-
(2017).Anapplicationofextremevaluetheorytolearninganalytics:Predictingcollaborationoutcomefromeye-trackingdata.JournalofLearningAnalytics,4(3),140–164.http://dx.doi.org/10.18608/jla.2017.43.8
ISSN1929-7750(online).TheJournalofLearningAnalyticsworksunderaCreativeCommonsLicense,Attribution-NonCommercial-NoDerivs3.0Unported(CCBY-NC-ND3.0)
140
An Application of Extreme Value Theory to Learning Analytics:
Predicting Collaboration Outcome from Eye-tracking Data
KshitijSharma
DepartmentofOperations,FacultyofBusinessandEconomicsUniversityofLausanne,Switzerland
ComputerHumanInteractioninLearningandInstructionSchoolofComputerandCommunicationSciences
ÉcolePolytechniqueFédéraledeLausanne,[email protected]
ValérieChavez-Demoulin
DepartmentofOperations,FacultyofBusinessandEconomicsUniversityofLausanne,Switzerland
PierreDillenbourg
ComputerHumanInteractioninLearningandInstructionSchoolofComputerandCommunicationSciences
ÉcolePolytechniqueFédéraledeLausanne,Switzerland
ABSTRACT: The statistics used in education research are based on
central trends such as themeanor standarddeviation,
discardingoutliers. This paper adopts another viewpoint that
hasemerged in statistics, called extreme value theory (EVT). EVT
claims that the bulk of normaldistribution is comprised mainly of
uninteresting variations while the most extreme valuesconvey more
information. We apply EVT to eye-tracking data collected during
onlinecollaborative problem solving with the aim of predicting the
quality of collaboration. Wecompare our previous approach, based on
central trends, with an EVT approach focused onextreme episodes of
collaboration. The latter provided a better prediction of the
quality ofcollaboration.
KEYWORDS: Eye-tracking, dual eye-tracking, extreme value theory,
computer
supportedcollaborativelearning,learninganalytics,collaborationquality
1 INTRODUCTION
This contribution borrows a framework from the field of
statistics called extreme value theory (EVT),which has been
developed for analyzing time series in domains such as finance and
environmentalsciences. We explore the relevance of EVT for learning
analytics, namely for analyzing
collaborativeinteractionsinaneducationalsetting.Forthesekindsofanalyses,statisticalmethodstraditionallyfocus
-
(2017).Anapplicationofextremevaluetheorytolearninganalytics:Predictingcollaborationoutcomefromeye-trackingdata.JournalofLearningAnalytics,4(3),140–164.http://dx.doi.org/10.18608/jla.2017.43.8
ISSN1929-7750(online).TheJournalofLearningAnalyticsworksunderaCreativeCommonsLicense,Attribution-NonCommercial-NoDerivs3.0Unported(CCBY-NC-ND3.0)
141
on the central tendencies (mean,median, and standard deviation).
Generally,we discardedwhatweconsidered to be outliers, which we
suspected might be due to measurement errors, cheating,
ormiscellaneous events foreign to the cognitive mechanisms under
scrutiny. Instead, EVT invites us
tofocusontheinteractionepisodes,whichdeviatefromthosecentraltendencies.Theshiftbetweenthesetwoapproaches,fromcentraltoextremes,isaccompaniedbyanothershift:theextremedatapointsdonotcorrespondtoanindividualsubjectorapairbuttosomespecifictimeepisodeswithinalongseriesoftimeeventsproducedbyeachindividualorpair.ThegoalofthispaperistodetermineifEVTcouldprovide
us with better discrimination among different levels of
collaboration quality compared
totraditionalmethods.Wethereforeapplybothmethodstothetimesseriesproducedbyeyetrackersandcomparetheresults.Sincewestudycollaboration,wesynchronizedtheeye-trackingdataproducedbyeachpeer(whatwecall“dualeye-tracking”).EVThasbeentraditionallyusedtoquantifyrareeventslikecentury
floods, avalanches, market crashes, or more recently terrorism
attacks. Outside of the
riskmanagementcontext,ithasnotbeenmuchdevelopedbecauseofthelackofraredata.Inthispaper,wepropose
the use and development of extreme value learning tools to explore
“rare data”
fromeducational“bigdata”experimentssuchaseye-trackingexperiments.
The paper is organized as follows: Section 2 describes the
nature of dual eye-tracking data
(DUET),followedinSection3byanintroductiontoEVT.Section4introducestheconceptthatbridgesDUETandEVTintwoways.Intheunivariateway,eachpairoftimeepisodesfromlearnersAandBissubstitutedbyameasureoftheirdifferences,whichproducesatimeseriesofsinglevalues.Inthebivariatemode,wetakeintoconsiderationthedynamiccouplingofthetwotimeseries.TherestofthepapercomparestheresultsproducedbyEVTtothoseresultingfromtraditionalapproaches.
2 EYE-TRACKING
Eye-trackingprovidesresearcherswithunprecedentedaccesstoinformationaboutusers’attention.Theeye-trackingdata
is rich in termsof temporal resolution.With theadventofeye-tracking
technology,the eye-tracking apparatus has become compact and easy
to use without sacrificing much of
itsecologicalvalidityduringthecontrolledexperiments.Previousresearchhadshownthateye-trackingcanbe
useful for unveiling the cognitive processes that underlie verbal
interaction and
problem-solvingstrategies.Weintroduceheresomekeyconceptsnecessarytounderstandthestudypresentedlater.
2.1 Fixations and Saccades
In a nutshell, gaze does not glide over visualmaterial in a
smooth continuousway but rather
jumpsaroundthestimulus:smallstopsaround200milliseconds,called“fixations,”arefollowedbylongjumps,called
“saccades.” It is hypothesized that information is collected only
during fixations. However, thedata analysis ismore complex.What if
the eyes stop after 180 or 170milliseconds? Can this still
beconsidered as a fixation? Eye-trackingmethods require different
thresholds to bedefined in order toprocessdata.Are these thresholds
the same for all subjects and for all tasks? Ifwe consider a
singlesubjectonasingletask,isthethresholdstableovertime?Isitthesameinthemiddleofthescreenoron
the periphery? Eye-tracking relies on the craft of “thresholding.”
Nüssli (2011) developed
-
(2017).Anapplicationofextremevaluetheorytolearninganalytics:Predictingcollaborationoutcomefromeye-trackingdata.JournalofLearningAnalytics,4(3),140–164.http://dx.doi.org/10.18608/jla.2017.43.8
ISSN1929-7750(online).TheJournalofLearningAnalyticsworksunderaCreativeCommonsLicense,Attribution-NonCommercial-NoDerivs3.0Unported(CCBY-NC-ND3.0)
142
optimization algorithms that systematically explore threshold
parameters in order to maximize
thequalityofproduceddata.Severalstudieshaveshownthatthelevelofexpertiseofanindividual(Ripoll,Kerlirzin,
Stein, & Reine, 1995; Abernethy & Russell, 1987; Charness,
Reingold, Pomplun, &
Stampe,2001;Reingold,Charness,Pomplun,&Stampe,2001)couldbedeterminedfromeye-trackingdatasincethewayonelooksatanX-RAY(Grant&Spivey,2003;Thomas&Lleras,2007)orapieceofprogrammingcode(Sharma,Jermann,Nüssli,&Dillenbourg,2012)revealsthewayoneunderstandsthesethings.Wewillnotdevelopthesefindingsinthispaperaswefocusoncollaborativesituations.Forinstance,withina
collaborative Tetris game, Jermann, Nüssli, and Li (2010) predicted
the level of expertise in a pair(expert–expert, novice–novice, or
expert–novice pair)with an accuracy of 75%. The core
relationshipbetweengazeandcollaborationresultsfromthegaze-dialoguecoupling.
2.2 Gaze-dialogue Coupling
Two eye-trackers can be synchronized for studying the gaze of
two persons interacting to solve
aproblemandforunderstandinghowgazeandspeecharecoupled.Meyer,Sleiderink,andLevelt(1998)showed
that the duration between looking at an object and naming it is
between 430 and 510milliseconds (eye–voice span). Griffin and Bock
(2000) found an eye–voice span of about
900milliseconds.ZelinskyandMurphy(2000)discoveredacorrelationbetweenthetimespentgazingatanobjectandthespokendurationthenameoftheobjectwasgivenaloud.Richardson,Dale,andKirkham(2007)proposedtheeye–eyespanasthedifferencebetweenthetimewhenthespeakerstartslookingat
the referred object and the time when listeners look at it. This
time lag was termed the “cross-recurrence” between the
participants. The average cross-recurrencewas found to be between
1,200and1,400milliseconds.JermannandNüssli(2012)appliedcross-recurrencetoapairprogrammingtask,enablingtheremotecollaboratorstoseetheiractionsonthescreen.Theauthorsfoundthatthecross-recurrence
levelswerehigherwhenselectionwasmutuallyvisibleonthescreen,whichrelatedtothecross-recurrenceofteamcoordination.
2.3 Quality of Interaction and Cross-recurrence Several authors
have found a relationship between the cross-recurrence of gazes and
the quality ofcollaboration.CherubiniandDillenbourg (2007) founda
correlationbetweengaze-recurrenceand
theperformanceofteamsinamapannotationtask.Inapeerprogrammingtask,JermannandNüssli(2012)found
higher gaze recurrence for pairs that collaborate well, as
estimated by theMeier, Spada,
andRummel(2007)qualitativecodingscheme.Inaconcept-maptask(Sharma,Caballero,Verma,Jermann,&Dillenbourg,2015;Sharma,Jermann,Nüssli,&Dillenbourg,2013)relatedcross-recurrencetohigherlearning
gains. In a collaborative learning task using tangible objects,
Schneider and Blikstein (2015)found that cross-recurrence is
correlated with the learning gains. In a nutshell, gaze is
coupledwithcognition,andsincegazeiscoupledwithdialogue,DUETmethodsconstituteapowerfultoolwithwhichto
quantitatively investigate the quality of collaboration. The
observed correlations do not
implycausality,butsomestudiesshowthatdisplayingthegazeofonepeertotheother,asadeicticgesture,increases
teamperformance (Duchowski et al., 2004;
Sharma,D’Angelo,Gergle,&Dillenbourg,
2016;Stein&Brennan,2004;VanGog,Jarodzka,Scheiter,Gerjets,&Paas,2009;VanGog,Kester,Nievelstein,
-
(2017).Anapplicationofextremevaluetheorytolearninganalytics:Predictingcollaborationoutcomefromeye-trackingdata.JournalofLearningAnalytics,4(3),140–164.http://dx.doi.org/10.18608/jla.2017.43.8
ISSN1929-7750(online).TheJournalofLearningAnalyticsworksunderaCreativeCommonsLicense,Attribution-NonCommercial-NoDerivs3.0Unported(CCBY-NC-ND3.0)
143
Giesbers, & Paas, 2009; Van Gog & Scheiter, 2010). More
importantly, these reported studies
havemostlybeenconductedusingANOVAs,correlation
tests,F-andt-testsandregressions,whichassumethat thedata
followanormaldistribution.Wewill showthat
thedistributiontailofeye-trackingdata(low frequency events) is
quite different from the tail of normal distribution. Specifically,
EVThypothesizes that the events that occur in the tail of a
distribution are more distinguishable
thanaveragebehaviour.Thenextsection,therefore,introducesthebasicsofEVT.
3 AN INTRODUCTION TO EXTREME VALUE THEORY
Extremeeventsaredefinedasthosehavinglowfrequencyandhighseverity(orimpact).EVTisabranchofstatisticsthatdealswithmodellingtheoccurrenceandmagnitudeofsuchevents.Forinstance,flood-wallsarenotbuiltforaverageeventsbutratherforrareandcatastrophicoccurrences.EVTforfinancialor
insuranceriskmanagement
looksatextremeeventsandconcentratesontheriskofsituationsthatmight
never have happened before (McNeil, Frey,& Embrechts, 2015).
Such events (market
crashes,insurancelosses,etc.)arerarebutverysevereforcompanies,hencetheneedtomodelthedeviationsfrom
thecentral tendencies inadifferentmanner.Actually,
thedistributionof financial time series isknown to be heavy-tailed.
Therefore, EVT methods aim to model the tail with concepts
describedhereafter. For a comprehensive introduction, see Coles
(2001), or see Chavez-Demoulin and
Davison(2012)forareviewofEVTforanalyzingtimeseries.
EVTisbasedonasymptoticresults.Therefore,thedatausedtomodeleventsisaverysmallsubsetofthewholedataset(usuallyabovethe90thor95thquantile).ThemainadvantagesofusingEVT1areasfollows:First,
it
isbasedonthemathematicalfoundationsthatforanycommondistributionF,wecancharacterizethetailofFandcanthereforeunderstandthegeneratingprocessofextremeeventsfromanyunderlyingdistributionF.Fcanbeanystandardcontinuousdistribution(normal,student,uniform,exponential,
gamma, etc.); hence, EVT imposes no strong assumption upon the data
generatingprocesses,unlikeANOVAs.Second,whenanalyzingthedependencestructurebetweentwosequencesof
extreme events, the bivariate EVT context does not impose a linear
shape of dependence ascorrelation requires (Sharma,
Chavez-Demoulin, & Dillenbourg, 2016). Third, even if the
theory
isestablishedforindependentandidenticallydistributedvariables,itcanbestraightforwardlyextendedtothe
stationary context— the contextwemeet ineye-trackingand
collaborative learning—or to
thenon-stationarycontext.Whyisdualeye-trackingastationarycontext?Thegazetime-seriesareinvariantof
temporal-shifts, i.e., ifweshift the timebya factor, thevariability
in thegazepatterns remain thesame.Moreover, the gaze data at time t
are not completely independent of where the person
waslookingattimet−1,i.e.,thereexistsanauto-correlationinthegazedata.Furthermore,wedescribetheadvantagesofEVTovergeneralmethodsusedinbehaviouralresearch:
• Advantageof EVToverparametricmodels that assumenormalityof
thedata:Aspreviouslymentioned,EVTdoesnotassumeanyunderlyingdistribution
thatgenerates thedata.That is,
1
Source:http://www.bioss.ac.uk/people/adam/teaching/OREVT/2007/node12.html
-
(2017).Anapplicationofextremevaluetheorytolearninganalytics:Predictingcollaborationoutcomefromeye-trackingdata.JournalofLearningAnalytics,4(3),140–164.http://dx.doi.org/10.18608/jla.2017.43.8
ISSN1929-7750(online).TheJournalofLearningAnalyticsworksunderaCreativeCommonsLicense,Attribution-NonCommercial-NoDerivs3.0Unported(CCBY-NC-ND3.0)
144
EVT can be applied to data from any standard continuous
distribution (normal, student,uniform,exponential,gamma,etc.).
• Advantage of EVT over parametric models applied on the
normalized data: EVT offers acomplementaryviewpoint to lookat
thedata,moreparticularly to lookat thetailof thedatadistribution.
This is justified because, often in the learning analytics context,
the tail of
thedistributionismoreinformativethanthebodyofdistribution.ThisisillustratedbytherealdataofFigure5.Inthatcontext,evenifthenormalityofthetransformeddatahold,theparametricmodelsappliedonthedatawouldnotbringmuchinformationbecausethereisnodependencestructuretoexploretheaveragevalues(thepointsseemtoberandomlyspread
inthemiddlequadrantoftheplotcontainingtheaveragevalues).Moregenerally,whenagroupofstudentsisinteracting
to accomplish a task, the upper tail of the joint distribution of
temporalconcentration (or lower tail of the joint distribution of
their spatial entropy, like in Figure
5)actuallyrepresentstheepisodesduringwhichthesubjectsaretogetherfocusedinahighlevelofcollaborativequality.Theaveragejointvaluesarelessinformative,probablycontainingothereffectsthancollaboration.Insuchcases,thecompetitiveperformanceofEVTapproachesoverparametricmodels,
appliedon thenormalizeddata, emerges from the fact that
EVTprovidesthecorrecttoolstolookattheextremesequencesofthedata.
•
AdvantageofEVTovernon-parametricmodels:Bothrelyonlyontheassumptionthatthedataarecontinuous.Manyofthenon-parametricmethodsusedinlearninganalyticsarehypothesistestingandprovideonevalue(thep-value),whichsummarizesthedata.Non-parametricformscanhandleonlylowdimensionalproblems,whichgoesagainsttheflowofbigdata.Ingeneral,inthe(non-stationary)timeseriescontext,thereismuchmoretogainfromdynamicparametricmodels
than from hypothesis testing. Because EVT is available for any
common continuousdistribution, it offers the advantages of
parametricmodels like relying on likelihood, allowingformal
inference, likelihood ratio-based hypothesis tests, and also takes
into account non-stationary nature in the case of time series and
covariate dependence. Note that
non-parametricmethodsintheEVTcontextarealsopossible.
3.1 Univariate Case
ClassicalEVTconsiderstwodifferentapproaches.Thefirstapproachprovidestheasymptoticbehaviourofthemaximum:
(1) whereX1,X2,…,Xn is an independentand identicallydistributed
randomsequencewithdistributionF.Supposethatwecanfindsequencesofrealnumbers{an>0}and{bn}
𝑎" > 0 suchthatthesequenceof
normalized(orstabilized)maximumM*n=(Mn–bn)/an𝑀"∗ =()*+),)
convergesindistribution.
A remarkable result states that the only possible distribution
for the maximum is the generalized
-
(2017).Anapplicationofextremevaluetheorytolearninganalytics:Predictingcollaborationoutcomefromeye-trackingdata.JournalofLearningAnalytics,4(3),140–164.http://dx.doi.org/10.18608/jla.2017.43.8
ISSN1929-7750(online).TheJournalofLearningAnalyticsworksunderaCreativeCommonsLicense,Attribution-NonCommercial-NoDerivs3.0Unported(CCBY-NC-ND3.0)
145
extremevalue(GEV)distribution:
(2)where−∞<μ<∞isthelocationparameter,σ>0isthescaleparameter,and−∞<ξ<∞istheshapeparameter.Thisresult
isequivalenttothewell-knowncentral limittheorem(whichprovidesa
limitingdistributionforthemeanofanyunderlyingdistribution)butforthemaximum.Concretely,inmodellingextremesofaseriesofobserveddatax1,x2,xq,wedividethedataintomblocksofn.Thisgivesusanobserved
series of block maximamn,1, mn,2, ..., mn,m on which we fit a GEV,
by maximum likelihoodestimation,andgetestimatedlocation(μ
̂),shape(σ ̂),andscale(ξ ̂)parameters.ThetoppanelsinFigure1 show an
example of the selection of extreme events using the
blockwise-maximamethod for
GEVmodelfitting.ThesecondclassicalEVTapproach(mathematicallyrelatedtothefirstone)characterizesthetailofanycontinuouscommondistributionFand
is referredtoasthepeaks-over-threshold (POT)approach.Moreprecisely,
itconsidersamodelfortheexceedancesabovesomehighthresholduthatdefinesthetailofthedistributionF.UnderthePOTapproachitcanbeshownthat:
• thenumberof exceedancesabove the thresholdu arises according
to aPoissonprocesswithparameterλ,andindependently,
•
theexceedancesizeW=X−ufollowsageneralizedParetodistribution(GPD):
(3)
definedon{w:w>0and 1 + 𝜉𝑤/˜𝜎 >0},where:
(4)
Essentially,parametersoftheGPD(thresholdexcesses)canbedeterminedbyGEV(blockmaxima).Theparameterξ,whichcontrolstheshapeofthetailofthedistributionF,isthesameforbothGPDandGEV.In
applications, the POT approach ismore flexible than the blockmaxima
approach and often
allowsmoredata(morethanjustoneperblock)andthereforeleadstolessuncertainty.Aswecanseeinthetop-leftpanelofFigure1(below),thenumberofpointsconsideredformodellingarethesameasthenumberofblocks.Ontheotherhand,thenumberofpointsinthebottom-leftpanelofFigure1islargerthan
when the POT method is used. Once we have determined the
appropriate threshold,
theparameterλofthePoissonprocessandtheGPDparameters˜σandξcanbeestimatedbymaximizingthe
likelihood function. The bottom panels in Figure 1 show an example
of the selection of
extremeeventsusingthePOTmethodfortheGPDmodelfitting.
-
(2017).Anapplicationofextremevaluetheorytolearninganalytics:Predictingcollaborationoutcomefromeye-trackingdata.JournalofLearningAnalytics,4(3),140–164.http://dx.doi.org/10.18608/jla.2017.43.8
ISSN1929-7750(online).TheJournalofLearningAnalyticsworksunderaCreativeCommonsLicense,Attribution-NonCommercial-NoDerivs3.0Unported(CCBY-NC-ND3.0)
146
Figure1:Topleft:arandomvariablesimulationandtheblockwise-maxima.Topright:thedensityplotforoneoftheblocks;theredpointsshowthemaximumvalueofeachblock.Bottomleft:thesamerandomvariableasinthetop-leftpanel,theredhorizontallineshowsthethresholdforthePOTmethod,theredpointsarethepoints-over-threshold.Bottomright:thedensityplotforthewhole
distribution;theredverticallineshowsthethresholdforthePOTmethodanddenotesthebeginningofthetailforthedistribution;theredcolouredareashowsthetail,whichcorrespondstothered
pointsinthebottom-leftpanel.
Themainpracticaluseofsuchfittedmodels(GEVblockmaximaorPOT)istheadequatecalculationofthe
extremequantile ofF, that is, thequantile at a very high
level.Using either theGEVor
POT,wecalculateavalue,whichhasaverylowprobabilityofbeingexceededinagiventimeperiod.Thisvalueiscalled
the “return value,” a name inspired by environmental data inwhich
the corresponding returnperiod question is in howmanymonths or
years can it be expected that a value of the time
seriesexceedsthesamevalueagain.Thereturnvalueissetataveryhighquantile,usually95%,whichmeansthat
there is only a 5% chance that a valuewill exceed the computed
return value. In Section 4,we
-
(2017).Anapplicationofextremevaluetheorytolearninganalytics:Predictingcollaborationoutcomefromeye-trackingdata.JournalofLearningAnalytics,4(3),140–164.http://dx.doi.org/10.18608/jla.2017.43.8
ISSN1929-7750(online).TheJournalofLearningAnalyticsworksunderaCreativeCommonsLicense,Attribution-NonCommercial-NoDerivs3.0Unported(CCBY-NC-ND3.0)
147
expose the calculation of the return level, and in Section 6 we
see that the return level is
actuallyeffectivefordeterminingcollaborativequality.
3.2 Bivariate Case
AnotherwayofmodellingcollaborationwithEVTistousethegazepatternsfromthetwoparticipantsinapairandanalyzethemasabivariatetimeseries.Givenabivariaterandomsample(X1,Y1),…(Xn,Yn),EVTaddressesthelimitingbehaviourofthecomponent-wisemaxima(M1,n,M2,n),thatis,therespectivemaximumofthesequences{Xi}and{Yi},i=1,…,nasin(1).
Theasymptotictheoryofbivariateextremesdealswithfindinganon-degeneratebivariatedistributionfunction(thatcantakemorethantwovalues)Gsuchthat,asn→∞
(5)
withsequencesal,n>0andbl,n∈R,l=1,2.Ifthelimit(5)exists2andGisanon-degeneratedistributionfunction,thenGhastheform:
(6)
ThefunctionA(ω)definedas0≤ω≤1istheso-calledPickandsdependencefunction.TheindependencecasecorrespondingtoG(z1,z2)=exp{−(1/z1+1/z2)},thePickandsfunctionA(ω),measuresthedeparturefromindependence.CompletedependencebetweenthetwoseriesisreflectedbyA(1/2)=0.5;whileatcompleteindependence,A(1/2)=1.
Whileanalyzingtheeye-trackingtimeseriesoftwopeers,themainpracticaluseofthebivariateEVTisto
measure extreme dependence, which is the probability of finding an
extreme event in one
timeseries,giventhatweobserveanextremeeventinthesecondtimeseries.Thetwoextremeeventsmustoccuratthesametime,asthetwodimensions
inthisbivariatespacearethetwogazetimeseriesforthe
twopeers.Thisprobability isquantifiedas the tail-dependencebetween
the two timeseries.Theclassical methods value typically used to
measure the dependence between the two series is
thecorrelationcoefficient.Thecorrelationcoefficientiscomputedatthecentraltendencies,whilethetail-dependence
is,as inthecaseofreturnvalues,computedataveryhighquantile.
InSection4,weusethree different extremal dependence measures as
complementary and interpretable ways
fordeterminingcollaborativequality.
4 CONCEPTS
ToapplyEVT toour researchquestion,predictingcollaborationquality
fromDUETtraces,weneedtodefineafewvariables.
2 To simplify the representation andwithout loss of
generality,we transform the data (Xi , Yi) to (Z1i , Z2i), i = 1,…,
nwithstandardFréchetmarginssothatPr(Zil≤z)=exp{−1/z}forallz>0andl=1,2.
-
(2017).Anapplicationofextremevaluetheorytolearninganalytics:Predictingcollaborationoutcomefromeye-trackingdata.JournalofLearningAnalytics,4(3),140–164.http://dx.doi.org/10.18608/jla.2017.43.8
ISSN1929-7750(online).TheJournalofLearningAnalyticsworksunderaCreativeCommonsLicense,Attribution-NonCommercial-NoDerivs3.0Unported(CCBY-NC-ND3.0)
148
4.1 Gaze Visual Agitation (VA) VA isdefinedas
thecoefficientofvariance (CoV)of the fixationduration.
Visualagitation foragiventimewindowtiscomputedasfollows:
(7)
In accordance with Richardson, Dale, and Tomlinson (2009), we
chose a time window size of twoseconds. The main reason for
analyzing the variance of the fixation duration and not the
fixationduration itself is the fact that the fixationduration is
task-dependent. For instance, in a visual
searchtask,thefixationdurationswill
inherentlybesmall,astheeyeswouldbeconstantlymovingtosearchthetargetobject,whereasinataskthatrequiresdeeperinformationprocessing,thefixationdurationsarehigher.Thetaskusedinourexperiment,drawingaconcept-maptask,liesinbetween:shortfixationdurationswhenpeerssearchforaconceptonthemapversuslongerfixationswhentheydiscussthelinkbetween
the two concepts. In order to keep various task episodes
comparable, we use the
scaledvarianceofthefixationduration.AlowvalueofVAwouldmeanrelaxedgazepatternswhileahighvaluecouldresultfromstressorfatigue.
4.2 Gaze Spatial Entropy (SE)
SEmeasuresthespatialdistributionofthegazeofeachpeer.TocomputeSE,wefirstdefinea100-pixel-by-100-pixelgridoverthescreenandwecomputeforeachpeertheproportionofgazefixationslocatedin
each grid cell (Figure 2). This results in a proportionality matrix
and the SE is computed as
theShannonentropyofthis2-dimensionalvector.Thespatialentropyisalsotask-independent,asitcanbecomputedforanytask,buttheinterpretationoftheentropyvaluesmightbedependentonthevisualstimuli.AlowvalueofSEwouldmeanthatthesubjectisconcentratingonafewelementsonthescreen,whileahighSEvaluewoulddepictawiderfocussize.
Figure2:Theprocessofcomputingentropy.Theimageontheleftshowstheexemplarconcept-mapandgazepatterns(greycirclesandarrows).Theimageontherightshowstheplacementofthegrid.
-
(2017).Anapplicationofextremevaluetheorytolearninganalytics:Predictingcollaborationoutcomefromeye-trackingdata.JournalofLearningAnalytics,4(3),140–164.http://dx.doi.org/10.18608/jla.2017.43.8
ISSN1929-7750(online).TheJournalofLearningAnalyticsworksunderaCreativeCommonsLicense,Attribution-NonCommercial-NoDerivs3.0Unported(CCBY-NC-ND3.0)
149
4.3 Return Levels: Univariate Extremes
Thereturnlevelisthequantileatahighlevel(above90%forexample)ofthedatadistribution.Whydowenotsimplycalculatethisquantilefromthedistributionofourentiredataset?Wecoulddothis,butsmall
discrepancies in the estimation of the body distribution would lead
to large errors in theestimation of the quantiles in the tail. The
POTmodel presented in Section 3 is
themathematicallycorrectwaytoestimatesuchhighquantilesandinpracticeleadstomoreaccurateestimation.TheEVTestimationalsobringsinformationabouthowheavyisthetailofthedistributionF;thatis,howlargearetheextremesthatdistributionFcangenerate?This
information isprovidedby thevalueof theshapeparameter ξ in (2) or
(3): as ξ becomes larger, the tail ofF becomesheavier.Wedonot
explore thisfeature further in this paper because as with any other
modelling approach, just from the set ofestimated parameters of
location µ, scale σ or ˜𝜎,, and shape ξ, it is cumbersome to
explain
andcomparethedifferentmodels.Hence,weusethereturnlevel,calculatedusingthemodelparameters,whichhasavaluableinterpretation.
As mentioned in Section 3, the return value (say, calculated at
the 95% quantile), symbolizes themeasureof the
(unseen)extremeeventwitha5%probability that theactual
(unseen)eventexceedsthisvalue.Inwhatfollows,wederivethereturnlevelcalculationfromthePOTmodelaboveathresholdu.
We recall that the underlying variable is denoted X and that the
exceedances occurrence arrivesaccording to a Poisson processwith
parameter λ, and the exceedance sizeW= X − u follows a
GPDdenotedasHin(3)withparameters( ,ξ).Forx>u,wehave:
Itfollowsthat
(8)
Hence,thereturnlevelxporextremequantileatthepercentilep(large)isthesolutionof
(9)
sothat,
(10)
-
(2017).Anapplicationofextremevaluetheorytolearninganalytics:Predictingcollaborationoutcomefromeye-trackingdata.JournalofLearningAnalytics,4(3),140–164.http://dx.doi.org/10.18608/jla.2017.43.8
ISSN1929-7750(online).TheJournalofLearningAnalyticsworksunderaCreativeCommonsLicense,Attribution-NonCommercial-NoDerivs3.0Unported(CCBY-NC-ND3.0)
150
Inanon-mathematicalway, the return levelxp is thevalueatwhich
theprobabilityofexceeding thisvalue is equal to 1 − p. We obtain
the estimated return level (10) by fitting the POT model to
theexceedance data, estimating the probability of exceeding the
threshold, Pr(X > u), using the
Poissonmodelandreplacingtheparameters
andξwiththeirmaximumlikelihoodestimates.
IsEVToverkill,orisitreallynecessarytoanalyzethetwovariablesthatwehavedefined,visualagitationandspatialentropy?Figure3usesQ–Qplotsforcomparingthedistributionofthesetwovariableswithanormaldistribution.Bothplotsshowaheavytailforlowfrequencyvaluesofspatialentropy(leftplot)and
visual agitation (right plot), respectively. This justifies the use
of sophisticated EVT methods toprocess these tails.We will
therefore compare the return levels calculated for the two
participants.Similarreturn
levelswoulddepictahigheramountoftemporalconcordance.
InSection6,wewillseethatcomparingreturnlevelsindeedprovidesanaccurate(andinterpretable)wayofdiscriminatinghighandlowcollaborationquality.
Figure3:Q–QplotsofSpatialEntropy(left)andVisualAgitation(right)definedinSection4.
4.4 Three Measures of Extremal Dependence: Bivariate
Extremes
Estimating dependence between the two partners in a pair’s
extremal behaviour provides
somecomplementaryinformationaboutthepeers’concordance.Wefirstintroducetheextremalcoefficient
θ=2Α(1/2) (11)whereA is thePickands functionmentioned
inSection3.Thus,θ∈ [1,2], and it
canbeconvenientlyinterpretedastheeffectivenumberofindependentseries;thecaseθ=2meansthatthetwoseriesareindependent
andwe therefore get complete independence. The case θ = 1means that
the
effectivenumberofindependentseriesis1,andthereforewegetcompletedependence.
The two other extremal dependence measures we consider come from
conventional
multivariateextremevaluetheory,characterizingtwoclassesofextremevaluedependence:asymptoticdependence
-
(2017).Anapplicationofextremevaluetheorytolearninganalytics:Predictingcollaborationoutcomefromeye-trackingdata.JournalofLearningAnalytics,4(3),140–164.http://dx.doi.org/10.18608/jla.2017.43.8
ISSN1929-7750(online).TheJournalofLearningAnalyticsworksunderaCreativeCommonsLicense,Attribution-NonCommercial-NoDerivs3.0Unported(CCBY-NC-ND3.0)
151
and asymptotic independence, which characterizes the behaviour
of variables as they
becomemoreextreme.Inthiscontext,weconsiderthecoefficientofextremaldependence
(12)
Thelimitvalueχ∈[0,1]isstrictlypositivewhenalargevalueofZ2leadstoanon-zeroprobabilityofaslargeasvalueZ1.
Inotherwords,χ is thetendencyforonevariabletobe
largegiventhattheother
islarge.Thismeansthattheonlypossibilityforasymptoticindependenceiswhenχ=0.Whenχ>0,thevariablesareasymptoticallydependent.Inthatcontext,wedefine,asasecondextremalcoefficient,theconditionalprobability
(13)
From this we see that χ̅ = 1 means perfect dependence between
the two series while χ̅ = 0 impliesindependence.The coefficientχ̅
is thereforeameasureofdependence for the classof
asymptoticallyindependentmodels.Inourcontext,χtellsusthelevelofasymptoticdependence,andχ̅tellsusaboutthestrengthoftheasymptoticdependence.Inpractice,as(12)and(13)arelimits,wesetavalueofztypicallyataveryhighquantilefor(12)andverylowonefor(13),referredtoasz×100percentilefor(12)andtakingthe(1−z)×100percentilefor(13),asshownintheresultsinSection6.
Figure4:Exampleillustratingthedeterminationofthecoefficientofextremaldependenceχandthestrengthofdependenceχ̅forthevisualagitationofapair.Thedashedlinesrepresentthe95%
confidenceintervalsforχandχ̅.Thetail-dependenceanditsstrengthisdeterminedbythevaluesatthehigherquantiles(typicallybetween95%and99%).Theredlinescorrespondto95%.
Figure4showsanexample illustrating thedeterminationof
thecoefficientofextremaldependenceχandthestrengthofdependenceχ̅
forthespatialentropyofapair.Whydowecalculateχandχ̅ forall
-
(2017).Anapplicationofextremevaluetheorytolearninganalytics:Predictingcollaborationoutcomefromeye-trackingdata.JournalofLearningAnalytics,4(3),140–164.http://dx.doi.org/10.18608/jla.2017.43.8
ISSN1929-7750(online).TheJournalofLearningAnalyticsworksunderaCreativeCommonsLicense,Attribution-NonCommercial-NoDerivs3.0Unported(CCBY-NC-ND3.0)
152
the quantiles? This is just an empirical method, and we are only
interested in the highest quantilevalues.
Again, isbivariateEVToverkill,or is it reallynecessary toanalyze
thevariables
thatwehavedefined,visualagitationandspatialentropy?Figure5showsthatthedependencestructurebetweenthespatialentropyofthetwopeersisfarfromlinear(forbothlowandhighcollaborativequalitypairs).Insuchacase,aPearsoncorrelationwouldleadtoerroneousconclusions.Thisleadstothedevelopmentofmoresophisticated
methods to adequately model dependence structure; see, for
instance, Sharma et al.(2017).
Figure5:Scatterplotsofspatialentropybetweenthepeerswithlow(leftpanel)andhigh(rightpanel)qualityofcollaboration.
5 EXPERIMENT
TheEVTframeworkpresentedaboveprovidesanewmethodforanalyzingthedualeye-trackingdata.Theresearchquestionwespecificallyaddress
isthefollowing:Doextremevaluesfromgazeepisodespredictthequalityofcollaborativelyproducedconceptmapsbetterthancentraltrends?
To answer this question, we conducted an experiment with 66
master’s students from ÉcolePolytechnique Fédérale de Lausanne who
participated in the present study. There were 20 femalesamong the
participants. The participants were each compensated with 30 Swiss
francs for
theirparticipationinthestudy.TheflowoftheexperimentisshowninFigure6.
Upontheirarrivalinthelaboratory,theparticipantssignedaconsentform.Thentheytookanindividualpre-testonthebasicsofneuronaltransmission.Thentheparticipants
individuallywatchedtwovideosabout“restingmembranepotential.”Next,theycreatedacollaborativeconcept-mapusingIHMCCMap
-
(2017).Anapplicationofextremevaluetheorytolearninganalytics:Predictingcollaborationoutcomefromeye-trackingdata.JournalofLearningAnalytics,4(3),140–164.http://dx.doi.org/10.18608/jla.2017.43.8
ISSN1929-7750(online).TheJournalofLearningAnalyticsworksunderaCreativeCommonsLicense,Attribution-NonCommercial-NoDerivs3.0Unported(CCBY-NC-ND3.0)
153
tools.3 Finally, they tookan individualpost-test.The
twovideoswere taken
from“KhanAcademy.”4,5Thetotallengthofthevideoswas17minutes.Itisworthmentioningthattheteacherwasnotphysicallypresentduringthevideos.Theparticipantscametothelaboratoryinpairs.Whilewatchingthevideos,the
participants had full control over the video playerwithout any time
constraint. The collaborativeconcept-map phase was 10–12 minutes
long. During that time participants could talk to each
otherwhiletheirscreensweresynchronized,
i.e.,peerswereabletoseeeachother’sactions.Boththepre-testandthepost-testcontainedtrue–falsequestions.
Figure6:Schematicrepresentationofthedifferentphasesoftheexperiment.
5.1 Quality of Collaboration
The final concept-map was compared with the concept-map created
by the two experts. The pairreceived a score using the following
rules: 1) one mark for each correct connection between
twoconcepts,2)onemarkforeachcorrectlabeloftheedgebetweentwoconcepts,3)halfamarkforeachpartiallycorrect
labelof theedgebetween twoconcepts.Thepairswere thendivided into
two levelsbased on the concept-map score using a median split. Why
do we consider this as a measure
ofcollaborationquality?ThereasonrestsintheworkofJermann,Mullins,Nüssli,andDillenbourg(2011),Jermann
and Nüssli (2012), and Kahrimanis, Chounta, and Avouris (2010), who
showed that theactions/task-basedoutcome isoften correlatedwith the
collaborationquality.Hence,ourassumptionabouthavingthecollaborativeproductqualityasaproxyofcollaborationqualityisgroundedinpreviousfindings.AsWiseandShaffer(2015)suggest,“...theoryplaysanever-morecriticalrole
inanalysis,”sousingthesesupportsfromtheliterature,wecanproceedwiththeaforementionedassumption.
3CMaptools4RestingMembranePotential-Part15RestingMembranePotential-Part2
-
(2017).Anapplicationofextremevaluetheorytolearninganalytics:Predictingcollaborationoutcomefromeye-trackingdata.JournalofLearningAnalytics,4(3),140–164.http://dx.doi.org/10.18608/jla.2017.43.8
ISSN1929-7750(online).TheJournalofLearningAnalyticsworksunderaCreativeCommonsLicense,Attribution-NonCommercial-NoDerivs3.0Unported(CCBY-NC-ND3.0)
154
6 RESULTS
6.1 Univariate Extremes
Recallthequestionweaddressinthispaper:DoesEVTrevealdifferencesthatcentraltrendsfailedtoreveal?
Figure7 shows thepipeline for dataprocessing. Let us beginwith
the central trends approach. Ifwecompare the difference in the
average levels of entropy of the peers, we observe no
significantdifferences between high- and low-quality pairs. An
ANOVA shows no significant difference in
theaverageentropydifferenceforthepeerswithhighand
lowcollaborationquality
(F[1,21.48]=0.01,p-value=.93,Figure8d).Thesamelackofdifferenceisfoundwiththevisualagitation(F[1,22]=1.73,p-value=.20,Figure8c).
Figure7:Thepipelineforunivariatedata-processing.
Now,wecomparethepreviousresultswiththoseprovidedbyEVT.Weestimatedthereturnlevel(10)atpercentilep.Tokeepenoughdata,wesetp=90
forvisualagitationandp=95
forspatialentropy.Thereasonforsettingp=90forvisualagitationistohaveenoughdatapointstofitaGEVorPOT.Thedifference
between peers in terms of return levels tells us about their
synchronicity. The differencebetween peers in return levels for
visual agitation is lower for high-quality pairs than for
low-qualitypairs(F[1,14.08]=4.92,p-value=.04,one-wayANOVAwithoutassumingequalvariances).Similarly,thedifference
between peers in return levels for spatial entropy is also lower
for high-quality
pairs(F[1,15.15]=8.39,p-value=.01,one-wayANOVAwithoutassumingequalvariances).Figures8aand8bshowthemeansandconfidenceintervalsforthedifferenceinthereturnlevelsforvisualagitationandspatial
entropy respectively. In otherwords, both for agitation andentropy,
the
extremesoccurwithhighersynchronicityforthehigh-qualitypairsthanforthelow-qualitypairs.
-
(2017).Anapplicationofextremevaluetheorytolearninganalytics:Predictingcollaborationoutcomefromeye-trackingdata.JournalofLearningAnalytics,4(3),140–164.http://dx.doi.org/10.18608/jla.2017.43.8
ISSN1929-7750(online).TheJournalofLearningAnalyticsworksunderaCreativeCommonsLicense,Attribution-NonCommercial-NoDerivs3.0Unported(CCBY-NC-ND3.0)
155
(a)Meansandconfidenceintervals(bluebars)forthedifferenceintheestimatedreturnlevels(10)at90percentileforvisualagitation,for
high-andlow-qualitypairs.
(b)Meansandconfidenceintervals(bluebars)forthedifferenceintheestimatedreturnlevels(10)at90percentileforspatialentropy,for
high-andlow-qualitypairs.
(c)Meansandconfidenceintervals(bluebars)forthedifferenceinthemeanvaluesforvisual
agitation,forhigh-andlow-qualitypairs.
(d)Meansandconfidenceintervals(bluebars)forthedifferenceinthemeanvaluesforspatial
entropy,forhigh-andlow-qualitypairs.
Figure8:Results:Univariateextremes
-
(2017).Anapplicationofextremevaluetheorytolearninganalytics:Predictingcollaborationoutcomefromeye-trackingdata.JournalofLearningAnalytics,4(3),140–164.http://dx.doi.org/10.18608/jla.2017.43.8
ISSN1929-7750(online).TheJournalofLearningAnalyticsworksunderaCreativeCommonsLicense,Attribution-NonCommercial-NoDerivs3.0Unported(CCBY-NC-ND3.0)
156
6.2 Bivariate Extremes
Weagaincomparethetwomethods:DoesEVTrevealdifferences(ofdependenciesamongpeers)thatcentraltrendsdidnot?
Letus startwith standard correlations. Ifwe compute the
correlationbetween the spatial entropyoftwopeers,we can see inboth
Figures 10c and10d, thatwe cannot learn anything from the
averagevalues (the body of the distribution), and the Pearson
correlation/linearmodel does notmake
sensehere.Thismightleadtofalseinterpretationsoftheunderlyingcollaborativeprocesses.
Let us now compare the EVT approach to the bivariate time
series. To estimate their
extremaldependence,westartbyestimatingtheextremalcoefficientθasin(11)betweenthevariablesforthetwopeers.Weobservethathigh-qualitypairshaveahigherdependenceforvisualagitationthanlow-qualitypairs(F[1,22]=6.07,p-value=0.02,Figure9a).Similarly,high-qualitypairshaveahigherlevelofdependence
in visual entropy than low-quality pairs,with the difference being
evenmore
significant(F[1,22]=7.65,p-value=0.01,Figure9b).Thescalesonthey-axesforFigures9aand9bareinverted.Aswe
mentioned in Section 4.4, complete dependence is reflected by θ =
1, whereas completeindependenceisreflectedbyθ=2.
Next,weestimatethelevelχdefinedin(12)andstrengthχ̅definedin(13)oftheextremaldependence.Weobserveahigherextremaldependence(calculatedatthe95%quantile)betweenthevisualagitationofpeersforpairswithhighcollaborationquality(F[1,22]=9.19,p-value=0.006,Figure11a).Moreover,weobserveanevenmoresignificantdifferenceinthestrengthoftheextremaldependence(calculatedat
the 95%quantile) in favour of the pairswith high collaboration
quality (F[1,22] = 11.71, p-value =0.002,Figure11c).
Regarding spatial entropy, we observe effects similar to visual
agitation. There is a higher extremaldependence (calculated at the
95% quantile) between the spatial entropy of peers with
highcollaborationquality(F[1,22]=6.31,p-value=0.01,Figure11b).Similartothecaseofvisualagitation,weobserveanevenmoresignificantdifferenceinthestrengthofextremaldependence(calculatedatthe95%quantile)forthepairswithhighcollaborationquality(F[1,22]=14.28,p-value=0.001,Figure11d).
There is a higher (χ) and stronger (χ)̅ (calculated at the
95%quantile) extremal dependence for bothvisual agitation (Figure
10a) and spatial entropy (Figure 10b) for the high-quality pairs
than the low-qualitypairs.Weobserveaclearseparation,
inthe2-dimensionalspaceofχandχ,̅betweenthehigh-and low-quality
pairs (with three and one exception for visual agitation and
spatial entropy,respectively). As we observe in the case of
temporal univariate return levels, the difference
ismoreevidentinthecaseofspatialentropythaninthecaseofvisualagitation.
-
(2017).Anapplicationofextremevaluetheorytolearninganalytics:Predictingcollaborationoutcomefromeye-trackingdata.JournalofLearningAnalytics,4(3),140–164.http://dx.doi.org/10.18608/jla.2017.43.8
ISSN1929-7750(online).TheJournalofLearningAnalyticsworksunderaCreativeCommonsLicense,Attribution-NonCommercial-NoDerivs3.0Unported(CCBY-NC-ND3.0)
157
(a)Meansandconfidenceintervals(bluebars)fortheestimatedextremalcoefficientθforVAofthe
participants,forhigh-andlow-qualitypairs.
(b)Meansandconfidenceintervals(bluebars)fortheestimatedextremalcoefficientθforSEofthe
participants,forhigh-andlow-qualitypairs.
Figure9:Bivariateextremes:Dependencemeasures.
(a)Coefficientχandstrengthχ̅ofextremal
dependenceforVAforhigh(redpoints)andlow(bluepoints)collaborationqualitypairs.
(b)Coefficientχandstrengthχ̅ofextremal
dependenceforSEforhigh(redpoints)andlow(bluepoints)collaborationqualitypairs.
-
(2017).Anapplicationofextremevaluetheorytolearninganalytics:Predictingcollaborationoutcomefromeye-trackingdata.JournalofLearningAnalytics,4(3),140–164.http://dx.doi.org/10.18608/jla.2017.43.8
ISSN1929-7750(online).TheJournalofLearningAnalyticsworksunderaCreativeCommonsLicense,Attribution-NonCommercial-NoDerivs3.0Unported(CCBY-NC-ND3.0)
158
Figure10:Results:Bivariateextremes,extremalcoefficient,andtaildependence.
(a)Meansandconfidenceintervals(bluebars)fortheestimated
levelofextremaldependenceχ inthe visual agitation of the
participants, for high-andlow-qualitypairs.
(b) Means and confidence intervals (blue bars)fortheestimated
levelofextremaldependenceχforspatialentropyoftheparticipants,forhigh-andlow-qualitypairs.
(c)SEvaluesforpeersinahigh-qualitypair.Thecorrelationdoesnotreflectthetruerelationship,asthereisnolinearrelationbetweentheSEvalues
forpeers.
(d)SEvaluesforpeersinalow-qualitypair.Thecorrelationdoesnotreflectthetruerelationship,asthereisnolinearrelationbetweentheSEvalues
forpeers.
Figure11.Results:Bivariateextremes,levels,andstrengthoftaildependence
-
(2017).Anapplicationofextremevaluetheorytolearninganalytics:Predictingcollaborationoutcomefromeye-trackingdata.JournalofLearningAnalytics,4(3),140–164.http://dx.doi.org/10.18608/jla.2017.43.8
ISSN1929-7750(online).TheJournalofLearningAnalyticsworksunderaCreativeCommonsLicense,Attribution-NonCommercial-NoDerivs3.0Unported(CCBY-NC-ND3.0)
159
(c)Meansandconfidenceintervals(bluebars)fortheestimatedstrengthofextremaldependenceχ̅inthevisualagitationoftheparticipants,forhigh-andlow-qualitypairs.
(d) Means and confidence intervals (blue bars)for the estimated
strength of extremaldependence χ̅ in the visual agitation of
theparticipants,forhigh-andlow-qualitypairs.
Figure11.Results:Bivariateextremes,levels,andstrengthoftaildependence.
7 DISCUSSION
DoesEVTprovideinterestingfindingscomparedtostatisticalmethodsbasedoncentraltrends?
Let us first address this question in the univariate context.
The comparison ofmean values of
visualagitationorspatialentropydidnotrevealanydifferencebetweenhigh-qualityandlow-qualitypairs.Onthecontrary,EVTrevealedthathigh-qualitypairshaveasignificantlysmallerdifferenceofreturnlevelsfor
both variables. This shows that during extreme episodes of
collaboration there exists a
higheramountof“togetherness”amongtheparticipantsinhigh-qualitypairs.
Thebivariatecontextisevenmoreinteresting.Thethreetaildependencecoefficientsweusedmeasuredependence
between the extremes of visual agitation and spatial entropy in a
time series. Morespecifically, from theextremal coefficientθwe
learn theeffectivenumberof independent series:
forhigh-qualitypairs,θ
̂≈1,meaningthatthetimeseriesofonepeer,forbothvariables,sufficestoexplain(ordescribe)theextremesoftheotherpeer.Thishighlightsanextreme“togetherness”incollaborationbetweenthetwoparticipantsofthepair.
The dependencemeasures χ and χ̅ play a role similar to the
Pearson correlation, but they avoid thedrawbacks of standard
correlation (not robust to outliers, restricted to linear
dependence structure,spoiledbyothereffectsaffecting thebodyof
thedistribution). Theextremaldependencemeasuresχandχ̅ focuson
theextremevaluesof the twovariables. Similarly to the
interpretationof correlation,largevaluesofχ
̂andχ̅̂indicateastrongdependencebetweentheirepisodesofhighVAandSE.Thefact
-
(2017).Anapplicationofextremevaluetheorytolearninganalytics:Predictingcollaborationoutcomefromeye-trackingdata.JournalofLearningAnalytics,4(3),140–164.http://dx.doi.org/10.18608/jla.2017.43.8
ISSN1929-7750(online).TheJournalofLearningAnalyticsworksunderaCreativeCommonsLicense,Attribution-NonCommercial-NoDerivs3.0Unported(CCBY-NC-ND3.0)
160
that the bivariate tail-dependence is higher and stronger for
the high-quality pairs confirms theunivariatefindings.
Usingthebivariatespaceformedbythesamegazemeasureforbothparticipants
inthepair(bothforVA and SE), we eliminate the need for grouping
(averaging or grouping the individualmeasures in
aregressionmodel)thepeermeasuresintopairvariables.
7.1 Why Does EVT Work?
One reason EVTworks is that, unlike standardmethods that suffer
from the difference between theassumed underlying distribution and
the actual distribution, EVT properly models the tail (of
anycommondistribution)using thecorrectmodel (POTorGEVblockmaxima).
Second,whenweuse theextreme episodes, we focus only on the moments
that might reflect the episodes during which
thecollaboratorsaremostlikelytobe“together.”Then,byfocusingonextremecollaborationepisodes,weremove
the noise that could have prevented classicalmethods from
differentiating the collaborationquality levels.This fact
isalsoevident inFigures10cand10d.Correlationdoesnot reflect
thecorrectrelationbetweentheSEforthetwoparticipants.
However,whycouldwenottakethetop5%quantileandperformanANOVAonthosevalues?Averysimple
answer is that the main assumption for ANOVA is that the values
should follow a
normaldistribution,anditismathematicallyproventhatthetailofanydistribution,whichisnormalinthecaseofANOVA,doesnotfollowthedistribution.Instead,itfollowstheGPD.Hence,itwouldbestatisticallywrongtoperformanANOVAonsuchvariables.CouldwesimplynormalizethedataandthenperformtheANOVA?Thiscouldleadtoaproblemaswecompletelyignoremanyotherpropertiesofdata(e.g.,skewandkurtosis)whilenormalizingthedata.Thus,keyaspectsofthedatagenerationprocessmightbehiddenor
removed.EVTprovidesamethod thatassumesnounderlyingdistribution
regarding thedata generating process, unlike other
classicalmethods. This removes the need to force the data
tofollowanygivenstatisticaldistribution.
7.2 When to use EVT?
EVT offers the correct way (in the sense that it is based on
mathematical foundations) to
analyzeabnormaldata(inthesenseofdatafarfromtheaveragevalues).TheEVTtheoryforthelargestvaluesor
peaks-over-threshold or bivariate case exposed in the paper is
available for any underlyingcontinuous distribution. It should be
used when analyzing the tail distribution (for any kind
ofcontinuousdistribution)asacomplementaryexplorationofthedata,orwhentraditionalmethodsfailorare
uninformative, either because the assumptions required by these
methods (like the linearmodel/Pearson correlation) based on linear
dependence between the two variables are violated ornearly violated
or because the average values on which all these (parametric or
non-parametric)methodsarebaseddonotcontaintherelevantinformationofinterest,beingthereforelesspredictive.For
example,when a student iswriting in a graphical table, the extreme
values of her time series ofwriting speed/pressure are her abnormal
sequences (in the sense of departure from her standard
-
(2017).Anapplicationofextremevaluetheorytolearninganalytics:Predictingcollaborationoutcomefromeye-trackingdata.JournalofLearningAnalytics,4(3),140–164.http://dx.doi.org/10.18608/jla.2017.43.8
ISSN1929-7750(online).TheJournalofLearningAnalyticsworksunderaCreativeCommonsLicense,Attribution-NonCommercial-NoDerivs3.0Unported(CCBY-NC-ND3.0)
161
measures)andrelatetoherepisodesofstress.Anotherexample,whenateacherlooksattheexamstoinfertheheterogeneityoftheclass,shecannotjustbesatisfiedbyarobustmeasureofthevariabilityofmarks.Shehas
tocarefully consider theworstand thebestmarks (theextremes)as the
limitsof theclass heterogeneity frame. Neglecting theworst and the
best would not onlymean neglecting
somestudents(whoprobablyhaveanimportantimpactontheclass)butalsoneglectingrelevantinformation.Furthermore,
while analyzing trace data (for example, click-streams), although
the theory is
notestablishedforthediscretecase,itistypicallyusedtocountvariables,likePoissonvariables,becauseoftheirapproximationbycontinuousdistribution.
8 CONCLUSION
Itiseasytounderstandthatastatisticalmodelthatpredictsariseinwaterlevelof5metreshasmoresocial
relevance thanamodel thatpredictsa riseof5centimetres. Ineducation,
thisapproach is lessintuitive. Typically, a teacher would care for
the average level of his class and try to cope with
itsheterogeneity. It is hence very counter-intuitive that EVT
reaches a higher discriminative power
thanmethodsbasedoncentraltrends.
Insciences,whatiscounter-intuitiveisalwaysinteresting.However,we
should not forget that the extreme values are not outliers but
extreme time episodes
duringcollaboration,whichislesscounter-intuitive.Ifateachermonitorsaclassroomwithseveralteams,(s)hewouldprobablybealsoattractedby“extreme”episodes;forinstance,whenpeersdonotspeakatallorwhentheyshoutateachother.Inourexperiment,therawdataisnotdialoguebutgazepatterns,andatthispointnothingprovesthatsimilarresultswouldbeobtainedwithotherbehavioural
traces.WedonotclaimthatEVTshouldreplaceotherstatisticalmethodsusedinlearninganalytics,butratherthatitexpandstherangeoftoolsavailabletolearningscientists.Byusingitacrossmultiplelearningcontexts,wewilllearnwhenandwhyitbringsmorediscriminativepowerthanmethodsbasedoncentraltrends.
REFERENCES
Abernethy,B.,&Russell,D.G.(1987).Therelationshipbetweenexpertiseandvisualsearchstrategyinaracquet
sport. Human Movement Science, 6(4), 283–319.
http://dx.doi.org/10.1016/0167-9457(87)90001-7
Charness,N., Reingold, E.M., Pomplun,M.,& Stampe,D.M.
(2001). Theperceptual aspectof
skilledperformanceinchess:Evidencefromeyemovements.Memory&Cognition,29(8),1146–1152.http://dx.doi.org/10.3758/BF03206384
Chavez-Demoulin, V., & Davison, A. C. (2012). Modelling time
series extremes. REVSTAT: StatisticalJournal,10(1),109–133.
Cherubini,M.,&Dillenbourg,P. (2007).Theeffectsofexplicit
referencing indistanceproblemsolvingover shared maps. In
Proceedings of the 2007 International ACM Conference on
SupportingGroupWork(GROUP’07),4–7November2007,SanibelIsland,FL,USA(pp.331–340).NewYork:ACM.http://dx.doi.org/10.1145/1316624.1316674
Coles, S. (2001). An introduction to statistical modeling of
extreme values. London:
Springer.http://dx.doi.org/10.1007/978-1-4471-3675-0
-
(2017).Anapplicationofextremevaluetheorytolearninganalytics:Predictingcollaborationoutcomefromeye-trackingdata.JournalofLearningAnalytics,4(3),140–164.http://dx.doi.org/10.18608/jla.2017.43.8
ISSN1929-7750(online).TheJournalofLearningAnalyticsworksunderaCreativeCommonsLicense,Attribution-NonCommercial-NoDerivs3.0Unported(CCBY-NC-ND3.0)
162
Duchowski,A.T.,Cournia,N.,Cumming,B.,McCallum,D.,Gramopadhye,A.,Greenstein,J.,…Tyrrell,R.A.
(2004).Visualdeicticreference inacollaborativevirtualenvironment.
InProceedingsofthe2004 Symposiumon Eye
TrackingResearch&Applications (ETRA ’04) 22–24March2004,
SanAntonio,TX,USA(pp.35–40).NewYork:ACM.http://dx.doi.org/10.1145/968363.968369
Grant, E. R., & Spivey, M. J. (2003). Eye movements and
problem solving guiding attention
guidesthought.PsychologicalScience,14(5),462–466.http://dx.doi.org/10.1111/1467-9280.02454
Griffin,Z.M.,&Bock,K.
(2000).Whattheeyessayaboutspeaking.PsychologicalScience,11(4),274–279.http://dx.doi.org/10.1111/1467-9280.00255
Jermann,P.,Mullins,D.,Nüssli,M.A.,&Dillenbourg,P.(2011).Collaborativegazefootprints:Correlatesof
interactionquality. InConnecting Computer-SupportedCollaborative
Learning to Policy andPractice:Proceedingsof the9th
InternationalConferenceonComputer-SupportedCollaborativeLearning(CSCL2011),4–8July2011,HongKong,China(Vol.1,No.EPFL-CONF-170043,pp.184–191).InternationalSocietyoftheLearningSciences.
Jermann, P., & Nüssli, M.-A. (2012). Effects of sharing text
selections on gaze cross-recurrence andinteractionquality
inapairprogrammingtask.
InProceedingsofthe2012ACMConferenceonComputer Supported Cooperative
Work (CSCW ʼ12), 11–15 February, Seattle, WA, USA
(pp.1125–1134).NewYork:ACM.http://dx.doi.org/10.1145/2145204.2145371
Jermann,P.,Nüssli,M.-A.,&Li,W.(2010).Usingdualeye-trackingtounveilcoordinationandexpertisein
collaborative Tetris. InProceedings of the 24thBCS Interaction
SpecialistGroupConference(BCS ’10), 6–10 September 2010, Dundee, UK
(pp. 36–44). Swindon, UK: BCS Learning &DevelopmentLtd.
Kahrimanis,G.,Chounta,I.A.,&Avouris,N.(2010).Studyofcorrelationsbetweenlogfile-basedmetricsof
interaction and the quality of synchronous collaboration.
International Reports on Socio-Informatics,7(1),24–31.
McNeil,A., Frey,R.,&Embrechts,P. (2015).Quantitative
riskmanagement:Concepts,
techniquesandtools.Princeton,NJ:PrincetonUniversityPress.
Meier, A., Spada, H., & Rummel, N. (2007). A rating scheme
for assessing the quality of computer-supported collaboration
processes. International Journal of Computer-Supported
CollaborativeLearning,2(1),63–86.http://dx.doi.org/10.1007/s11412-006-9005-x
Meyer, A. S., Sleiderink, A.M., & Levelt,W. J. (1998).
Viewing and naming objects: Eyemovementsduring noun phrase
production. Cognition, 66(2), B25–B33.
http://dx.doi.org/10.1016/S0010-0277(98)00009-2
Nüssli,M.-A.(2011).Dualeye-trackingmethodsforthestudyofremotecollaborativeproblemsolving.PhDThesis,ÉcolePolytechniqueFédéraledeLausanne.
Reingold,E.M.,Charness,N.,Pomplun,M.,&Stampe,D.M.(2001).Visualspaninexpertchessplayers:Evidence
from eye movements. Psychological Science, 12(1),
48–55.http://dx.doi.org/10.1111/1467-9280.00309
Richardson, D. C., Dale, R., & Kirkham,N. Z. (2007). The art
of conversation is coordination
commongroundandthecouplingofeyemovementsduringdialogue.PsychologicalScience,18(5),407–413.http://dx.doi.org/10.1111/j.1467-9280.2007.01914.x
-
(2017).Anapplicationofextremevaluetheorytolearninganalytics:Predictingcollaborationoutcomefromeye-trackingdata.JournalofLearningAnalytics,4(3),140–164.http://dx.doi.org/10.18608/jla.2017.43.8
ISSN1929-7750(online).TheJournalofLearningAnalyticsworksunderaCreativeCommonsLicense,Attribution-NonCommercial-NoDerivs3.0Unported(CCBY-NC-ND3.0)
163
Richardson, D. C., Dale, R., & Tomlinson, T. M. (2009).
Conversation, gaze coordination, and beliefsabout visual context.
Cognitive Science, 33(8), 1468–1482.
http://dx.doi.org/10.1111/j.1551-6709.2009.01057.x
Ripoll, H., Kerlirzin, Y., Stein, J.-F., & Reine, B. (1995).
Analysis of information processing, decisionmaking, and visual
strategies in complex problem solving sport
situations.HumanMovementScience,14(3),325–349.http://dx.doi.org/10.1016/0167-9457(95)00019-O
Schneider,B.,&Blikstein,P.(2015).Comparingthebenefitsofatangibleuserinterfaceandcontrastingcasesasapreparationforfuturelearning.InProceedingsofthe11thInternationalConferenceonComputer
Supported Collaborative Learning(CSCL 2015), 7–11 June 2015,
Gothenburg,Sweden.InternationalSocietyoftheLearningSciences.
Sharma,K.,Caballero,D.,Verma,H., Jermann,P.,&Dillenbourg,P.
(2015). LookingATversus lookingTHROUGH: A dual eye-tracking study
inMOOC context. Proceedings of the 11th InternationalConference on
Computer Supported Collaborative Learning(CSCL 2015), 7–11 June
2015,Gothenburg,Sweden.InternationalSocietyoftheLearningSciences.
Sharma,K.,Chavez-Demoulin,V.,&Dillenbourg,P.(2017).Non-stationarymodelingoftail-dependenceoftwosubjects’concentration.(Toappear)AnnalsofAppliedStatistics.
Sharma,K.,D’Angelo,S.,Gergle,D.,&Dillenbourg,P.(2016).VisualaugmentationofdeicticgesturesinMOOCvideos.InProceedingsofthe12thInternationalConferenceoftheLearningSciences(ICLS’16),20–24June2016,Singapore.ISLS.http://dx.doi.org/10.22318/icls2016.28
Sharma,K.,Jermann,P.,Nüssli,M.-A.,&Dillenbourg,P.(2012).Gazeevidencefordifferentactivitiesinprogram
understanding. In Proceedings of the 24th Annual Conference of
Psychology ofProgramming Interest Group. London, UK, November
21–23, 2012.https://infoscience.epfl.ch/record/184006
Sharma, K., Jermann, P.,Nüssli,M.-A.,&Dillenbourg, P.
(2013).Understanding collaborativeprogramcomprehension: Interlacing
gaze and dialogues. In Proceedings of the 10th
InternationalConference on Computer-Supported Collaborative
Learning (CSCL 2013), 15–19 June 2013,Madison, WI, USA.
International Society of the Learning
Sciences.https://infoscience.epfl.ch/record/184007
Stein,R.,&Brennan,S.E.(2004).Anotherperson’seyegazeasacueinsolvingprogrammingproblems.InProceedings
of the 6th International Conference onMultimodal Interfaces (ICMI
’04) 13–15October 2004, State College, PA, USA (pp. 9–15). New
York: ACM.http://dx.doi.org/10.1145/1027933.1027936
Thomas, L. E., & Lleras, A. (2007). Moving eyes and moving
thought: On the spatial compatibilitybetween eye movements and
cognition. Psychonomic Bulletin & Review, 14(4),
663–668.http://dx.doi.org/10.3758/BF03196818
VanGog,T.,Jarodzka,H.,Scheiter,K.,Gerjets,P.,&Paas,F.(2009).Attentionguidanceduringexamplestudy
via the model’s eye movements. Computers in Human Behavior, 25(3),
785–791.http://dx.doi.org/10.1016/j.chb.2009.02.007
VanGog,T.,Kester,L.,Nievelstein,F.,Giesbers,B.,&Paas,F.
(2009).Uncoveringcognitiveprocesses:Differenttechniquesthatcancontributetocognitive
loadresearchandinstruction.Computers
-
(2017).Anapplicationofextremevaluetheorytolearninganalytics:Predictingcollaborationoutcomefromeye-trackingdata.JournalofLearningAnalytics,4(3),140–164.http://dx.doi.org/10.18608/jla.2017.43.8
ISSN1929-7750(online).TheJournalofLearningAnalyticsworksunderaCreativeCommonsLicense,Attribution-NonCommercial-NoDerivs3.0Unported(CCBY-NC-ND3.0)
164
inHumanBehavior,25,325–331.http://dx.doi.og/10.1016/j.chb.2008.12.021VanGog,
T., & Scheiter, K. (2010). Eye tracking as a tool to study and
enhancemultimedia learning.
LearningandInstruction,20(2),95–99.http://dx.doi.org/10.1016/j.learninstruc.2009.02.009Wise,A.,&Shaffer,D.W.(2015).Whytheorymattersmorethaneverintheageofbigdata.Journalof
LearningAnalytics,2(2),5–13.http://dx.doi.org/10.18608/jla.2015.22.2Zelinsky,G.J.,&Murphy,G.L.(2000).Synchronizingvisualandlanguageprocessing:Aneffectofobject
name length on eye movements. Psychological Science, 11(2),
125–131.http://dx.doi.org/10.1111/1467-9280.00227