It's Always April Fools' Day! On the Difficulty of Social Network Misinformation Classification via Propagation Features

0.32 MB·

by  Mauro Conti

It’s Always April Fools’ Day! On the Difficulty of Social Network Misinformation
Classification via Propagation Features
MauroConti,DanieleLain, WalterQuattrociocchi
RiccardoLazzeretti,GiulioLovisotto IMTSchoolforAdvancedStudies,Lucca
UniversityofPadua
Abstract Nielsen,R.K.2015). ItisnotasurprisethenthattheOxford
7
Dictionary in 2016 elected Post-truth as word of the year.
1 Given the huge impact that Online Social Networks (OSN)
Thedefinitionreads:
0 hadinthewaypeoplegetinformedandformtheiropinion,
2 they became an attractive playground for malicious entities “Relatingtoordenotingcircumstancesinwhichobjec-
thatwanttospreadmisinformation,andleveragetheireffect. tivefactsarelessinfluentialinshapingpublicopinion
n
Infact,misinformationeasilyspreadsonOSNandisahuge thanappealstoemotionandpersonalbelief”.
a
threatformodernsociety, possiblyinfluencingalsotheout-
J Severalstudiespointedouttheeffectsofsocialinfluence
comeofelections,orevenputtingpeople’slifeatrisk(e.g.,
6 online(Centola2010;FowlerandChristakis2010;Quattro-
spreading “anti-vaccines” misinformation). Therefore, it is
1 ciocchi, Caldarelli, and Scala 2014; Salganik, Dodds, and
of paramount importance for our society to have some sort
Watts 2006). Results reported in (Kramer, Guillory, and
of“validation”oninformationspreadingthroughOSN.The
I] need for a wide-scale validation would greatly benefit from Hancock 2014) indicate that emotions expressed by others
S automatictools. onFacebookinfluenceourownemotions, providingexper-
. Inthispaper,weshowthatitisdifficulttocarryoutanauto- imentalevidenceofmassive-scalecontagionviasocialnet-
s
c maticclassificationofmisinformationconsideringonlystruc- works. As a result of disintermediated access to informa-
[ turalpropertiesofcontentpropagationcascades.Wefocuson tion and of algorithms used in content promotion, commu-
structural properties, because they would be inherently dif- nication has become increasingly personalized, both in the
1 ficult to be manipulated, with the the aim of circumventing way messages are framed and how they are shared across
v classificationsystems. Tosupportourclaim,wecarryoutan social networks. Selective exposure has been shown to fa-
1 extensiveevaluationonFacebookpostsbelongingtoconspir-
vortheemergenceofecho-chambers—polarizedgroupsof
2 acy theories (as representative of misinformation), and sci-
2 entific news (representative of fact-checked content). Our like-minded people where users reinforce their world view
4 findings show that conspiracy content actually reverberates withinformationadheringtotheirsystemofbeliefs(DelVi-
0 inawaywhichishardtodistinguishfromtheonescientific cario et al. 2016). Confirmation bias, indeed, plays a piv-
1. content does: for the classification mechanisms we investi- otal role in informational cascades (Quattrociocchi, Scala,
0 gated,classificationF1-scoreneverexceeds0.65duringcon- and Sunstein 2016; Bessi et al. 2015b; Zollo et al. 2015a;
7 tent propagation stages, and is still less than 0.7 even after Sunstein 2002). Recent works (Bessi et al. 2015a; Zollo et
1 propagationiscomplete. al.2015a)showedthatattemptstodebunkfalseinformation
: arelargelyineffective. Inparticular,discussiondegenerates
v
1 Introduction when the two polarized communities interact with one an-
i
X other (Zollo et al. 2015b). OSN users therefore tend to se-
AnincreasingnumberofpeoplegetinformedonOnlineSo-
lect information that are consistent with their beliefs (even
r cial Networks (OSN) (Newman, N. and Levy, D.A.L. and
a if containing false claims) and propagate it to like-minded
Nielsen,R.K.2015). However,asOSNalloweveryuserto
friends, and to ignore information dissenting with their be-
post content, which propagates among users through viral
liefs. Thisconfirmsthatmisinformationisahugethreatfor
processes,theseplatformsbecameattractivetargetsformis-
modernsociety. Notonlyitcanputpeople’slifeatrisk, as
information creators. Moreover, the hyperconnected world
inthecaseof“anti-vaccines”misinformation,itisalsostart-
and increasing complexity of reality create a scenario in
ingtobeusedagainstpoliticalopponents,suchasintheUS,
which viral processes on OSN are driven by confirmation
duringthe2016electoralcampaign(Silvermanetal.2016)
bias, eliciting the proliferation of unsubstantiated rumors
and during Election Day (Rogers, K. and Bromwich, J.E.
and hoaxes all the way up to conspiracy theories (Bessi et
2016).
al. 2015b; 2015a). News stories undergo the same popu-
Rising attention to the spreading of fake news and un-
larity dynamics as other forms of online contents (such as
substantiated rumors led researchers to investigate many
selfiesandcatpictures)(Newman,N.andLevy,D.A.L.and
of their aspects, from the characterization of conversation
Copyright(cid:13)c 2017,AssociationfortheAdvancementofArtificial threads (Backstrom et al. 2013), to the detection of bursty
Intelligence(www.aaai.org).Allrightsreserved. topics on microblogging platforms (Diao et al. 2012), to
the disclosure of the mechanisms behind information dif- mation, and posts belonging to scientific news as repre-
fusionfordifferentkindsofcontents(Romero,Meeder,and sentativeoffact-checkedcontent. Ourfindingshighlight
Kleinberg 2011). Spreading of misinformation also moti- the complexity of creating automatic solutions to mis-
vatedmajorcorporationslikeGoogleandFacebooktopro- information classification. Indeed, structural features of
vide solutions to the problem (First Draft News Coalition contentpropagationdonotallowustoobtainnotableim-
2016). However,giventheamountofcontentpostedevery- provements from a random guess baseline: F -score for
1
dayonOSN(e.g.,Facebookreports1.18billiondailyactive the classification mechanisms we investigated is always
users on September 2016 (Facebook 2016)), effective fact- lowerthan0.65duringcontentpropagationstages,andis
checking would greatly benefit of automatic classification stilllessthan0.7evenafterpropagationiscomplete.
tools, that would possibly not require human intervention.
Moreover, classification of fake news and misinformation
Outline. Section 2 overviews related work in news and
should ideally use properties that misinformation creators
misinformation propagation in OSN. Section 3 formally
can not manipulate. Considering in the classification, for
models and defines content propagation in Facebook. Sec-
example, the trustworthinessofnews domains, orthe topic
tion 4 presents our experiment design and the features we
ofcontent,couldleadtoanarmsracewithcreatorsoffalse
extractfromcontentpropagationcascades. Then,Section5
news.Instead,thetopologyofpropagationcascades,andthe
evaluates our classification and, finally, Section 6 critically
patternsofusers’interactionwithcontent,areoutsideofthe
discusses our results, summarizes the paper and delineates
domainofour“adversaries”,andaremuchmoredifficultto
futurework.
bemanipulated.
Inthiswork,weinvestigatedetectionofviralprocessesby
2 RelatedWork
comparingdiffusionofpostsfromscientificandconspiracy
pagesontheItalianFacebooknetwork. Theformerdiffuse Several studies moved towards the spreading of rumors
scientificknowledge, wheredetailsaboutthesources(such and behaviors on online social networks, challenging both
as authors and funding programs) are easy to access. We their structural properties and their effects on social dy-
thereforeselecttheirpostsasrepresentativeoffact-checked namics(Moreno,Nekovee,andPacheco2004;Doerr,Fouz,
content. The latter aim at diffusing what is neglected by andFriedrich2012;Seo,Mohapatra,andAbdelzaher2012;
“manipulated” mainstream media. Specifically, conspiracy Borge-Holthoefer et al. 2013; Cozzo et al. 2013; Borge-
theoriestendtoreducethecomplexityofrealitybyexplain- Holthoefer, Rivero, and Moreno 2012). In (Ugander et
ingsignificantsocialorpoliticalaspectsasplotsconceived al. 2012), authors find that the probability of contagion is
bypowerfulindividualsororganizations. Sincethesekinds tightly controlled by the number of connected components
ofargumentscansometimesinvolvetherejectionofscience, inanindividual’scontactsneighborhood,ratherthanbythe
alternativeexplanationsareinvokedtoreplacethescientific actualsizeoftheneighborhood.In(CentolaandMacy2007)
evidence. Forinstance, peoplewhorejectthelinkbetween researchers show that, although long ties are relevant for
HIVandAIDSgenerallybelievethatAIDSwascreatedby spreading information about an innovation or social move-
the US Government to control the African American pop- ment, they are not sufficient with respect to the social re-
ulation. We therefore select their posts as representative of inforcement necessary to act on that information. A key
misinformation. factor in identifying true contagion in social network is to
distinguish between peer-to-peer influence and homophily:
Contribution. Thecontributionsofthispaperarethefol- in the first case, a node influences or causes outcomes to
lowing: itsneighbors,whereasinthesecondone,dyadicsimilarities
between nodes create correlated outcome patterns among
• Weshowthatautomaticfact-checkingwithclassification
neighbors that could mimic viral contagions even without
techniques employing only structural features of content
directcausalinfluence(McPherson,Smith-Lovin,andCook
propagationcascades(featuresthatarerobusttoattacker’s
2001). The study presented in (Aiello et al. 2012) reveals
manipulation) does not suggest to bring usable results.
that there is a substantial level of topical similarity among
Given the grain of our data, we design classifiers that
users which are close to each other in the social network,
leverages topological properties of content propagation
suggesting that users with similar interests are more likely
cascades, and properties of the evolution over time of
to be friends. In (Aral, Muchnik, and Sundararajan 2009)
users’ interactions with content. A set of classifiers op-
authors develop an estimation framework to distinguish in-
erateswithevolutionpropertiestoclassifycontentduring
fluenceandhomophilyeffectsindynamicnetworksandfind
its early propagation stage, another set of classifiers op-
thathomophilyexplainsmorethan50%oftheperceivedbe-
erateswithmorefeaturestoclassifycontentthatalready
havioralcontagion.In(Bakshyetal.2012)theanalysisfaces
propagated. Indeed, being able to issue warnings about
the role of social networks and exposure to friends’ activ-
possiblefakenewsasearlyaspossible,andretroactively
ities in information resharing on Facebook. Once having
flag such news can be useful, in the fight against OSN
isolated contagion from other confounding effects such as
misinformation.
homophily,authorsclaimthatthereisaconsiderablyhigher
• We evaluate our classifiers on a well-known dataset of chancetosharecontentswhenusersareexposedtofriends’
FacebookpostsfromItalianpages. Weusepostsbelong- resharing.Allthesecontributionsstrivetounderstandthein-
ing to conspiracy theories as representative of misinfor- nermechanismofrumorspreadingandtoeventuallypredict
massivediffusionprocesses,i.e. cascades. Cascadesrecur- weuselacksinformationaboutthelikeandfollowrelation-
renceandpredictionhasbeenshapedin(Chengetal.2014) ships,thatwethereforecannotconsider.
and(Chengetal.2016).
PotentialPropagationGraph. Beforeformallymodeling
3 Methods
the spreading of content on Facebook, we describe some
In this section, we first report and describe the employed fundamental concepts of the OSN. We recall that, in our
dataset (Section 3.1). We then give some necessary back- model, only pages can post new content on the social net-
ground knowledge, and present our reference model and work.Henceforth,werefertothepagethatoriginallyposted
definition of content propagation mechanisms of Facebook somecontentastheseedpage. Instead,usersfindnewcon-
(Section3.2). tentbylookingattheirtimeline,wheretheyseerecentcon-
tent posted by the pages they like, and content that their
3.1 Dataset friendsrecentlyinteractedwith. Theycantheninteractwith
such content through the means of resharing, liking, and
We employ a well-known dataset of posts shared by Ital-
commenting it. However, all of these interactions can hap-
ian Facebook users (Bessi et al. 2015a). This dataset con-
pen directly on the original content, or on some types of
tainspostspublishedby73publicFacebookpages:34pages
interactions of users’ friends. For example, a user observ-
that publish scientific content (e.g., press releases of peer-
ingacommentorreshareofapostbyoneofhisfriendscan
reviewed articles), and 39 pages that publish conspiracy
decidetointeractdirectlywiththeoriginalpost,orwithhis
theories-relatedcontent(e.g. newworldorder, chemtrails).
friend’sinteractionitself. Inthefirstcase,user’sinteraction
Additionally, thedatasetcontainsinformationaboutthein-
looks exactly the same and it is impossible to understand
teractionofuserswiththeseposts,andusers’ego-networks
whether the content was found thanks to his friend’s inter-
(i.e., the list of users that are their friends, when such list
action,ordirectlyontheseedpage. Infact,differentlyfrom
is public). Additionally, for a set of posts, the dataset pro-
related work (Friggeri et al. 2014), the dataset we employ
vides information about the propagation cascade of such
doesnotprovidethistypeofinformation. Wethereforetake
content, generated by users’ reshares, and subsequent re-
aconservativeapproach,sayingthatthecontentpotentially
sharesfromtheirfriends. Thispropagationfromoneuserto
propagated to the user both from the seed page and from
otheruserscanhappenmultipletimes,formingacascadeof
any of his friends that interacted with the content, without
resharing. Usinginformationfromthedataset,weextracted
any distinction. On the other hand, in the second case, it
112141non-emptypropagationcascades,89491forconspir-
becomes clear that the user found the content thanks to his
acy and 22650 for science, respectively. We underline that
friend. Wethereforesaythatthecontentpropagated tothe
the dataset is obtained by using the Facebook Graph API,
userfromhisfriend.
andcontainsonlypublicinformation. Hence,timestampsof
We now formalize the above observations. Let P be the
resharesandcommentsareavailable,buttimestampsoflike
setofcontentspostedonFacebookbyseedpages. Foreach
interactionsarenot.
post p ∈ P, created at some time t by a seed page, at a
generic subsequent point in time t+δ we define a poten-
3.2 BackgroundandDefinitions
tial propagation graph Gp (cid:104)Vp ,Ep (cid:105), where Vp is
t+δ t+δ t+δ t+δ
We now present our reference formal representation of the set composed by the seed page, and by the users that
Facebook’sfriendshipgraph,andthepotentialpropagation interacted with p during the time interval [t,t+δ]. Fi-
graphgeneratedbycontentpostedonthesocialnetwork. nal potential propagation graph Gp(cid:104)Vp,Ep(cid:105) is the graph
formed considering all interactions with p on the timespan
of the analysis (as new interactions with old content are
Facebook Friendship Graph. We model Facebook rela- always possible on Facebook). Two nodes v ,v ∈ Vp
tionshipsasagraphG(cid:104)V,E(cid:105),thatwecallFacebookfriend- 1 2 t+δ
are connected by an undirected potential propagation edge
shipgraph, whereV isthesetofnodesthatrepresententi- ep(v ,v ) ∈ Ep if either (i) v or v already interacted
ties, namely user accounts and page accounts. We assume 1 2 t+δ 1 2
withpand∃e|e(v ,v ) ∈ E (thatis,v andv arefriends
two main differences among these two types of entity: (i) 1 2 1 2
on Facebook), or (ii) v or v is the seed page. Therefore,
pages can post new content on the OSN, while users can 1 2
anedgee(v ,v )∈Ep indicatesthatthecontentppoten-
onlyinteractwithsuchcontentbyliking,commenting,and 1 2 t+δ
tiallypropagatedeitherfromv tov ,orfromv tov . We
resharingit; and(ii)userscanestablishfriendshiprelation- 1 2 2 1
associatetoeachedgee∈Ep twodifferentproperties:
ships with other users, while pages cannot. Indeed, two t+δ
users v ,v ∈ V are connected by an edge e(v ,v ) ∈ E 1. Time: the time when the interaction between v
1 2 1 2 1
if they are friends on Facebook. Pages are not connected andv tookplace;
2
byanyedge,astheydonothaveproperfriendshiprelation-
2. Type: either “like”, “comment”, “reshare”, or
ships. ThismodelisasimplificationofhowFacebookactu-
”friendship”, depending on the type of interaction
ally works, because users can post new content, and pages
betweenv andv .
and users are linked by like relationships. Similarly, users 1 2
can follow other users, without having any friend relation- Hereafter in the paper, we will refer to these properties for
ship with them. However, for this work, we do not focus an edge e ∈ Ep with e.property name (for example,
t+δ
on new content generated by users. Moreover, the dataset e.time).
Figure 1 depicts an example of this propagation model. 1. EarlyStage,meaningthatclassificationofthetype
We represent a simple Facebook friendship graph in Fig- ofposthappensasearlyaspossibleduringitsprop-
ure 1a, where edges represent friendship relations. Nodes agationphase;
v ,...,v are OSN users, while node s represents a public
1 4 2. Final Stage, meaning that classification happens
page posting content on the platform. We suppose v and
2 whenthepostalreadystoppedpropagating(within
v interacted with a given content posted by s, and v re-
4 3 the considered timespan of the employed dataset),
shared the post from v . Node v did not interact with the
2 1 anditscascadeiscomplete.
content, hence is not present on the potential propagation
These two scenarios are of particular interest, both from
graphofthepost,thatwerepresentinFigure1b.There,with
a research and a practical points of view. Indeed, be-
edges (s,v ),(s,v ),(v ,v ) we represent possible propa-
2 4 2 4 ing able to issue automated warnings as early as possible
gationpathsofthecontent: eitherv orv couldhaveseen
2 4 aboutpotentialmisinformationcouldhelpinreducingtheir
thecontentfromtheseedpage,orfrompreviousinteraction
spread(Friggerietal.2014). Ontheotherhand, stillbeing
of their friend. Node v only has an edge (v ,v ), as we
3 2 3 able to perform such a classification when the post already
knowthattheinteractionhappenedthankstov ,andthere-
2 propagatedmightserveasawarninginordertopreventits
fore the content propagated from this node. Additionally,
furtherdiffusion. Moreover,retroactivelyflaggingoldposts
we highlight that some possible propagation edges, repre-
aspotentiallyfakecouldhelptraininguserstodiscriminate
sentedwithadashedline,correspondtofriendshipedgesof
betweenfact-checkedanddubiousinformation,amajordi-
thefriendshipgraph.
rectioninthefightagainstmisinformation(FirstDraftNews
Coalition2016).
Toinvestigatethesetwoscenarios,wesetuptwodifferent
experiments, modeled as binary classification tasks, where
the two classes are conspiracy and science. We describe
s s
theminthefollowing.
v v
4 4 Early Stage Classification. In this scenario, we have ac-
v cess only to users’ interactions with a post up to a certain
1 v2 v2 time t + δ, after it has been created at time t, (e.g., up
to after two hours after creation). With this scenario, we
simulatehowOSNssuchasFacebookcouldtrytocontinu-
ouslydetectmisinformationcontentduringitspropagation.
v v
3 3 Therefore,foreachpostp,givenitscreationtimet,weex-
tract features from the partial propagation graph Gp , and
t+δ
fromtheevolutionofpropertiesofthepropagationgraphbe-
(a)FacebookFriendshipGraph (b)PotentialPropagationGraph
forethecurrenttimet+δ(discussedfurtherinSection4.2).
Weuse30minutessteps,asthisprovedtobeagoodtrade-
Figure1: SampleFacebookfriendshipgraph, andpotential
off between the granularity of the analysis and the number
propagation graph of some content. A dashed line repre-
of intervals to consider. We stop at two days (2880 min-
sentspotentialpropagationedgesthatalsocorrespondstoa
utes, or 96 time steps) after the publication time of a post,
friendship edge of the friendship graph. Node v1 does not
because we observed that most of the interactions happen
interactwiththecontent,andisthereforenotpresentinthe
in this period (> 95%). Finally, we compare the perfor-
potentialpropagationgraph.
manceofdifferentwell-knownclassifiers, namelyRandom
Forests (Ho1995),LinearDiscriminantAnalysis(Izenman
2013),andMulti-LayerPerceptron(Haykin1998),foreach
4 Experiments δ ∈[30,60,...,2880],inacross-validationscheme.
Inthissection,wedescribethedetailsoftheexperimentand
Final Stage Classification. In this scenario, we suppose
of the analysis performed on the dataset. We first describe
that a post already stopped its propagation, and no new in-
the experimental design (Section 4.1). We then thoroughly
teractionswithithavebeenobservedforalongtime.Wede-
report and motivate the different features that we extracted
scribethefinalpropagationgraphGp ofpostpwithasetof
frompropagationcascades(Section4.2).
high-levelandtopologicalfeatures,thatwedescribeinmore
detailinSection4.2. Wethencomparetheperformanceof
4.1 ExperimentalDesign
differentclassifiers,inacross-validationscheme.
The aim of our experiments is to show that it is difficult to
4.2 FeatureExtraction
discriminatebetweenconspiracytheories,andfact-checked
scientific news, by using only the propagation graph of the Toextractinformationfromthedifferentpropagationgraphs
post, as it would be difficult to manipulate by misinforma- GpandGp,weidentifythreepossiblecategoriesoffeatures:
t
tioncreators. Inparticular, weevaluatethisintwospecific (i)high-levelpropertiesofthecontentpropagation,(ii)topo-
momentsofcontentpropagation: logical properties of the propagation graphs, (iii) evolution
properties of the content propagation. We use feature sets Fronczak, and Hołyst 2004), also known as Wiener index,
(i),(ii),and(iii)consideringtheevolutionaftertwodays,in givesusindicationsoftheviralityofthecontent,intermsof
theFinalStageClassificationscenario. Weonlyusefeature distance of propagation from the seed page. Long cascad-
set (iii), in the Early Stage Classification scenario, because ingnews,resharedmanytimesfrominteractingfriends,will
theotherfeaturescharacterizeonlyGp,butnotGp.Wesum- exhibitalongeraveragepathlengththannewswhoseinter-
t
marize these different features in Table 1, along with their actions happened mostly from the seed node. Finally, the
formaldefinitions,anddescribetheminthefollowing. diameter ofagraph,definedasthelongershortestpathbe-
tweenanypairofnodesofthegraph,indicatesthespreading
High Level Properties. These features represent high-
distanceofposts.
levelpropertiesofthecompletepropagationcascaderepre-
sentedbyGp(cid:104)Vp,Ep(cid:105). EvolutionProperties. Thesefeaturesrepresentevolution
Somefeaturesrepresentverygeneralpropertiesrelatedto propertiesovertimeofthepostppropagation,fromitscre-
theviralityofthecontent. Thesegeneralpropertiesarethe ationtimettoasubsequentpointt+δ. Tocomputethese
lifetimeofthecascade,measuredasthedistanceinminutes features, we construct the propagation graphs at different
from the first to the last captured interaction with the con- time steps G ,...,G . We then calculate the value of
t+30 t+δ
tent; the size of the cascade in terms of number of nodes threeofourhigh-levelfeaturesforeachgraphateachtime
(userswhointeractedwiththecontent);thenumberoftotal step: (i) Friendships Ratio, (ii) Size, and (iii) Interactions
interactions; andthetimerequiredforthecascadetoreach Ratio. Foreachofthesehigh-levelfeature,weobtainatime
its 90% total interactions (referred to as 90% interactions series
time). v ,v ,...,v ,
1 2 δ/30
Otherfeaturesderivedfromhighlevelpropertiesattempt
on which we compute a set of well-known statistical mea-
to capture different possible types of interaction with the
sures that represent the evolution over time of the se-
content. Friendships ratio is defined as the proportion of
ries(Wiens,Horvitz,andGuttag2012). Namely,thesesta-
edges whose type is “friendship” over the total number of
tistical measures are the mean, standard deviation, linear
edgesandrepresentsthenumberoftimes,inproportion,that
weighted mean and quadratic weighted mean, average ab-
the post potentially propagated among friends, rather than
solutechange,andmaximumoftheseries. Thesemeasures
directly from the seed node. Indeed, if no friends of users
capturetheevolutionofthetimeseriesuptothecurrenttime
interact with some content, its potential propagation graph
and,especially,donotrequireustoknowthewholetimese-
only contains edges with the seed page. Interactions ratio,
ries,animportantpropertyforourEarlyStageclassification
instead,representstheaverageexposuretointeractionsfrom
experiment. We decided to derive time-series only for the
friendsofuserswiththecontent. Sinceverticesareinteract-
threehigh-levelfeatureslistedabove. Indeed,wearguethat
ingusers,andedgesarepotentialinteractions,highervalues
the other features either have no temporal properties (e.g.,
of this metric mean lower exposure (little interaction with
lifetime,timetoreach90%interactions),evolveinsimilaror
the content by one’sfriends). These features are motivated
predictableways(e.g.,diameter),ordescribebehaviorsthat
bytheobservationthatitispossiblethattheusers’fruition
arealreadycapturedbytheselectedtimeseries. Moreover,
ofdifferenttypesofcontentisdifferent,withsometypesof
evolutionpropertiesoftimeserieswouldbeespeciallyused
contentbeinginteracteddirectlyfromthesource,andother
in Early Stage classification. Social networks need to per-
types of content relying more on word-of-mouth propaga-
form feature extraction on this scenario at every time step:
tion(Romero,Meeder,andKleinberg2011).
featuresderivedfromtopologicalpropertiesaretoocompu-
Topological Properties. We also select as features some tationallyexpensive(e.g.,diametercalculationrunsinmore
well-knownpropertiesofthetopologicalstructureofgraphs. than quadratic complexity w.r.t. the number of vertices) to
These properties are commonly used to learn information becontinuouslycalculatedforeachnewpost.
about graph structures, and have been applied in solving
problems such as link analysis and prediction (Al Hasan 5 Results
et al. 2006), and especially in cascade and virality predic-
Inthissectionwepresenttheresultsofourexperiments. In
tion(Hong,Dan,andDavison2011;Chengetal.2014).The
particular, we first discuss evaluation metrics and baseline
averagevertexdegreefeaturerepresentstheaveragenumber
(Section5.1). WethenreporttheresultsontheEarlyStage
of possible propagation edges for the content at any given
Classification scenario (Section 5.2), and the results on the
hop. Higher values of this metric indicate the presence of
FinalStageClassificationscenario(Section5.3).
interacting users greatly exposed to the content, or able to
influencemanyoftheirsocialfriends. Theglobalclustering
5.1 EvaluationMetrics
coefficient (Holland and Leinhardt 1971), a measure of the
density of connections of graphs, is another indication of As usual in binary classification, the classification base-
whether the possible propagation paths are generated from line is the performance of a random classifier on the data:
interactionsbetweenfriends,ordirectlywiththeseedpage. withoutanyinformationregardingthepropagationgraph,it
Assortativity coefficient, defined as the degree correlation guesseseitherscienceorconspiracywithequalprobability,
betweenpairsoflinkednodes(Newman2002),canmeasure as a coin flip. The goal of our experiments is to show that
how friends influence each others in interacting with con- structuralfeaturesdonothelpmoresophisticatedmodelsin
tentonthesocialnetwork. Averagepathlength(Fronczak, improvingthebaselineperformance.
FeatureName Description
HighLevelProperties
Size Numberofedgesofthegraph|Ep|
FriendshipsRatio Proportionofedgeswhosetypeis“friendship”: |{e∈Ep|e.type=“friendship”}|/|Ep|
InteractionsRatio Numberofverticesoverthenumberofedges: |Vp|/|Ep|
Lifetime Timepassedbetweenpostcreationtimeandthelastinteractiontime: max e.time−t
e∈Ep
90%InteractionsTime Timerequiredforthecontenttoreachits90%numberofinteractions
TopologicalProperties
AverageVertexDegree Averagepossiblepropagationpathsofthepost: 2·|Ep|/|Vp|
ClusteringCoefficient Ratioofconnectedtripletsofnodes
AssortativityCoefficient Degreecorrelationbetweenpairsoflinkednodes
AveragePathLength Averagelengthoftheshortestpaths
Diameter Lengthofthelongestshortestpathbetweenanypairofvertices
EvolutionProperties
Mean 1 (cid:80)nv
n i i
LinearWeightedMean 2 (cid:80)ni·v
n(n+1) i i
QuadraticWeightedMean 6 (cid:80)ni2·v
n(n+1)(2n+1) i i
StandardDeviation (cid:113)(cid:80)ni(vi−v¯)2,wherev¯isthemeanofv
n
AverageAbsoluteChange 1 (cid:80)n−1|v −v |
n i i i+1
Maximum max v
i
i
Table1: DescriptionandformaldefinitionoffeaturesweuseonourFinalStageandEarlyStageclassificationexperiments.
Unfortunately, our dataset is highly imbalanced (com- withthebaseline.
posedby89491newsforconspiracy,andonly22650news Toperform(ii),weundersampledthemostfrequentclass
forscience). Withsuchimbalance,standardevaluationmet- (i.e.,conspiracy).Wethereforeextractedexactly22650con-
rics(suchasprecision,accuracy,andrecall)canbemislead- spiracysamplesfromthefulldataset,andcreatedasubsam-
ing, because they do not account for the uneven class fre- plewithperfectlybalancedclasses. Toaccountforpossible
quencies. Even computing averages of these metrics using biases caused by the undersampling, we repeated the pro-
weights based on the class frequencies does not fit our in- cess several times, and averaged the outcomes. Using per-
tentionsofconsistentlycomparingourresultswiththefixed fectly balanced datasets, we can evaluate precision, recall,
baseline. accuracy,andF -scorevaluesofourclassifiers. Indeed,the
1
To deal with this imbalance, we performed two distinct value of these metrics for a random classifier on balanced
experiments: (i) we consider only metrics that take imbal- dataisexactly0.5,whichweuseasbaseline.
ance into account and use the full dataset, and (ii) we con- Hereafter,ifnotdifferentlyspecified,themetricsthatre-
sidermeaningfulmetricswithbalanceddataset,obtainedun- quire a positive class (such as precision, recall, and F -
1
dersamplingtheoriginaldataset. score) use conspiracy as the positive class, and science as
To perform (i), as metrics we use Area Under Receiver thenegativeone.
OperatingCurve(AUC)(HanleyandMcNeil1982),andCo-
5.2 EarlyStageClassification
hen’s Kappa (Cohen 1968) (scaled into the interval [0, 1]).
Indeed,thevalueofthesemetricsforarandomclassifieris To evaluate this scenario, we recall that we used a total
exactly 0.5, which we use as baseline. In this way, we can of 18 features. We then followed the methodology ex-
use the full dataset and still be able to compare our results plained in Section 4.1, and evaluated three separate clas-
sifiers (namely, Linear Discriminant (LD), Random Forest 0.70
(RF), Multi-Layer Perceptron (MLP)) at every 30 minutes
timestep.Asdiscussed,wefirstevaluatedAUCandCohen’s 0.65
Kappa of these classifiers on the full, unbalanced dataset.
0.60
We then evaluated precision, recall, and accuracy of these
classifiersonundersampledbalanceddatasets.
e0.55
Figure2reportstheAUCandtheCohen’sKappametrics, ur
s
on a 5-fold cross-validation scheme, at different time steps ea0.50
m
after the original post is published. We can see that evolu- 1-
tionpropertiesofthepropagationgraphdonothelpsignif- F0.45
icantly our classifiers. Indeed, the curves are close to the
0.40
random classification baseline (dotted horizontal line), and
LD
farfromperfectclassification(i.e.,1.0). Cohen’sKappalies 0.35 RF
almost exactly on the random classification baseline, sug- MLP
gestingthattheclassificationperformanceisalmostrandom. 0.30
10 20 30 40
Using AUC, it is possible to set good classification thresh- Time (Hours)
olds, slightly improving the performance. MLP performs
slightlybetterthanRFandLD,andwecanobservethatre- Figure3: EarlyStageClassificationscenario, classification
sultsdonotchangesignificantlyafterafewhoursofanaly- F1-score,asafunctionofelapsedtime.
sis. However the outcome remains unsatisfying (i.e., AUC
under0.6foreachclassifierateachtimestep).
Figure 4 reports average AUC and Cohen’s Kappa on a
0.70 5-foldcross-validationscheme,onthefulldataset. Thedot-
ted horizontal line shows the baseline performance of ran-
0.65 dom classification. Indeed, we observe that our classifiers
donotsignificativelyimprovethebaseline,withthemetrics
0.60
remaining below 0.75. LD performance is poorer than RF
andMLP,probablyduetothesimplicityoftheclassification
0.55
tool.
e
alu0.50
V
1.0
Cohen's Kappa-LD
0.45 Area Under Curve
AUC-LD
Cohen's Kappa
Cohen's Kappa-RF
0.40
0.8
AUC-RF
0.35 Cohen's Kappa-MLP
AUC-MLP
0.6
0.30
0 10 20 30 40 50 e
Time (Hours) alu
V
Figure2: EarlyStageClassificationscenario,AUCandCo- 0.4
hen’sKappa,asafunctionofelapsedtime.
0.2
Figure3reportstheF -scoreobtainedontheundersam-
1
pled balanced dataset, on a 5-fold cross-validation scheme,
averaged over ten repetitions of the undersampling proce-
0.0
dureafterwards.Figure3furtherconfirmsthefindingshigh- LD RF MLP
lighted in Figure 2: classifiers are not able to leverage the
Figure4: FinalStageClassificationscenario,AUCandCo-
evolution properties to discriminate between science and
hen’sKappa.
conspiracy. In fact the F -score curve, after a slight incre-
1
ment during the first 10 hours, stabilizes relatively close to
thebaseline, andremainsbelow0.65throughoutthewhole In Figure 5 we report the ROC curve, on a 5-fold cross-
48hourstimespan. validation scheme, on the full dataset. From Figure 5 we
canobservethatnoclassifiercanreachagoodtradeoffbe-
5.3 FinalStageClassification
tween true positive rate and false positive rate. Indeed, the
Toevaluatethisscenario,werecallthatweusedatotalof28 curves are relatively close to the baseline (diagonal dotted
features with three distinct classifiers: Linear Discriminant line), meaning that, as the decision threshold changes, lots
(LD),RandomForest(RF),Multi-LayerPerceptron(MLP). ofsamplesaremisclassified.
Again, we first evaluated the unbalanced dataset, and then Table 2 reports several metrics computed on the under-
evaluatedtheundersampledbalanceddatasets. sampledbalanceddataset. Resultsareaveragedona5-fold
1.0 sification. These results highlight the necessity of includ-
ing content-related features, or polarization metrics, in fu-
ture analysis (i.e., whether particular users and their echo
0.8 chambers are more polarized towards one type of content).
Unfortunately, misinformation creators can easily control
e
at content-relatedfeatures,inordertoavoidalgorithmicdetec-
e R0.6 tion. Moreover,userpolarizationcanbeclearlyunderstood
v
siti from past users’ behaviors, but it takes time to understand
o
P polarization of new users. Hence, automatic detection of
e 0.4
u fakenewsremainsanopenchallenge.
Tr
0.2 LD
Future Work. The employed dataset has some limita-
RF
tions,boundtotheFacebookAPI:(i)itonlycontainsFace-
MLP
book public information; (ii) it does not contain the times-
0.0
0.0 0.2 0.4 0.6 0.8 1.0 tamp of likes, one of the most common interactions; and
False Positive Rate
(iii) it does not always provide information about whether
Figure 5: Final Stage Classification scenario, ROC curves. interactionwithcontenthappenedbecauseofinteractionsof
For visual clarity, we linearly interpolated the points in the user’sfriends. Wewouldliketoanalyzefiner-graineddata,
curves, and plotted markers at each 0.02 step in the false thattakesthesefactorsintoaccount,becauseitcouldleadto
positiverate. improvedresults.
In our analysis, we identified a set of features and used
them in well-known classifiers. Our experiments were ex-
cross-validationscheme,andthenovertenrepetitionsofthe tensive, butnotcomplete. However, weexpectthattheuse
undersampling procedure. These metrics show that even if of different models would not provide significant improve-
theclassifiersareabletoimprovethebaselineslightly,they ment. Thisclaimneedstobefurthervalidated,possiblyalso
cannotreachgoodperformance. usingmorerecentdatasets.
In the future, we also plan to test different methods for
Classifier Precision Recall Accuracy F score misinformation classification, based on user polarization
1
andcontent-relatedfeatures,toinvestigatewhetherthesein-
LD 0.578 0.523 0.570 0.549 formationcouldhelppropagationproperties,andovercome
thedifficultyofthisproblem.
RF 0.654 0.742 0.675 0.695
MLP 0.659 0.688 0.665 0.672
References
Table2: FinalStageClassificationscenario,performanceof Aiello,L.;Barrat,A.;Schifanella,R.;Cattuto,C.;Markines,
theclassifiers. B.; and Menczer, F. 2012. Friendship prediction and ho-
mophily in social media. ACM Transactions on the Web
(TWEB)6(2):1–33.
6 Conclusions Al Hasan, M.; Chaoji, V.; Salem, S.; and Zaki, M. 2006.
Earlydetectionofmisinformationplaysacrucialroleinso- Linkpredictionusingsupervisedlearning. InInternational
Conference on Data Mining (SDM): Workshop on Link
cial networks. In this paper, we analyzed the difficulty of
discerning conspiracy posts from scientific posts on Face- Analysis,Counter-TerrorismandSecurity.
book. Wefocusedonusingonlystructuralfeaturesofcon- Aral, S.; Muchnik, L.; andSundararajan, A. 2009. Distin-
tentpropagation,becausetheycannotbeeasilymanipulated guishinginfluence-basedcontagionfromhomophily-driven
bymisinformationcreators. Ourresultsshowthatmisinfor- diffusionindynamicnetworks. ProceedingsoftheNational
mationclassificationduringitsearlypropagationstagewith AcademyofSciences(PNAS)106(51):21544–21549.
thesefeaturesisunsuccessful,suggestingthatthespreading
Backstrom, L.; Kleinberg, J.; Lee, L.; and Danescu-
dynamics captured by our features are independent on the
Niculescu-Mizil,C. 2013. Characterizingandcuratingcon-
typeofcontent. Furthermore,evenconsideringthecascade
versation threads: Expansion, focus, volume, re-entry. In
attheendofcontentpropagationdoesnothelp: alsointhis
International Conference on Web Search and Data Mining
case,theimprovementprovidedbyaclassifieroverrandom
(WSDM). ACM.
coinflipsisnegligible.
Bakshy, E.; Rosenn, I.; Marlow, C.; andAdamic, L. 2012.
OurfindingssuggestthatinFacebookusersinteractwith
Theroleofsocialnetworksininformationdiffusion. InIn-
different types of content in similar ways, reinforcing the
ternational Conference on World Wide Web (WWW), 519–
hypothesis of echo chambers (Del Vicario et al. 2016). In-
528. ACM.
side these chambers, strongly polarized by topic (Bessi et
al. 2015b), content propagation exhibits very similar struc- Bessi,A.;Coletto,M.;Davidescu,G.;Scala,A.;Caldarelli,
turalproperties,thatarethereforelessusefulincontentclas- G.; and Quattrociocchi, W. 2015a. Science vs conspiracy:
Collective narratives in the age of misinformation. PLOS Fronczak, A.; Fronczak, P.; and Hołyst, J. 2004. Aver-
ONE10(2):1–17. age path length in random networks. Physical Review E
Bessi, A.; Petroni, F.; DelVicario, M.; Zollo, F.; Anagnos- 70(5):056110.
topoulos, A.; Scala, A.; Caldarelli, G.; and Quattrociocchi, Hanley, J., andMcNeil, B. 1982. Themeaninganduseof
W.2015b.Viralmisinformation:Theroleofhomophilyand theareaunderareceiveroperatingcharacteristic(roc)curve.
polarization. In International Conference on World Wide Radiology143(1):29–36.
Web(WWW)Companion. ACM.
Haykin,S.1998.NeuralNetworks:AComprehensiveFoun-
Borge-Holthoefer, J.; Meloni, S.; Gonc¸alves, B.; and dation. PrenticeHallPTR,2ndedition.
Moreno, Y. 2013. Emergence of influential spreaders
Ho,T.1995.Randomdecisionforests.InInternationalCon-
in modified rumor models. Journal of Statistical Physics
ference on Document Analysis and Recognition (ICDAR),
151(1-2):383–393.
volume1,278–282. IEEE.
Borge-Holthoefer, J.; Rivero, A.; and Moreno, Y. 2012.
Holland, P., and Leinhardt, S. 1971. Transitivity in Struc-
Locating privileged spreaders on an online social network.
tural Models of Small Groups. Small Group Research
PhysicalReviewE85(6):066123.
2(2):107–124.
Centola,D.,andMacy,M. 2007. Complexcontagionsand
Hong,L.; Dan,O.; andDavison,B. 2011. Predictingpop-
the weakness of long ties. American Journal of Sociology
ularmessagesintwitter. InInternationalConferenceCom-
113(3):702–734.
paniononWorldWideWeb(WWW). ACM.
Centola,D.2010.Thespreadofbehaviorinanonlinesocial
Izenman,A. 2013. Lineardiscriminantanalysis. InModern
networkexperiment. Science329(5996):1194–1197.
MultivariateStatisticalTechniques.Springer. 237–280.
Cheng,J.;Adamic,L.;Dow,P.;Kleinberg,J.;andLeskovec,
Kramer,A.;Guillory,J.;andHancock,J. 2014. Experimen-
J. 2014. Cancascadesbepredicted? InInternationalCon-
tal evidence of massive-scale emotional contagion through
ferenceonWorldWideWeb(WWW). ACM.
social networks. Proceedings of the National Academy of
Cheng,J.;Adamic,L.;Kleinberg,J.;andLeskovec,J.2016.
Sciences(PNAS)111(24):8788–8790.
Do cascades recur? In International Conference on World
WideWeb(WWW),671–681. ACM. McPherson,M.;Smith-Lovin,L.;andCook,F. 2001. Birds
ofafeather: Homophilyinsocialnetworks. AnnualReview
Cohen,J. 1968. Weightedkappa: Nominalscaleagreement
ofSociology415–444.
provisionforscaleddisagreementorpartialcredit. Psycho-
logicalBulletin70(4):213. Moreno,Y.;Nekovee,M.;andPacheco,A.2004.Dynamics
of rumor spreading in complex networks. Physical Review
Cozzo, E.; Banos, R.; Meloni, S.; and Moreno, Y.
E69(6):066130.
2013.Contact-basedsocialcontagioninmultiplexnetworks.
PhysicalReviewE88(5):050801. Newman, N. and Levy, D.A.L. and Nielsen, R.K. 2015.
Reuters Institute digital news report. http://www.
DelVicario,M.; Bessi,A.; Zollo,F.; Petroni,F.; Scala,A.;
digitalnewsreport.org/survey/2015/. [On-
Caldarelli, G.; Stanley, H.; and Quattrociocchi, W. 2016.
line;accessed10-January-2017].
Thespreadingofmisinformationonline. Proceedingsofthe
NationalAcademyofSciences(PNAS)113(3):554–559. Newman,M. 2002. Assortativemixinginnetworks. Physi-
calReviewLetters89(20):208701.
Diao, Q.; Jiang, J.; Zhu, F.; and Lim, E. 2012. Finding
bursty topics from microblogs. In Annual Meeting of the Quattrociocchi, W.; Caldarelli, G.; and Scala, A. 2014.
Association for Computational Linguistics: Long Papers - Opiniondynamicsoninteractingnetworks: Mediacompeti-
Volume 1 (ACL), 536–544. Association for Computational tionandsocialinfluence. ScientificReports4.
Linguistics.
Quattrociocchi, W.; Scala, A.; and Sunstein, C. 2016.
Doerr, B.; Fouz, M.; and Friedrich, T. 2012. Why rumors Echo chambers on facebook. Available at SSRN: https:
spread so quickly in social networks. Communications of //ssrn.com/abstract=2795110.
theACM55(6):70–75.
Rogers, K. and Bromwich, J.E. 2016. The hoaxes,
Facebook. 2016. Company info. http://newsroom. fake news and misinformation we saw on election day.
fb.com/company-info/. [Online; accessed 10- http://www.nytimes.com/2016/11/09/us/
January-2017]. politics/debunk-fake-news-election-day.
First Draft News Coalition. 2016. About First Draft. html. [Online;accessed10-January-2017].
https://firstdraftnews.com/about/. [Online; Romero, D.; Meeder, B.; and Kleinberg, J. 2011. Dif-
accessed10-January-2017]. ferences in the mechanics of information diffusion across
Fowler, J., and Christakis, N. 2010. Cooperative behav- topics: Idioms, political hashtags, and complex contagion
ior cascadesin human social networks. Proceedings ofthe ontwitter. InInternationalConferenceonWorldWideWeb
NationalAcademyofSciences(PNAS)107(12):5334–5338. (WWW),695–704. ACM.
Friggeri, A.; Adamic, L.; Eckles, D.; and Cheng, J. 2014. Salganik,M.;Dodds,P.;andWatts,D. 2006. Experimental
Rumor cascades. In AAAI Conference on Weblogs and So- study of inequality and unpredictability in an artificial cul-
cialMedia(ICWSM). turalmarket. Science311(5762):854–856.
Seo,E.;Mohapatra,P.;andAbdelzaher,T. 2012. Identify- oftheNationalAcademyofSciences(PNAS)109(16):5962–
ingrumorsandtheirsourcesinsocialnetworks.InSPIEDe- 5966.
fense, Security, and Sensing, 83891I–83891I. International Wiens,J.;Horvitz,E.;andGuttag,J.2012.Patientriskstrat-
SocietyforOpticsandPhotonics. ificationforhospital-associatedc.diffasatime-seriesclas-
Silverman, C.; Strapagiel, L.; Shaban, H.; Hall, sificationtask. AdvancesinNeuralInformationProcessing
E.; and Singer-Vine, J. 2016. Hyperpartisan Systems(NIPS)467–475.
Facebook pages are publishing false and mislead- Zollo, F.; Bessi, A.; Del Vicario, M.; Scala, A.; Cal-
ing information at an alarming rate. https: darelli, G.; Shekhtman, L.; Havlin, S.; and Quattrocioc-
//www.buzzfeed.com/craigsilverman/ chi, W. 2015a. Debunking in a world of tribes. arXiv.
partisan-fb-pages-analysis. [Online; accessed preprint,http://arxiv.org/abs/1510.04267.
10-January-2017].
Zollo,F.;Novak,P.K.;DelVicario,M.;Bessi,A.;Mozeticˇ,
Sunstein,C. 2002. Thelawofgrouppolarization. Journal I.;Scala,A.;Caldarelli,G.;andQuattrociocchi,W. 2015b.
ofPoliticalPhilosophy10(2):175–195. Emotional dynamics in the age of misinformation. PLOS
ONE10:1–22.
Ugander, J.; Backstrom, L.; Marlow, C.; and Kleinberg, J.
2012. Structuraldiversityinsocialcontagion. Proceedings