An Evaluation of Information Visualizati on inAttention-Limited Envir onments

JacobSomervell, D. ScottMcCrickard , Chris North, Maulik ShuklaDepartment of Computer Science

Vir ginia Polytechnic Institute and StateUniversityBlacksburg, VA 24061-0106USA

jsomerve@vt.edu,mccricks@cs.vt.edu,north@cs.vt.edu,mshukla@vt.edu

ABSTRACTPeopleoftenneedtoquickly accessormaintainawarenessofsecondary informationwhile busywith other primarytasks.Informationvisualizations provide rapid,effective accesstoinformation,but aregenerally designedto be examined byusersas the primary focus of their attention. The goal ofthis researchis to discover how to designinformationvisu-alizationsintended for theperipheryandto understandhowquickly andeffectively peoplecaninterpret informationvi-sualizationswhile they are busily performing other tasks.We evaluatedhow several factorsof a visualization(visualdensity, presencetime,andsecondarytasktype)impactpeo-ple’s abilities to continue with a primary taskandto com-pletesecondarytasksrelatedto thevisualization. Ourresultssuggestthat,with relaxed time pressure,reducedvisual in-formationdensityanda singlewell-definedsecondary task,people caneffectively interpret visualizationswith minimaldistractionto theirprimarytask.Keywords: Informationvisualization,informationdensity,dual-tasksituations,empirical evaluation,peripheraldisplays

INTRODUCTIONPeopleneedinformation. Many decisionsandactionsarebasedon information gathered from a variety of sources.The weatherdictateswhat people wear and whethertheycarry umbrellas. Stockpricesinfluenceinvestments. Traf-fic informationhelpsdecidewhich routesto take andwhichnot to take. With the advent of the internetand wirelesstechnology, theseandotherinformationarereadilyavailableoncomputerdesktops,cell phones,handheldcomputers,in-dashvehicle displays,andelsewhere.

With theavailability of this informationcomestheproblemof presenting it in aneffectivemanner. Thefield of informa-tion visualization investigatesmethods for addressingthisproblemusinggraphical representationsthatcaptureandre-flect important aspectsof the information [3, 23]. Infor-mationvisualizations canenable usersto quickly assimilatelarge amounts of data,andempirical evaluation hasled toimproveddesignsover time [4].

However, the evaluation of information visualizationshasfocusedalmostexclusively on situationsin which usersex-plore theinformationin a visualizationastheironly task.Inreality, usingavisualizationis quiteoftennotaperson’ssoleor primary task. Computerusershave long usedvisualiza-tionssuchasemail toolsandsystemloadmonitors to keeptrackof informationwhile performingothertasks.Today, asinformationfrom chattools,Webalertsystems,stocktrack-ers,scoretickers,buddy lists, schedules,andothersourcesinvadeour desktops,it is important to understandhow bestto communicatethis informationin aneffectivemannerwithminimal negative impacton theuser’sothertasks.

Complex informationvisualizationsoftenrequiresignificantattentionto analyze,andtheir presencecananddoesaffectotherwork. Thispaperexplorestheuseof informationvisu-alizationsassecondary displays(peripheral visualizations).In general, a person’s attentionwill befocusedonsomepri-marytask,but at timesit maybenecessaryto divert partialattentiontoasecondary taskthatinvolvesgatheringinforma-tion fromavisualization. Thismayoccurthrough peripheralvision or shifts in visual focus, but theprimary focusof at-tentionshouldremainontheprimary task.Hence,only lim-ited attentioncanbedevotedto thesecondary visualizationtask.For example, a student maywantto work on a collab-orative assignment while watchingfor chatmessagesfromhis colleagues,or an investmentprofessionalmay want tomonitor stockpriceswhile sending email to her clients,orthe driver of a vehiclemay want to look at mapdirectionswhile driving.

In order to designinformation visualizations intended forsecondary tasks,more understandingis neededabout theutility of visualizationin multiple-tasksituations.It is sus-pectedthat suchvisualizations are distracting, but little isknownaboutthedegreetowhichit distractsusersandwhetheruserscanovercomethesedistractions andinterpret the pe-ripheral visualizations. Similar to standalone informationvisualizations, we expect thatperipheral visualizations willhavesomebenefitsin termsof userperformancefor assimi-lating information.However, we alsoexpectthatthedesignof peripheralvisualizationswill needto bedifferentthanthatof standalone visualizations.For example, typically a goalin informationvisualization designis to maximizevisualin-formationdensity[25]. But in peripheral visualizations,in-creasedvisual densitymay result in additional distractionanddecreaseduserperformance.

In thisresearch, weconduct controlledempirical evaluationsthatexamine theimpacts,bothpositive andnegative,of theuseof informationvisualization asa secondary taskunder avarietyof conditions.

Weattemptto addresstwo primary questionswith thiswork:� How quickly and effectively can people interpret an

information visualization while busily performing othertasks?Thatis, we wantto learnwhether peoplecanpar-tially switchfrom their primary taskto thesecondary vi-sualizationtaskwhenaninformationvisualizationis pre-sented.We examine people’s performanceon the visu-alizationtasksandthe distractionon their primary taskduring andafter the presentation of a peripheral visual-ization.� How can peripheral visualizations be designedto re-

duce distraction while maintaining awareness? Forexample, a visualizationthat contains moredatapointshasthepotential to bettershow clustersandtrendsin thedata,but it seeminglybecomesmoredifficult to quicklyfocus on individualdatapointsandcancausedistraction.In searchof a winning combinationor tradeoff analysis,we examine several attributesof information visualiza-tionsincluding visualinformationdensity, presencetime,andtasktype.

In answeringthesequestions,we hopeto establishguide-lines for the presentation of information visualizations inthe periphery. This work represents the initial stagesof alengthyseriesof studiesto achieve this goal. This researchhasthe potential for long rangeimpact in many domains.For example, studieshave shown that integratedin-vehiclesystemsdo decreasethe attentionof the driver to the driv-ing task,but do communicateinformationmoreeffectivelythannon-integrated systems[9]. Effective methods for thetiming, placement, andrepresentationof informationin in-vehicleinformationsystemscould impactsafetyissuesandhelppreventseriousaccidents.

RELATED WORKMany of theguidelines weusedin defining ourexperimentsstemmedfrom early researchon perception in user inter-faces. Someof the earliestevaluations examined the per-ceptibility andreadability of rapidserialvisualpresentations(RSVPs)of letters,strings,andwords [10, 14]. More re-cently, researchershavebeenexamining theeffectivenessofgraphical displayswhenpresentedfor short times,focusingon changesin visual featureslike color andorientation[20,24,6]. Othersconsideredtheeffectsof visualattributessuchastexture, luminence,dimensionality, andmotionin thevi-sualdisplay[8, 12, 13]. This work required participants toquickly interpretcomplex visualdisplays,resultingin guide-linesfor theuseof color in displaydesign.

While perception playsa partin theunderstandingof infor-mationin theperiphery, alsoimportantis theability to tran-sition attentionbetweentasksquickly and easily. All thepreviouslymentionedevaluationsconsideredtheviewing ofdisplaysasthesoletaskof theuser. However, in multi-taskenvironments, userswould be balancing attention. In re-centyears,severalresearchteamshaveexaminedtheeffects

of displaying informationor attractingattentionto displaysin the periphery while the useris attempting someprimarytask [2, 7, 18, 19, 22]. For thesestudies,the researchersconducteddual-taskexperimentsin which participants per-formed somecentraltaskswhile various typesof displaysshowed different typesand amounts of information. Thisinformationwasusedin answeringquestions or performingsecondary tasks. In general,the displays in the peripheryweretextual [7, 18,19] or simplegraphical [2, 22] displays.

Our researchfollows a similar experimentaldesign, but dif-fersin thatwearefocusingnotontextual orsimplegraphicaldisplaysbut on visualizations that usemany factors(color,shape,position)tocommunicateinformation.As isexplainedin the next section,we usedprior resultsas guidelines inidentifying reasonable boundsfor ourexperimentaldesign.

EXPERIMENTSIn conductingtheexperiments,weexaminehow variousfac-torsaffecttheability to assimilateinformationfrom displaysin theperiphery. Specifically, we focus on threefactors:vi-sual information density, visualizationpresencetime, andthe type of task the userwantsto accomplish with the in-formation.

Little work hasbeendoneto assessthevariouseffectsvisualinformationdensity may have on information assimilation,particularly in multi-tasksituations. We speculatethat therecommendeddensity may dependon the useof the data.For example, displayingmany datapointsmaybebeneficialfor recognizing patterns in thedata,while displaying fewerdatapointsmay bemorehelpful for determining a specificvalueor datum. If oneis trying to locatea singletargetin adisplay, fewer distracterswould bebest. We areconcernedwith how the visual densityof the peripheral informationmaycausedistractionandaffect primary taskperformance.

Wealsowanttodeterminewhateffect,if any, thepresenceofvisualizations may have on primary taskperformance. Forperipheraldisplays,presencetime becomes importantwhenit may interrupt or distractfrom a primary task. Determin-ing limits andrecommendations for presencetime is partic-ularly important in safetycritical systemssuchasindustrialmachinery, monitoring stations,andvehicleoperation. It isdesirable to only show theimportantperipheralinformationfor anamount of timethatwill not interferewith theprimarywork task,yetenable thesecondary tasks.

Pilot Stud yTo betterunderstandhow people interactwith peripheralvi-sualizationsin attention-limitedenvironments,weconductedapilot studyin whichparticipantsplayedavideogame(pri-mary task)while simultaneously attempting to interpret aninformationvisualization(secondary task). 45 participantsperformedthreeroundsof theexperiment,with eachrounddisplaying a visualizationof a differentlevel of visualden-sity. In eachround, participants werefirst given four ques-tionsthatthey wouldanswerusingthevisualization. Partic-ipantsplayedthevideogamefor 30 seconds, thenthevisu-alizationappearedfor 30secondswhile thegamecontinued,andthenthe visualization disappeared for the final 10 sec-onds of the game. After eachround, participants thenan-

Figure1: Sampleinformation visualizationsusedin the experiment.Figure � showsa low densityvisualization, Figure�

showsahigh densityvisualization representingthesamedistrib ution of data. For themain experiment,eachparticipantsawonly onevisualization(either high or low density)andansweredonequestionper round. Questionsthat participantsansweredincluded “In which quadrant is the redcircle” and “In which quadrant arethe yellow objectsclustered”. Notethat for any quadrant, the high density visualization contains sixteentimes asmany yellow objectsasthe low density.

sweredthefour questionsabout thevisualization. Thegameinvolvedcatching droppingblockswith a ‘catcher’, by usingtheleft andright arrow keys. We felt thatthis wouldrequireconstantattention,thusmakingit possibleto observe whenparticipantsweredistracted.

We varied the densityof the visualizations betweenhigh,medium, andlow, with eachvisualization representingthesamedatadistribution. That is, correspondingquadrantsofthethreevisualizationscontainedthesamepercentageof ob-jectsrelatingto thequestions thatparticipantsmustanswer.Thehighdensityvisualizationhad320objects,medium had80,andlow had20. Visualizationscontained coloredshapeobjects,with threedifferentshapesandfour differentcolors.We chose20 objectsasa minimumdensitybecauseit wasfeasibleto determine patterns and clusteringat that level,but not with fewer. We expectedthat increasingthe num-berof objectswouldbea benefit to thesetasks(determiningpatternsand identifying clusters)and a detriment to othertasks(finding or counting theoccurrenceof certainobjects).Figure1 shows thehighandlow densityvisualizations.

The four questionsasked during eachround were: identi-fying thequadrantof thevisualizationcontaining anobjectof a certaincolor andshape,counting occurrencesof certinobjects,determining the patternformedby certainobjects,andidentifying the quadrant containing a clusterof certainobjects.Sincetherewerefour questionsto answerpervisu-alization,wedecidednotto requireparticipantsto attempttomemorize thequestions before beginning thegame. Hence,thequestionsweredisplayedonscreenduringtheentireround.We alsodecided not to require participants to perform ex-tra manualinteractions to indicateanswersto thequestions

while playing the game. So participantsneededto remem-bertheir answersasthey discoveredthemduring theround.After thegamewasover, thequestionswereagain presentedalongwith four possibleanswersfor eachquestionin amul-tiple choiceformat.

To betterunderstandhow participantsbehaved, we usedanon-head-mounted eye tracking systemwith somepartici-pantsto monitor when,where, andfor how longthey lookedat the various regions of thescreen(game,questions, visu-alization). We alsocollecteduserperformance datafor thegame andcorrectnessdatafor thequestions.

Lessons learnedIn constructing thepilot study, ourexpectationwasthatpar-ticipantswouldspendmostof theirtimelookingatthegame.However, afterthevisualizationappeared,participantslookedat both the questions and the visualizationmore than thegame. Basedon eye-tracking data,participantson averagespent13percent of thetotalfixationtimelookingattheques-tions. This seemedto impact their game performance, astheir block-catchingratedroppedby 10% after the visual-izationwasdisplayed. Furthermore, thehigh memory loadof rememberingtheanswersto thequestionsuntil theendofthe gamemay have negatively impactedtheir performanceonthesecondarytask.Thesebehaviorsarenot typicalof theinteractions we intended to model. Hence,in our mainex-periment,wedecidedto presentonly onequestionperroundsothatparticipantscouldeasilymemorize thequestionandremembertheir answer. This avoidstheneedto displaythequestiononscreenduringthegameandvisualization.

Anotherinterestingobservationstemmedfromthevisualiza-

tion presencetimesusedin thepilot study. In thepilot study,wechoseto leavethevisualizationon-screenfor 30secondswith the expectationthat participantswould only look at itfor asmallportionof thetime. However, participantsrepeat-edly lookedat thevisualization andat thequestionsfor theentireon-screentime. Hence,in the main experiment,wechoseto limit thevisualizationpresencetimeto gainabettersenseof how quickly participantscouldeffectively processthevisualization.

Webasedourupperboundonasafetyregulationestablishedby the Societyof Automotive Engineers. It statesthat in-vehiclenavigation systemactivities shouldrequirea maxi-mumof 15 seconds to complete,with eyes-off-the-roadac-counting for approximately 60 to 75 percent of that time[11]. For one test group, we choseto presentthe visual-ization for eight seconds, just under the bound definedbythis guideline.

As a lower limit, we consideredprior work in which par-ticipantswereaskedto identify trendsin dataasquickly aspossiblein single-tasksituations[12]. In this work, partic-ipantstypically completed the tasksin lessthana second.For our othertestgroup, we choseto have thevisualizationpresentfor onesecondsincethedual-tasksituationis moredifficult. We hopedto determine if it is feasiblefor peopleto find answersto the secondary-taskquestionsin as littleasonesecondor if it requiredasmuchaseightseconds(ormore).

Experi mental desi gnThis 2 (time) X 2 (density)X 2 (tasktype)experimentwasdesignedto determine relative performanceon tasksin adual-tasksetting.Twenty-eightundergraduatestudents froma computer scienceclassparticipatedin theexperiment forclasscredit. Participants performedsix rounds of playingavideogame andansweringquestions aboutthevisualizationthat appeared. As in the pilot study, the primary taskusedin this experimentwasa simplevideogame.Theonly dif-ferencewasa fasterdroprateof blocks thatmadethegameslightly moredifficult.

Thesecondary taskconsistedof answering asinglequestionabout informationthatwasdisplayed in a visualizationthatappearedwhile playingthegame.Thequestionsaskedpar-ticipantsto notein which quadrant (upper left, upperright,lower left, lower right) of the visualization a targetwaslo-cated.The target waseithera singleitem (e.g. redsquare)or a clusterof items(e.g. greenobjects).As in mostvisu-alizations,the picture contained various outlier objects. Ineachround, participantsviewed either a high or low den-sity visualization.High densityvisualizationscontained320objectsandlow densityvisualizations contained20 objects.Thesemockvisualizationsweredesignedto mimic commoninformationvisualizations suchasthe Spotfirestarfield[1]or mapsof landmarks. While the lack of real underlyingdatamayhave madethetaskmoredifficult, we believe thatit wasnecessaryto ensureuniform understandingby all par-ticipants.

Eachround startedwith thepresentation of thequestionthatthe participant would answerusing the visualization. The

Figure2: Gameand visualization seenby participants inthe experiment. The visualization was only presentforeither oneor eight seconds.Before eachround, partici-pants were given a questionthat they usedthe visualiza-tion to answer.

questionwas then removed andparticipants thenplayed asimplegameasin thepilot study. After 15 seconds of play-ing thegame,thevisualizationappearedon thescreen.In-corporatedin thevisualizationwastheanswerto the targetquestion. The visualizationremained visible for eitheroneor eightseconds,dependingon thetestgroup, andthendis-appeared. Participants then playedthe gamefor an addi-tional10seconds. Thetargetquestionthenreappearedalongwith 4 multiple choiceanswers.SeeFigure2 for a screen-shotof theexperimentalsetup.

Theexperimentincludedthreeindependentvariables:

Time 1 or 8 seconds timepicturewaspresentDensity low or high low=20objects,high=320Question singleor cluster find singleor a cluster

The time the visualizationwaspresentvariedbetweenpar-ticipants:eitheroneor eightseconds. Theexpectationwasthatparticipants would become acclimatedto thetime. Thevisualizationdensity(low = 20 objects,high = 320objects)andquestiontype (find singleitem, find cluster)werebothwithin-subjectsvariables. Eachparticipant saw both highandlow densityvisualizationsandeachsaw both typesoftarget question.

Theexperimentalsetupincludedfour testgroups. Thefirsttestgroup hada visualizationpresencetime of onesecond,andthey saw high thenlow densityvisualizations. Thesec-ond testgroup differed in that they saw low densitybeforehigh densityvisualizations. The third test group was thesameasgrouponeexceptthey saw thevisualizationfor eightseconds. Groupfourwaslikegroup twowith aneightsecondvisualizationpresencetime. For eachround, all participantsplayedthesamegame,saw thesamevisualization, andtriedto answerthesamequestion.Theonly things thatvariedina given round werethedensityof thevisualizationandthetime thevisualizationwasvisible.

Figure 3: Average performance (after the visualizationappeared) for the 1 secondconditions, basedon high vslow densityand singlevscluster questiontype. Both den-sity andquestionresultin significantdiffer encein perfor-mance.

To measureprimary taskperformancewemeasuredtheper-cent of blockscaughtboth for the time before the visual-ization appeared and for the time period after it appeared(including while it wasvisible). We refer to this asperfor-mance. The expectation wasthat presentingandremovingthe visualizationmaybe disruptive to the participants. Forthe secondary task performancewe measuredthe correct-nessrate for answeringthe questions. We refer to this ascorrectness.

RESULTSTheresultsof thisdual-taskexperimentincludemeasuresofperformanceontheprimary taskaswell asmeasuresoncor-rectnessin thesecondary task. We compareddifferent con-ditions usinga paired-samplet-test. Analyzing thesemea-suresseparatelyallows us to examine the issuesdescribedpreviously: effect of visualizationpresenceongameperfor-mance,effectof visualizationdensity on informationassim-ilation, andeffect of visualizationdensity on gameperfor-mance. We expectedto find that the presence of the visu-alizationwould impactperformance on the game. We alsoexpectedthatlocatinga singleitem will beeasierin thelowdensityvisualizations,andlocatinga clusterof objectswillbe easierin the high densityvisualizations.The followingsectionssummarizetheresultsof theexperiment.

PerformanceWe found no main effect betweenperformancebeforethevisualizationappeared andafter the visualizationappearedfor eithertheone-second group, ������� ����������� , ������������� ,( ����� �!� �#"�$ , �&%'�(��������� ), or the eight-second group,�������)�*��� +�$-, , �.�/����+��#$ , ( � � ����� ��+#� , � % ����� ���� ).This is somewhatunexpectedbecausein thepilot study, thepresenceof thevisualization resultedin over 10%differencein performance. Comparing performanceon rounds withhigh density visualizationsto rounds with low densityvi-sualizationsindicatesa main effect in the onesecondcon-ditions, �����0��1�324+5� $�� , �6�(�!� ��+�" , with low densityvi-sualizations( � �7�!� ����$ , 8:9;�<��� �#"!� ) yielding betterperformanceover high densityvisualizations( �;�=�!� �#�#� ,8:9>�?�!� �-,-� ). No maineffect on performancefor densitywasfound in the eight secondcondition, ���@���)�A2B�!� ���#� ,

Figure 4: 2x2x2 representation of average performanceafter visualization has appeared. Dark arr ows indicatestatistical significanceand the dir ection of which groupperformed better. Gray arr ows indicate marginal signif-icance.

�C�6���D,�+#� , ( �&E=�'��� �#��" , ��FC�'��� ���0� ). Whenwe com-paredperformanceon roundswith thesecondary taskques-tion to locatea singleobject to roundswith the secondarytaskto locatea clusterof objects,we found a maineffect intheonesecondcondition, ���@���G�H+5� $!�0� , �I�J�!� ����� , withperformancehigherwhenlocatingasingleobject( �K�L��� � ,8:9?�.�!� ����$ ) ascomparedto locatingacluster( �M�.�!� ��,�+ ,8:9N�'��� ��,�� ). Therewasno maineffect for questiontypein theeightsecondcondition, ���@���4�=2B�!� "���, , �&�O��� �#�!� ,( �QPR�L�!� ����� , �&S��L��� ���T� ). SeeFigures3 and4 for a rep-resentationof meanperformanceafter thevisualizationhasappeared.

Correct nessTo examinewhether theorder in which participants saw thevisualizationsaffectedcorrectness,wecomparedthemwithinasingletimecondition. Wefoundnomaineffectoncorrect-nessfor the two orderings, with ���U��V�W��� $-,5� , �X�?��� ����$( � EZY\[]F�^ �_��� � , � F�^`[]EZY �N�����a$�� ) for the onesecondcondition and ���U��Q�b2B�����a$�� , �(�c�!� ����$ ( � E Y\[dF!^ ���� "�+#" , � F�^e[]E Y �f�!� "��#+ ) for the eight secondcondition.As expected, we did find a main effect for time, ���@���g�24�5��+#��+ , �/�A��� �#�#��+ , with thosein the eight secondcon-dition ( � �c��� "�$ , 8h9i�c��� �#��� ) answeringmore ques-tionscorrectly thanthosein theonesecondcondition( �M�������,5� , 8:9_�'����+���� ). Comparing correctnesson high den-sityvisualizationsto low densityvisualizationsrevealsamaineffect in theonesecondcondition, �����0��:�j24+�� � , �g�C��� ��+�,( �&E>�k�!� $-�#+ , ��F=�>��� �#" ) as well as in the eight sec-ondcondition, �����0��l�(24+�� ����, , �m�'�!� �!�" ( � E �H��� �#� ,� F �i�#� � ), with people answeringmore questionscor-rectly with low densityvisualizations.Comparing correct-nesson ‘find singleitem’ questions to correctnesson ‘findclusterof items’ questionsrevealsa maineffect in boththeoneandeight secondgroups. In the onesecondgroup wehave �����0��&�b24+5��+5�0" , �n�o��� ��$-� , with more questionsansweredcorrectlyon ‘find cluster’ questions( �p�H��� �#" ,8:9<�f�!� ����+ ) thanon ‘find single item’ questions( � ���� $���+ , 8:9i�c�!� +#� ). In the eight secondgroup we have�����0�� �/24+5� �#��, , �q�L�!� �!�" , with morequestions answered

Figure 5: Secondarytask correctnessbasedon visual-ization density and question type. There is a significantdiffer encein correctnessin both the 1 and 8 secondcon-ditions.

Figure 6: 2x2x2 representation of average correctnessbasedon density, question type and time. Dark arr owsindicate statistical significancewith the arr owhead indi-cating the higher correctnessrate. Gray arr ows indicatemarginal significance.

correctly on ‘find cluster’questions ( �r���#� � , 8:9(�/��� � )thanon ‘find single item’ questions ( � �_�!� ����� , 8:9K����s��#� ). SeeFigures5 and6 for a representationof correct-nessbasedon densityandquestiontype. We alsowantedto examine whetherdensityaffectedcorrectnessfor differ-entquestion types.For theonesecondcondition with ‘findsingleitem’ questionswe find a marginal effect for density,�������t�c24+�� ��" , �/�(��� ����, , with people answeringmorequestionscorrectlywith thelow densityvisualizations( �<���� ��$�� ) thanwith thehighdensityvisualizations( �<�L��� ��+ ).The samecomparisonin the eight secondgroup produceda main effect, ���@���I�i24+5� $���� , �f�7�!� ��+�" , with morequestionsansweredcorrectlywith low densityvisualizations( � �b�#� � , 8:9;�u�!� � ) than high densityvisualizations( �<�m���D,�� , 8:9?�.��� �#� ). Nomaineffectwasfound for den-sity in answering‘find cluster’questions in eithertheoneoreightsecondconditions, with �����0��)�A2B��� $���� , �.�'��� �#���( ��Ec�>�!� �#$�� , �&F'�3���D,-�$ ) for the one secondgroup.Participants answered all ‘find cluster’ questions correctly(100%) in theeightsecondcondition for bothdensities.SeeFigure7 for arepresentationof correctnessbasedondensity,within a question.

Figure 7: Corr ectnessbasedon density for the single-item question. There is a significant differ encewithinboth the 1 and 8 secondgroups.

DISCUSSION

PerformanceSomewhat surprisingly, introducingthesecondary taskpro-ducednodiscernibledifferenceonprimarytaskperformance.Increasingthevisualizationpresencetime from onesecondto eightsecondsdoesnot seemto effect performance.Thisconstrastswith thepilot study, wherevisualizationpresencetime waslonger(30 seconds) andtheprimary taskwaseas-ier. We suspectthat the large number of secondary tasksrequiredof theparticipantsin thepilot wastheprimary con-tributing factorto this result.

Furthermore,thevisualdensityhada significantimpactonperformancein theonesecondconditionsafterthevisualiza-tion appeared.Peoplewereableto catchmorefalling blockswhentherewas a low densityvisualizationon the screen,ascomparedto whentherewasa highdensityvisualization.Perhaps the low densityvisualizationwas lesscognitivelydemanding andsimply distractedthe usersless. However,in theeightsecondcondition therewasnodiscerniblediffer-encein performancefor eitherhighor low densityvisualiza-tions. Thelonger time spanseemedto provide enough timefor participants to analyze thevisualization.

We found that the questiontype impactedperformance intheonesecondcondition. Peoplecaughtmoreblocks whenthey weretrying to locateasingleobjectin thevisualizationthanwhenlocatingclustersof objects. Perhapslocatingasingleitem is not asdifficult aslocatinga clusterof objects,or perhapsparticipants simplydid not try ashard– notethatthey answeredlessthanhalf of thequestionscorrectly.

As with density, we seethatquestion typeproducesno dis-cernible differencein performance in theeightsecondcon-dition. Hence,theremust be someform of time pressureon theparticipants whenthevisualizationis visible for onesecond. We seesignificantnegative impactsonperformancewith bothdensity andquestiontype,only in theonesecondtestgroups. This shorttime couldhave elevatedparticipantstresslevels,causingthedegradationin performance.

Obviously thesecondary taskcouldbedonein roughly onesecond, sohaving thevisualizationvisible for eightsecondsmay seemlike overkill. We expectedthat peoplewould

probably look at the visualizationseveral timesduring theeightsecondconditionandthiswouldreduceperformanceinthosecases.Togetasenseof this,wehadahandful of partic-ipantsperform theexperimentwhile having theireyemove-mentsrecordedwith a reflectiveeye-trackingsetup.As sus-pected,participants lookedat thevisualizationseveraltimesin theeightsecondconditionsandonly onetime in theonesecondcondition. The duration of the fixationson the vi-sualizationwerealsoindicative of thebehavior. In the onesecondcondition thefixation wasbetween0.5 and0.9 sec-onds. In theeightsecondcondition thefixation timesweremuchlarger: 1.5to 3.0seconds.

CorrectnessIn considering correctnesson thesecondary task,it appearsthathaving thevisualizationvisible for a longer time is ben-eficial. We alsoseethat visualizationdensityplaysan im-portant role in answeringthequestions. In boththeoneandeightsecondconditions we seesignificantlyhighercorrect-nessratesfor thelow densityvisualizations. This is interest-ing aswe expectedto seemixedperformancewith density,basedon thetypeof question theparticipants wereanswer-ing. We expectedpeopleto perform betteron findingsingleitemswith low densityvisualizations,andbetteron findingclusterswith highdensityvisualizations.

Looking furtherat questiontype,we seethatin general par-ticipantsperformedbetterwhenlocatingclustersof objectscomparedto locatinga singleobject. Hereagainthis couldbeattributedto pre-attentiveprocessing.Thetaskof locatinga clusterof objectsbasicallyrequired participants to locaterepeatedinstancesof a target color in the display. Shapewasnota factorfor clustersasit waswhenlocatinga singleobject.

In addition, in the onesecondcondition, we discoveredatradeoff betweenthe primary and secondary tasksfor thetwo questiontypes. For single-itemquestions,userperfo-manceimproved but correctnessdegraded. Whereas,forclusterquestions,userperformancedegradedbutcorrectnessimproved. For clusterquestions, the investmentwasworth-while, sincetheimprovement in correctnessis muchgreaterthanthedecreasein performance. Perhapsclusterquestionscausedmoredistractionbydrawingmoreof theparticipants’attentionawayfromtheprimarytask,but to thebenefitof thesecondary task. Or, perhapsparticipants simply gave up onsingle-itemquestionsandmaintainedfocuson the primarytask.In any case,thismightindicatethatin timepressuresit-uationspeople have a fixedamount of attentionalresourcesthatmustbetradedoff betweentasks,especiallyfor visuallyintensive tasksasin ourexperimentalsetup.

CONCLUSIONS AND FUTURE WORKWe conductedthis researchto explore how quickly andef-fectively peoplecan interpret an information visualizationwhile performingothertasks.We considereda primary taskthatdemandshighattentionfromtheuserin orderto empha-sizethe taskswitchingrequired in dual-tasksituations.Weusedabstractinformation displaysin the secondarytasks,andaskedparticipantsto answerquestionsabout thevisual-ization.

We wantedto examine how several factorsof thevisualiza-tion would affect a person’s ability to interpret themin anattention-limited environment. In this work, we focusedonvisual informationdensity, visualizationpresencetime,andsecondary tasktype asattributesto vary. Eachof theseat-tributeshasparticular relevanceto thedisplayof visualiza-tionsin theperiphery.

The following list outlines the most interestingfindings ofourstudies:� Peripheral visualizations can be intr oduced without

hindering primary taskperformance.Theprimary taskweconsideredrequiredconstantandconsistentattention,andit appearedthatparticipantswereableto allocatethisattentionfor visualizationspresentedfor eitheroneoreightseconds.This is important as it suggestthat people areabletoperformnon-trivial dualtaskswith someefficiency,andgiveshopefor peripheral visualizationdesignfor sim-ilar scenarios suchashighway driving. It is important tonotethat if theprimaryor secondary taskrequiredmorethought and reasoning, then performancemight be af-fectedby a visualization, aswasseenin the pilot studyandprior work that considereddocumentediting as theprimarytask[18].

� Inter preting complexvisualizationswithin onesecondin a dual-task scenariocannotbe doneeffectively, butwith relaxed time constraints can be very effective.Despiteprior work thatsuggeststhatthepresenceof cer-tain visual attributescanbe recognized in well underasecond[12], we found that in dual-tasksituationspartic-ipantsperform poorly whenonly shown a visualizationfor onesecond.However, whengiveneightseconds, theycompletedthetasksalmostperfectly. Thelonger durationgivesusersfreedom to choosewhento taskswitch,suchasat timeswhentheir primary tasksituationis momen-tarily stableandrequireslessattention.

� Lower densitydisplayscanresult in performancethatis as good or better than high density displays in adual-task scenario. We foundthis to betruebothwhenparticipants werefinding single items andfinding clus-tersof items. Note that in our experiments,the cluster-basedtaskwasfairly simple:participantsweretold thataclusterof informationexistedandthey merelyneededtoidentify where. However, astasksbecomemorenumer-ousanddifficult, participants aremorelikely to becomedistractedfrom the primary taskaswe saw in our pilotstudy. Thatpresentsauniquechallengeof breaking downa visualizationtaskinto a seriesof sub-tasksthatcanbecompletedindependentlywith low densityvisualizationsand reassembledmentally. This variant of the “chunk-ing” problem, initially studiedby Herb Simon[21], hasimportant ramifications for the domainof peripheral in-formation visualization.

� Finding clustersof visually similar itemsis easierthanlocating a single item. Locatinga clusterof itemsof asinglecolor resultedin morecorrect answersthanlocat-ing a singlecoloredshape.This resultdirectly supportsprior work by Pomerantz thatsuggestswhendealingwith

separable dimensions (suchasshapeandcolor), dividedattentiontaskswould take longer[20]. Recallthat locat-ing thesingleitem involvedbothcolor andshape,whichmakesit adividedattentiontask.

Our work hasfocusedon presentingperipheral informationto people while they arebusy performing someother taskthat requires significantamounts of attention. We focusedon the factorsof visual information densityand presencetime,with informationrepresentation thenext logical factortoconsider. ResearchersincludingClevelandandMackinlayhave experimentallyestablishedvisualorder-of-precedencerulesfor standalonevisualizations[5, 17], but correspondingrulesfor visualizationsin theperipheryareneeded. In addi-tion, if the conceptof informationchunking with low den-sity visualizationsis tosupport increasingquantitiesof infor-mation,thennew low-effort peripheralinteractionstrategieswill needto be explored to enableperipheral informationnavigation with minimal distraction.

A better understandingof the effects of visualizations assecondary displayswill impactthe increasingdevelopmentof desktopinformationmanagementtools. Computer usershave long usedvisualizations like email tools and systemloadmonitorsto keeptrackof informationwhile performingothertasks.As systemslike Letizia [16] provide userswithadditional informationonourdesktopto helpwith browsingandcommunicating, it is becoming increasingly necessaryto identify methods for effectively communicatingthis in-formationwith minimal disturbanceto othertasks.

With furtherstudy, weseethiswork impactingoff-the-desktopsituationsaswell, suchasdisplaysin factoriesandvehicles.In thesesituations,good guidelines for developing visual-izationsassecondary displaysshift from beingbeneficialtointerpret thevisualizationperipherally to beingessentialtodoso.Priorstudieshavelookedat theuseof iconsandothersimplevisualdisplaysin in-vehicle systems[15]. However,astheinformationavailablewhile driving increases,design-ersmustbe readywith safe,effective methods for commu-nicatingit to drivers.

REFERENCES[1] ChristopherAhlberg, ChristopherWill iamson,andBen

Shneiderman.Dynamicqueries for information exploration: Animplementation andevaluation. In Proceedings of the ACMConference on Human Factors in Computing Systems (CHI ’92),pages619–626,Monterey, CA, May 1992.

[2] Lyn Bartram, Colin Ware,andTomCalvert.Moving icons: Detectionanddistraction. In Proceedings of the IFIP TC.13 InternationalConference on Human-Computer Interaction (INTERACT 2001),Tokyo, Japan, July 2001.

[3] StuartK. Card, JockD. Mackinlay, andBenShneiderman.Readingsin Information Visualization. MorganKaufman, 1999.

[4] ChaomeiChenandMary Czerwinski. Empiricalevaluation ofinformatin visualizations: An introduction. International Journal ofHuman-Computer Studies, 35(3),2000.

[5] Willi amS.ClevelandandRobertMcGill . Graphical perception:Theory, experimentation,andapplication to thedevelopmentofgraphical models.Journal of the American Statistical Association,79(387),September1984.

[6] Andrew Csinger. Thepsychology of vision.Technical report,Departmentof Computer Science,University of Brit ish Columbia,1992.

[7] Mary Czerwinski, EdwardCutrell, andEric Horvitz. Instantmessagingandinterruption: Influence of tasktypeon performance.In Proceedings of OzCHI 2000, December2000.

[8] JamesT. EnnsandRonald A. Rensink. Preattentive recovery ofthree-dimensional orientation from line drawings.PsychologicalReview, 98:335–351,1991.

[9] S.H. Fairclough,M. C. Ashby, andA. M. Parkes. In-vehicledisplays,visualworkload, andusability evaluation. In A. G. Gale,I. D. Brown, C. M. Haslegrave,H. W. Kruysse,andS.P. Taylor,editors,Vision in Vehicles, volume4, pages245–254. Amsterdam:North Holland,1993.

[10] K. I. Foster. Visualperception of rapidly presentedwordsequencesofvaryingcomplexity. Perception and Psychophysics, 8:215–221,1970.

[11] Paul Green. The15-second rule for driver information systems.InProceedings of the ITS America Ninth Annual Meeting, pagesCD–ROM, Washington, DC, 1999.

[12] ChristopherG. Healey, KelloggS.Booth, andJamesT. Enns.High-speed visualestimation usingpreattentive processing.ACMTransactions on Human Computer Interaction, 3(2):107–135,1996.

[13] ChristopherG. Healey andJamesT. Enns.Largedatasetsat a glance:Combiningtexturesandcolorsin scientific visualization. IEEETransactions on Visualization and Computer Graphics,5(2):145–167,1999.

[14] JamesF. Juola, NicklasJ.Ward,andTimothy McNamara. Visualsearchandreading of rapidserial presentationsof letter strings,words,andtext. Journal of Experimental Psychology: General,111(2):208–227,1982.

[15] JohnD. Lee,CherCarney, StevenM. Casey, andJohnL. Campbell.In-vehicle displayiconsandother information elements: Preliminaryassessmentof visual symbols.Technical Report FHWA-RD-99-196,UnitedStatesFederal Highway Administration, December 1999.

[16] HenryLieberman.Autonomousinterfaceagents. In Proceedings ofthe ACM Conference on Human Factors in Computing Systems (CHI’97), pages67–74,Atlanta,GA, March1997.

[17] JockD. Mackinlay. Automating thedesignof graphical presentationsof relational information. ACM Transactions on Graphics,5(2):110–141,April 1986.

[18] Paul P. Maglio andChristopherS.Campbell. Tradeoffs in displayingperipheral information. In Proceedings of the ACM Conference onHuman Factors in Computing Systems (CHI 2000), April 2000.

[19] D. ScottMcCrickard,RichardCatrambone,andJohnT. Stasko.Evaluating animationin theperiphery asa mechanismformaintaining awareness.In Proceedings of the IFIP TC.13International Conference on Human-Computer Interaction(INTERACT 2001), pages148–156,Tokyo, Japan,July 2001.

[20] JamesR. Pomerantz. Perceptualorganization in informationprocessing. In Michael Kubovy andJamesR. Pomerantz, editors,Perceptual Organization, chapter 6. LawrenceErlbaumAssociates,1981.

[21] HerbertA. Simon.How big is a chunk?Science, 183:482–488,1974.

[22] JacobSomervell, Ragavan Srinivasan,OmarVasnaik,andKimWoods.Measuring distraction andawarenesscausedby graphicalandtextual displays in theperiphery. In Proceedings of the 39thAnnual ACM Southeast Conference, Athens,GA, March 2001.

[23] RobertSpence. Information Visualization. Addison-Wesley, 2001.

[24] AnneTreismanandStephen Gormican. Featureanalysisin earlyvision: Evidence from searchasymmetries.Psychological Review,95:15–48, 1988.

[25] EdwardR. Tufte.Envisioning Information. GraphicsPress,Cheshire,CT, 1990.