Fake News Spreaders Profiling Through Behavioural
Analysis
Notebook for PAN at CLEF 2020
Matteo Cardaioli1,2 , Stefano Cecconello1 , Mauro Conti1 , Luca Pajola1 , and Federico
Turrin1
1
Department of Mathematics, University of Padua, Padua
2
GFT Italy, Milano, Italy
{matteo.cardaioli, stefano.cecconello, conti, pajola, turrin}@math.unipd.it
Abstract The growth of social media and the people interconnection led to the
digitalization of communication. Nowadays the most influential politicians or sci-
entific communicators use the media to disseminate news or decisions. However,
such communications media can be used maliciously to spread the so-called fake-
news in order to polarise public opinion or to deny scientific theories. It is there-
fore important to develop intelligent and accurate techniques in order to identify
the spreading of fake-news. In this paper, we describes the methodology regarding
our participation in the PAN@CLEF Profiling Fake News Spreaders on Twitter
competition. We propose a supervised Machine-Learning (ML) based framework
to profile fake-news spreaders. Our method relies on the combination of Big Five
personality and stylometric features. Finally, we evaluate our framework detec-
tion capabilities and performance with different ML models on a tweeter dataset
in both English and Spanish languages.
1 Introduction
Social Networks, such as Twitter, Facebook, or Instagram, are nowadays the main
source of information for millions of people. The success of Social Networks is mainly
due to their usability and their free access to every kind of information [20]. However,
Social Networks lack in content control, and this allowed the emergence of several ma-
licious phenomena such as the aggregation of terrorist groups [27] and the spreading
of hate messages [2] or fake-news messages [17]. Due to their impact, in recent years,
the analysis of online platforms as Online Social Networks (OSNs), blogs, and forums
attracted a wide area of security researchers. Different tasks can be found, from bot
detection in OSNs [1, 18] to hate speech detection [5, 13].
Among the different Social Network analysis, fake-news detection is increasingly
attracting the researcher and industrial attention due to the impact of such a phenomenon.
In [11], the authors investigated on the fake-news spreading on Twitter during the 2016
U.S. presidential election showing dramatic results in term of fake-news exposition and
Copyright c 2020 for this paper by its authors. Use permitted under Creative Commons Li-
cense Attribution 4.0 International (CC BY 4.0). CLEF 2020, 22-25 September 2020, Thessa-
loniki, Greece.
sharing. To limit the spreading of fake news, it is also important to identify in advance
users more likely believe in them.
These results highlight the importance of developing accurate techniques to identify
and prevent fake-news spreading. In this context, ML plays an important role because
it can help to classify fake-news in an automatic manner, that otherwise would be im-
possible for a human operator. Common approaches for the identification of fake-news
rely on traditional methods of features extraction such as: Bag of Words, TF-IDF, sty-
lometry; but also novel methods such as fact-checkers automatic detectors [24].
An unexplored field related to fake-news detection is the identification of fake news
spreaders, which is the goal of PAN@CLEF Profiling Fake News Spreaders on Twitter
competition [21]. In this work, we investigate how behavioural features (i.e., personal-
ity, stylometry) can discriminate a fake-news spreader.
We summarise the contribution of the paper as follows:
– We present and implement a framework for profiling fake-news spreaders by ex-
ploiting behavioural features (i.e., personality, stilometry).
– We compare and evaluate the classification performances of different supervised
ML models in a Twitter feed dataset.
The remainder of the paper is organised as follows. Section 2 provides an overview
of the related work. Sections 3 and 4 outline respectively our methodology we used
to identify the fake-news tweets and the evaluation approach we implemented. Finally,
Section 5 concludes the paper.
2 Related works
The analysis of phenomena in social networks can be very complex and thus standard
Natural Language Processing (NLP) techniques might not be very effective. For profil-
ing tasks [23] (e.g., gender, race) researchers started to look in other disciplines (e.g.,
psychology) to enrich tasks’ feature space; for example, common trends involve tech-
niques that use “emotions” or “behaviours”.
Emotion-based profiling. In online platforms such as social networks, users interact
with each other through posts, comments, and messages, sharing ideas, feelings, and
emotions. Such features can be used to understand better both cultural-insights and hu-
man behaviours. For example, Kusen et al. [19] analyze users’ emotions during different
events and found that they tend to conform to the emotional valence of the respective
real-world event; however, users also tend to react with positive messages during neg-
ative events, exhibiting a strong shifted emotion. Recently, several works show that
emotional-based features can improve the efficacy of tasks such as clickbait identifica-
tion [8] and fake-news detection [9].
Behavioural-based profiling. Temporal patterns can be a powerful tool to profile phe-
nomena. For example, behavioural-based techniques are a consolidated standard in bio-
metric [26]. In the context of online platforms security, behavioural features are used
to detect bots [3, 14, 16]. Chu et al. [3] use behavioural biometrics such as mouse and
keystroke dynamics to discriminate bots from humans. Similarly, Hall et al. [14] to de-
tect bots activities in Wikipedia developed an algorithm that leverages on the fact that
bots patterns are often distinct from human patterns. In massively multiplayer online
role playing games (MMORPG) bots are often used to obtain games items that can be
exchanged with real money. Kang et al. designed a detection approach to solve such a
problem by identifying behavioural patterns that are unique for bots in MMORPG [16].
3 Methods
This study aims to detect Twitter user profiles that are keen to be spreaders of fake-news.
Our analysis wants to retrieve information that can provide insights into the attitude to
spread fake-news, rather than identifying potential fake-news among the users’ tweets.
Our approach is therefore focused on the extraction of behavioural characteristics of the
user such as personality and writing style.
3.1 Dataset description
The provided dataset contains 60000 tweets (100 per user) derived from 600 distinct
Twitter accounts. For each user profile are also reported: the language, a unique author
ID, and a binary label that defines the class (the information regarding which label cor-
responds to each class is not provided by the organizers due to GDPR reasons). The
users are equally distributed among two languages: 300 English profiles and 300 Span-
ish profiles. Furthermore, the dataset contains an equal number of fake-news spreader
and not fake-news spreader profiles. To guarantee better profile anonymisation, sensible
information contained in the original tweets has been obfuscated using some keywords.
In particular, the following keywords were used: “user”, “rt” (re-tweet), “hashtag”,
“URL”. The keywords are always preceded by a hashtag and are always capitalised.
3.2 Pre-processing and Feature Extraction
In this study, we wanted to use a multidisciplinary approach, with the aim of exploring
the impact of uncommon features on identification of fake news spreaders. We focused
our analysis on two main behavioural characteristics: writing style and personality.
Writing Style Features Stylometry features have been used to solve several tasks [7].
Guided by their popularity, we decide to include 10 stylometric features that summarise
the writing style of the users. In particular, we evaluated:
– Diversity score describes how tweets are novel between each other. We use the def-
inition given in [25], and it is defined as the average diversity of the user tweets.
Given a list of tweets C, the diversity of the tweet ti is defined as 1−max(Jac(xi , xj )),
where i 6= j and j ∈ [0, ..., |C|]. “Jac” is the Jaccard similarity of two sets.
– Readability score is the average of the “Flesch reading ease” of the tweets. The
score is calculated with the python library “Textstat”3 . This particular metric is
defined over a multiple set of languages, such as English and Spanish.
3
https://pypi.org/project/textstat/
– Hash avg is the average hashtags per tweet.
– Usr avg is the average of mentions per tweet.
– Url avg is the average of URLs per tweet.
– Retweet is the average retweets per tweet.
– Lower is the average lowercase characters per tweet.
– Upper is the average uppercase characters per tweet.
– Punctuation is the average of punctuation characters per tweet.
– Alpha is the average of alphabetical characters per tweet.
Personality Features To extract personality features, we firstly performed a pre-processing
on the dataset. We removed all the keywords used to obfuscate sensible information
(e.g, #URL, #RT). We then merged all the tweets, creating a unique corpus for every
user. We used Watson Personality Insights - IBM 4 to retrieve personality information
from written text. In particular, giving as input the corpus and the language (i.e., “Span-
ish” or “English”). The output of IBM Watson consists of a JSON containing: Needs,
Values, and Big Five personality characteristics. For each of them, the service provides
a percentile score. The higher this score is, the greater is also the presence of the specific
personality trait for the user. Needs describe at a high level those aspects of a product
that are likely to resonate with the author of the input text. Values describe motivating
factors that influence the author’s decision-making. The Big Five model [15, 22], or
OCEAN, is one of the most famous in the study of personality. According to this the-
ory, personality can be divided into five independent traits. Below, we provide a brief
definition for each trait, as reported in [12].
– Agreeableness is a person’s tendency to be compassionate and cooperative toward
others.
– Conscientiousness is a person’s tendency to act in an organised or thoughtful way.
– Extraversion is a person’s tendency to seek stimulation in the company of others.
– Emotional Range, also referred to as Neuroticism or Natural Reactions, is the extent
to which a person’s emotions are sensitive to the person’s environment.
– Openness is the extent to which a person is open to experiencing different activities.
In Figure 2 we compare the average values of the Big Five percentiles for both Class
0 and Class 1. We can notice the following: i) on average, Class 0 tends to have lower
Agreeableness and Extraversion values and ii) on average, Class 0 tends to have higher
Emotional Range values.
Many studies on the Big Five present a two-level hierarchy, with the five domains
discussed above assuming more specific traits, called "facets" at a second level [6]. In
particular, each Big Five trait can assume six facets as described in [4]. The total number
of personality features are 54.
Figure 1 depicts a sunburst chart reporting the percentile scores returned by Watson
APIs. The chart reports the best score for each personality characteristics (Big Five,
Needs, and Values). For Needs and Values, the chart reports the score for the features
provided by Watson (respectively 12 and 5 values). For Big Five, a supplementary level
reports the percentile for each one of the five characteristics, followed by the same score
for each facet.
4
https://www.ibm.com/watson
Adventurousness (86%)
s (89%)
)
interest
(65%
)
%)
8%
Sel
(86
(9
lity
ng
f-ex
ion
%)
Artistic
gi
a
n
pre
tion
89
nat
)
le
Pr
1%
al
t(
ssi
Stab
Structu
agi
ac
h
(1
Emo
c
-c
lle
on
tic
ng
ity
Im
vi
te
ali
ility
or
ri
(89
In
t )
th
ty
ts 3%
re (1%)
Lo
Au
%)
Li
en
(6
(14%
be (5
ve
ss
2%
rt Ope em e
(4
y nn v n
ie us
)
Id (2 es
3%
)
s
ea 9%
(92
ch io
ut
%)
l A )
)
Ha (3 ) a 7%
rm 3% C s(
Exc on ) n es )
item y( ur
e
tif
ul 56%
ent 2% ct ) Du ss (
) line
Co
(29 e r
ru
%
%) Ord
ns
St
ds
Curio )
(1
(22%
cie
sity ( line
(17
nt i o
6%) iscip
e
Self-d
%)
Ne
usnes
Closeness
(3%)
cy (8%)
Self-effica
s
Challenge (2%)
Bi g 5
8%) Activity level (27%)
Valu
Cons
ence (1
nscend
Self-tra
O
3%)
( 9 2 e ss
sion
Assert
erv
p
(3 %
( 3
es
nt iveness
e
eme
ati
nn
(27%)
%
ve r
anc %) Che
)
%)
)
on
-enh (44
tra
) erfu
Self
(73
m % l nes
ns 11
Ex
i ) s (3
d o e( % %)
He n g ( 3 Ex
cit
c ha i o n Ou em
to t Em
tg en
ss va ot s oin t-s
eek
er es
ion
ne ns
al r
e ab
l en G
re g ing
en (76 ang
p o %)
e Agre
ga (3 (60
O C ) 1%
)
(8 % ri %)
7%
Al
tr ou )
Co
(5
sn
)
ui
6%
es
ss
sm
op
s
Unc
re
%)
Trust (12%)
(6
(4
e
Mo
(2
st
ra
2
(73
ss
9%
%
to
)
t
om
de
2%)
)
io
4%
ne
)
Symp
le
Fiery (90%)
n
ion
sty
us
pro
ib
9
(2
rry (8
pt
cio
ly (
rat
4%
(65
e
mis
sc
ns
a
cho
de
)
thy (8
Su
%
co
ing
to wo
mo
)
lf-
lan
Se
Im
(24
8%)
Me
Prone
%)
Figure 1. Example sunburst chart reporting the percentile scores returned by the API provided by
Watson Personality Insights.
4 Evaluation
In this section, we discuss the process implied to build our model. The pipeline of our
approach is defined as follows:
– Pre-processing. Special token cleaning (for personality features only).
– Feature extraction. Stylometric and personality features. The final set of features
consists of the concatenation of both stylometric and personality features, with the
addition of the typing language (for a total of 64 features). This approach allow us
to implement cross-language models.
– Feature selection. The dimensionality of data is reduced by applying a KBest fea-
ture selection algorithm.
– Model. Simple binary classification models (e.g., Random Forest).
Class 0 Class 1
Conscientiousness
Extraversion Openness
0.2 0.4 0.6 0.8 1.0
Agreeableness Emotional
Figure 2. Radar chart reporting the average Big Five scores for “Class 0” and “Class 1” classes.
Different models are evaluated to discriminate between fake-news spreaders and
non-spreaders: Decision Tree (DT), K-Nearest Neighbour (KNN), Support Vector Clas-
sifier (SVC), and Random Forest (RF). For the model selection, we used a repeated
nested-cross-validation (100 times for 80-20 random splits) with 5-fold inner cross-
validation, applied to the aforementioned pipeline. This approach is used to assess the
robustness of the prediction for every ML model among different splits. According to
Table 1 we selected RF classifier that showed the highest average accuracy and the
lowest variability in nested-cross-validation. Finally, we trained the RF classifier on the
whole dataset, using the same pipeline with a 5-fold cross-validation. To obtain the best
RF’s hyperparameters a 5-fold cross-validation is applied on the same parameter grid
shown in Table 1, obtaining n_estimators = 25 and max_depth = 4. Finally, RF is
trained on the whole dataset, using these hyperparameters.
Model Parameter Values Training Accuracy Test Accuracy
DT max_depth [2, . . . , 5] 0.75 ± 0.05 0.65 ± 0.05
KNN n_neighbors [3, . . . , 10] 0.81 ± 0.03 0.71 ± 0.04
kernel [‘rbf’. ‘linear’]
SVC c [10−2 , 10−1 , 1, 102 ] 0.94 ± 0.08 0.73 ± 0.04
γ [10−3 , 10−2 , 10−1 , 1, 10]
n_estimators [25, 50, 100, 200]
RF 0.87 ± 0.03 0.74 ±0.03
max_depth [3, 4, 5]
Table 1. Training and Test accuracy of the nested validation for the different models.
5 Conclusions
In this work, we describe the approach we used for the PAN@CLEF Profiling Fake
News Spreaders on Twitter challenge. Our proposed framework relies on the combina-
tion of linguistic features with psychological traits with simple machine learning meth-
ods (e.g., Random Forest). Our method reached the following accuracy: 0.6750 (En-
glish corpus), 0.7150 (Spanish corpus), and 0.6950 (average between the two corpus).
Our work shows the feasibility of using psychological features to determine whether
a user is a fake news spreader or not. The intuition on using psychological behaviours
in the fake news domain is also confirmed by the recent work proposed by Giachanou
et al. [10]5 , where the authors combines linguistic patterns with personality scores to
distinguish between fake news spreaders and checkers. We believe that future research
directions could benefit from this, enlarging the feature space of the data.
References
1. Chavoshi, N., Hamooni, H., Mueen, A.: Debot: Twitter bot detection via warped correlation.
In: 2016 IEEE 16th International Conference on Data Mining (ICDM). pp. 817–822 (2016)
2. Chetty, N., Alathur, S.: Hate speech review in the context of online social networks.
Aggression and violent behavior 40, 108–118 (2018)
3. Chu, Z., Gianvecchio, S., Koehl, A., Wang, H., Jajodia, S.: Blog or block: Detecting blog
bots through behavioral biometrics. Computer Networks 57(3), 634 – 646 (2013)
4. Costa Jr, P.T., McCrae, R.R.: The Revised NEO Personality Inventory (NEO-PI-R). Sage
Publications, Inc (2008)
5. Davidson, T., Warmsley, D., Macy, M., Weber, I.: Automated hate speech detection and the
problem of offensive language. In: Proceedings of the 11th International AAAI Conference
on Web and Social Media. pp. 512–515. ICWSM ’17 (2017)
6. DeYoung, C.G., Quilty, L.C., Peterson, J.B.: Between facets and domains: 10 aspects of the
big five. Journal of personality and social psychology 93(5), 880 (2007)
7. Feng, S., Banerjee, R., Yejin, C.: Syntactic stylometry for deception detection. vol. 2, pp.
171–175 (01 2012)
8. Ghanem, B., Rosso, P., Rangel, F.: An Emotional Analysis of False Information in Social
Media and News Articles. ACM Transactions on Internet Technology (TOIT) 20(2), 1–18
(2020)
9. Giachanou, A., Rosso, P., Crestani, F.: Leveraging emotional signals for credibility
detection. In: Proceedings of the 42nd International ACM SIGIR Conference on Research
and Development in Information Retrieval. p. 877–880. SIGIR’19, Association for
Computing Machinery, New York, NY, USA (2019)
10. Giachanou, A., Ríssola, E., Ghanem, B., Crestani, F., Rosso, P.: The Role of Personality and
Linguistic Patterns in Discriminating Between Fake News Spreaders and Fact Checkers, pp.
181–192 (06 2020)
11. Grinberg, N., Joseph, K., Friedland, L., Swire-Thompson, B., Lazer, D.: Fake news on
twitter during the 2016 us presidential election. Science 363(6425), 374–378 (2019)
12. Grippa, F., Leitão, J., Gluesing, J., Riopelle, K., Gloor, P.: Collaborative Innovation
Networks. Springer (2018)
5
This paper is published after the time of our submission.
13. Gröndahl, T., Pajola, L., Juuti, M., Conti, M., Asokan, N.: All you need is “love”: Evading
hate speech detection. In: Proceedings of the 11th ACM Workshop on Artificial Intelligence
and Security. p. 2–12. AISec ’18, Association for Computing Machinery, New York, NY,
USA (2018)
14. Hall, A., Terveen, L., Halfaker, A.: Bot detection in wikidata using behavioral and other
informal cues. Proc. ACM Hum.-Comput. Interact. 2(CSCW) (Nov 2018)
15. John, O.P., Srivastava, S., et al.: The big five trait taxonomy: History, measurement, and
theoretical perspectives. Handbook of personality: Theory and research 2(1999), 102–138
(1999)
16. Kang, A., Jeong, S.H., Mohaisen, A., Kim, H.K.: Multimodal game bot detection using user
behavioral characteristics. SpringerPlus 5 (12 2016)
17. Koohikamali, M., Sidorova, A.: Information re-sharing on social network sites in the age of
fake news. Informing Science 20 (2017)
18. Kudugunta, S., Ferrara, E.: Deep neural networks for bot detection. Information Sciences
467, 312 – 322 (2018)
19. Kušen, E., Strembeck, M., Cascavilla, G., Conti, M.: On the influence of emotional valence
shifts on the spread of information in social networks. In: Proceedings of the 2017
IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining
2017. p. 321–324. ASONAM ’17, Association for Computing Machinery, New York, NY,
USA (2017)
20. Potthast, M., Gollub, T., Wiegmann, M., Stein, B.: TIRA Integrated Research Architecture.
In: Ferro, N., Peters, C. (eds.) Information Retrieval Evaluation in a Changing World.
Springer (Sep 2019)
21. Rangel, F., Giachanou, A., Ghanem, B., Rosso, P.: Overview of the 8th Author Profiling
Task at PAN 2020: Profiling Fake News Spreaders on Twitter. In: Cappellato, L., Eickhoff,
C., Ferro, N., Névéol, A. (eds.) CLEF 2020 Labs and Workshops, Notebook Papers.
CEUR-WS.org (Sep 2020)
22. Rothmann, S., Coetzer, E.P.: The big five personality dimensions and job performance. SA
Journal of Industrial Psychology 29(1), 68–74 (2003)
23. Shu, K., Wang, S., Liu, H.: Understanding user profiles on social media for fake news
detection. In: 2018 IEEE Conference on Multimedia Information Processing and Retrieval
(MIPR). pp. 430–435. IEEE (2018)
24. Vo, N., Lee, K.: Learning from fact-checkers: Analysis and generation of fact-checking
language. In: Proceedings of the 42nd International ACM SIGIR Conference on Research
and Development in Information Retrieval. pp. 335–344 (2019)
25. Wang, K., Wan, X.: Sentigan: Generating sentimental texts via mixture adversarial
networks. In: Proceedings of the Twenty-Seventh International Joint Conference on
Artificial Intelligence, IJCAI-18. pp. 4446–4452. International Joint Conferences on
Artificial Intelligence Organization (7 2018)
26. Yampolskiy, R.V., Govindaraju, V.: Behavioural biometrics: A survey and classification. Int.
J. Biometrics 1(1), 81–113 (Jun 2008)
27. Zech, S.T., Gabbay, M.: Social network analysis in the study of terrorism and insurgency:
From organization to politics. International Studies Review 18(2), 214–243 (2016)