My Phone and Me: Understanding People’s Receptivity to Mobile Notifications AbhinavMehrotra VeljkoPejovic JoVermeulen UniversityofBirmingham UniversityofLjubljana UniversityofCalgary UniversityCollegeLondon Slovenia Canada UnitedKingdom [email protected] UniversityofBirmingham [email protected] UnitedKingdom [email protected] RobertHendley MircoMusolesi UniversityofBirmingham UniversityCollegeLondon UnitedKingdom UnitedKingdom [email protected] [email protected] ABSTRACT varietyofinformationsuchasemailmessages,socialnetwork Notifications are extremely beneficial to users, but they of- events and birthday reminders. Notifications are at the core tendemandtheirattentionatinappropriatemoments. Inthis of this information awareness, as they use audio, visual and paper we present an in-situ study of mobile interruptibility hapticsignalstosteertheuser’sattentiontowardsthenewly- focusing on the effect of cognitive and physical factors on arrivedinformation. the response time and the disruption perceived from a noti- Notificationsareextremelybeneficialtotheusers: however, fication. Throughamixedmethodofautomatedsmartphone at the same time, they are a cause of potential disruption, loggingandexperiencesamplingwecollected10372in-the- since they often require users’ attention at inopportune mo- wild notificationsand474questionnaireresponsesonnotifi- ments. Indeed,previousstudieshavefoundthatinterruptions cationperceptionfrom20users. Wefoundthattheresponse atinopportunemomentscanadverselyaffecttaskcompletion time and the perceived disruption from a notification can be time[11, 12, 25], leadtohightaskerrorrate[8]andimpact influencedbyitspresentation,alerttype,sender-recipientre- theemotionalandaffectivestateoftheuser[5,7].Also,users lationshipaswellasthetype,completionlevelandcomplex- mightgetannoyedwhentheyreceivenotificationspresenting ity of the task in which the user is engaged. We found that informationthatisnotusefulorrelevanttotheminthecurrent even a notification that contains important or useful content context[13]. Atthesametime,studieshaveshownthatusers cancausedisruption. Finally,weobservethesubstantialrole cannotignoretheirsmartphonesforalongtime,becausethey ofthepsychologicaltraitsoftheindividualsontheresponse startfeelingstressedandanxiousaboutmissingimportantin- timeandthedisruptionperceivedfromanotification. formationuntiltheyfinallypickupthephonetocheckforany newnotifications[26].Thistensionisexacerbatedbythefact AuthorKeywords that individuals deal with hundreds of notifications in a day, MobileSensing;Notifications,Interruptibility, someofwhicharedisruptive[23] Context-awareComputing. Previous studies have investigated the user’s receptivity to ACMClassificationKeywords mobilenotifications[15,29,32]. AsdefinedbyFischer[15], H.1.2.ModelsandPrinciples:User/MachineSystems;H.5.2. receptivity encompasses a user’s reaction to an interruption InformationInterfacesandPresentation(e.g. HCI):UserIn- and their subjective experience of it. For instance, users terfaces might quickly respond to a notification when they are idle, but they can still get annoyed because of the content of the INTRODUCTION notification. Previous studies have shown that the user’s re- Smartphones enable a new form of effortless information ceptivitytoanotificationisdeterminedby: (i)howinterest- awareness. Throughouttheday,asmartphoneuserreceivesa ing,entertaining,relevantandactionableitscontentisforthe user [15]; (ii) the type of application that triggers it – com- munication applications are considered as the most impor- Permissiontomakedigitalorhardcopiesofallorpartofthisworkforpersonalor tant[32];(iii)timecriticalityandsocialpressure[29]. Atthe classroomuseisgrantedwithoutfeeprovidedthatcopiesarenotmadeordistributed forprofitorcommercialadvantageandthatcopiesbearthisnoticeandthefullcita- sametime,somestudieshaveproposedvariousmechanisms tiononthefirstpage. Copyrightsforcomponentsofthisworkownedbyothersthan to infer opportune moments, i.e., moments in which a user ACMmustbehonored.Abstractingwithcreditispermitted.Tocopyotherwise,orre- quicklyand/orfavorablyreactstoanotification[14,23,28]. publish,topostonserversortoredistributetolists,requirespriorspecificpermission and/[email protected]. In order to infer interruptibility these studies have used ma- CHI’16,May07-12,2016,SanJose,CA,USA. chine learning classifiers provided with different contextual ©2016ACM.ISBN978-1-4503-3362-7/16/05...$15.00. DOI:http://dx.doi.org/10.1145/2858036.2858566 factors including user’s transitions between activities [18], engagementwithamobiledevice[14],timeofday,location andactivity[28]aswellasnotificationcontent[23]. However, none of these studies have deployed the proposed mechanisms in a real world scenario with in-the-wild notifi- (a) (b) cations of a regularly used application. The key reason be- hindthisisthefactthattheaccuracyofthesemechanismsis still lower than the user’s expectations. In a real world sce- nario,theuserswouldnotacceptasystemthatmightdeferor stop an important notification. Previous studies have shown thatusersarewillingtotoleratesomeinterruption,inorderto notmissanyimportantinformation[20]. Webelievethatin- terruptibilitymanagementsystemsfailtoachieveaveryhigh accuracyinpredictingtheopportunemomentbecausethereis stillalackofunderstandingconcerningthefactorsinfluenc- ing the user’s receptivity to mobile notifications in different physicalandcognitivesituations. (c) (d) In order to bridge this gap, in this work we conduct an in- situ study to collect objective and subjective data about mo- bilenotifications. WedesignedanddevelopedMyPhoneand Me (Figure 1), an application that uses a novel experience samplingmethod(ESM)approachtouncoverthefactorsand Figure1. MyPhoneandMeapplication: (a)mainscreen, (b)phone motivations impacting the user’s reaction and sentiment to- usagestatistics,(c)applicationusagestatistics,(d)dailynotifications. wards a notification. Through My Phone and Me, we col- lected10372notifications,474responsesfortheESMques- tionnairesand11personalitytestresultsfrom20users.Using userstendtoclickhighlydisruptivenotificationsiftheycon- this data, we investigate users’ interaction with mobile noti- tainvaluableinformation. Whileusersareawareofnotifica- fications in different physical and cognitive contexts. More tions even when their phone is in silent mode, our analysis specifically,thekeycontributionsofthisworkaretheinves- showsthatthealertmodalityhasasignificantimpactonthe tigationof: time taken by the users to view the notification. Finally, we observe the substantial role of psychological traits on how a the impact of a notification’s alert modality on the user’s personreactstoamobilenotification,callingforhighlyper- • abilitytoperceiveanotificationalert; sonalizedinteractionbetweenasmartphoneanditsuser. theimpactofthealertmodality, sender-recipientrelation- • ship, presentationofanotification, theongoingtasktype, REASONING ABOUT USERS’ RECEPTIVITY TO MOBILE completionlevelandtaskcomplexityontheresponsetime; NOTIFICATIONS theimpactofthesender-recipientrelationship,andtheon- An interruption tries to steer a user from an ongoing task • goingtask’stype,completionlevelandcomplexityonthe to the secondary task signaled by it [8]. As suggested by perceiveddisruption; Clark [10], users can respond to an interruption in four pos- the role of the sender-recipient relationship, notification sibleways: (i)handleitimmediately;(ii)acknowledgeitand • contentandtheperceiveddisruptionontheuser’sdecision agree to handle it later; (iii) decline it (explicitly refusing to toacceptordismissanotification; handleit);(iv)withdrawit(implicitlyrefusingtohandleit). the impact of the user’s personality on the perceived dis- A user can respond to mobile notifications in a fairly • ruptionandresponsetimetoanotification. different way as compared to an in-person interruption. For communication-related interruptions, for example, users The findings of our study are wide-ranging, and may have might perceive more disruption from an in-person interrup- a direct impact on the way future notification management tionthanfromamobilenotificationsbecauseofthepresence mechanismsareconstructed. First,weobservethatasender- ofaninterrupterintheformercase. Mobilenotificationsen- recipient relationship, notification priority and an ongoing ableflexibilityinthewayaninterruptionishandledbecause task’s type and complexity influence the response time for ofthelackofthephysicalpresenceofthesenderandtheasyn- the notification, but there is no impact of the ongoing task chronousnatureofmobilemessagingcommunication1.Thus, completion level on the response time. Moreover, our re- theexactmomentofhandlinganinterruptioncanbenegoti- sultsshowthattherecipient’srelationshipwiththesenderof atedandtherecipientcandecidewhenandhowtoattendtoa a notification, the ongoing task’s type, completion level and notification. complexity influence the perceived disruption. Our findings implythatthehigherthelevelofdisruptionperceivedbythe 1Certainsocialnormsandexpectationsfromthesenderside,how- user the higher the probability of the notification being dis- ever, constrain the flexibility that the receiver has in reacting to a missed. Fromourresults,wealsoobservethat,nevertheless, message[30]. (c1) Seen Decision Time Time (c2) (a) (b) Figure2. ThethreetimemeasurementsofanotificationcapturedbytheMyPhoneandMeapplication. Thetimeofnotificationarrival(a),thetime whenanotificationisseen(b),andthetimewhentheuseraccepted(c1)ordismissed(c2)anotification.Thetimedifferencebetween(a)and(b)isseen timeandthetimedifferencebetween(b)and(c1orc2)isthedecisiontime. However,thisflexibilityintroducesmanyotherissues. First, notificationsinthenotificationbarareseenwhentheuserun- notifications can go unnoticed when a user does not register locksthephone. Incaseanotificationarriveswhentheuser analert. Second,usuallynon-persistentnotificationsmaybe is already using the phone (i.e., the phone is unlocked), the forgottenabout–auserridingabicycle,mightdecidetoat- seentimeofthisnotificationwouldbecomputedaszero. We tend to a notification once they arrive at the destination, yet term the time from the notification arrival until the moment forgettodoso. Finally,althoughdesignedtosignalaninter- thenotificationwasreacteduponastheresponsetimeforthe ruptionbutnotinterruptthemselves,mobilenotificationscan notification.Forouranalysis,webreaktheresponsetimeinto still induce unnecessary disruption to a user’s routine. For twointervals: instance,thedisruptioncanhappenwhenauserdecidestoat- Seentime(ST)–timefromthenotificationarrivaluntilthe tendtoanotificationimmediately,despitebeinginthemiddle • timethenotificationwasseenbytheuser. of another task, only to find that the notification is about an unrelatedpromotionaloffer.Moreover,adisruptionmayhap- Decisiontime(DT)–timefromthemomentausersawa • penevenifanotificationisnotattendedto,asthethoughtof notification until the time they acted upon it (by clicking, a lingering notification may interfere with the user’s current launchingitscorrespondingapporswipingtodismiss). taskperformance[33]. Weexaminethewayinterruptiontiming, withrespecttothe Inthisstudy,weinvestigatethefactorsinfluencingtheuser’s primary task, determines the user’s response to the notifica- response to a mobile notification, where the response is de- tion. Moreover,weareinterestedinthewayuserstriagedis- finedbythetimetakentoregisterandreacttoanotification, ruptive notifications. Can users quickly discern when noti- andthewayinwhichthenotificationishandled(i.e.,clicked fications are disruptive? We hypothesize that humans might ordismissed).Moreover,weinvestigatetheuser’smotivation still attend to a notification, even if they know that the pri- for being self-disruptive by clicking the disruptive notifica- mary task is going to be disrupted. For example, in their tions. studyofWhatsAppnotifications,Pielotetal. [30]showhow, due to an inner pressure raised by social expectations, users Ourassumptionisthattheresponsetimeforanotificationand quicklyrespondtoinstantmessaging(IM)communicationor thedisruptionperceivedbytheuserareinfluencedbythedif- frequentlychecktheirphones,inducingselfinterruptionsjust ferentaspectsofthenotificationaswellastheuser’scontext. inordertosatisfythesocialexpectations.Inourwork,weare TocapturethismeasurewedevelopedanAndroidexperience lookingbeyondjustIMnotificationsandinvestigatetheway samplingmethod(ESM)applicationthatmonitorstheactual anydisruptivemessageishandled. ThroughourESMstudy notificationsusersreceiveontheirphone,recordstheirreac- weidentifythemotivationbehindreactingtoadisruptiveno- tiontonotificationsandthenqueriestheuserstoidentifytheir tificationandthereasoningandtheexternalfactorsthatlead rating of the disruption caused by the notification. We aug- totheexactreaction. mentthiswithquestionsaboutthemotivationforhandlinga notification in a particular way. Further, our ESM question- We aim for a comprehensive investigation of interruptibility naires ask the user to provide data on the type, complexity fromauserperspective,thuscomparingtheeffectofdifferent and the completion level of the ongoing task and the user’s aspects of a notification on its response time and disruptive- relationshipwiththesender. Finally,wecollectparticipants’ ness. Finally, we investigate the potential role of individual personalitytraitmeasuresattheendoftheexperiment. psychological traits on how users perceive and react to dis- ruptivenotifications. First, we investigate the ability of users to adjust their re- sponsetimestoanotification, andseehowquicklytheycan triagedifferentnotificationsindifferentsituations. Asshown DATACOLLECTION inFigure2,wetakethreetimemeasurementsforeachnotifi- In order to investigate the nature of disruptive notifications cation: thetimeofnotificationarrival(a),thetimewhenthe andfactorsthatdeterminetheuser’sreceptivitytomobileno- notificationisseen(b), andthetimewhentheuseraccepted tifications in different physical and cognitive situations, we (c1) or dismissed (c2) the notification. Note that in order to conductedanin-situfieldstudy. Morespecifically,wedevel- detectthemomentatwhichanotificationisseen,weusethe opedanAndroidappcalledMyPhoneandMe–anAndroid unlockeventofthephoneandassumethatallnewlyavailable experiencesamplingmethod(ESM)applicationthatcollects informationaboutin-the-wildnotifications,users’interaction Group Features Question Options Time Arrival,seenandtheremovaltimeofanotification. Did you notice the alert (i)Yes,andIdecidedtocheckmyphoneim- Notification (e.g., vibration, sound, Whetherthenotificationwasclickedordismissed. mediately. (ii) Yes, but I was already using response flashing LED) for this myphone. (iii)Yes, butIignoredthealert. Notification notification when it first Senderapplicationandthetitleofanotification. (iv)No,Ididn’tnoticethealert. details arrived? Signalsusedbyanotificationtoalerttheuser: sound, vi- (i)Idecidedtoimmediatelyclickit.(ii)Ide- Alerttype brate,andflashingLED. cided to dismiss it because it didn’t require How did you handle the Physicalactivity, location, presenceofsurroundingsound, anyfurtheraction. (iii)Idecidedtodismiss notificationwhenyoufirst Context WiFiconnectivity,proximitytothephone,surroundinglight itbecauseitwasnotrelevantoruseful. (iv)I sawit? data intensity. Thisdataiscollectedonarrivalandremovalofa decidedtoreturntoitlater.(v)Other(descrip- notificationfromthenotificationbar. tive). Table1.DescriptionoffeaturesfromtheMyPhoneandMedataset. (i)Thesenderisimportant.(ii)Thecontentis important.(iii)Thecontentisurgent.(iv)The Select all factors that content is useful. (v) I was waiting for this made you decide to notification.(vi)Theactiondemandedbythe click/dismiss the notifica- senderdoesnotrequirealotofeffort.(vii)At with them in natural situations (while they are performing tion. thismoment,Iwasfree.(viii)Other(descrip- their day-to-day activities), and the physical and cognitive tive). contextdetails. (i) Partner (ii) Immediate family (chil- dren, parents) (iii) Extended family The My Phone and Me application uses Android’s Notifica- (nieces/nephews, cousins, aunts/uncles) tionListenerService[1]toaccessnotificationsandGoogle’s What best describes your (iv) Friend (v) Acquaintance (vi) Superior ActivityRecognitionAPI [3]andESSensorManager[22]to relationshiptothesender? at work (vii) Colleague (viii) Subordinate at work (ix) Client (x) Service provider obtain the context information. Table 1 lists the groups of (xi) Sender is not a person (xii) Other features captured by the application. The collected context relationship(descriptive). data has not been explored for the analysis presented in this Please describe what the Descriptiveresponse. paper. To infer the user’s response to a notification, the My notificationwasabout. PhoneandMeapplicationcheckswhethertheapplicationthat Pleasedescribewhatactiv- ityyouwereinvolvedwith triggeredthenotificationwaslaunchedaftertheremovaltime Descriptiveresponse. whenyoureceivedtheno- ofthatnotification. Weareawarethatsomenotificationsare tification. dismissedbecausetheydonotrequireanyfurtheraction. For (i) Starting a new task/activity. (ii) In the When the notification ar- this reason, we capture seen time and use the difference be- middle of a task/activity. (iii) Finishing a rived,Iwas: tweenseentimeandremovaltimetounderstandhowlongit task/activity.(iv)Notdoinganything. Thetask/activityIwasdo- takesfortheusertoreadandreacttoanotification. Five-level Likert scale rating between ing when the notification "stronglydisagree"and"stronglyagree". Tocollectsubjectivedatafromusers,theMyPhoneandMe arrivedwascomplex. I found the notification Five-level Likert scale rating between applicationtriggersfourquestionnairesinaday. Aquestion- disruptive. "stronglydisagree"and"stronglyagree". naireistriggeredonlywhenanotificationishandled;itcon- Table2.Questionsandtheiroptionsfromquestionnairetriggeredbythe tainsquestionsaboutwhythenotificationwasclickedordis- MyPhoneandMeapp. missed by presenting a screenshot of that notification. The application triggers a questionnaire for a randomly selected notification in every four hours time window between 8.00 veloped by Goldberg [16]. A notification to take this test is amand8.00pmandthelastquestionnaireatarandomtime triggered once the user has responded to 28 questionnaires. between8.00pmand10.00pm. Theapplicationdidnottrig- A user can also take the test at any time by clicking on the ger any questionnaire after 10pm so that the participants do personalitytestbuttonpresentintheapplication’sactionbar. not feel annoyed at responding to the surveys late at night. The application automatically used the local time zones be- cause it relies on the phone’s time. Moreover, if the user is busy,thequestionnairecanbedismissedbysimplyswipingit RecruitmentoftheParticipants fromthenotificationbarandnoquestionnaireisshowntothe The My Phone and Me application was published on the userforthenext30minutes. Google Play Store from 12th August 2015. It was installed by74participantswithoutanymonetaryincentive. Asshown A questionnaire comprises seven multiple-choice and two inFigure1,MyPhoneandMetellstheusersabouttheirad- free-response questions. The list of questions and their op- diction to the phone. It allows users to check statistics on tions are shown in Table 2. Since we ask the users to enter their phone usage and interruptions. The application visual- thefreeformtextfortwoquestions,itcouldincreasetimeto izesauser’sphoneactivitiesbasedondifferentcriteria,such respond to a questionnaire and may become a source of an- astheirhourlyphoneusage(Figure1C),hourlyusageofin- noyance. Therefore, the application allows the users to dic- dividualapplications(Figure1D)andhowmuchtheyinter- tatetheresponsestothesequestions. Theseanswersarethen actwithnotifications(Figure1B).Webelievethatdisplaying convertedtotextusingAndroid’sSpeechRecognizerAPI[2]. this information has a minimal interference with users’ ac- Additionally,theMyPhoneandMeapplicationaskstheusers tual behavior for interacting with notifications, but provides totakeapersonalitytestbasedonthe50itemBig-FiveFac- a valuable functionality in order to make the users keep the torMarkersfromtheInternationalPersonalityItemPool,de- applicationinstalledontheirphones. Inordertoensureprivacycompliance,theMyPhoneandMe application goes through a two-level user agreement to ac- Thekeyfindingsofthissectionare: cess the user’s notifications. Firstly, the user has to give ex- Users are aware of the notification alerts even when plicit permission as required by the Android operating sys- • the phone is in silent mode. However, seen time is tem. Secondly, the application shows a list of information fastestwhenthephoneisinvibratemodeandslowest that is collected and asks for user consent. Moreover, we forsilentnode. show the original content of a notification to the user along Notificationsareseenfastestwhentheuseriscommut- withthequestionnaireinordertoavoidanyrecallbiasinthe • ingandslowestwhenidle. databutwedonotcollectthenotificationcontentforprivacy reasons. User’sattentivenessincreases(reducingtheseentime) • withtheincreaseinthecomplexityofanongoingtask. DATASET The decision time is higher for the notifications from The data collection was carried out for around two months, • lessfrequentlycontactedsenders. duringwhichwecollected19494notifications,611responses High-prioritynotificationsgetquickerresponse. forthequestionnaire(comprisingasetofninequestionslisted • inTable2)and11personalitytestresults(50item-basedBig- Five Factor Markers by Goldberg [16]) from 74 users who The Role of Alert Modality in Perceiving a Notification installed the My Phone and Me application. Many users stoppedrespondingtothequestionnairesafterafewdaysand Alert some did not respond at all. Therefore, we select a subset Anotificationcanalerttheuserbymeansofvibration,sound of the data for the analysis and include data of users who and/or flashing LED. In order to investigate how users per- responded to at least 14 questionnaires. There are 20 users ceive alerts with different alert modalities, we used the re- whosatisfiedthisconstraint. So, ourfinaldatasetcomprises sponsesprovidedbytheusersforQ1(Didyounoticethealert of 10372 notifications, 474 questionnaire responses and 11 (e.g., vibration, sound, flashing LED) for this notification personalitytestresults. Additionally, duringthesetupphase whenitfirstarrived?).Accordingtoourdataset,whentheno- we asked participants to enter their age and gender: in our tifications (with which the questionnaires were linked) were datasetthereare11maleand9femalesagedbetween19and triggeredtheuser’sphonewasfor25.54%ofthetimesinthe 50 years old. However, we do not ask them to provide any silentmode,21.50%vibratemode,41.94%soundmodeand otherdemographicinformation. 11.03%soundwithvibrate. As we are primarily using the questionnaire responses, we Usersreportedthattheymissednotificationalertsfor14.63%, comparedtheclickrate(i.e., percentageofnotificationsthat 15.38%,23.75%,21.05%oftimestheirphonewasinsilent, are clicked out of total notifications) of the overall notifica- vibrate, sound, and sound with vibrate mode respectively. tionswiththenotificationsthatwerelinkedtoquestionnaires. Thisprovidesevidencethatwhenthephoneisinsilentmode Theclickrateforoverallnotificationsis62.52%,andforno- usersarestillawareofthenotificationalerts. tifications linked with questionnaires is 70.04%. Note that a notification is considered to be clicked either when it is WhatFactorsInfluencetheSeenTime? clickedonthenotificationbarorwhenitscorrespondingap- Weinvestigatetheroleofalertmodality,senderandtheongo- plicationislauncheddirectly. ingtasktype,complexityandcompletionlevel,ininfluencing theseentimeofanotification. We are aware that our dataset has potential limitations that stem from the inherent nature of an in-the-wild study. The TheImpactofAlertModalityonSeenTime dataset remains unbalanced because it is not possible to ob- In order to perform this analysis, from our dataset of 10372 tain an equal number of questionnaire responses from all notificationsweuseallthenotificationsofthe20userswho users for each test category. For instance, there is a very respondedtoatleast14questionnaires. However,weignore smallchancethatourapplicationtriggersaquestionnairefor the notifications that arrived when the user was already en- eachtypeofsenderfromtherecipientssocialcircle. Further- gagedwiththephonebecausewecouldnotcalculatetheseen more,inpractice,ausermightnotevenreceivenotifications timeofthesenotifications. Thisleavesuswith4929notifica- fromeachofthesendertypesduringtheperiodofthestudy. tions. Aone-wayAnalysisofVariance(ANOVA)oftheseen Therefore, we use the data from 20 users who responded to timewascarriedoutforeachalertmodality.Theresultsshow atleast14questionnaires,i.e.,theminimumnumberofques- thatthealertmodalityhasanimpactontheseentimeofnoti- tionnaires that were answered by users in this set. We also fications,withF(3,4925)=26.41, p<0.001. ATukeypost- makeanhypothesisofdataindependence(i.e.,datainstances hoctest(bysettingthe↵=0.05)revealedthattheseentime are obtained from different users) which might not hold in isstatisticallysignificantlyhigherforsilentnotifications(av- realityandcanbetestedonlyinacontrolledsetting. erage 7.332 mins). The seen time for the notifications alert- ing withvibrate onlymode is thelowest (average 3minutes UNDERSTANDINGRESPONSETIME and21second). Soundonlyandsoundwithvibratenotifica- Inthissectionweinvestigatetheeffectofdifferentfactorson tionsarethesecond(average5minutesand57seconds)and theseenanddecisiontimeofanotification. third (average 4 minutes and 50 seconds) most quickly seen byusers. Quiteinterestingly,arecent15-userstudybyPielot etal.[29]alsofoundthatnotificationstendtobeseenfaster WhatFactorsInfluencetheDecisionTime? whenthephoneisinthevibratemode. Here,weconfirmthis Weanalyzetheeffectofthetype,complexityandcompletion finding, but also point to the above missed notification per- level of the ongoing task, and the sender type on the time a centageinthesilentmode(14.63%)andshowthatsettingthe usertakestodecidehowtoreacttoanotification.Wefindthat phonetosilentdoesnothelpinescapinginterruptions. neither of these factors have a statistically significant effect onthedecisiontimeofnotificationswiththeexceptionofthe TheImpactofOngoingTaskTypeonSeenTime sender. To investigate the impact of the ongoing task on the no- A one-way repeated measures ANOVA of the decision time tification’s seen time, we require the type of task that wascarriedoutforeachsendertype.Theresultsshowthatthe the users were involved with when the notification arrived. sender type has an impact on the seen time of notifications We classified the information users provided through ESM with F(10,212) = 2.429, p = 0.00936. A Tukey post-hoc questionnaires about the ongoing task into the following test(bysettingthe↵=0.05)revealedthatoutofthe11sender six categories: work, communication, traveling, mainte- types(showninTable2),notificationsfrompartnerleadtothe nance/personal,leisureandidle. Theclassificationwasdone fastestdecisiontime(meanDTis3.315s),followedbyimme- manually, by two coders who initially disagreed on five en- diatefamilymemberswithanaveragedecisiontimeof4.891 tries.Twocommonlabelsforthesewerefoundafteradiscus- seconds. Ontheotherhand,notificationsfromextendedfam- sionwith threeother coders. Note thatourapp allowsusers ily members and service providers have the largest decision toskipthestepofprovidingtheinformationonthequestion time, 11.93 and 8.146 seconds respectively. There was no about their current task by selecting the “Prefer not to say" statistically significant difference in the decision time of the option. Insuchcases,wediscardtheentryfromouranalysis notifications from other senders. These results demonstrate oftheeffectoftheongoingtaskoninterruptibility. that notifications are quickly handled when they are sent by A one-way ANOVA of the seen time is carried out for each thecloserelativesoftheuser. Inothercasesuserstakemore task type. The results show that the ongoing task type has time in reading the content before deciding how to handle animpactontheseentimeofnotifications,withF(5,217) = it. Wehypothesizethatthisbehaviorstemsfromthecontent 2.963, p = 0.013. ATukeypost-hoctest(bysettingthe↵as ofnotificationsfromclosefriendsorfamilymembers,which 0.05) reveals that the seen time is the lowest when the noti- mightbemorepredictable,andapartofadailyroutine(e.g., ficationsarrivewhiletheuseriscommunicating(average47 “pickkidsfromschool").Ontheotherhand,theusershaveto seconds)andhighestwhiletheuserisidle(average9minutes spendmoretimeonthenotificationsfromlessfrequentlycon- and30seconds). Othertasktypesdonothaveastatistically tactedsources,asthecontentmaybelessfamiliartothem. significant effect on the seen time of notifications and have anaverageseentimeof5minutes45seconds. Asshownin arecentstudy[31]notificationsaremorewelcomewhenre- TheRoleofNotificationPresentation cipientsarebored. However,ourresultsshowthatwhilethe Inourdataset,2953(outof7795)notificationswerereceived usersmightbewillingtoacceptmorenotificationswhenidle, whentheuserwasengagedwiththephone.Outofthese2953 thetimeneededtoattendtosuchnotificationsmightbehigher notifications,860areso-called“low-priority"while2093are comparedtothetimeneededtoattendtoanotificationwhile “high-priority”notifications[4]. Here,ahigh-prioritynotifi- auserisbusy. cationisaforegroundnotificationthatgetsinthewayofthe user’s ongoing activity and the user cannot perform any ac- tiontogetitoutofthewaywithoutclickingordismissingit TheImpactofOngoingTaskComplexityonSeenTime To analyze the effect of ongoing task complexity, we first (e.g. Vibermessages). Alow-priorityonesimplyappearson encode the reported task complexity, which was reported the notification bar without getting in the way of the user’s as a value on the Likert scale (Strongly disagree=1, Some- ongoingactivity(e.g. Gmailnotifications). whatdisagree=2,Neutral=3,Somewhatagree=4andStrongly We investigate the effect of the notification presentation on agree=5)tothequestion“Thetask/activityIwasdoingwhen the response time (i.e., the sum of seen time and decision thenotificationarrivedwascomplex". TheSpearman’srank time)ofanotification.Theresultofatwosamplet-testshows correlation coefficient is computed to evaluate the relation- thatthereisastatisticallysignificanteffectofnotificationpri- shipbetweenthecomplexityofanongoingtaskandtheseen ority on the response time, t(2951) = 17.694, p < 0.001, timeofanotification. Theresultsshowthatthereisaweak, withhigh-prioritynotificationsgettingquickerresponsethan negative correlation between the two variables, ⇢ = 0.183, low-prioritynotifications. Themeanresponsetimeforhigh- � p = 0.005. Thus, the increase in the seen time of notifi- prioritynotificationsis11.94sversus25.91sforlow-priority cations is correlated with the decrease in rating of ongoing notifications. task’scomplexity. Webelievethatthiscorrelationexistsbe- cause the users become more alert while performing a com- plextaskandthus,quicklyperceivetheinterruptions. Onthe WHYANOTIFICATIONBECOMESDISRUPTIVE other hand, when the users are not performing any complex In this section we investigate the effect of different factors task,theybecomelessattentivetotheinterruptions. Finally, on the perceived disruption. Since the perceived disruption wefoundthatfactorssuchasthecompletionleveloftheon- wasmeasuredwitha5-pointLikertscale,weencodethere- goingtaskandthesendertypedonothaveastatisticallysig- sponsesas:Stronglydisagree=1,Somewhatdisagree=2,Neu- nificanteffectontheseentimeofnotifications. tral=3,Somewhatagree=4andStronglyagree=5. comingfromtheirfamilyandfriends[15],andthatthemore Thekeyfindingsofthissectionare: “distant" the sender is, the less likely it is that a notification willbeclickedon[23]. Resultsfromourstudycomplement Perceiveddisruptionincreaseswiththeincreaseinthe • thiswiththefindingthattheperceiveddisruptionvarieswith complexityofanongoingtask. thesenderofanotification. Notifications are perceived as most disruptive if they • arrivewhentheuserisinthemiddleoforfinishinga TheRoleofOngoingTaskType task,andleastdisruptiveiftheuserisidleorstartinga Aone-wayANOVAofthereporteddisruptioniscarriedout newtask. for each type of ongoing task (see Table 2). The results Messages from subordinates and system messages show that the type of task that the user is engaged with on • (where the sender is not a person) are considered as the arrival of a notification has a significant impact on the most disruptive. Whereas, extended family members disruption the user perceives when the notification arrives, areconsideredastheleastdisruptive. F(5,380) = 13.03, p < 0.001. A Tukey post-hoc test (by setting↵=0.05)revealedthattheperceiveddisruptionisthe highest when the user is working and the lowest while the TheRoleofOngoingTaskComplexity user is idle. After work, traveling and then leisure are the We investigate whether the complexity of an ongoing task tasks where the users perceive the highest level of disrup- is associated with the perceived disruption reported by the tion. When the users are not idle, they perceive least dis- users. AKendall’sTaucorrelationcoefficientwascomputed ruptionwhilecommunicatinganddoingapersonalormain- toassesstherelationshipbetweentheongoingtaskcomplex- tenancetask. Sincethecommunicationcaninvolvenotifica- ityandperceiveddisruption. Wefoundastrong,positivecor- tions themselves, e.g. two mobile users exchanging What- relation between the two variables, R⌧ = 0.477, p < 0.001. sAppmessages,theaboveresultisnotsurprising. Asshown This demonstrates that the users are likely to get more dis- inarecentstudy[31],usersarereceptivetoinformationwhen rupted by a notification that arrives when they are engaged theyarebored. Ourresultsareinlinewiththesefindingsin inanintricatetaskandlessdisruptedwhentheyareperform- showingthatperceiveddisruptionislowestwhentheuseris ing a simple task. In our preliminary analysis [27] we have idle. foundthatwhenusersareengagedincomplextaskstheyalso expressmoreofanegativesentimenttowardsinterruptions. UNDERSTANDING THE ACCEPTANCE OF NOTIFICA- TIONS TheRoleofOngoingTaskCompletionLevel Inthissectionweinvestigatethefactorsthatmaketheusers Aone-wayANOVAofthereporteddisruptionwascarriedout accept(click)ordismissanotification. foreachclassoftaskcompletionlevel(starting,inthemiddle, finishingandnotdoinganything). Theresultsshowthatthe completionlevelofanongoingtaskhasasignificantimpact Thekeyfindingsofthissectionare: on the disruption perceived by the users from the notifica- Likelihood of the acceptance of a notification de- tions, F(3,451) = 19.43, p < 0.001. A Tukey post-hoc test • creaseswiththeincreaseintheperceiveddisruption. (bysetting↵as0.05)revealsthattheperceiveddisruptionis thehighestwhentheuseriscurrentlyinvolvedinatask. The Disruptivenotificationsareacceptedbecausetheycon- • perceiveddisruptionisthelowestwhentheuserisstartinga tainusefulinformation. taskoridleandthereisnostatisticallysignificantdifference betweenthesegroups. Theseresultsshowthattheperceived Procedure disruption when the user is highly engaged in a task is very Throughthequestionnaires,weaskedtheusersthereasonfor highnotonlyfromthedesktopnotification, asdiscussedfor clicking/dismissinganotification(seeTable2). Ifanotifica- examplein[13,24],butalsofromthemobilenotifications. tion(linkedwiththequestionnaire)isclickedbytheuser,we askthemtoselectallfactorsthatmadethemdecidetoclick TheRoleofSender the notification, otherwise, we ask them to select the factors Wecomputeaone-wayANOVAofthereporteddisruptionfor that made them decide to dismiss the notification. We pro- each type of sender (see Table 2). According to the results, vide a predefined list of seven and six options for clicking F(10,444) = 3.987, p < 0.001,thetypeofsenderhasasig- (seeTable3)anddismissing(seeTable4)thenotification. In nificantimpactonthedisruptionperceivedbytheusersfrom addition,thereisaboxforopen-endedanswersincaseusers the notifications. A Tukey post-hoc test (by setting the ↵ as donotfindanappropriateanswerintheprovidedlist. 0.05)revealsthattheperceiveddisruptionishighestwhenthe sender is not a person or is a subordinate at work (no statis- In Table 3 and Table 4 we calculate the percentage of times tically significant difference between these two groups) and eachfactorwasreportedasareasonforclickinganddismiss- the lowest when the sender is an extended family member. ing the notifications. Since, users may select more than one Moreover, colleagues and service providers are the second option,thetotalcountpercentageinthetableaddsuptomore mostdisruptivesendergroups. Thereisnosignificantdiffer- than 100%. According to these responses, the users mostly encebetweentheothergroups. Previousstudiesshowedthat accept notifications when they are free, but also the impor- usersexpressanegativesentiment towardsthemessagesnot tance of the sender and the usefulness of the content make Option Count(%) Option Count(%) Senderisimportant 31.546 Senderisimportant 25.926 Thecontentisimportant 27.129 Thecontentisimportant 33.333 Thecontentisurgent 14.511 Thecontentisurgent 20.370 Thecontentisuseful 31.546 Thecontentisuseful 35.185 Iwaswaitingforthisnotification 15.773 Iwaswaitingforthisnotification 11.111 Theactiondemandedbythesenderdoesnotrequire Theactiondemandedbythesenderdoesnotrequire 20.189 16.667 alotofeffort alotofeffort Atthismoment,Iwasfree 37.224 Atthismoment,Iwasfree 18.519 Table3.Userresponseaboutwhytheyaccept(click)notifications. Table5.Userresponseaboutwhytheyacceptdisruptivenotifications. Option Count(%) Estimated Variable Std.Error tvalue pvalue Senderisnotimportant 19.565 Coefficient Thecontentisnotimportant 40.580 Extroversion 0.017481 0.005694 3.070 0.0278* Thecontentisnoturgent 43.478 Agreeableness -0.012833 0.005387 -2.382 0.0630 Thecontentisnotuseful 38.406 Conscientiousness 0.005942 0.004420 1.344 0.2366 Theactiondemandedbythesenderdoesrequirea Neuroticism 0.008659 0.004042 2.142 0.0851 3.623 lotofeffort Openness -0.003114 0.005369 -0.580 0.5870 Atthismoment,Iwasbusy 19.565 N=11 Table4.Userresponseaboutwhytheydismissnotifications. R2=0.737 F(5,5)=2.802(p=0.01413) Table6. Resultsoflinearregressionwiththeaveragedisruptionasa themacceptanotification. Ontheotherhand,usersavoidat- dependentvariableandthepersonalitytraitsasindependentvariables. tendingtonotificationsthatdonotcontainimportant,urgent orusefulcontent. Theseresponsesdemonstratethatthevalue showsthat104outof474notifications(withwhichtheques- ofcontentisusedfordecidingwhethertoclickordismissa tionnaires were linked) were reported as disruptive. These notification. Moreover, the users very rarely state that they arethenotificationsforwhichtheusersomewhatandstrongly werebusyandthushadtodismissanotification. Thiscould agreedthattheyperceiveddisruptionfromthesenotifications. indicate that the users give precedence to a notification over theprimarytask,butonlyifthecontentisvaluable. However,54%ofthesedisruptivenotificationswereaccepted (clicked)bytheusers,regardlessofthefactthattheycaused DisruptiveNotificationsareLikelytobeDismissed disruption. To investigate the reason for this, we checked We examine the impact of the disruption caused by the no- users’responsesaboutthefactorsthatmadethemclickthese tifications on their likelihood of being accepted. In order to notifications. Table5showsthepercentageoftimeseachfac- quantifythis,weencodedtheresponseforperceiveddisrup- torwasreportedbytheusersforacceptingthedisruptiveno- tion with the following values: Strongly disagree=1, Some- tifications. Asuserswereallowedtoselectmorethanoneop- whatdisagree=2,Neutral=3,Somewhatagree=4andStrongly tion,thesumofthepercentagesinthetableaddsuptomore agree=5. In order to detect the acceptance of a notification, than100. "Contentisimportant"and"Contentisuseful"are we check whether it was clicked by the user. In case it was themostdominantreasonprovidedbytheusersforclicking dismissed,wecross-validatetheuser’sresponsefortheques- thedisruptivenotifications. Thistellsusthateventhenotifi- tion How did you handle the notification? If the user re- cationscontainingimportantorusefulcontentcancausedis- spondedthatIdecidedtodismissitbecauseitdidn’trequire ruption.Wesuspectthatthesenotificationsmaycontainvalu- anyfurtheraction,wemarkthisnotificationasaccepted. Fi- ableinformation,buttheywerenotrelevantatthemomentof nally, we use 0 to indicate that the notification is dismissed delivery. However, our study remains limited to make such and1foranacceptednotification. conclusions and provide an understanding about why users perceivedisruptionfromusefulnotifications. We fit a logistic regression model to estimate the effect of perceived disruption on the likelihood of the acceptance of notifications. ThemodelwasstatisticallysignificantX2(1) = DOESPERSONALITYMATTER In this section we investigate the role of personality on the 48.3, p < 0.001. The results indicate the likelihood of the reported disruption, seen time and decision time of notifica- acceptance of a notification decreases by 0.581 times (95% tions. We computed the score for the five personality traits confidence interval limits for the slope were [0.497, 0.675]) (i.e., the so-called Big Five: Extroversion, Agreeableness, for a unit increase in the perceived disruption (based on the Conscientiousness, Neuroticism and Openness) for each of 5-pointLikertscale). the 11 users who fully completed the exit questionnaire that However,thecoefficientofdeterminationforthefittedmodel includesthe50questionsrelatedtopersonalitytraits.Forthis isnothigh(R2 =0.1434),whichimpliesthatnotonlythedis- computation,weusedthescoringinstructionsthatcomewith ruptionperceivedbytheuser,butalsootherfactorsinfluence thepersonalitytest[16]. theuser’sdecisiontoacceptanotification. ImpactonReportedDisruption Whyaredisruptivenotificationsaccepted? Toquantifytherelationshipbetweenthefivepersonalitytraits As discussed above, the disruption perceived by the user and thedisruption perceived bythe users fromnotifications, makesanotificationmorelikelytobedismissed. Ourdataset we fit a linear regression model with the average disruption Estimated Estimated Variable Std.Error tvalue pvalue Variable Std.Error tvalue pvalue Coefficient Coefficient Extroversion 0.44042 0.14926 2.951 0.0319* Extroversion 0.45350 0.13446 3.373 0.0198* Agreeableness 0.25281 0.14122 1.790 0.1334 Agreeableness 0.29340 0.12722 2.306 0.0692 Conscientiousness -0.42050 0.11586 -3.629 0.0151* Conscientiousness -0.26102 0.10438 -2.501 0.0544 Neuroticism -0.39663 0.10597 -3.743 0.0134* Neuroticism -0.36975 0.09547 -3.873 0.0117* Openness -0.05059 0.14074 -0.359 0.7339 Openness 0.08584 0.12679 0.677 0.5284 N=11 N=11 R2=0.9007 R2=0.9035 F(5,5)=9.073(p=0.01511) F(5,5)=9.366(p=0.01411) Table 7. Results of linear regression with the average seen time as a Table8.Resultsoflinearregressionwiththeaveragedecisiontimeasa dependentvariableandthepersonalitytraitsasindependentvariables. dependentvariableandfivepersonalitytraitsasindependentvariables. eralizedacrossgroupsofuserswhosharethesamepersonal- as a dependent variable, and the five personality traits as in- itytraits. dependent variables. Here, the average disruption is com- putedasameanofdisruptionreportedbytheuserthroughthe IMPLICATIONS questionnaires. Allresponseswereencodedwiththefollow- ingvaluesStronglydisagree=1, Somewhatdisagree=2, Neu- DeferringNotifications tral=3, Somewhat agree=4, and Strongly agree=5. Table 6 Previous studies show that users perceive more disruption shows the parameters of the fitted linear regression model. fromnotificationswhenengagedinintricatetasksandforthe Theresultsshowthattheextroversionpersonalitytraitsignif- firsttimeweconfirmthisformobilesettingsinaquantitative icantlyaffectstheaverageperceiveddisruptionandthatextro- way.Thus,inordertobenefitusers,theOSshouldoffermore vertsaremoreinclinedtobedisruptedbyanotification. The flexibilitytothemforsettingthebusymomentssothatonly high value for the coefficient of determination (R2 = 0.737) time-criticalnotificationscouldbetriggered. Thiswouldpo- shows that the value of average disruption reported by the tentially allow interruptibility management (IM) systems to users is highly influenced by the personality of users, albeit learnpatternstopredicttheuser’sengagementwithcomplex thiscouldbeaconsequenceofoursmallsamplesize. tasksandprioritizeinterruptionsaccordingly. ImprovingNotificationPresentation ImpactonNotification’sSeenTimeandDecisionTime Wefoundthatusersbecomemoreattentiveatbusymoments, We then investigate whether the personality traits influence butarelikelytoperceivemostnotificationsasdisruptiveand theseenanddecisiontime. Inordertoperformtheseanaly- dismiss them. However, disruptive notifications tend to be ses,wecompute: accepted if they contain useful content. The presentation of 1. the average seen time of notifications for each user: the notificationsummariescouldbeadaptedtohelpusersquickly average time taken by a user to view a notification. It is decidewhethertoclickordismissnotifications,e.g. byhigh- computed as the mean of the seen time of all notification lighting notifications from priority contacts or with priority receivedbytheuser. contentthatislearntovertimebyanIMsystem. 2. theaveragedecisiontimeofnotificationsforeachuser:the BuildingaPersonality-DependentInterruptibilityModel averagetimetakenbyausertoclick/dismissafterviewing We observed that the perceived disruption, seen and deci- anotification. Itiscomputedasthemeanofdecisiontime siontimeareinfluencedbytheuser’spersonalitytraits. This ofallnotificationsreceivedbytheuser. demonstrates the potential to take the personality trait into account in interruption models and, for example, generalise We first fit a linear regression model with the average (per interruptibility models across groups of users who share the user)seentimeasadependentvariable,andthefiveperson- same personality traits. These findings can be exploited to alitytraitsasindependentvariables.Theparametersofthefit- designmoreeffectivemachinelearningalgorithmsforintel- tedlinearregressionmodelareshowninTable7. Theresults ligentnotifications. demonstrate that the time in which a notification is viewed bytheusersissignificantlyinfluencedbytheirExtroversion, RELATEDWORK Conscientiousness and Neuroticism personality traits. We Multitasking is fundamental in workplaces, where task fitanotherlinearregressionmodelwiththeaveragedecision switchinghappenseveryfewminutes[17],butalsointhepri- timeasadependentvariableandthefivepersonalitytraitsas vatesphere,whereanincreasingnumberofpersonalcomput- independentvariables. Theparametersofthefittedlinearre- ing devices mediate the flow of data, be it of entertainment, gression model are shown in Table 8, and demonstrate that social or informative nature. Unfortunately, multitasking is thedecisiontimeforanotificationissignificantlyinfluenced seldom seamless, since the limited amount of human atten- bytheuser’sExtroversionandNeuroticismpersonalitytraits. tion is sought by a range of competing tasks/themes. The Theneedforindividualmodelsofhumaninterruptibilityhas disruptiveness of interruptions was analyzed by Miata and beenidentifiedearlier,bothinthedesktopofficesetting[19], Norman [24], who were among the first to note and explain aswellaswithmobilesmartphoneusers[23]. Theabovere- its variability with the context and particularly their align- sultsshowpotentialfortheinterruptibilitymodelstobegen- ment with respect to the primary task a user is working on. Toexplainwhytheinterruptionsaredisruptive,Altmannand behigheveniftheperceiveddisruptionlevelishigh. Onthe Trafton propose the Memory for Goals model that explains otherhand, thisselfperceptionmightbethemostimportant how users’ intention move the necessary mental state of the factorthatdeterminestheuser’slongtermsentimenttowards problembetweentheforegroundandthebackgroundoftheir notifications. When it comes to our ESM sampling, despite attention and how such state deteriorates when kept in the beingaslightaspossible(weaskonlyuptofourESMques- background [6]. The importance of the problem state held tionnairesperdayfromeachuser),theyincreasethenumber inthememorycorrespondstothecomplexityoftheprimary ofnotificationsauserseesduringthedatacollectionperiod. task and, consequently, recovering after a complex task is Thedensityofnotificationsnegativelyimpactsthesentiment moredemandingthanifaroutinetaskisinterrupted[9]. towards individual notifications [28]. However, we believe that in our case the impact is equally distributed among no- In this paper, to the best of our knowledge we are the first tifications, andconsequently, thatthefindingsabouttherole toinvestigatetheroleofthetaskcomplexityoninterruptibil- of different factors still hold. Finally, while the work is the ity in the mobile context. Our results confirm the theory of first to our knowledge to uncover the role of individual psy- AltmannandTraftonandwefindthatinterruptingacomplex chological traits on mobile interruptibility, it is important to task remains disruptive in the mobile setting. However, the notethatwerelatedthetraitswiththereported interruptibil- natureofinterruptionsinourstudyisfundamentallydifferent ity. Therefore, our finding that extrovert people report to be fromtheabovework–ourusersreceivenotificationssignal- more disturbed by notifications than others, should be inter- ing interruptions, and are not “forced" into the interruption, pretedunderthefactthatgeneralreportingaboutoneselfmost perse. Thisallowsustoinvestigatemoresubtlephenomena, likelydependsontheperson’slevelofextroversion. suchastherelationshipbetweentheprimarytaskandthetime toregisteranotification. Oneofourkeyfindingsisthatusers CONCLUSIONS workingonmorementallydemandingtasksneedlesstimeto In this paper we have presented a study of mobile interrupt- notice a notification. We hypothesize that the “high alert" ibility,specificallyconcentratingontheidentificationoffac- stateinwhichauseriswhenworkingonacomplextask[21] tors that make an interruption disruptive and the impact on alsoleadstoamoreagilereactiontoanotification. theresponsetimetoanotification. Thecontributionsofthis Multitasking theories are build upon data acquired in highly study are threefold. First, we have confirmed the validity of controlled environments. In reality, however, mobile users somepastdesktopinterruptibilitystudiesinamobilesetting. engage in unconstrained communication with their friends Second,forthefirsttime,wehaveinvestigatedtheroleofno- andfamily,moveaboutindifferentsurroundingsandgetno- tification presentation, sender-recipient relationship and per- tifications from a range of applications. Shirazi et al. [32] sonalityformodellinginterruptibility. Finally,ourworkcon- show that there is a high variation in the way a notification firmsfindingsfromrecentinterruptibilitystudies. is handled depending on the application with which it is as- Through a mixed method of automated smartphone logging sociated. Personalcommunicationapplications,forexample, andESMsamplingwehaveobtainedadatasetofin-the-wild are preferred and attended to faster, than applications asso- notificationsandESMreportsonnotificationperceptionfrom ciated with utilities and tools. Personal interests, relevance 20users.Wehaveanalysedthedatatoshowthattheresponse and actionability of the content are additional qualifiers that time of a notification in the mobile environment is not only impact a user’s reaction to a notification [15]. In our recent influenced by an ongoing task’s type, completion level and workweuncoverthatthecontentandthesender-recipientre- task complexity, but also by the notification’s alert modal- lationship play a significant role in the decision to accept a ity,presentationandsender-recipientrelationship.Ourresults notification [23]. In this study, we further refine the role of haveshownthatthepresentationofanotificationanditsalert the sender and find that the messages from extended family type, as well as the type, completion level and complexity membersareperceivedastheleastdisruptive. Moreover,the ofataskwithwhichtheuserisengaged,allimpacttheseen contentisoneofthemainreasonsforacceptingthenotifica- time. Moreover, the relationship with the sender influences tionseventhoughtheyinterferewiththeuser’scurrenttask. the user’s decision on accepting a notification or not. The data also reveals how the sentiment (i.e., perceived disrup- LIMITATIONS tion) towards a notification varies with the type, completion Most limitations of the work presented in this paper stem level and complexity of an ongoing task and the recipient’s from our decision to collect data in the wild, with the min- relationshipwiththesender. Finally,differentpeopleexhibit imum amount of intervention from our users. For example, different reactions and we observe a substantial role of the whenitcomestothecomputationoftheseentimeofanoti- individual psychological traits on how a person reacts to a fication, remotely, we can only detect if a user unlocked the mobilenotification. phoneandassumethatallnotificationswereseen. Wecannot detecttheprecisetimewhenauserstartsreadingasummary ACKNOWLEDGEMENTS of a message from the notification bar. Moreover, in case This work was supported through the EPSRC Grants a notification arrives when the user is already engaged with “MACACO:Mobilecontext-AdaptiveCAchingforContent- thephone,weassumethattheuserhasseenthenotification. Centric networking” (EP/L018829/2) and “UBhave: Further, since our users are not confined to a laboratory, we ubiquitous and social computing for positive behaviour are limited to self-reported level of disruption from a notifi- change”(EP/I032673/1). cation. In reality, the impact on the primary task need not