to handle the network.
Main core decomposition, centrality measures and community detection
In our analysis we employed several techniques coming from the network science toolbox [
], namely
decomposition, community detection algorithms and centrality measures. We used
Python package to
perform all the computations.
-core [
] of a graph G is the maximal connected sub-graph of G in which all vertices have degree at least
Given the
-core, recursively removing all nodes with degree
allows to extract the
(k+ 1)
-core; the main core is the
non-empty graph with maximum value of
-core decomposition can be employed as to uncover influential nodes in a
social network [8].
Community detection is the task of identifying communities in a network, i.e. dense sub-graphs which are well separated
from each other [
]. In this work we consider Louvain’s fast greedy algorithm [
], which is an iterative procedure
that maximizes the Newman-Girvan modularity [
]; this measure is based on randomizations of the original graph as
to check how non-random the group structure is.
A centrality measure is an indicator that allows to quantify the importance of a node in a network. In a weighted directed
network we can define the In-strength of a node as the sum of the weights on the incoming edges, and the Out-strength
as the sum of the weights on the out-going edges. Betweenness centrality [
] instead quantifies the probability for a
node to act as a bridge along the shortest path between two other nodes; it is computed as the sum of the fraction of
all-pairs shortest paths that pass through the node. PageRank centrality [
] is traditionally used to rank webpages in
search engine queries; it counts both the number and quality of links to a page to estimate the importance of a website,
assuming that more important websites will likely receive more links from other websites.
Time series analysis
In our experiments, we carried out a trend analysis of time series concerning users’ activity, topics contained in
disinformation articles and the number of interconnections between different outlets.
In statistics, a trend analysis refers to the task of identifying a population characteristic changing with another variable,
usually time or spatial location. Trends can be increasing, decreasing, or periodic (cyclic). We used the Mann-Kendall
statistical test [
] as to determine whether a given time series showed a monotonic trend. The test is non-parametric
and distribution-free, e.g. it does not make any assumption on the distribution of the data. The null hypothesis H0, no
monotonic trend, is tested against the alternative hypothesis
that there is either an upward or downward monotonic
trend, i.e. the variable consistently increases or decreases through time; the trend may or may not be linear. We used
mkt Python package.
The multiple testing (or large-scale testing) problem occurs when observing simultaneously a set of test statistics, to
decide which if any of the null hypotheses to reject [
]. In this case it is desirable to have confidence level for the
whole family of simultaneous tests, e.g. requiring a stricter significance value for each individual test. For a collection
of null hypotheses we define the family-wise error rate (FWER) as the probability of making at least one false rejection,
(at least one type I error). We used the classical Bonferroni correction to control the FWER at
by strengthening
the threshold of each individual testing, i.e. for an overall significance level
simultaneous tests, we reject the
individual null hypothesis at significance level α/N .
Ethics statement
We do not need ethical approval as data was publicly available and collected through Twitter Streaming API; we do not
infringe Twitter terms and conditions of use. The same holds for data relative to Facebook, which was obtained using
netvizz application in accordance with their terms of service.
Results and discussion
Assessing the reach of Italian disinformation
Sources of disinformation
To understand the reach of different disinformation outlets, we first computed the distribution of the number of articles
and tweets per source. We observed, as shown in Fig 3, that a few websites dominate on the remaining ones both in
terms of activity and social audience.
In particular, with approximately 200k tweets (over 50% of the total volume) and 6k articles (about
of the total
number), "" stands out on all other sources; this outlet spreads disinformation spanning several subjects,
from immigration to health-care and conspiratorial theories, and it runs campaigns against fact-checkers as well as
labeling its articles with false "fact-checking" labels as to deceive readers.
Interestingly, two other uppermost prolific sources such as "" and "" do not
receive the same reception on the platform; the former has stopped its activity on March and the latter is literally–it
translates as "All the immigrants crimes"–a repository of true, false and mixed statements about immigrants who
committed crimes in Italy.
We can also recognize three websites associated to public Facebook pages that have been recently banned after the
investigation of Avaaz NGO, namely "", "" and "", as they were "regularly spreading
fake news and hate speech in Italy" violating the platform’s terms of use [26].
We further computed the distribution of the daily engagement (the ratio
no.articles published/no.tweets shared
per day) per each source, noticing that a few sources exhibit
a considerable number of social interactions in spite of fewer associated tweets, compared to uppermost "".
We show the time series for the daily engagement of Top-10 sources, which account for over 95% of total tweets, in Fig
4. We can notice in particular that "" exhibits remarkable spikes of engagement w.r.t to a very small number
of total tweets compared to other outlets, whereas "" shows a suspiciously large number of shares in the
month preceding the elections (and after the release of Avaaz report).
We excluded "" from this analysis as it was added only at the end of April (we collected around
30k associated tweets and less than 1000 articles); official magazine of "CasaPound" (former) neo-fascist party–with
style and agenda-setting that remind of Breitbart News–it exhibits a daily engagement of over 200 tweets, exceeding all
other websites .
Figure 3:
A (Top).
The distribution of the total number of shared articles per website.
B (Bottom).
The distribution of
the total number of associated tweets per website. We show Top-11 (which account for over 95% of the total volume of
tweets), and we aggregate remaining sources as "Others".
As elections approached, we were interested to understand whether there were particular trends in the daily reception of
different sources. Focusing on Top-10 sources (except "") we performed a Mann-Kendall test to
assess the presence of an upward or downward monotonic trend in the time series of (a) daily shared tweets and (b)
daily engagement. Taking into account Bonferroni’s correction, the test was rejected at
α= 0.05/10 = 0.005
; both
(a) and (b) exhibit an upward trend for "" alone, whereas the remaining sources are either stationary or
monotonically decreasing. As this outlet strongly supported euro-skeptical positions (and often gave visibility to many
Italian representatives of such arguments) we argue that in the run-up to the European elections its agenda became
slightly more captivating for the social audience.
User activity
For what concerns the underlying community of users sharing disinformation, we first computed the distribution of the
number of shared tweets and unique URLs shared per number of users, noticing that a restricted community of users
is responsible for spreading most of the online disinformation. In fact, approximately 20% of the community (
users) accounts for more than 90% of total tweets (
330k), in accordance with similar findings elsewhere [
Among them, we identified accounts officially associated to 18 different outlets (we manually looked at users’ profile
description and usernames); they overall shared 8310 tweets.
We also distinguished five classes of users based on their generic activity, i.e. the number of shared tweets containing an
URL to disinformation articles: Rare (about 9.5k users) with only 1 tweet; Low (about 8k users) with more than 1 tweet
Figure 4: Daily engagement for Top-10 sources (ranked according to the total number of shared tweets). The Mann-
Kendall test (upward trend at significance level 0.005) was accepted only for "".
and less than 10; Medium (about 3k users) with a number of tweets between 11 and 100; High (about 500 users) with
more than 100 tweets but less than 1000; Extreme (exactly 20 users) with more than 1000 shared tweets. About 1 user
out of 5 shared more than 10 disinformation articles in five months.
As shown in Fig 5A, we can notice that a minority of very active users (the ensemble with High and Extreme activity)
accounts for half of the deceptive stories that were shared, and over 3/4 of the total number of tweets was shared by less
than 4 thousand users (Medium,High and Extreme activity).
We overall report 21,124 active (20 of which are also verified), 800 deleted, 124 protected and 112 suspended accounts.
Verified accounts were altogether involved in 5761 tweets, only 18 of which in an "active" way, i.e. a verified account
actually authored the tweet. We observed that they were mostly called in with the intent to mislead their followers,
adding deceptive content on top of quoted statuses or replies.
Next we inspected the distribution of the number of users concerning their re-tweeting activity, i.e. the fraction of
re-tweets compared to the number of pure tweets; as shown in Fig 5B this is strongly bi-modal, and it reveals that users
sharing disinformation are mostly "re-tweeters": more than 60% of the accounts exhibit a re-tweeting activity larger
than 0.95 and less than 30% have a re-tweeting activity smaller than 0.05. This shows that a restricted group of accounts
is presumably responsible for conveying in the first place disinformation articles on the platform, which are propagated
afterwards by the rest of the community.
We computed the distribution of some user profile features, namely the count of followers and friends, the number
of statuses authored by users and the age on the social platform (in number of months passed since the creation date
to May 2019). We report that users sharing disinformation tend to be quite "old" and active on the platform–with an
average age of 3 years and more than a thousand authored statuses. We were able to gather information via Twitter API
only for active and non-protected users.
We further inspected recently created accounts, noticing that approximately a thousand user was registered during the
collection period, i.e. the last six months; they show similar distributions of aforementioned features compared to
older users. Overall (see Fig 5B) they mostly pertain to active classes (Medium and High) and they account for 15%
(around 18k tweets) of the total volume of tweets considered–which lowers to approximately 288k tweets excluding
those authored by non-active, suspended and protected accounts. Furthermore, about a hundred exhibit abnormal
activities, producing more than 10k (generic) tweets in the period preceding the elections and directly sharing more
Figure 5:
A (Top).
A breakdown of the total volume of tweets according to the activity of users. Fractions of users
created in the six months before the elections are indicated with lighter shades; these account respectively for 0.18%
(Rare), 0.6% (Low), 2.04% (Medium) and 2.98% (High) of total tweets.
B (Center).
The distribution of the number of
users per retweeting activity. C (Bottom). The distribution of daily tweets shared by recently created users.
than 10 disinformation stories each. We performed a Mann-Kendall test to the time series of daily tweets shared by
such users (see Fig 5C), assessing the presence of a monotonically increasing trend (at significance level
α= 0.05
The main referenced source of disinformation is "" with more than 60% (circa 12k tweets) of the total
number of shared stories. An activity of this kind is quite suspicious and could be further investigated as to detect the
presence of "cyber-troops" (bots, cyborgs or trolls) that either attempted to drive public opinion in light of up-coming
elections (via so-called "astroturfing" campaigns [
]) or simply redirected traffic as to generate online revenue through
advertisement [1, 2, 3].
Figure 6: A stacked-area chart showing the distribution of different topics over the collection period. The daily coverage
on themes related to Immigration/Refugees and Europe/Foreign is stationary, whereas focus on subjects related to
Crime/Society and Politics/Government is monotonically increasing towards the elections (end of May 2019).
The agenda-setting of disinformation
Theme analysis
For what concerns the main themes covered by different disinformation outlets, relative to the resulting audience on
Twitter, we based our analysis on the first level of agenda-setting theory [
], which states that news media set the public
importance for objects based on the frequency in which these are mentioned and covered. In the case of disinformation
news an agenda-setting effect could occur as a result of the rise in the coverage, even if some audience members are
aware that news are false [
]. We focused on the prevalence of titles, which were shared at least once, as they usually
pack a lot of information about their claims in simple and repetitive structures [
]; besides, the exposure (such as the
presence alone of misleading titles on users’ timelines) could affect ordinary beliefs and result in resistance to opposite
arguments [29] and an increased perceived accuracy of the content, irrespective of its credibility [46].
We avoided automatic topic modeling algorithms [
] as they are not suitable for small texts, and we employed a
dictionary-based text-analysis, an approach which is largely used for testing communication theories such as agenda
setting and selective exposure in big social media data [
]. Therefore we manually compiled a list of keywords
associated to five distinct topics namely: Politics/Government (PG), Immigration/Refugees (IR), Crime/Society (CS),
Europe/Foreign (EF), Other (OT). Keywords were obtained with a data-driven approach, i.e. inspecting Top-500 most
frequent words appearing in the titles, and taking into account relevant events that occurred in the last months. We
provide Top-20 keywords for each topic in Table 1.
In particular, PG refers to main political parties and state government as well as the main political themes of debate.
IR includes references to immigration, refugees and hospitality whereas CS includes terms mostly referring to crime,
minorities and national security. Finally EF contains direct references to European elections and foreign countries. It is
worth mentioning that the most frequent keyword was "video", suggesting that a remarkable fraction of disinformation
was shared as multimedia content [49].
We computed the relative presence of each topic in each article by counting the number of keywords appearing in the
title and accordingly assessed their distribution across tweets over different months. We can observe in Fig 6 that the
discussion was stable on controversial topics such immigration, refugees, crime and government, whereas focus on
European elections and foreign affairs was quite negligible throughout the period, with only a single spike of interest at
Politics/Government Immigration/Refugees Europe/Foreign Crime/Society Other
salvini immigrati euro rom video
italia profughi europa milano anni
pd clandestini ue casa contro
italiani profugo fusaro bergoglio foto
m5s ong diego morti vuole
italiana porti meluzzi mafia può
italiano migranti libia bambini vogliono
milioni africani macron roma parla
lega immigrato soros donne byoblu
sinistra islamici francia bruciato via
casapound imam francesi confessa niccolò
maio seawatch gilet falsi casal
soldi nigeriani gialli bus vero
guerra nigeriana europee choc ufficiale
cittadinanza nigeriano germania figli bufala
prima islamica tedesca case anti
raggi africano mondo chiesa sta
governo stranieri notre famiglia grazie
renzi chiusi dame magistrato casarini
zingaretti sea francese polizia farli
Table 1: Top-20 keywords associated with each topic.
the beginning of January corresponding to the quarrel between Italian and France prime ministers. We also performed
Mann-Kendall test to assess the presence of any monotonic trends in the daily distribution of different topics; we
rejected the test for
α= 0.05/5=0.01
for IR and EF whereas we accepted it for the remaining topics, detecting the
presence of an upward monotonic trend in CS and PG, and a downward monotonic trend in OT.
In the observation period, the disinformation agenda was well settled on main arguments supported by leading parties,
namely "Lega" and "Movimento 5 Stelle", since 2018 general elections; this suggests that they might have profited from
and directly exploited hoaxes and misleading reports as to support their populist and nationalist views (whereas "Partito
Democratico" appeared among main targets of misinformation campaigns); empirical evidence for this phenomenon
has been also widely reported elsewhere [
]. However, the electoral outcome confirmed the decreasing trend of
"Movimento 5 Stelle" electoral consensus in favor of "Lega", which was rewarded with an unprecedented success.
Differently from 2018 [
] we in fact observed one main cited leader: Matteo Salvini ("Lega" party). This is consistent
with a recent report on online hate speech [
], contributed by Amnesty International, which has shown that his activity
(and reception) on Twitter and Facebook is 5 times higher than Luigi Di Maio (leader of "Movimento 5 Stelle"); not
surprisingly, his main agenda focuses (negatively) on immigration, refugees and Islam (which generated most of online
interactions in 2018 [
]), which are also the main objects of hate speech and controversy in online conversations of
Italian political representatives overall.
It appears that mainstream news actually disregarded European elections in the months preceding them, focusing on
arguments of national debate [
]; this trend was also observed in other European countries according to FactCheckEU
], claiming that misinformation was not prominent in online conversations mainly because European elections are
not particularly polarized and are seen as less important compared to national elections. We believe that this might have
affected the agenda of disinformation outlets, which are in general susceptible to traditional media coverage [
], thus
explaining the focus on different targets in their deceptive strategies.
Usage of hashtags
Among most relevant hashtags shared along with tweets–in terms of number of tweets and unique users who used them
(see Fig 7)–a few indicate main political parties (cf. "m5s", "pd", "lega") and others convey supporting messages for
precise factions, mostly "Lega" (cf. "salvininonmollare", "26maggiovotolega"); some hashtags manifest instead active
engagement in public debates which ignited on polarizing and controversial topics (such as immigrants hospitality,
vaccines, the Romani community and George Soros). We also found explicit references to (former) far-right party
Figure 7: Top-10 hashtags per number of shared tweets (blue) and unique users (orange).
Figure 8: The cloud of words for Top-50 most frequent hashtags embedded in the users’ profile description.
"CasaPound" and the associated "Altaforte" publishing house, as well as some disinformation websites (with a
remarkable polarization on "criminiimmigrati" which was shared more than 5000 times by only a few hundred
Table 2: List of Top-10 users according to different centrality measures, namely In-strength, Out-Strength, Betweenness
and PageRank; we indicate with a cross nodes that do not belong to the main K-core (k=47) of the network.
Rank In-Strength Out-Strength Betweenness PageRank
1napolinordsud ×Filomen30847137 IlPrimatoN IlPrimatoN
2RobertoPer1964 POPOLOdiTWlTTER matteosalvinimi matteosalvinimi
3razorblack66 laperlaneranera Filomen30847137 Sostenitori1 ×
4polizianuovanaz ×byoblu byoblu armidmar
5Giulia46489464 IlPrimatoN a_meluzzi Conox_it ×
6geokawa petra_romano AdryWebber lauraboldrini ×
7Gianmar26145917 araldoiustitia claudioerpiu pdnetwork ×
8pasqualedimaria ×max_ronchi razorblack66 libreidee ×
9il_brigante07 Fabio38437290 armidmar byoblu
10 AngelaAnpoche claudioerpiu Sostenitori1 ×Pontifex_it ×
We also extracted hashtags directly embedded in the profile description of users collected in our data, for which
we provide a cloud of words in Fig 8. The majority of them expresses extreme positions in matter of Europe and
immigration: beside explicit references to "Lega" and "Movimento 5 Stelle", we primarily notice euro-skeptical (cf.
"italexit", "noue"), anti-Islam (cf. "noislam") and anti-immigration positions (cf. "noiussoli", "chiudiamo i porti")
and, surprisingly enough, also a few (alleged) Trump followers (cf. "maga" and "kag"). The latter finding is odd but
somehow reflects the vicinity of Matteo Salvini and Donald Trump on several political matters (such as refugees and
national security). On the other hand, we also notice "facciamorete", which refers to a Twitter grassroots anti-fascist
and anti-racist movement that was born on December 2018, as a reaction to the recent policies in matter of immigration
and national security of the Italian establishment.
Principal spreaders of disinformation
Central users in the main core
In order to identify most influential nodes in the diffusion network, we computed the value of several centrality measures
for each account. We show in Table 2 the list of Top-10 users according to each centrality measure, and we also indicate
whether they belong or not to the main K-core of the network [
]; this corresponds to the sub-graph of neighboring
nodes with degree greater or equal than
k= 47
, which is shown in Fig 9. We color nodes according to the communities
identified by the Louvain modularity-based community algorithm [
] run on the original diffusion network (over 20k
nodes and 100k edges).
Although we expect centrality measures to display small differences in their ranking, we can notice that the majority of
nodes with highest values of In-Strength, Out-Strength and Betweenness centralities also belong to the main K-core of
the network; the same does not hold for users which have a large PageRank centrality value. A few users strike the eye:
1. matteosalvinimi
is Matteo Salvini, leader of the far-right wing "Lega" party; he is not an active spreader
of disinformation, being responsible for just one (true) story coming from disinformation outlet "lettoquo-" (available at
which was shared over 1800 times. He is generally passively involved in deceptive strategies of malicious
users who attempt to "lure" his followers by attaching disinformation links in replies/re-tweets/mentions to his
2. a_meluzzi
is Alessandro Meluzzi, a former representative of centre-right wing "Forza Italia" party (whose
leader is Silvio Berlusconi); he is a well-known supporter of conspiracy theories and a very active user in the
disinformation network, with approximately 400 deceptive stories shared overall.
Accounts associated to disinformation outlets, namely
with "",
with "",
with "" and
A manual inspection revealed that most of the influential users are indeed actively involved in the spread of dis-
information, with the only exception of
who is rather manipulated by other users, via men-
tions/retweets/replies, as to mislead his huge community of followers (more than 2 millions). The story shared by
Matteo Salvini underlines a common strategy of disinformation outlets identified in this analysis: they often publish
Figure 9: The main K-core (
k= 47
) of the re-tweeting diffusion network. Colors correspond to different communities
identified with the Louvain’s algorithm. Node size depends on the total Strength (In + Out) and edge color is determined
by the source node.
simple true and factual news as to bait users and expose them to other harmful and misleading content present on the
same website.
Besides, we notice in the ranking a few users who are (or have been in the past) target of several disinformation
campaigns, such as lauraboldrini
(Laura Boldrini),
("Partito Democratico" party) and
(Papa Francesco). We also report a suspended account (
), a protected one (
) and
a deleted user (pasqualedimaria).
In addition, we investigated communities of users in the main K-core–which contains 218 nodes (see Fig 9)–and we
noticed systematic interactions between distinct accounts. We manually inspected usernames, most frequent hashtags
and referenced sources, deriving the following qualitative characterizations:
community corresponds to "Lega" party official accounts:
whereas the third account, noipersalvini, belongs to the same community but does not appear in the core.
community represents Italian far-right supporters, with several representatives of CasaPound (former)
party (including his secretary
who does not appear in the core), who obviously refer to
"" news outlet.
community is strongly associated to two disinformation outlets, namely ""
) and "" (
); the latter was one of the pages identified in Avaaz
report [26] and deleted by Facebook.
community is associated to the euro-skeptical and conspiratory outlet "" (
and it also features Antonio Maria Rinaldi (
), a well-known euro-skeptic economist who has just
been elected with "Lega" in the European Parliament.
community corresponds to the community associated to ""
(TuttICrimin) disinformation outlet.
the remaining
) and
) represent different groups of very active "loose cannons" who do not exhibit a clear affiliation.
Eventually, we employed Botometer algorithm [
] as to detect the presence of social bots among users in the main
core of the network. We set a threshold of 50% on the Complete Automation Probability (CAP)–i.e. the probability of
an account to be completely automated–which, according to the authors, is a more conservative measure that takes into
account an estimate of the overall presence of bots on the network; besides, we computed the CAP value based on the
language independent features only, as the model includes also some features conceived for English-language users.
We only detected two bot-like accounts, namely
, respectively with probabilities 58%
and 64%, that belong to the same Purple community. A manual check confirmed that the former habitually shares
random news content (also mainstream news) in an automatic flavour whereas the latter is the official spammer account
of "" disinformation outlet. We argue that the impact of automated accounts in the diffusion of malicious
information is quite negligible compared to findings reported in [8], where about 25% of accounts in the main core of
the US disinformation diffusion network were classified as bots.
Dismantling the disinformation network on Twitter
Similar to [
], we performed an exercise of network dismantling analysis using different centrality measures, as
to investigate possible intervention strategies that could prevent disinformation from spreading with the greatest
We first ranked nodes in decreasing order w.r.t to each metric, plus the core number–the largest
for which the node
is present in the corresponding
-core–and the In and Out-degree, which exhibited the same Top 10 ranking as their
weighted formulation (Strengths), but they do entail different results at dismantling the network. Next we delete them
one by one while tracking the resulting fraction of remaining edges, tweets and unique articles in the network.
We observed that eliminating a few hundred nodes with largest values of Out-Degree promptly disconnects the network;
in fact these users alone account for 90% of the total number of interactions between users. For what concerns the
number of tweets sharing disinformation articles, the best strategy would be to target users with largest values of
In-Strength who, according to our network representation, are likely to be users with a high re-tweeting activity; in
fact, confirming previous observations, a few thousand nodes account for more than 75% of the total number of tweets
shared in the five months before the elections. However, as shown in Fig 10, it is more challenging to prevent users to
be exposed from even a tiny fraction of disinformation articles, as the network exhibits an almost linear relationship
between the number of users disconnected and the corresponding number of remaining stories; as such the spread of
malicious information would be completely prevented only blocking the entire network.
Interconnections of deceptive agents
To investigate existing connections between different disinformation outlets and other external sources, we first analyzed
the network of websites with a core decomposition [
], obtaining a main core (
k= 14
) which contains 35 nodes as
a result of over 75,000 external re-directions via hyperlinks (shown in Fig 11A). Over 99% of the articles includes a
hyperlink in the body. We may first notice frequent connections between distinct disinformation outlets, suggesting the
presence of shared agendas and presumably coordinated deceptive tactics, as well as frequent mentions to reputable
news websites; among them we distinguish "IlFattoQuotidiano", which is a historical supporter of "Movimento 5
Stelle", and conservative outlets such as "IlGiornale" and "LiberoQuotidiano" which lean instead towards "Lega". We
also observe that most of the external re-directions point to social networks (Facebook and Twitter) and video sharing
Figure 10: Results of different network dismantling strategies w.r.t to remaining unique disinformation articles in the
network. The x-axis indicates the number of disconnected accounts and the y-axis the fraction of remaining items in the
websites (Youtube); this is no wonder given that disinformation is often shared on social networks as multimedia
content [
]. In addition, we inspected nodes with the largest number of incoming edges (In-degree) in the original
network, discovering among uppermost 20 nodes a few misleading reports originated on dubious websites (such as
""), flagged by fact-checkers but that were not included in any blacklist. We believe that a more
detailed network analysis could reveal additional relevant connections and we leave it for future research.
Furthermore, we focused on the sub-graph composed of three particular classes of nodes, namely Russian (RU)
sources, EU/US disinformation websites and our list of Italian (IT) outlets; we manually identified notable Russian
sources ("RussiaToday" and "SputnikNews" networks) and we resorted to notable blacklists to spotlight other EU/US
disinformation websites–namely "", "dé", the list compiled by Hoaxy [
] and references to
junk news in latest data memos by COMPROP research group [16, 17, 14, 18].
The resulting bipartite network–we filtered out intra-edges between IT sources to better visualize connections with the
"outside" world–contains over 60 foreign websites (RU, US and EU) and it is shown in Fig 11B.
We observe a considerable number of external connections (over 500 distinct hyperlinks present in articles shared
more than 5 thousand times) with other countries sources, which were primarily included within "",
"" and "". Among foreign sources we encounter several well-known US sources
("", "" and "" to mention a few) as well as RU ("", ""
and associated networks in several countries), but we also find interesting connections with notable disinformation outlets
from France ("" and ""), Germany (", Spain ("")
and even Sweden ("" and ""). Besides, a manual inspection of a few articles revealed that
stories often originated in one country were immediately translated and promoted from outlets in different countries
(see Fig 12). Such findings suggest the existence of inter-connected deceptive strategies which span across several
countries, consistently with claims in latest report by Avaaz [
] which revealed the existence of a network of far-right
and anti-EU websites, leading to the shutdown of hundreds of Facebook pages with more than 500 million views just
ahead of the elections. Far-right disinformation tactics comprised the massive usage of fake and duplicate accounts,
recycling followers and bait and switch of pages covering topics of popular interest (e.g. sport, fitness, beauty).
It is interesting that Facebook decided on the basis of external insights to shutdown pages delivering misleading content
and hate speech; differently from the recent past [
] it might signal that social media are more willing to take action
Figure 11: Two different views of the network of websites; the size of each node is adjusted w.r.t to the Out-strength,
the color of edges is determined by the target node and the thickness depends on the weight (i.e. the number of shared
tweets containing an article with that hyperlink).
A (Left).
The main core of the network (
k= 14
); blue nodes are
Italian disinformation websites, green ones are Italian traditional news outlets, red nodes are social networks, the
sky-blue node is a video sharing website and the pink one is an online encyclopedia.
B (Right).
The sub-graph of
Russian (orange), EU (olive green), US (violet) and Italian (blue) disinformation outlets.
against the spread of deceptive information in coordination with findings from third-party researchers. Nevertheless, we
argue that closing malicious pages is not sufficient and more proactive strategies should be followed [26, 3].
In order to check the relevance of inter-connections with websites of different countries, we applied a simple degree
preserving randomization [
] to the network depicted in Fig 11B and tested whether the percentages of links respectively
towards EU, US and RU were significantly different from the mean value observed in the random ensemble (obtained
re-wiring the network for 1000 times). We thus performed a Z-test at
α= 0.05/3
, rejecting the null hypothesis in
all cases; in particular the number of RU and US connections are higher than expected whereas the number of EU
connections is lower.
Finally, we performed a Mann-Kendall test to see whether there was an increasing trend, towards the elections, in the
number of external connections with US and RU disinformation websites; we rejected it at α= 0.05/2=0.0025.
We studied the reach of Italian disinformation on Twitter for a period of five months immediately preceding the European
by analyzing the content production of websites producing disinformation, and the characteristics of
users sharing malicious items on the social platform. Overall, thousands of articles–which included hoaxes, propaganda,
hyper-partisan and conspiratorial news–were shared in the period preceding the elections. We observed that a few
outlets accounted for most of the deceptive information circulating on Twitter; among them, we also encountered a
few websites which were recently banned from Facebook after violating the platform’s terms of use. We identified
a heterogeneous yet limited community of thousands of users who were responsible for sharing disinformation. The
majority of the accounts (more than 75%) occasionally engaged with malicious content, sharing less than 10 stories
each, whereas only a few hundred accounts were responsible for (the spreading) of thousands of articles (see Fig 5).
We singled out the most debated topics of disinformation
by inspecting news items and Twitter hashtags.
We observed that they mostly concern polarizing and controversial arguments of the local political debate such as
immigration, crime and national safety, whereas discussion around the topics of Europe global management had a
Figure 12: An example of disinformation story who was published on a Swedish website ("") and then
reported by an Italian outlet (""). Interestingly, this news is old (July 2018) but it was diffused again in the
first months of 2019.
negligible presence throughout the collection period; the lack of European topics was also reported in the agenda of
mainstream media.
Then we identified the most influential accounts in the diffusion network resulting from users sharing disinformation
articles on Twitter
, so as to detect the presence of active groups with precise political affiliations. We discovered
strong ties with the Italian far-right and conservative community, in particular with "Lega" party, as most of the users
manifested explicit support to the party agenda through the use of keywords and hashtags. Besides, a common deceptive
strategy was to passively involve his leader Matteo Salvini via mentions, quotes and replies as to potentially mislead his
audience of million of followers. We found limited evidence of bot activity in the main core, and we observed that
disabling a limited number of central users in the network would considerably reduce the spread of disinformation
circulating on Twitter, but it would immediately raise censorship concerns.
Finally, we investigated inter-connections within different deceptive agents
, thereby observing that they repeat-
edly linked to each other websites during the period preceding the elections. Moreover we discovered many cases where
the same (or similar) stories were shared in different languages across different European countries, as well as U.S. and
This analysis confirms that disinformation is present on Twitter and that its spread shows some peculiarities in terms of
themes being discussed and of political affiliation of the key members of the information spreading community. We
are aware that disinformation news in Italy have a higher share on Facebook than Twitter and that the use of Twitter
in Italy as a social channel is limited compared to other social platforms such as Facebook, WhatsApp or Instagram.
Therefore similar studies on other social media platforms will be needed and beneficial to our understanding of the
spread of disinformation.
F.P. and S.C. are supported by the PRIN grant HOPE (FP6, Italian Ministry of Education). S.C. is partially supported
by ERC Advanced Grant 693174. The authors are very grateful to Hoaxy developers team at Bloomington Indiana
Pierri F, Ceri S. False news on social media: a data-driven survey. ACM SIGMOD Record Vol 48 Issue 2 (June).
Allcott H, Gentzkow M. Social media and fake news in the 2016 election. Journal of Economic Perspectives.
Lazer DMJ, Baum MA, Benkler Y, Berinsky AJ, Greenhill KM, Menczer F, et al. The science of fake news.
Science. 2018;359(6380):1094–1096. doi:10.1126/science.aao2998.
Ferrara E. Disinformation and social bot operations in the run up to the 2017 French presidential election. First
Monday. 2017;22(8).
Bastos MT, Mercea D. The Brexit botnet and user-generated hyperpartisan news. Social Science Computer
Review. 2019;37(1):38–54.
Vosoughi S, Roy D, Aral S. The spread of true and false news online. Science. 2018;359(6380):1146–1151.
Shao C, Ciampaglia GL, Varol O, Yang KC, Flammini A, Menczer F. The spread of low-credibility content by
social bots. Nature communications. 2018;9(1):4787.
Shao C, Hui PM, Wang L, Jiang X, Flammini A, Menczer F, et al. Anatomy of an online misinformation network.
PLOS ONE. 2018;13(4):1–23. doi:10.1371/journal.pone.0196087.
Grinberg N, Joseph K, Friedland L, Swire-Thompson B, Lazer D. Fake news on Twitter during the 2016 U.S.
presidential election. Science. 2019;363(6425):374–378. doi:10.1126/science.aau2706.
Bovet A, Makse HA. Influence of fake news in Twitter during the 2016 US presidential election. Nature
Communications. 2019;10(1):7. doi:10.1038/s41467-018-07761-2.
Pierri F, Piccardi C, Ceri S. Topology comparison of Twitter diffusion networks reliably reveals disinformation
news. arXiv. 2019;.
[12] Henley J. How populism emerged as an electoral force in Europe. The Guardian. 2018;.
Dennison S, Zerka P. The 2019 European election: How anti-Europeans plan to wreck Europe and what can be
done to stop it. European council on foreign relations. 2019;.
Howard PN, Bradshaw S, Kollanyi B, Bolsolver G. Junk News and Bots during the French Presidential Election:
What Are French Voters Sharing Over Twitter In Round Two?;.
Stella M, Ferrara E, De Domenico M. Bots increase exposure to negative and inflammatory content in online
social systems. Proceedings of the National Academy of Sciences. 2018;115(49):12435–12440.
Hedman F, Sivnert F, Howard P. News and political information consumption in Sweden: Mapping the 2018
Swedish general election on Twitter; 2018.
Kollanyi B, Howard PN. Junk news and bots during the German parliamentary election: What are German voters
sharing over Twitter; 2017.
Marchal N, Kollanyi B, Neudert LM, Howard PN. Junk News During the EU Parliamentary Elections: Lessons
from a Seven-Language Study of Twitter and Facebook. 2019;.
Commission E. Tackling online disinformation; 2019. Available from:
digital-single-market/en/tackling-online- disinformation.
Nielsen RK, Newman N, Fletcher R, Kalogeropoulos A. Reuters Institute Digital News Report 2019. Report of
the Reuters Institute for the Study of Journalism. 2019;.
Del Vicario M, Gaito S, Quattrociocchi W, Zignani M, Zollo F. News consumption during the Italian referendum:
A cross-platform analysis on facebook and twitter. In: 2017 IEEE International Conference on Data Science and
Advanced Analytics (DSAA). IEEE; 2017. p. 648–657.
Vicario MD, Quattrociocchi W, Scala A, Zollo F. Polarization and fake news: Early warning of Potential
misinformation targets. ACM Transactions on the Web (TWEB). 2019;13(2):10.
Giglietto F, Iannelli L, Rossi L, Valeriani A, Righetti N, Carabini F, et al. Mapping Italian News Media Political
Coverage in the Lead-Up to 2018 General Election. Available at SSRN:
AGCOM. News vs Fake nel sistema dell’informazione. Report available at:
10179/12791486/Pubblicazione+23-11-2018/93869b4f-0a8d- 4380-aad2-c10a0e426d83?
version=10. 2018;.
Cantarella M, Fraccaroli N, Volpe R. Does Fake News Affect Voting Behaviour? Available at SSRN:
https://ssrncom/abstract=3402913. 2019;.
Avaaz. Far Right Networks of Deception. Available at:
20Network%20Deception%2020190522pdf. 2019;.
Shao C, Ciampaglia GL, Flammini A, Menczer F. Hoaxy: A Platform for Tracking Online Misinformation.
In: Proceedings of the 25th International Conference Companion on World Wide Web. WWW ’16 Companion.
Republic and Canton of Geneva, Switzerland: International World Wide Web Conferences Steering Committee;
2016. p. 745–750. Available from:
Hui PM, Shao C, Flammini A, Menczer F, Ciampaglia GL. The Hoaxy misinformation and fact-checking diffusion
network. In: Twelfth International AAAI Conference on Web and Social Media; 2018.
Zollo F, Bessi A, Del Vicario M, Scala A, Caldarelli G, Shekhtman L, et al. Debunking in a world of tribes. PloS
one. 2017;12(7):e0181821.
Rieder B. Studying Facebook via data extraction: the Netvizz application. In: Proceedings of the 5th annual ACM
web science conference. ACM; 2013. p. 346–355.
Hagberg A, Swart P, S Chult D. Exploring network structure, dynamics, and function using NetworkX. Los
Alamos National Lab.(LANL), Los Alamos, NM (United States); 2008.
[32] Barabási AL. Network science. Cambridge university press; 2016.
Batagelj V, Zaversnik M. An O(m) algorithm for cores decomposition of networks. arXiv preprint cs/0310049.
[34] Fortunato S, Hric D. Community detection in networks: A user guide. Physics reports. 2016;659:1–44.
Blondel VD, Guillaume JL, Lambiotte R, Lefebvre E. Fast unfolding of communities in large networks. Journal
of statistical mechanics: theory and experiment. 2008;2008(10):P10008.
Girvan M, Newman ME. Community structure in social and biological networks. Proceedings of the national
academy of sciences. 2002;99(12):7821–7826.
[37] Freeman LC. A set of measures of centrality based on betweenness. Sociometry. 1977; p. 35–41.
Page L, Brin S, Motwani R, Winograd T. The PageRank citation ranking: Bringing order to the web. Stanford
InfoLab; 1999.
Mann HB. Nonparametric tests against trend. Econometrica: Journal of the Econometric Society. 1945; p.
[40] Kendall MG. Rank correlation methods. Griffin. 1948;.
[41] Efron B, Hastie T. Computer age statistical inference. vol. 5. Cambridge University Press; 2016.
Ratkiewicz J, Conover MD, Meiss M, Gonçalves B, Flammini A, Menczer FM. Detecting and tracking political
abuse in social media. In: Fifth international AAAI conference on weblogs and social media; 2011.
McCombs ME, Shaw DL, Weaver DH. New directions in agenda-setting theory and research. Mass communication
and society. 2014;17(6):781–802.
Vargo CJ, Guo L, Amazeen MA. The agenda-setting power of fake news: A big data analysis of the online media
landscape from 2014 to 2016. New Media & Society. 2018;20(5):2028–2049. doi:10.1177/1461444817712086.
Horne BD, Adali S. This just in: fake news packs a lot in title, uses simpler, repetitive content in text body, more
similar to satire than real news. arXiv preprint arXiv:170309398. 2017;.
Zajonc RB. Mere exposure: A gateway to the subliminal. Current directions in psychological science.
Blei DM, Ng AY, Jordan MI. Latent dirichlet allocation. Journal of machine Learning research. 2003;3(Jan):993–
Vargo CJ, Guo L, McCombs M, Shaw DL. Network issue agendas on Twitter during the 2012 US presidential
election. Journal of Communication. 2014;64(2):296–316.
Wang P, Angarita R, Renna I. Is this the era of misinformation yet: combining social bots and fake news to
deceive the masses. In: Companion Proceedings of the The Web Conference 2018. International World Wide Web
Conferences Steering Committee; 2018. p. 1557–1561.
International A. Il Barometro dell’odio - Elezioni europee 2019. Available at:
cosa-facciamo/elezioni-europee/. 2019;.
[51] Conti N. Elezioni europee, ma poca Europa. La Repubblica. 2019;.
[52] FactCheckEU. Good news and bad news after election week-end. 2019;.
[53] McCombs M. Setting the agenda: Mass media and public opinion. John Wiley & Sons; 2018.
Davis CA, Varol O, Ferrara E, Flammini A, Menczer F. Botornot: A system to evaluate social bots. In: Proceedings
of the 25th International Conference Companion on World Wide Web. International World Wide Web Conferences
Steering Committee; 2016. p. 273–274.
Maslov S, Sneppen K. Specificity and stability in topology of protein networks. Science. 2002;296(5569):910–913.
ResearchGate has not been able to resolve any citations for this publication.
Full-text available
In recent years, malicious information had an explosive growth in social media, with serious social and political backlashes. Recent important studies, featuring large-scale analyses, have produced deeper knowledge about this phenomenon, showing that misleading information spreads faster, deeper and more broadly than factual information on social media, where echo chambers, algorithmic and human biases play an important role in diffusion networks. Following these directions, we explore the possibility of classifying news articles circulating on social media based exclusively on a topological analysis of their diffusion networks. To this aim we collected a large dataset of diffusion networks on Twitter pertaining to news articles published on two distinct classes of sources, namely outlets that convey mainstream, reliable and objective information and those that fabricate and disseminate various kinds of misleading articles, including false news intended to harm, satire intended to make people laugh, click-bait news that may be entirely factual or rumors that are unproven. We carried out an extensive comparison of these networks using several alignment-free approaches including basic network properties, centrality measures distributions, and network distances. We accordingly evaluated to what extent these techniques allow to discriminate between the networks associated to the aforementioned news domains. Our results highlight that the communities of users spreading mainstream news, compared to those sharing misleading news, tend to shape diffusion networks with subtle yet systematic differences which might be effectively employed to identify misleading and harmful information.
Full-text available
Over the last decade, the erosion of trust in public institutions and traditional media sources have been proceeding in parallel. As recent developments in media consumption have led to a proliferation of politically charged online misinformation, it is no wonder that many have been questioning whether the spread of fake news has affected the results of recent elections, contributing to the growth of populist party platforms. In this work, we aim to quantify this impact by focusing on the causal effect of the spread of misinformation over electoral outcomes in the 2018 Italian General elections. We exploit the presence of Italian and German linguistic groups in the Trento and Bolzano/Bozen autonomous provinces as an exogenous source of variation, assigning individuals into distinct filter bubbles each differently exposed to misinformation. To do so, we introduce a novel index based on text mining techniques to measure populism. Using this approach, we analyse the social media content of each party and their leaders over the course of the electoral campaign for the 2013 and 2018 elections. We then collect electoral and socio-demographic data from the region and, after constructing a proxy for exposition to misinformation, we measure the change in populist vote across the two groups in-between the two general elections, using a combination of difference-indifference and two-stages-least-squares inference methods. Our results indicate that misinformation had a negligible and non-significant effect on populist vote in Trentino and South Tyrol during the Italian 2018 general elections.
Full-text available
The massive spread of digital misinformation has been identified as a major threat to democracies. Communication, cognitive, social, and computer scientists are studying the complex causes for the viral diffusion of misinformation, while online platforms are beginning to deploy countermeasures. Little systematic, data-based evidence has been published to guide these efforts. Here we analyze 14 million messages spreading 400 thousand articles on Twitter during ten months in 2016 and 2017. We find evidence that social bots played a disproportionate role in spreading articles from low-credibility sources. Bots amplify such content in the early spreading moments, before an article goes viral. They also target users with many followers through replies and mentions. Humans are vulnerable to this manipulation, resharing content posted by bots. Successful low-credibility sources are heavily supported by social bots. These results suggest that curbing social bots may be an effective strategy for mitigating the spread of online misinformation.
Full-text available
Significance Social media can deeply influence reality perception, affecting millions of people’s voting behavior. Hence, maneuvering opinion dynamics by disseminating forged content over online ecosystems is an effective pathway for social hacking. We propose a framework for discovering such a potentially dangerous behavior promoted by automatic users, also called “bots,” in online social networks. We provide evidence that social bots target mainly human influencers but generate semantic content depending on the polarized stance of their targets. During the 2017 Catalan referendum, used as a case study, social bots generated and promoted violent content aimed at Independentists, ultimately exacerbating social conflict online. Our results open challenges for detecting and controlling the influence of such content on society.
Full-text available
The dynamics and influence of fake news on Twitter during the 2016 US presidential election remains to be clarified. Here, we use a dataset of 171 million tweets in the five months preceding the election day to identify 30 million tweets, from 2.2 million users, which contain a link to news outlets. Based on a classification of news outlets curated by, we find that 25% of these tweets spread either fake or extremely biased news. We characterize the networks of information flow to find the most influential spreaders of fake and traditional news and use causal modeling to uncover how fake news influenced the presidential election. We find that, while top influencers spreading traditional center and left leaning news largely influence the activity of Clinton supporters, this causality is reversed for the fake news: the activity of Trump supporters influences the dynamics of the top fake news spreaders.
In the past few years, the research community has dedicated growing interest to the issue of false news circulating on social networks. The widespread attention on detecting and characterizing deceptive information has been motivated by considerable political and social backlashes in the real world. As a matter of fact, social media platforms exhibit peculiar characteristics, with respect to traditional news outlets, which have been particularly favorable to the proliferation of false news. They also present unique challenges for all kind of potential interventions on the subject. As this issue becomes of global concern, it is also gaining more attention in academia. The aim of this survey is to offer a comprehensive study on the recent advances in terms of detection, characterization and mitigation of false news that propagate on social media, as well as the challenges and the open questions that await future research on the field. We use a data-driven approach, focusing on a classification of the features that are used in each study to characterize false information and on the datasets used for instructing classification methods. At the end of the survey, we highlight emerging approaches that look most promising for addressing false news.
Users’ polarization and confirmation bias play a key role in misinformation spreading on online social media. Our aim is to use this information to determine in advance potential targets for hoaxes and fake news. In this article, we introduce a framework for promptly identifying polarizing content on social media and, thus, “predicting” future fake news topics. We validate the performances of the proposed methodology on a massive Italian Facebook dataset, showing that we are able to identify topics that are susceptible to misinformation with 77% accuracy. Moreover, such information may be embedded as a new feature in an additional classifier able to recognize fake news with 91% accuracy. The novelty of our approach consists in taking into account a series of characteristics related to users’ behavior on online social media such as Facebook, making a first, important step towards the mitigation of misinformation phenomena by supporting the identification of potential misinformation targets and thus the design of tailored counter-narratives.
Finding facts about fake news There was a proliferation of fake news during the 2016 election cycle. Grinberg et al. analyzed Twitter data by matching Twitter accounts to specific voters to determine who was exposed to fake news, who spread fake news, and how fake news interacted with factual news (see the Perspective by Ruths). Fake news accounted for nearly 6% of all news consumption, but it was heavily concentrated—only 1% of users were exposed to 80% of fake news, and 0.1% of users were responsible for sharing 80% of fake news. Interestingly, fake news was most concentrated among conservative voters. Science , this issue p. 374 ; see also p. 348
This study presents an analysis of the online media coverage in the run up of 2018 Italian General election. We illustrate how immigration, corruption and privileges of the elite – also related to a certain rhetoric on the inability of the state to protect the rights of the needy – were in fact the most salient topics throughout the months before and during the election. Both topics were largely central in both Salvini’s League and Di Maio’s Five Stars Movement agenda. Nevertheless, the leaders most frequently cited in online news articles were Matteo Renzi (PD) and Silvio Berlusconi (Forza Italia). By deep diving into the contents of main leaders media coverage and respective social media engagement we document the centrality of stories unravelling around leaders’ legal issues, alleged collusion and scandals. While the media clusters emerging from the network analysis clearly resembles the tripartite structure of the contemporary Italian politics articulated in centre-left, centre-right and Five Stars Movement, the different weight and articulation within each cluster clearly describe the strengths of M5S and centre-right (largely dominated by the League) and the weakness of the centre-left. While our analysis depicted a clear profile of the actors and topics that catalyzed the highest social media interactions and thus attention, we also illustrate a number of strategies employed by different communities to amplify the reach of contents aligned with own worldview while reframing negative coverage through comments. Overall, explicitly partisan and hyper-partisan sources catalyzed a significant share of the social media interactions performed by the online audience in the lead up of the election. The analysis includes the evaluation and mapping of the Italian media landscape from several perspectives and is based on large-scale data collection of online news articles published on the web, shared and interacted on Facebook and Twitter.
The twenty-first century has seen a breathtaking expansion of statistical methodology, both in scope and in influence. ‘Big data’, ‘data science’, and ‘machine learning’ have become familiar terms in the news, as statistical methods are brought to bear upon the enormous data sets of modern science and commerce. How did we get here? And where are we going? This book takes us on an exhilarating journey through the revolution in data analysis following the introduction of electronic computation in the 1950s. Beginning with classical inferential theories - Bayesian, frequentist, Fisherian - individual chapters take up a series of influential topics: survival analysis, logistic regression, empirical Bayes, the jackknife and bootstrap, random forests, neural networks, Markov chain Monte Carlo, inference after model selection, and dozens more. The distinctly modern approach integrates methodology and algorithms with statistical inference. The book ends with speculation on the future direction of statistics and data science.