Content uploaded by Iouliana Litou
Author content
All content in this area was uploaded by Iouliana Litou on Apr 27, 2017
Content may be subject to copyright.
A Model for Identifying Misinformation
in Online Social Networks
Sotirios Antoniadis1, Iouliana Litou2(B
), and Vana Kalogeraki2
1Nokia Solutions and Networks Hellas A.E., Athens, Greece
sotiris.antoniadis@nsn.com
2Deptartment of Informatics,
Athens University of Economics and Business, Athens, Greece
{litou,vana}@aueb.gr
Abstract. Online Social Networks (OSNs) have become increasingly
popular means of information sharing among users. The spread of news
regarding emergency events is common in OSNs and so is the spread
of misinformation related to the event. We define as misinformation
any false or inaccurate information that is spread either intentionally
or unintentionally. In this paper we study the problem of misinformation
identification in OSNs, and we focus in particular on the Twitter social
network. Based on user and tweets characteristics, we build a misin-
formation detection model that identifies suspicious behavioral patterns
and exploits supervised learning techniques to detect misinformation.
Our extensive experimental results on 80294 unique tweets and 59660
users illustrate that our approach effectively identifies misinformation
during emergencies. Furthermore, our model manages to timely identify
misinformation, a feature that can be used to limit the spread of the
misinformation.
1 Introduction
Online Social Networks (OSNs) have evolved into major means of communica-
tion and information spreading. They enumerate over 1.61 billion users, which
corresponds to 22% of the world’s population. However, one major challenge is
that the information communicated through the network is not always credible.
Previous studies confirm the existence of spam campaigns in OSNs [1][2]. Spam
campaigns are organized attempts towards spreading false or malicious content
through the coordination of accounts or other illicit means in the network. It is
estimated that, among the messages published on Twitter, 1% of the messages
is spam while 5% of the accounts are spammers1.
Twitter2has evolved as one of the most popular microblogging services. Users
publish short messages (tweets) of at most 140 characters and follow any other
S. Antoniadis—Part of this work was performed when this author was at Athens
University of Economics and Business.
1http://digital.cs.usu.edu/ kyumin/tutorial/www-tutorial.pdf
2https://twitter.com/
c
Springer International Publishing Switzerland 2015
C. Debruyne et al. (Eds.): OTM 2015 Conferences, LNCS 9415, pp. 473–482, 2015.
DOI: 10.1007/978-3-319-26148-5 32
474 S. Antoniadis et al.
registered users to receive status updates. Twitter offers to users the opportu-
nity to report a tweet as spam,compromised or abusive. Other filters to detect
spam (e.g the number of followers in regard to followees, random favorites and
retweets etc.) are also used. Still, tweets containing misinformation regarding
an emergency event may not be identified based solely on the aforementioned
mechanisms.
The credibility of images propagated in the network has been the focus of
recent work. Zubiaga and Ji [3] and Gupta et al. [4] focus on the credibility
of images propagated in the network during emergency events but not in the
content of the information. The works closest to ours is that of Castillo et al.
[5]andXiaetal.[6]. Both works use supervised learning and Bayesian Net-
work classification to identify credible information propagated in the network.
Castillo et al. [5] cluster the instances to newsworthy or chats and later perform
credibility analysis on the newsworthy clusters. Xia et al. [6]proposeamodelto
detect an emergency event and identify credible tweets. As we illustrate in our
experimental evaluation, our approach performs better than both approaches in
identifying misinformative tweets with over 14% higher accuracy.
In this work we suggest a methodology for identifying and limiting misin-
formation spread in OSNs during emergency events by identifying tweets that
are most likely to be inaccurate or irrelevant to an event. Our work makes the
following contributions:
– We present a novel filtering process for identifying misinformation during
emergencies that is fast and effective. The filters are identified based on an
extensive analysis conducted on a large dataset of users and tweets related to
the emergency event of Hurricane Sandy. As our experimental results illus-
trate, the filtering process extracts over 81% of the misinformative tweets,
while identifying over 23% of the tweets that contain misinformation.
– We employ a number of supervised learning algorithms that very effectively
classify credible or misinformative tweets. Based on the features we propose,
classification techniques achieve weighted average accuracy of 77%.
– Our experiments suggest that without considering propagation of tweets,
our classification methodology achieves 77.8% weighted average accuracy,
offering the ability to timely limit the spread of false news before cascading.
Furthermore, the filtering process and the classification algorithms perform
in less than 2 seconds, making our methodology appropriate for real-time
applications.
2 Problem Description, Parameters and Methodology
Several studies reveal that news spread faster in the network of Twitter compared
to traditional news media [7,8]. Yet, a fundamental challenge is the quality of
content published in OSNs. Distinguishing between credible and inaccurate infor-
mation regarding emergency events is important, since misguidance or inability
to timely detect useful information may have critical effects.
A Model for Identifying Misinformation in Online Social Networks 475
Fig. 1. Example of misinformation for Hurricane Sandy.
Objective: The objective of our work is to detect misinformation related to
emergency events in the Twitter social network. We define as misinformation
any false or inaccurate information that is spread either intentionally or uninten-
tionally. An example of misinformation concerning the event of Sandy hurricane
is presented in Figure 1.
Our Approach: Our approach for solving the problem of misinformation iden-
tification follows 3 discrete steps: (i) Given a set of tweets Trelated to an event
and a set of users Uthat published at least one tweet t∈T, we conduct an
extensive analysis on characteristics of tweets and users who published them. Our
analysis focuses on a number of features and combinations of them and assists in
identifying abnormal behaviors of users and characteristics of tweets. (ii) Based
on the findings of the analysis, we extract extreme or suspicious behaviors and
exploit them to filter tweets that are more likely to constitute misinformation.
(iii) We then apply a series of learning algorithms implemented on Weka [9]to
identify misinformative tweets, using supervised learning techniques.
2.1 Parameters
Tweet Features: Each tweet t∈Tis represented as a feature vector Itthat
includes information about the tweet and the user who published it. Thus,
each tweet t∈Tis characterized by the following information: (i) Number
of characters - words: Short messages may not contain useful information, while
long messages may cover unrelated topics. (ii) Number of favorites - retweets
- replies: The popularity of a tweet may be an indication of its content. We
expect that tweets of interest will be cascaded in the network and thus be more
476 S. Antoniadis et al.
retweeted or favorited. (iv) Number of mentions - hashtags - URLs - media:
Features related to the structure of the tweet are considered to draw conclusions
about the quality of the tweet.
User Features: For each user u∈Uthat published at least one tweet t∈Twe
consider the following characteristics: (i) Number of followers - followees: Tru s t -
worthy users such as news agencies are expected to have many followers [7], while
spammers may have more followees. We define as followees the number of users an
account follows. (ii) Followers-Followees Ratio (FF-Ratio): We compute the FF-
Ratio of a user uas FF Ratio =followers(u)/(followers(u)+followees(u)).
(iii) Total tweets - Tweets during the event: We suspect illegitimate users may
be more active for a short time (e.g. the time of the event), thus we also consider
the number of tweets users publish. (iv) Days Registered: The days the user is
registered in the network before publishing a tweet. Recent users have greater
chances of being spammers in contrast to older ones.
Additional Features: Finally, for each tweet we extract the following set of fea-
tures: (i) URLs to Tweets (UtT) - Media to Tweets (MtT): For tweets published
by a user we compute the ratio of tweets containing URLs and Media separately.
We suspect that users frequently publishing URLs are candidate spammers, while
media may be irrelevant to the event. (ii) Followers to Replies (FtR) - Retweets
(FtRt) - Favorited (FtFav): Less popular tweets may indicate disapproval from
followers. Therefore we consider the ratio of followers to the features indicating
the popularity. (iii) Average Tweets per Day (ATpD): The average number
of tweets published by user uthat is registered da(u) days in the network is
computed as AT pD =t(u)/d a(u), where t(u) is the total number tweets u
published. (iv) Positive / Negative / Average Sentiment: We use SentiStrength
[10] to extract the positive and negative sentiment rate of the tweet text and
compute the average sentiment.
3 Data Analysis
In order to evaluate the performance of our approach for detecting misinforma-
tion during emergency events we used a dataset of tweets related to the Sandy
Hurricane, a major emergency event that unfolded in 2012, from October 22 to
November 2, and severely affected the area of New York City 3. Tweets related
to the event were collected based on the keywords “sandy” and “hurricane”, as
described in [3]. We then use the findings of the analysis to decide which values
constitute a possible indication of misinformation.
Analysis of User Characteristics: In Figures 2and 3we present the number
of tweets published by users, both total and during the event. We split the
number of tweets in buckets of 100, i.e., bucket 0 contains the number of users
that published 1 to 99 tweets in total. The Power Low distribution shown in the
Figures is in accordance to findings of Bagrow et al. in [7]. Most of the users
3http://en.wikipedia.org/wiki/Hurricane Sandy
A Model for Identifying Misinformation in Online Social Networks 477
published tweets related to the event with a frequency of 60 to 1000 seconds,
while there are users that published more than one tweet per minute. The number
of users’ followers and followees are presented in Figures 4and 5respectively.
The trend is similar for both connection types, with the majority of users having
few followers and followees. The peak in the number of users that have up to
2000 followees is due to Twitter policy that limits the users to follow up to 2000
users and is later differentiated based on the followers to followees fraction. In
Figure 6we also present the FF-Ratio. The FF-Ratio approaches a Gaussian
distribution with most users having an average ratio of around 0.5, meaning
that they have equal amount of followers and followees, although we can observe
another peak from 0.9 to 1.
Analysis of Tweet Characteristics: In Figure 7we present an analysis of the
number of words contained in a tweet. The majority of the tweets include 20 to 120
characters and 5 to 20 words. We further consider retweets, favorites and replies
of a tweet to determine its popularity. As observed in Figures 8and 9,thenum-
ber of retweets and favorites follows a power law distribution. Finally, most of the
tweets have fewer than 20 replies, but after a point the tweets containing more than
Fig. 2. User total tweets. Fig. 3. User Sandy Tweets. Fig. 4. User Followers.
Fig. 5. User Followees. Fig. 6. User FF-Ratio. Fig. 7. Words in tweets.
Fig. 8. Retweets received by
tweets.
Fig. 9. Favorites received by
tweets.
Fig. 10. Replies received by
tweets.
478 S. Antoniadis et al.
Fig. 11. Tweets with
mentions.
Fig. 12. Tweets with
URLs.
Fig. 13. Tweets with
hashtags.
Fig. 14. Tweets with
media.
20 replies rises (Figure 10). Figures 11 through 14 depict the number of mentions,
URLs, hastags and media presented in a tweet. Most of the tweets contain at most
one mentions and the majority of the tweets related to the event contain no link,
while there are tweets with over two links. 70% of the tweets contain no hashtags,
while the number of tweets containing a media is restricted to less than 10%.
4 Experimental Evaluation
We evaluated the performance of our approach on 80294 tweets related to hur-
ricane Sandy from 59660 users. In the first set of experiments we focus on
estimating the performance of the filtering process. By applying the filters of
Table 1, 12955 tweets are returned. We manually annotate a sample of 4000
randomly selected tweets among them. Humorous, irrelevant and deleted tweets
and accounts are considered as misinformation (assuming they are reported or
deleted due to violations [11]). For 176 of the tweets we could not draw conclu-
sions. Out of the remaining 3824 tweets, 898 constitute misinformation. Since
tweets are randomly selected, we conclude that over 23% of the filtered tweets
are indeed misinformation. We also annotated 4000 random tweets among those
that did not meet the filtering criteria. For the 3559 tweets that could be clas-
sified, 212 are identified as misinformation, i.e., less than 6%. Overall, 1110 out
the total 7383 labelled tweets constitute misinformation. The filtering process
captures 898, yielding recall values of over 81%.
Supervised Learning: We exploited a set of different supervised learning
algorithms implemented on Weka [9] to evaluate the performance of information
identification on the set of features we considered. We use 10-fold cross validation
for evaluating the classification results. The labelled dataset of the 3824 filtered
tweets is used as input to Weka. In Table 2we present the classification results.
Weighted Average F1 measure indicates that Bootstrap Aggregating has the
best performance. Regarding average precision, Random Forest achieves better
results, with 0.792 average precision.
Table 1. Filters applied during the filtering process.
Wor d s ≥30 Characters == 140 Favorite s ( ≥2&& ≤10) (≥1100)
Hashtags ≥4Mentions ≥4Retweets ((≥2&& ≤10) (≥1000))
Media ≥2Followees (≤10 ≥100000) Followers (≤10 ≥200000)
Replies ≥11 URLs ≥3Followers/Followees≥30000
Event tweets ≥7Total Tweets ≥500000 Interval (≤300sec ≥70000sec)
A Model for Identifying Misinformation in Online Social Networks 479
Table 2. Summary of Classification using Supervised Learning Algorithms.
Precision Recall F-Measure Precision Recall F-Measure
Bayes Network J48
Credible 0,845 0,834 0,839 0,825 0,889 0,856
Misinformation 0,480 0,500 0,490 0,516 0,385 0,44
Weig h t ed Avg. 0,759 0,755 0,757 0,752 0,771 0,758
k-Nearest Neighbors Random Forest
Credible 0,800 0,951 0,869 0,821 0,958 0,884
Misinformation 0,586 0,225 0,325 0,699 0,318 0,438
Weig h t ed Avg. 0,750 0,781 0,741 0,792 0,808 0,779
Adaptive Boosting Bootstrap Aggregating
Credible 0,839 0,888 0,863 0,828 0,931 0,877
Misinformation 0,549 0,447 0,493 0,622 0,372 0,466
Weig h t ed Avg. 0,771 0,784 0,776 0,780 0,799 0,780
Real-Time Misinformation Identification: The values of retweets, favorites
and replies are unknown at the time the tweet is published. Thus, to evaluate the
performance of our approach in timely detecting misinformation we conducted
another set of experiments ignoring the above attributes and features related to
them. The results for Bootstrap Aggregating and Random Forest are presented in
Table 3. The table shows that precision and recall drop slightly. Still, weighted
average precision is over 0.77, indicating the approach is appropriate for real
time misinformation identification. The filtering process requires just 963ms and
adding the execution times of the algorithms, less than 2 seconds are needed to
efficiently extract tweets containing misinformation, proving that the method is
efficient under real-time constraints.
Table 3. Classification with features known at run time.
Precision Recall F-Measure Precision Recall F-Measure
Random Forest Bootstrap Aggregating
Credible 0,812 0,949 0,875 0,816 0,951 0,879
Misinformation 0,631 0,286 0,394 0,655 0,301 0,412
Weig h t ed Avg. 0,770 0,793 0,762 0,778 0,799 0,769
Execution time: 0.41 sec 0.89 sec
5 Related Work
Castillo et al. [5] aim at automatically detecting credibility of information in
the network of Twitter. They use a number of features related to tweets and
supervised learning to distinguish between newsworthy or false news and later
perform credibility analysis. Gupta et al. [12] present TweetCred, an extension
of the previous work that enables users feedback. Xia et al. [6] also study the
480 S. Antoniadis et al.
problem of information credibility on Twitter after the event is detected and
relevant tweets are retrieved. Bosma et el. [13] suggest a framework for spam
detection using unsupervised learning. Anagnostopoulos et al. [14] study the
role of homophily in misinformation spread. McCord and Chuah [15] are using
traditional classifiers to detect spams on Twitter. Stringhini et al. in [11] aim at
identifying spammers in social networks. Identifying spammers on the network
of Twitter is also the objective of Benevenuto et al. in [16]. They extract features
of the account that may be indication of spamming behavior and use SVM learn-
ing model to verify their approach. Zubiaga and Ji in [3] and Gupta et al. [4]
focus on the credibility of images propagated in the network during emergency
events. They consider a number of features related to the image and the tweet.
Budak et al. [17] address the problem of misinformation spread limitation by
performing an extensive study on influence limitation. Faloutsos [18] developed
a botnet-detection method and a Facebook application, called MyPageKeeper,
that quantifies the presence of malware on Facebook and protects end-users.
Ghosh et al. [19] examine suspended accounts on Twitter and investigate link
farming and finally discourage spammers to acquire large number of following
links. Thomas et al. [1] identify the behaviors of spammers by analyzing tweets
of suspended users in retrospect. Mendoza et al. [20] focus on cascades of tweets
during emergency events and study the propagation of rumours. They conclude
that this defers from the propagation of news tweets and it is possible to detect
rumors by aggregating analysis on the tweets. Liu et al. [21]proposeahybrid
model that utilizes user behavior information, network attributes and text con-
tent to identify spams.
6 Conclusions
In this work we presented a methodology for identifying misinformation on social
networks during emergency events. As we illustrate in our experiments our app-
roach manages to correctly identify misinformation achieving accuracy of up to
77%. The filtering process suggested in this work identifies over 81% of misin-
formative tweets. Our approach is fast and effective and timely identifies misin-
formation, offering the ability to limit the spread in the network.
Acknowledgment. This research has been co-financed by the European Union (Euro-
pean Social Fund ESF) and Greek national funds through the Operational Program
“Education and Lifelong Learning” of the National Strategic Reference Framework
(NSRF) - Research Funding Program:Thalis-DISFER, Aristeia-MMD, Investing in
knowledge society through the European Social Fund, the FP7 INSIGHT project and
the ERC IDEAS NGHCS project.
A Model for Identifying Misinformation in Online Social Networks 481
References
1. Thomas, K., Grier, C., Song, D., Paxson, V.: Suspended accounts in retrospect: an
analysis of twitter spam. In: Internet Measurement Conference, pp. 243–258 (2011)
2. Gao, H., Hu, J., Wilson, C., Li, Z., Chen, Y., Zhao, B.Y.: Detecting and charac-
terizing social spam campaigns. In: ACM Conference on Computer and Commu-
nications Security, pp. 681–683 (2010)
3. Zubiaga, A., Ji, H.: Tweet, but verify: Epistemic study of information verification
on twitter (2013). CoRR, vol. abs/1312.5297
4. Gupta, A., Lamba, H., Kumaraguru, P., Joshi, A.: Faking sandy: characterizing
and identifying fake images on twitter during hurricane sandy. In: Ser. WWW 2013
Companion (2013)
5. Castillo, C., Mendoza, M., Poblete, B.: Predicting information credibility in time-
sensitive social media. Internet Research 23(5), 560–588 (2013)
6. Xia, X., Yang, X., Wu, C., Li, S., Bao, L.: Information credibility on twitter in
emergency situation. In: Chau, M., Wang, G.A., Yue, W.T., Chen, H. (eds.) PAISI
2012. LNCS, vol. 7299, pp. 45–59. Springer, Heidelberg (2012)
7. Bagrow, J.P., Wang, D., Barabasi, A.-L.: Collective response of human populations
to large-scale emergencies (2011). CoRR, vol. abs/1106.0560
8. Guy, M., Earle, P., Ostrum, C., Gruchalla, K., Horvath, S.: Integration and dis-
semination of citizen reported and seismically derived earthquake information via
social network technologies. In: Cohen, P.R., Adams, N.M., Berthold, M.R. (eds.)
IDA 2010. LNCS, vol. 6065, pp. 42–53. Springer, Heidelberg (2010)
9. Weka. http://www.cs.waikato.ac.nz/ml/weka/
10. Thelwall, M., Buckley, K., Paltoglou, G., Cai, D., Kappas, A.: Sentiment in short
strength detection informal text. J. Am. Soc. Inf. Sci. Technol. 61(12), 2544–2558
(2010)
11. Stringhini, G., Kruegel, C., Vigna, G.: Detecting spammers on social networks. In:
ACSAC, pp. 1–9 (2010)
12. Gupta, A., Kumaraguru, P., Castillo, C., Meier, P.: Tweetcred: real-time credibility
assessment of content on twitter. In: Aiello, L.M., McFarland, D. (eds.) SocInfo
2014. LNCS, vol. 8851, pp. 228–243. Springer, Heidelberg (2014)
13. Bosma, M., Meij, E., Weerkamp, W.: A framework for unsupervised spam detec-
tion in social networking sites. In: Baeza-Yates, R., de Vries, A.P., Zaragoza, H.,
Cambazoglu, B.B., Murdock, V., Lempel, R., Silvestri, F. (eds.) ECIR 2012. LNCS,
vol. 7224, pp. 364–375. Springer, Heidelberg (2012)
14. Anagnostopoulos, A., Bessi, A., Caldarelli, G., Vicario, M.D., Petroni, F.,
Scala, A., Zollo, F., Quattrociocchi, W.: Viral misinformation: The role of
homophily and polarization (2014). CoRR, vol. abs/1411.2893
15. McCord, M., Chuah, M.: Spam detection on twitter using traditional classifiers.
In: Calero, J.M.A., Yang, L.T., M´armol, F.G., Garc´ıa Villalba, L.J., Li, A.X.,
Wang, Y. (eds.) ATC 2011. LNCS, vol. 6906, pp. 175–186. Springer, Heidelberg
(2011)
16. Benevenuto, F., Magno, G., Rodrigues, T., Almeida, V.: Detecting spammers on
twitter. In: CEAS (2010)
17. Budak, C., Agrawal, D.: Abbadi, A.E.: Limiting the spread of misinformation in
social networks. In: WWW, pp. 665–674 (2011)
18. Faloutsos, M.: Detecting malware with graph-based methods: traffic classification,
botnets, and facebook scams. In: WWW (Companion Volume), pp. 495–496 (2013)
482 S. Antoniadis et al.
19. Ghosh, S., Viswanath, B., Kooti, F., Sharma, N.K., Korlam, G., Benevenuto, F.,
Ganguly, N., Gummadi, P.K.: Understanding and combating link farming in the
twitter social network. In: WWW, pp. 61–70 (2012)
20. Mendoza, M., Poblete, B., Castillo, C.: Twitter under crisis: can we trust what we
rt? In: Proceedings of the First Workshop on Social Media Analytics, ser. SOMA
2010, pp. 71–79. ACM, New York (2010)
21. Liu, Y., Wu, B., Wang, B., Li, G.: Sdhm: a hybrid model for spammer detection in
weibo. In: 2014 IEEE/ACM International Conference on ASONAM, pp. 942–947,
August 2014
