PLEASE DO NOT DISTURB
Cover design by Siniz H.K. Kim
Printed by GildePrint Drukkerijen, Enschede Published by Martijn H. Vastenburg
ISBN/EAN: 978-90-9022393-3
© 2007 Martijn H. Vastenburg
PLEASE DO NOT DISTURB
modeling user experiences for considerate home products
Proefschrift
ter verkrijging van de graad van doctor aan de Technische Universiteit Delft,
op gezag van de Rector Magnificus prof.dr.ir. J.T. Fokkema, voorzitter van het College van Promoties,
in het openbaar te verdedigen op maandag 17 december 2007 om 15:00 uur door
Martijn Hugo VASTENBURG
Dit proefschrift is goedgekeurd door de promotor: Prof.dr. H. de Ridder
Toegevoegd promotor: Dr.MSc. D.V. Keyson
Samenstelling promotiecommissie: Rector Magnificus, voorzitter
Prof.dr. H. de Ridder, Technische Universiteit Delft, promotor
Dr.MSc. D.V. Keyson, Technische Universiteit Delft, toegevoegd promotor Prof.dr. P.J. Stappers, Technische Universteit Delft
Prof.dr.ir. J.H. Eggen, Technische Universiteit Eindhoven
Univ.-Prof.Mag.dr. A. Ferscha, Johannes Kepler Universität, Linz, Oostenrijk Prof.dr.ir. A. Nijholt, Universiteit Twente
Prof.dr. R. Vertegaal, Queens University, Canada
TABLE Of CONTENTS
1 Introduction 7
1.1 Motivation 8
1.2 Terminology 10
1.3 Modeling user experiences 11
1.4 Research goals 13
1.5 Approach 14
2 Considerate home notification systems: a field study
of acceptability of notifications in the home 19
2.1 Related work 20 2.2 Initial model 23 2.3 User study 25 2.4 Results 30 2.5 Discussion 35 2.6 Conclusions 38 2.7 Future work 39
3 Considerate home notification systems: a user study
of acceptability of notifications in a living room laboratory 43
3.1 Expected results 47
3.2 User study 48
3.3 Results 53
3.4 Discussion and conclusions 57
3.5 Future work 59
4 A user experience-based approach to home atmosphere control 63
4.1 Related work 64
4.2 Case study 66
4.3 Comparative user study 75
4.4 Personalization of the atmosphere model 80
4.5 Reflections on design and approach 81
5 Measuring user experiences of prototypical autonomous
products in a simulated home environment 85
5.1 Methodology 87
5.2 Results 90
5.3 Discussion and conclusions 93
5.4 Future work 95
6 General findings and future directions 99
6.1 General findings 99
6.2 Modeling user experiences 101
6.3 Methodological issues 102
6.4 Take home lesson 103
6.5 Future directions 104
Summary 107
Samenvatting 111
References 115
Acknowledgements 121
About the author 123
1
INTRODUCTION
Advances in sensor technology, home-networking technology, miniaturization, and modeling and reasoning software have made it possible for everyday products to sense their environment, anticipate user needs, and act appropriately. With today’s technology, energy can be saved by automatically adapting a predefined heating schedule to the living patterns in the home. Products that are context-aware and take initiative, rather than products that solely react to user input, could soon enter the homes of ordinary people. In a matter of years, the home could be filled with context-aware, networked and pro-active devices. This vision of pervasive technology that has been unobtrusively embedded into the environment has been termed, amongst others, ubiquitous computing [Weiser, 1991] and ambient intelligence [Aarts, Marzano, 2003].
In recent years, much research has been conducted towards realizing this vision of ambient intelligence. Whereas in the early days, most of the research was technology-oriented, nowadays increasingly the focus is on users and user experience. Aarts and Marzono presented five challenges towards creating ambient intelligence: embedding hardware, sensing, personalization, adaptation and anticipation. As argued by Aarts and Marzano, the enabling technology for ambient intelligence is now available. In view of the five challenges, the focus is now shifting to user-oriented challenges: personalization, adaptation and anticipation.
1.1 MOTIvATION
The home thermostat used to be a prototypical example of simple-to-use products, with only one single rotating knob for setting the desired temperature (Figure 1, left). Today’s high-end home control displays, on the other hand, tend to be showcases of complex multifunctional products that can be used to control not only the in-home climate, but also the alarm system, air quality and music (Figure 1, right). Similarly, walkmans, telephones and remote controls used to be examples of one-function-per-device. Nowadays their functions are often combined in a single “mobile phone” one-function-per-device. Do these new functions and multifunctional products make life easier?
Industry has recognized the usability problems caused by the buildup of features and functions into devices. A simple and elegant remedy for usability problems would be stripping all surplus functions. Simplicity can be considered the primary motivation behind information appliances, i.e., small devices mainly intended to perform one single task [Raskin, 2000; Weiser, 1991; Norman, 1999]. A well-known example is the Apple iPodTM . The redesign of the portable music player has resulted in a minimalistic and attractive product, which has attracted a huge customer group. The iPodTM answers a latent user need to carry extensive music libraries, thereby diminishing the need to make a selection of songs for each trip, while keeping the user interface as simple as possible [Jones, 2005].
Whereas in many cases functionality can be removed, as with the iPodTM, there are situations in which the added functionality results in higher product value. Industry-wide initiatives have started to counter the usability problems of domestic appliances. DiamondHelp, for instance, is a proposed non-proprietary framework for linking home appliances and providing a central large display [Rich et al, 2006]. In the
case of DiamondHelp, the central user interface enables task-level access to functions of the individual products. Furthermore, by embedding predefined task-models into the products, the central DiamondHelp interface could anticipate user goals and collaborate with the users, especially with complex features that are only occasionally used.
By clustering functionality into task-level constructs, as with DiamondHelp, the abstraction level of user-product interaction can be increased. As a result, users are enabled to communicate with the product at a level that matches their (task-oriented) needs and desires. Task analysis is a commonly used approach in the design of human computer interaction towards improving the usability of user interfaces [Kirwan, Ainsworth, 1992; Benyon et al., 2005; van der Veer et al., 1996; Horvitz et al., 1998]. Without any doubt, the use of task analysis has improved the usability and improved the user performance on tasks [DeKoven, 2004]. One could however argue the usefulness of rigid task models in a user experience-oriented setting. Will users for example enjoy their coffee better when they can make the coffee as efficient as possible? A coffee machine with a single button might do the job. Research has shown, on the other hand, that the process of making coffee, including for example the sound of grinding coffee beans, can contribute to the appreciation of the resulting cup of coffee [Özcan, van Egmond, 2006; van Egmond, 2007]. Apparently, designers need to deal with factors other than efficiency and usability alone [Norman, 2003].
The shift from user interface design to user experience design started several years ago. Experience design covers the design of (the user experience of) products based on the consideration of an individual’s or group’s needs, desires, beliefs, knowledge, skills, experiences, and perceptions. Rather than looking at the product by itself, experience design considers products in their context of use, i.e., the product in relation to relevant human factors (information on the user, the user’s social environment, and the user’s tasks) and the physical environment (location, infrastructure, and physical conditions) [Schmidt, Beigl, Gellersen, 1999].
user-product interaction at the level of desired experiences, rather than in terms of tasks and functions. Existing tools and techniques that deal with human experience, originally developed for use in the field of for example psychology, could be used to create products that consider user experiences.
1.2 TERMINOLOGy
When discussing technology that takes into account the state of the user in adapting its behavior, the word “considerate” has been used before. The term “considerate computing” has been coined by Wayt Gibbs, a Scientific American journalist who interviewed Eric Horvitz and Roel Vertegaal on the notion of attentive user interfaces. In his article [Wayt Gibbs, 2005], considerate computing denotes devices that consider user attention in adapting their behavior. For example, a considerate car might notice that the driver is looking away at the moment when attention is needed, and the car might activate an alarm sound that draws back the attention to the car ahead.
When used in this thesis, the word “considerate” is not limited to user attention. Considerate home products is being used as a generic label that denotes products in the home that aim to optimize product behavior by considering the user and the context of use. To illustrate this kind of products, one could make an analogy with a butler. Similar to a butler, a considerate home product might create a relaxing atmosphere when you come home stressed after a hard day’s work. Furthermore, a considerate home product might decide to postpone messages until the stress-levels have decreased.
In this thesis, considerate home products are closely linked to models of user experiences. User experience can be approached in many ways. For example, user experience of a product can be viewed from a business perspective: businesses can make money by creating experiences, and products could be regarded as a means towards creating experiences [Pine, Gillmore, 1999]. Throughout the thesis, however, user experience refers to the feelings and emotions that the user has when interacting with a product (Figure 2).
User Motivation Resources Mental state Knowledge Attitudes Expectations User experience during interaction System Products Objects Services People Infrastructure Involved in interaction Context Physical context Social context Temporal context Task context Perception Interaction
For example, a user might feel disturbed (Figure 2: “mental state”) when receiving a not-so-urgent message (Figure 2: “interaction”) in the middle of a conversation (Figure 2: “social context”). In other words, user experience is considered a momentary emotion that can be evaluated in various ways, for example with psycho-physiological measurements. User experience is regarded as a subjective, personal feeling during interaction with a system, whereas usability should be regarded as an objective product attribute [Roto, 2007].
1.3 MODELING USER ExPERIENCES
To get a better understanding of what “modeling user experiences” stands for, an example of how experience models could be embedded into a home atmosphere control system is given below. In a traditional setting, as shown in Figure 3, home control products are unaware of user experiences. If, for example, a user wants to create a romantic atmosphere in the living room, the user could make an intimate combination of lights and music. It is up to the user, time and time again, to make the translation from the desired atmosphere to settings of volume, lights, music genre, etcetera.
When dealing with a small number of products, manual control is not a problem. If, however, the number of products in use is high, creating a desired atmosphere setting can become a cumbersome task. A central control interface could simplify the confi guration process. A home atmosphere controller can act as a central remote control for creating home atmospheres (Figure 4). To create a romantic atmosphere, a user can make a combination of settings, which can be stored for simple future access.
Figure 3 Traditional view of user-product interaction
light 1 ... light n music home atmosphere atmosphere set functions and features desired atmosphere set functions user set functions user experience
Figure 4 Centralized controller for products in the home
Low-dimensional models can be constructed of user experiences for specifi c applications. As an example, Chapter 4 describes the atmosphere model. The atmosphere model, as shown in Figure 5, describes how a group of users has experienced a set of predefi ned home atmospheres.
The atmosphere model is being used by the atmosphere controller to enable user-product interaction in terms of desired atmospheres rather than functions and features (Figure 6). The capture-arrow represents the process of measuring and modeling user experiences.
User experiences vary between users and between situations. Rather than creating different models for each user and each situation, one might try to create models that describe experiences for a set of users or a set of situations. Throughout the thesis, such user-independent models are referred to as generic models. In the case of the atmosphere model, individual users can personalize the generic model by giving feedback on generic settings.
set desired atmosphere
light 1 captur e light n music ... home atmosphere atmosphere set considerate atmosphere controller
set desired atmosp
desired atmosphere captur e s user user experience considerate atmosphere model
Figure 6 Models of experiences embedded in a centralized controller for home products
The applicability of models of experiences is not restricted to improving the user-product interaction. These models can also be used by autonomous user-products to assess the cost and value of autonomous product actions, as will be discussed in Chapter 2, 3 and 5. In Figure 7, “desired atmosphere” has been replaced by “desired experience”, and “atmosphere” has been replaced by “stimulus”, since the fi gure depicts a generic view of user-product interaction. In short, models of experiences can be used to bridge user needs and product actions.
1.4 RESEARCh GOALS
The primary goal underlying this thesis is to fi nd out how user experiences of home products can be measured and modeled, and to fi nd out how these models of experiences can be operationalized in considerate home products. The focus will be on context-sensitive appliances that control the presentation of information and media in the living room.
Throughout multiple user studies, methodological issues related to studying user experiences of prototypical smart home products will be considered. Design research on prototypical smart home products faces a methodological challenge, in terms of user experiences of autonomous prototypes in a controlled way, while preserving the sense of realistic experiences. To preserve ecological validity, the methods, materials and setting of the experiment need to approximate the real-life situation that is under study [Shadish, Cook, Campbell, 2002; Brewer, 2000]. Furthermore, it can be diffi cult to preserve external validity, i.e., generalize results from laboratory studies. Advantages and disadvantages of testing in an artifi cial home setting versus testing in the fi eld are discussed.
Figure 7 Generic view of user-product interaction with a considerate product captur e stimulus set sense considerate product desired experience captur e user environment user experience considerate model
1.5 APPROACh
The implicit interactions framework was found suitable as a framework for positioning the individual studies described in this thesis (Figure 8). The rationale behind the framework: interaction designers should not only decide what type of information exchange is needed, but also the manner in which it should take place. The framework consists of two dimensions: attentional demand and initiative. Attentional demand ranges from background to foreground. Foreground interactions take place when users are communicating with a device in an explicit and conscientious manner. Background interactions, on the other hand, take place when users are less conscientiously oriented towards making commands or monitoring output; implicit interaction can be regarded as “hidden” or “natural” user actions that trigger background tasks. The second dimension, initiative, ranges from reactive to pro-active device behavior. Initiative level is defined by who initiates an interaction; in a reactive setting, devices react to user actions, in a pro-active setting the device itself takes initiative.
Traditional interaction design has focused on explicit interactions in a reactive setting; traditionally, products react to user actions, and user-product interaction is explicit. Ju and Leifer argue that, as products move out of the tradional setting towards being partners of the users, the interaction designer should be able to select between implicit interaction and explicit interaction, and between background and foreground interaction, based on the situation at hand. Eventually, interactive devices might be more helpful if they could do what people do - and expect from each other [Reeves, Nass, 1996].
Towards envisioning how models of experiences might be used in considerate home products, four studies have been conducted. Each study covers a different segment of the interaction space as defined by the implicit interactions framework. The first two studies are focused on pro-active interaction, i.e., the right-hand segments in Figure 8. A typical example of pro-active products is a notification device that displays alerts and notifications. A prototype home notification system was constructed and studied in the first two user studies.
The home notification system as described in Chapter 2 is an example of a pro-active device that interacts with users in the foreground (Figure 8: upper right segment). When, however, many messages are communicated to users in the foreground, users might be overwhelmed with notifications. Human attention could soon be the most precious resource in a computerized environment [Garlan et al., 2002; Horvitz, 1999; Vertegaal, 1997]. Towards integrating an intelligent notification system into everyday life, an understanding is needed of how people experience notifications during everyday activities, and how notifications affect these activities. A field study was conducted in the homes of ten participants to find out how participants experience notifications in the home, and how acceptability of notifications is related to their engagement in activities in terms of concentration, physical activity, social interaction, and urgency of the activity, and to the perceived message urgency.
To create a better understanding of how acceptability of notifications is influenced by the level of intrusiveness of the presentation, and how acceptability depends on timing in relation to user activities, a second study on notifications was conducted in a living room laboratory (Chapter 3). Participating couples spent an evening in the living room laboratory, and were asked to evaluate 30 notifications. Notifications were presented alternately in the foreground and in the background. This second study covered both right-hand segments of the implicit interactions framework; the home notification system started interaction by getting users attention, and the presentation switched between background and foreground.
Whereas the first two studies focused on pro-active systems, most traditional products are reactive by nature; they respond to user actions. To understand how modeling user experiences might help the design of reactive products, a home atmosphere control system was constructed and studied, as described in Chapters 4 and 5.
The case of explicit interaction with the atmosphere control system is studied first (Figure 8: upper left segment). In general, a combination of explicit interaction and reactive system behavior can be found in devices where the interaction is part of a primary task actively controlled by the user. The goal of the third study, as described in Chapter 4, was to demonstrate how a model of experiences can be integrated into a potentially complex home atmosphere control system, while considering a wide range of possible user needs. The complex control problem of creating home atmospheres using light, music, and projected wall-art can be reduced by focusing on desired experiences, rather than product functions and features. A case study is described in which subjective interpretations of living room atmospheres were measured and embedded into a prototype display system.
people are willing to delegate control to a pro-active home atmosphere control system. A user study with ten participating couples was conducted in the living room laboratory. Previously distributed diaries were used to create a realistic set of user activities and predefined system settings for each participating couple. The underlying question of this study was to find out if the perceived user control, predictability of system actions, and the perceived system complexity affect the willingness to delegate control to pro-active systems.
whAT’S NExT
foreground
background
Chapter 2: Home Notifications (static presentation)proactive
reactive
2
CONSIDERATE hOME NOTIfICATION
SySTEMS: A fIELD STUDy Of
ACCEPTABILITy Of NOTIfICATIONS IN
ThE hOME
Information and communication technology can help people stay up to date with events in the world. Traffic updates are sent to commuters using voice messages on mobile phones, new mail is announced using auditory signals on computers, and washing machines use irritating beeps to indicate the laundry is ready. Increasingly, people at home are connected to networked information services [den Hartog et al., 2004]. Medicine reminders, burglar alarms, and weather and news update services all notify users in their homes of possibly interesting events. These notifications can be helpful and appreciated, but they can also be inconvenient and distract the user. Since the number of information services present in everyday life appears to be growing, people might soon be overwhelmed with notifications.
To avoid overwhelming users with notifications at unwanted moments, notification systems need to be made aware of their environment. Ideally, notification systems should sense the state of its users and their environment, reason about the value of the notification message content, and decide the best time and form for presenting messages. An understanding of how the acceptability of notifications is influenced by contextual factors is needed to design considerate notification systems.
In the domain of task-oriented work environments, many results are available on how interruptions affect people and how systems can optimally choose timing and modality of notifications, as summarized in section 2.1. Both the objective impact of interruptions on task performance and the subjective acceptability of interruptions have been studied. However, it is not known if the results of these studies also apply to the home environment; the acceptability of notifications in the home could differ significantly from the work environment. To create a considerate mechanism for scheduling and presenting notifications in the home, we need to know what factors influence the acceptability of notifications by the user in the home. As a first step, the present study concentrates on two factors: engagement in activities and message urgency.
The original publication is available online at www.springerlink.com.
This chapter is organized as follows. Section 2.1 describes related work. In section 2.2, the initial model of acceptability of notifications is described, including the expected results. The present study incorporates research methodologies for on-line registration of user experiences under natural circumstances. The resulting field study design, in which a laptop with notification and questionnaire software was placed in the houses of ten participants, is described in section 2.3. Section 2.4 describes the key findings, while the remainder of the article is used to discuss the results and future steps.
2.1 RELATED wORk
CONSIDERATE COMPUTING
Notification systems provide access to, and draw user attention to, information secondary to the current user activities [McCrickard, Chewar, 2003]. Because the primary activity is interrupted by the notification, task performance may decrease. Prior studies in the area of interruptibility1 and notification systems have in common the goal of increasing task performance [e.g., McCrickard et al., 2003; McFarlane, 1998, 1999]. Typical application domains are air traffic control and office work [McFarlane, Latorella, 2002].
Since human attention is a scarce resource, each notification message can be considered a potential threat to task performance. Attention can be viewed as a constrained resource that can be traded for some utility [McCrickard, Chewar, 2003; McCrickard et al., 2003; Horvitz, 1999]. The attentive user interface paradigm [Shell, Selker, Vertegaal, 2003; Vertegaal, 2003; Horvitz, Kadie, et al., 2003] aims at avoiding overloading the user by adapting system behavior based on the sensed user attention focus. Attentive user interfaces generally calculate the cost in terms of user attention and the benefit in terms of subjective or objective performance factors, in order to predict acceptability and select the optimal timing of the interruptions. The term “considerate computing” has been coined by Wayt Gibbs, a Scientific American journalist, based on the notion of attentive user interfaces. In his article [Wayt Gibbs, 2005], considerate computing denotes devices that consider user attention in adapting their behavior.
The cost of notifications in terms of user attention can be reduced by adapting the presentation of messages to the user state. Presentation in the users’ periphery minimizes the impact of interruptions on ongoing activities [McCrickard, Chewar, 2003; Maglio, Campbell, 2000]. In the case of aware notification systems, non-urgent messages could be presented in the periphery of the user, while urgent messages could be presented in the foreground.
The cost of interruptions can also be reduced by adapting the timing of notifications. The cost in between tasks is lower, because supposedly people may be between evaluation of the last activity and formation of a new goal [Miyata, Norman, 1986]. In a study on notification systems for mobile devices, scheduling of messages was linked to transitions in physical activity, under the assumption that changes in physical activity can be used as an indicator of user activity switches [Ho, Intille, 2005]. According to them, notifications that were delivered at activity transitions were generally more easily accepted by the participants.
MEASUREMENT Of IMPACT
To assess and model the acceptability of notifications, a mechanism is needed to measure acceptability. Traditionally, studies of interruptions are based on objective measurements of effects of interruptions on task performance. More recently, subjective measurements have been used to measure acceptability of interruptions: video tagging [Horvitz, Apacible, 2003; Kern, Schiele, 2006], rating scales [Ho, Intille, 2005], and self-reports by sticking up fingers [Hudson et al., 2003]. As a next step, interruptibility might soon be based on physiological measurements [Chen, Hart, Vertegaal, 2007].
Video tagging was used to study the interruptibility of office workers [Horvitz, Apacible, 2003]. Participants performed five 1-hour sessions in their offices. Sessions were taped on video, and system events were captured. After each session, subjects were asked to tag and assess the video. As a major advantage of post hoc video tagging, the setup does not interfere with user activities, resulting in more natural and realistic user behavior. It might however be difficult for participants to rate their interruptibility after the session, since users would have to recall situations based on the video.
Kern and Schiele (2006) used video tagging to study mobile interruptibility. A series of 94 realistic everyday-life interruptions were captured on short videos using an actor. A group of 24 subjects were asked to annotate the video clips. The experiment focused on individual differences in interruptibility, therefore the researchers wanted all participants to rate the same situations. Results indicated for example that in judging interruptibility, women are more likely to consider their social context than men. Although video tagging made it possible to collect multiple user ratings for a single situation, it is not clear whether the participants were able to relate to the videos and judge the situations accordingly.
Subjects had to hold up fingers to indicate the rating. This way the disturbance caused by the alerts and responses was minimized. The subjects were asked to give an in-situ self-report after each alert (“beeper study”).
In these studies, subjective data on the impact of interruptions were collected and used to construct computer models that help improve the coordination and presentation of interruptions. These computer models consider not only task performance, but a whole range of factors that the users themselves found relevant. Subjective measures seem appropriate for an exploratory study in the home environment; a range of relevant factors can easily be measured.
NOTIfICATIONS IN ThE hOME
Notifications are not restricted to work environments, they also occur in the home. Interruptions have been studied before in the home environment [e.g., Tran, Calcaterra, Mynatt, 2005], with a focus on task performance (preparing punch in the kitchen). It is however hard, if not impossible, to express the effect of interruptions and notifications solely in terms of task performance. As an example, a mobile phone will play a low-battery warning regardless of the current context. If the low-battery is empty in the middle of the night, the warning could result in a disturbed night’s sleep [Picard, 2000]. Should the effect of this interruption be modeled only in terms of a decreased “performance” in sleeping? In the home, apparently, other factors than task performance come into play [Haines et al., 2007; Howard, Kjeldskov, Skov, 2007].
2.2 INITIAL MODEL
The goal of the user study described here is to gain insight into attentional, social, and urgency aspects relevant to the acceptability and preferred timing of notifi cations in a living room setting. A rather extensive initial model for predicting the best time to present messages in a considerate home notifi cation system will be used to improve our understanding of the factors underlying acceptability. The knowledge gained will be used to refl ect on this initial model and to simplify the model wherever possible.
The initial model represents a cost-benefi t tradeoff for notifi cations (Figure 1). Users are expected to gain benefi ts in terms of the value of notifi cation messages, and to experience cost in terms of interrupted user activities. The model roughly resembles the cost benefi t mechanism described by McCrickard and Chewar (2003). However, whereas McCrickard and Chewar present a generic model covering all factors relevant to attention cost and utility benefi t of notifi cations, the model used in the present study is focused on how people experience notifi ciations of messages. To get a better understanding of the mechanism, we asked users to rate their engagement in current activities, the urgency of the message, and the acceptability and preferred timing of the notifi cation.
In the model as shown in Figure 1, engagement in activities indicates the involvement of users in their current activities. Engagement is measured using subjective ratings of concentration level, physical activity level, social interaction level, and urgency of current activities. Message urgency is a subjective rating of the urgency of the notifi cation message, which needs to be judged independent of the current user activities. Acceptability of the notifi cation and preferred timing are subjective ratings of the general acceptability and preferred timing considering the message and the user activities at the time of interruption.
Figure 1 Initial model of acceptability of notifi cations. The subjective general acceptability and preferred timing are linked to both user activity related factors and message urgency
engagement in activities concentration
level
interruptibility
acceptability of the notification + preferred timing
message value physical activity
level
physical activity social interaction
level
social interaction urgency of current
activities
In the perspective of the taxonomy of McFarlane (1998), there are eight factors underlying human interruptibility. The variation in interruptions used in our present study was restricted. The source of interruption is fixed to a computer; the method of coordination is set to scheduled interruptions, the method of expression is fixed to a plain depiction of the message on a computer screen, and the channel of conveyance is set to a computer screen. Accordingly, variation in acceptability ratings in our study is a consequence of four factors: (1) the individual characteristic of the person receiving the interruption, (2) the meaning of the interruption (i.e., the type of interruption, for example an alarm), (3) the human activity changed by the interruption, and (4) the effect of the interruption (the impact of the interruption, e.g., start a new activity).
ExPECTED RESULTS
The acceptability of notifications is expected to be positively related to the user rated urgency of the notification message, but negatively related to the engagement of users in their activities (Figure 2). These expectations are in line with the conceptual framework as presented by McCrickard and Chewar (2003), in which the optimal presentation form is linked to the goal that users want to achieve based on the notification messages. Following McCrickard and Chewar, a more urgent message is expected to lead to a higher perceived benefit, and consequently to a higher acceptability of the notification. A higher level of engagement of users in their activities is expected to lead to a higher perceived cost, and consequently to a lower acceptability.
Figure 2 Expected acceptability of notifications. Acceptability is expected to be positively related to the user rated urgency of the messages, and negatively related to the engagement in user activities. Moreover, linear combinations are assumed
low high
2.3 USER STUDy
Data was collected over 30 sessions (10 participants x 3 sessions). In each session, 12 notification messages were scheduled, for a total of 360 scheduled notifications.
Participants. Ten subjects (six women, four men) participated in the study, age ranged from 25 to 56 (mean age 33 years). Participants were selected based on their home situation, being not living alone and no children at home. All participants were employed; nine out of ten had an academic degree.
Procedure. Test subjects participated at home. A laptop and a webcam were installed in the living room of the participant. The laptop was located on top of a table, in such a way that the screen of the laptop was visible from most positions in the room. The laptop was used to activate notification messages based on scenarios, as described in “Notification scenarios and activation”, and to present the questionnaires. The webcam was used to log motion activity, and to capture the people present in the room at the time of interruption. Participants could delete the webcam pictures before returning the laptop at the end of the experiment. A microphone was used to log audio activity. The experimenter left the scene after placing the equipment and instructing the participants; participants were told how to start and end each session, they were asked to do whatever they would do on a normal evening, and to treat all messages as authentic messages.
Participants selected three evenings within one week for the experiment. Participants were asked to do whatever they would do regularly, so user activity was not a controlled condition. Since the study took about 18 hours per participant in total, a natural dispersion in user engagement was expected. Notifications were given approximately two times per hour. When a notification was activated, a bell sound was played to attract user attention, and the first part of the questionnaire had to be filled in at the laptop. Then, the notification message was immediately shown, and the second part of the questionnaire was presented. Participants were instructed to fill in the questionnaires themselves; partners were not allowed to do so. The bell sound and volume were not varied during the study. The study started when the subject arrived home from work or around 16:00 at a non-working day, and ended at bed-time.
Table 1 Overview of the notification messages that were used in the experiment. The messages, originally in Dutch, were defined and classified by a panel of three product designers
Classification ID Notification message
low-urgency (L) L1 L2 L3 L4 L5 L6 L7 L8 L9 L10 L11 L12
A good program is about to start on TV. There’s a new package in the mail box.
Don’t forget to get some bread out of the freezer for breakfast. Rent needs to be paid this week.
Don’t forget to create a shopping list for the visit tomorrow. Toilet paper is empty.
The newspaper has arrived.
Sports news flash: Holland has won the EC match against Sweden: 1-0. To save energy, the thermostat should be set lower.
Don’t forget to water the plants.
The light in the attic has not been switched off. A new email from “De Volkskrant” has arrived.
medium-urgency (M) M1 M2 M3 M4 M5 M6 M7 M8 M9 M10 M11 M12
Garbage will be collected tonight.
Don’t forget to double-lock the front door.
The washing machine has finished, please put the laundry in the drier. The battery of your mobile phone is almost empty.
The radiator in the attic needs to be switched off. The plants in the garden need water.
The video tape needs to be returned to the video rental shop today. Coffee is ready.
It’s raining, don’t forget to close the window in the attic. Elections are today.
Today is mother’s birthday, don’t forget to give her a call. The 08:30 appointment tomorrow has been cancelled; you can set the alarm 30 minutes later.
high-urgency (H) H1 H2 H3 H4 H5 H6 H7 H8 H9 H10 H11 H12
Granddad has had an emergency admission to the hospital. You need to be at work 30 minutes earlier tomorrow. The roof is leaking.
Someone is touching your car.
A burglar might have entered the study.
The iron has not been switched off (beware of fire). Eggs are ready, please switch off the stove. Do not forget to take your medications now. Smoke has been detected in the shed.
Lock your windows and close the doors - chemical leakage close by. Please turn the tap off; bathtub is completely filled.
Each single message might result in entirely distinct user ratings when presented to different participants or in different situations, based on the message structure, style, phraseology and by relationships between messages. In recognizing individual differences and subjective factors, emphasis was placed on creating a set of messages that would produce a wide spectrum of levels of perceived urgencies, from low-urgency to high-urgency across subjects, rather than attempting to create a set of messages that would be equally perceived by all subjects. In analyzing the results, message urgency was therefore based on perceived urgency rather than induced urgency.
Notification scenarios and activation. Notification scenarios were created, one scenario per participant per day. The scenarios stated the order of the notification messages and the scheduling of the interruptions. The order of 36 pre-defined messages was randomized for each participant, using a combination of 4 low-urgency, 4 medium-urgency and 4 high-urgency messages for each day.
Notifications were only activated after motion was detected by the webcam, thereby reducing the chance of presenting notifications when no participant was present. Half of the notifications were activated immediately after motion was detected, and the rest of the notifications were activated five minutes after motion was detected. The five minute delay, which was scheduled randomly, was used to distribute the moment of interruption in relation to motion activity. An extra delay of 10 up to 30 minutes was scheduled between notifications.
Questionnaire. A questionnaire was used to collect subjective data. The questionnaire consists of three parts. After hearing the notification bell, participants were asked to rate their engagement in activities (Figure 3). Then, the notification message was shown (Figure 4). In part 2, participants were asked to rate the message, without considering the current activities (Figure 5). In part 3, participants were asked to rate the acceptability and the preferred timing of the notification, considering the message and the activities at the time of the bell. All questionnaire items are directly related to the model depicted in Figure 1.
Preferred timing was presented using two sub-questions: participants had to indicate if they wanted the message to be shown at all, and, if the answer was positive, users were asked to indicate the preferred timing on a scale from now to much later.
Figure 4 Presentation of the notifi cation message (originally in Dutch). The message was shown after participants had rated their current activities
Figure 5 Questionnaire part 2 and 3 (originally in Dutch). Participants were asked to rate the urgency and value of the message in part 2. In part 3, participants were asked to rate the acceptability and preferred timing of the notifi cation
2.4 RESULTS
fACTORS Of ACCEPTABILITy
To understand the factors underlying acceptability and preferred timing, and to consider the interrelationships between the items of the initial model as depicted in Figure 1, a factor analysis was conducted on the results of the questionnaire. A total number of 231 completed questionnaires were collected in the field study. Table 2 shows the results of the factor analysis with principal components using SPSS 14TM with Varimax rotation and Kaiser Normalization, all component loadings <0.20 are suppressed.
The four components with Eigenvalues greater than 1 that emerged from the factor analysis were labeled as message urgency (C1), user engagement in activities (C2), social interaction (C3) and physical activity (C4), based on the factor loadings as depicted in Table 2. These four components explained a cumulative percentage of variance of 79%.
Variation in general acceptability (Q8) could be explained using only the components message urgency (C1) and user engagement in activities (C2). The factor analysis shows high factor loadings of message urgency (Q6), message value (Q7) and general acceptability (Q8) on component 1, which is consistent with our expectations; a positive correlation between message urgency, message value and general acceptability is expected. Also, as expected, general acceptability is negatively influenced by user engagement in activities (C2).
Table 2 Rotated component matrix using Varimax rotation, component loadings <0.20 suppressed. The four components that emerged from the factor analysis were labeled as C1: message urgency, C2: user engagement in activities, C3: social interaction and C4: physical activity
Question Component C1 C2 C3 C4 User engagement in activities Q1 Concentration .80 Q2 Physical activity .92 Q3 Social interaction -.23 .35 .49 -.30 Q4 Urgency of activities .74 .40 Q5 Interruptibility -.85
Message urgency Q6 Message urgency .94
Q7 Message value .94
Acceptability Q8 General acceptability .86 -.33
Preferred timing Q9a Timing A .78 .53
Q9b Timing B .21 .88
General acceptability
Figure 6 shows the acceptability of notifications plotted against message urgency (C1) and user engagement in activities (C2) that were shown relevant to the acceptability in the factor analysis. Each of the 231 items in the graph represents a single interruption of a single subject. The notifications are labeled by acceptability level as rated in Q8, reduced to three levels (low=0/1, medium=2/3, high=4/5). Messages that are rated highly urgent, as shown by the horizontal axis, tend to be highly acceptable. Medium-urgency messages tend to be moderately acceptable, and low-Medium-urgency messages tend to be unacceptable. The factor analysis also showed a negative factor loading of acceptability (Q8) on user engagement in activities (C2). This relation can also be seen in the figure; acceptability is rated higher for low levels of engagement in activity than for high levels.
The experimental outcome resembles the expected outcome as depicted in Figure 2. Using linear discriminant analysis in SPSSTM, the data set was classified into three clusters based on the ratings on Q8. Based on these clusters, 84.0% of the cases could be correctly classified. The clustering was accurate for the high-acceptability cases (90.7% correct), while 76.7% of the medium-acceptable cases were correctly classified, and 81.0% of the low-acceptable cases. The dashed discriminant lines depicted in Figure 6 are roughly parallel, which suggests that general acceptability can be described by means of a simple linear model. Therefore, multiple linear regression was applied to investigate a possible linear relationship between Q8 (general acceptability) and components C1 and C2.
Figure 6 Subjective acceptability ratings plotted against the acceptability-related components from the factor analysis. Messages with high urgency tend to be highly acceptable (“+”), the acceptability of medium and low-urgency messages was more difficult to predict. The three acceptability classes separated in the figure by the dashed lines result from a linear discriminant analysis:
The multiple linear regression showed that a significant proportion (84%) of the variance in general acceptability could be accounted for by a linear combination of C1 (message urgency) and C2 (user engagement in activities): general acceptability=1.58C1 - 0.60C2 + 2.74 (R2=0.84, F
(2,228)=604.3, p<0.001). Figure 7 shows the subjective acceptability ratings plotted against the acceptability related components from the factor analysis, combined with the linear model.
Table 3 Cross-tabulation count of general acceptability (Q8) and preferred timing (Q9). The table shows the relation between preferred timing and acceptability
Preferred timing (Q9) Total now (0) (1) (2) (3) (4) much later (5) never General Acceptability (Q8) of the notification message 0 (not acceptable) 0 0 1 1 0 2 22 26 1 0 2 2 3 0 3 13 23 2 1 2 3 6 1 1 5 19 3 8 2 10 3 1 1 2 27 4 22 12 0 2 1 0 0 37 5 (very acceptable) 52 4 0 1 1 0 0 58 Total 83 22 16 16 4 7 42 190
Figure 7 Multiple linear regression of subjective acceptability ratings against the acceptability-related components from the factor analysis. General acceptability can be described using a simple linear equation: Q8=1.58C1 - 0.60 C2 + 2.74 (R2=0.84, F
PREfERRED TIMING
The questionnaire results show participants wanted to see all high-acceptable messages immediately, medium-acceptable messages should be postponed, and low-acceptable messages should not be presented at all. Considering the scheduling issue, we asked participants two questions: (1) did you want to see the message (Q9a), and if so, (2) what would be the best time to present the message (Q9b). The results of Q9a and Q9b are combined in Table 3; negative answers to Q9a are represented by a “never” score in the table. In case of a positive answer to Q9a, the subjective preferred timing is listed in the table, ranging from now to much later. A significant relationship was found between acceptability (Q8) and preferred timing (Q9) (χ2=225.64, df=30, p<0.001). This relation suggests that a lower acceptability results in a higher desire to postpone, or even to skip, notification messages, and vice versa.
SUBjECTIvE RATINGS Of MESSAGE URGENCy
The subjective scores on message urgency range from low to high (Figure 8), which reflects the effort of the panel of product designers in creating a diverse and plausible set of notification messages. The horizontal axis shows the induced message urgency, i.e., the message urgency which was pre-classified by the panel. The vertical axis shows the message urgency component (C1), i.e., the perceived message urgency scores of the participants. Participants were consistent in their urgency ratings for highly-urgent messages. A higher degree of variation was observed for low-urgency and medium-urgency messages. Alarm messages were rated highest on urgency, including “Somebody is touching your car.” and “The bath is running over.”.
A one-way ANOVA indicated a significant effect of induced message urgency on perceived message urgency (F(2,228)=123.3, p<0.001), suggesting that the classification of the panel and the ratings participants are similar. Percieved message urgency for the high-urgency messages (M=0.94; SD=0.53) was significantly higher (t=-11.3; p<0.005) than for the medium-urgency messages (M=-0.30; SD=0.80). Likewise, the perceived urgency for the medium-urgency messages was significantly higher (t=-3.37; p<0.005) than for the low-urgency messages (M=-0.73; SD=0.75).
ENGAGEMENT AND ACTIvITIES
No significant relation was found between user activities and interruptibility. Participants were instructed to enter all activities at the time of interruption. Examples of typical activities people were engaged in can be seen in Table 4. Only in 4 out of 231 cases, multiple activities were mentioned. Similar activities (watching TV, on the phone) appear in varying degrees of interruptibility. Consequently, no one-to-one relation between user activities and interruptibility can be defined.
In the experiment, the user activities were not controlled. A variation in activities, and consequently in user engagement in activities (C2) levels, is expected. A normality test confirmed that the distribution of C2 levels resembles a normal distribution (Kolmogorov-Smirnov, sign. 0.25); the variation in user engagement in activities in the experiment resembled a normal distribution.
Table 4 Typical user activities clustered by interruptibility
Interruptibility (Q5) Activities
HIGH watching TV, just entered the room, watching commercial break, finishing phone conversation, closing the window
MEDIUM cooking, using computer, watching TV, listening to music, brushing teeth, cleaning the house
2.5 DISCUSSION
GENERAL ACCEPTABILITy
High-urgency messages were found to be acceptable, no matter what. Based on the results of the pilot study, this dominating effect of message urgency was expected. Whereas the effect of user engagement in activities was not clear in the pilot study, the present study does show that acceptability of low-urgency and medium-urgency messages may be improved by creating a system that is aware of user activities, and that adapts the presentation and timing to the activity context. Based on the results of the present study, one might conclude an effective way to predict acceptability of notifi cations is to consider only message urgency and user engagement in activities.
Figure 9 presents an updated model of acceptability being a simplifi ed version of the initial model (Figure 1). Physical activity level, social interaction level and urgency of user activities did not correlate to acceptability and preferred timing; therefore these factors have been removed from the model. Concentration level (Q1) has been generalized to attention level. Message urgency (Q6) and message value (Q7) were highly correlated; these factors have been combined in the updated model.
PREfERRED TIMING
The preferred timing for presenting messages appeared to be directly related to acceptability (Table 3). Highly-acceptable messages were requested to be shown immediately, while non acceptable messages were to be postponed or not shown at all. Apparently the preferred timing depends on acceptability: immediate interruptions are accepted for highly acceptable messages only; low and medium acceptable messages should be presented at a later point in time.
For 42 out of 190 notifi cations, participants indicated they did not want the message to be shown at all. Based on the questionnaire data and the exit interviews, two possible explanations come to mind. First of all, the notifi cation system in the experiment did not adapt the presentation style to the content of the message. A realistic notifi cation system might adapt the presentation style to the messages; non-urgent messages could be presented in a non-obtrusive style. Since adaptation was not possible in the experiment, participants might have selected messages not to be presented at all.
Figure 9 Updated model of acceptability of notifi cations. The subjective acceptability and preferred timing are linked to the attention level and the perceived message urgency. The bold arrows indicate message urgency to be the primary indicator of acceptability and preferred timing attention
level
acceptability of the notification + preferred timing value cost
A second explanation might stem from the experimental setup. Some participants indicated they found it hard to project themselves into the suggested situations. For example, after seeing the message “Reminder: waste paper will be collected tonight”, one participant said waste paper was not collected in his neighborhood at all, therefore he rejected the message.
REfLECTION ON ExPERIMENT DESIGN
In the pilot study, no significant relation between user engagement and acceptability of notifications was found. In the present experiment, the length of the experiment was extended from 1 to 3 evenings per participant, the questionnaire was redesigned by asking users to first rate their degree of engagement before presenting a notification message, and participants were reminded to consider both context and the notification message when assessing acceptability.
The effect of extending the experiment from 1 to 3 evenings per participant can be seen in Table 5 showing a linear regression between Q8 (acceptability) and components C1 (message urgency) and C2 (engagement) per evening. The table shows that the effect of C2 on Q8 decreases in time, from -0.71 on day 1 to -0.51 on day 3, whereas the effect of C2 remains significant. A possible explanation for this reduction in time could be that user engagement was boosted on the first night because of the novelty of the system. On day 2 and 3, when participants get acquainted with the system, the novelty effect diminishes. The changes in user ratings in time underline the need for longitudinal studies, and confirm our choice to extend the experiment.
Table 5 Linear prediction equations for general acceptability based on multiple linear regression, per day and overall. Q8 (general acceptability) could be accounted for by C1 (message urgency) and C2 (user engagement in activities)
In measuring the effect of mental activity load on acceptability, the design of the questionnaire was crucial. In the pilot study, notification messages were shown before users were asked to rate their mental activity load. Also, the presentation of messages was varied; new messages were signaled using alternately a shrill and a soft sound. In that study, we were unable to measure the effect of mental activity load, probably because the notification message biased people in rating their mental activity load.
In the present experiment participants had to rate their activities before the notification message was shown, the presentation style was fixed, and participants were reminded to consider both context and the notification message when assessing acceptability. Whereas in the pilot study no correlation was found between Q8 and C2, in the present study a significant negative correlation (overall: r=-0.326, n=231, p<0.005, day 1 only: r=-0.262, n=77, p<0.05) was found. The redesign of both the questionnaire and the reminder enabled measuring the relation between engagement in activities and general acceptability.
METhODOLOGICAL ISSUES
In general, user ratings in short-term user studies with prototypical technology can be influenced by several artifacts that result from the nature of the study. These artifacts could –in this specific case- be solved by using a realistic system with real messages for a longer period of time. Short-term user studies with prototypical technology, such as the present study, may however guide the development of systems that can be used for longitudinal studies in terms of problem understanding and model construction.
Artifacts that might have influenced the results of the present study include:
The number of notifications was set to an average of two messages per hour. A •
realistic notification system would activate messages based on their availability, which might lead to many notifications in a short time span. Consequently, an oversupply of notifications might result in lower acceptability ratings.
Notifications were only given in the vicinity of the messaging system. Therefore, •
the range of activities in which the user could have been engaged at the moment of interruption was by definition limited. This may have reduced the influence of user engagement on acceptability.
Although asked to treat all messages as authentic messages, participants knew the •
2.6 CONCLUSIONS
The present study reveals that in the home setting user state and context are secondary predictors of acceptability of notifications; message urgency is found to be the primary predictor of acceptability. A cost-benefit approach towards predicting acceptability, in which acceptability is based on the value of the notification message and the cost of interrupting the user, is shown to be a workable approach (Figure 9).
At first sight, it might seem logical to discard sensor-based context-aware systems, and focus on prediction mechanisms for perceived message urgency. However, in order to be able to predict perceived urgency, a system does need to be aware of the context. It therefore seems important to study how contextual cues can help predict the perceived urgency of messages. For example, when people are on the phone, they tend to be highly engaged in their activity, and consequently the perceived urgency of messages tends to be lower. Low-urgency and medium-urgency messages could then be postponed till after the conversation.
In the present study, the acceptability of notifications was examined in the home living context. In measuring the effect of notifications in a realistic environment using realistic user activities, a major challenge is to avoid influencing user behavior. Rather than studying acceptability of notifications in an artificial lab environment using artificial user activities, participants in the present study could do whatever they usually did, and they could experience the notifications in a realistic and natural setting. The questionnaire design proved to be essential for measuring acceptability; by asking participants to rate the user state before showing the notification message, the effect of the message on perceived user engagement in activities could be assessed. Furthermore, the study shows the need for longitudinal user studies, since changes in user experiences –due to for example product novelty- cannot be captured in short, one day experiments.
2.7 fUTURE wORk
A major challenge in the development of future aware home notification systems will be to predict perceived urgency of messages. While in the present study subjective measures were asked directly to the users, an automated system will have to base predictions on objective measures. Perceived message urgency might be related to the message (message structure, phraseology, relationship between messages), the context (user activities, state of the environment) and the user (user values, user state). Additional user studies are needed to capture perceived message urgency and to create personalized prediction models. Studies on mobile interruptibility have shown that profiles for prototypical users can shorten the learning time of a notification system [Kern, Schiele, 2006]; the use of prototypical user profiles for urgency prediction could be studied in a home environment.
The present study was restricted to short-term acceptability. A home notification system could also consider long-term effects when assessing acceptability of interruptions. As an example, think of prevention of repetitive strain injury (RSI). To prevent RSI, a typist should pause regularly to remove tension. The short-term acceptability of interruptions in the primary typing task tends to be very low; people do not like to pause typing. However, in the long run the pauses prevent RSI, resulting in a high acceptability of the pauses. Similarly, a home notification system could for instance induce interruptions in order to reduce stress levels.
Given a system which utilizes the level of message urgency to manage notifications, one could consider using different ambient displays for messages depending upon the classified level of urgency (Chapter 4). A system could display all low urgency messages via a non-intrusive interface in the background, for example a display next to a door. The high urgency messages could be communicated via an attention-demanding alert. The medium urgency messages could then be classified by an intelligent system in order to select the best interface and intrusion level. Studies are planned to investigate the effect of presentation on the acceptability of notifications. Ideally, these studies should be conducted in a realistic setting with real messages over a longer period of time.
In envisioning the ideal home notification system, the goal might be not only to improve acceptability. Rather than creating a system that presents notifications in a way that is acceptable, it might be better to present notifications that people enjoy or feel helped by. Even though the present study contributes to the process of making systems considerate of its users, further studies are needed towards extending the scope of consideration.
whAT’S NExT
The fi eld experiment as described in Chapter 2 showed that people generally want to be informed of urgent messages as soon as possible, wheres non-urgent messages should not be presented at all. Based on the user experiences that were measured in the fi eld, a simple cost-benefi t model of acceptability was constructed.
To improve the acceptability of non-urgent messages, a better understanding was needed of how acceptability of notifi cations is infl uenced by the level of intrusiveness of the presentation, and how acceptability depends on timing in relation to user activities. As described in Chapter 3, an additional user study on acceptability of notifi cations was conducted in a living room laboratory. Ten participant couples were asked to engage in everyday home activities, and to subjectively rate factors that were expected to infl uence acceptability. Notifi cations were presented alternately in the foreground and in the background; this second study on home notifi cations covers both right-hand segments of the implicit interactions framework (fi gure above). The chapter suggests how acceptability of notifi cations could be improved by adjusting the level of intrusiveness of the presentation to message urgency: urgent messages should be presented intrusively, non-urgent messages should be postponed until people are less engaged in their activities.
foreground
background
Chapter 3: Home Notifications
3
CONSIDERATE hOME NOTIfICATION
SySTEMS: A USER STUDy Of
ACCEPTABILITy Of NOTIfICATIONS IN
A LIvING ROOM LABORATORy
An increasing number of products in the home is competing for the user’s attention [den Hartog et al., 2004]. Email notifiers, medicine reminders, washing machines, mobile phones, instant messengers and many other notification generators push information to their users, even though the user-perceived value of some of the messages can be questioned. Considering the growing number of information providers, and the increasing number of messages across products, users might soon be overwhelmed with notifications. As early visionaries of ubiquitous computing, Weiser and Brown (1997) recognized the need for calmness; when computers are all around, these systems need to be designed “so that the people being shared by the computers remain serene and in control”. The shift from functional use and performance to meaningful presence of technology has also been recognized by Hallnäs and Redström (2002); they emphasize the need to design products that co-exist with users and with other products in their life-world. Unfortunately, relatively little knowledge is available on how people experience notifications in the home, even though a large body of knowledge exists in terms of technology that is needed to build intelligent products that are considerate of the user as well as the context of use [e.g., Horvitz et al., 1998; Garlan et al., 2002; Altosaar et al., 2006]. Towards integrating an intelligent notification system into everyday life, an understanding is needed of how people experience notifications during everyday activities, and how notifications affect these activities.
This chapter will be submitted to the International Journal of Human-Computer Studies.
