This chapter presented two studies for investigating the factors influencing users’ atten- tiveness and receptivity to mobile notifications.
The first study focussed on mobile interruptibility, specifically on the identification of factors that make an interruption disruptive and the impact on the response time to a notification. The contributions of this study are twofold. First, we have confirmed the validity of some past desktop interruptibility studies [CCH00b, CCH00a, CCH01] to show that ongoing task’s complexity and completion level influences the perceived inter- ruption in a mobile setting. Second, for the first time, we have investigated the role of notification presentation, sender-recipient relationship and emotional states for modelling interruptibility. More specifically, the key contributions are listed in Boxes 3.1, 3.2 and 3.3. Through a mixed method of automated smartphone logging and ESM sampling we have
obtained a dataset of in-the-wild notifications and ESM reports on notification perception from 20 users. We have analysed the data to show that the response time of a notification in the mobile environment is not only influenced by an ongoing task’s type, completion level and task complexity, but also by the notification’s alert modality, presentation and sender-recipient relationship. Our results have shown that the presentation of a notifica- tion and its alert type, as well as the type, completion level and complexity of a task with which the user is engaged, all impact the seen time. Moreover, the relationship with the sender influences the user’s decision on accepting a notification or not. Finally, the data also reveals how the sentiment (i.e., perceived disruption) towards a notification varies with the type, completion level and complexity of an ongoing task and the recipient’s relationship with the sender.
In the second study we have performed a causality analysis between users’ emotional states and their interaction with notifications. We collected 5118 responses to question- naires for logging users’ emotional states (activeness, happiness and stress) from 28 users over a period of 20 days. First of all, using a non-parametric correlation test (Kendall’s Rank) we have shown that the users’ activeness level has a significant association with the seen and decision time of notifications that arrive in the next hour. Then, we have conducted an in-depth causality analysis considering a variety of contextual variables as confounders to show that in stressful situations people become more attentive that results in the reduction of notification response time.
CHAPTER 4
PREDICTING OPPORTUNE MOMENTS TO
DELIVER NOTIFICATIONS
4.1
Overview
Users’ interaction with mobile notifications is indeed extremely complex and depends on numerous aspects. Fortunately, some of the aspects can be captured, and above all, quantified, by means of the embedded sensors with which modern mobile phones are equipped. These sensors allow a mobile application to collect information about users’ day-to-day activities [CMT+08, MPM14], preferences [XLL+13], and the surround- ing environment [LPL+09]. Past studies have investigated the use of some of the sensed information to infer opportune moments for interruptions, i.e., moments when a user quickly and/or favourably reacts to a notification [FGB11, PM14b]. More specifically, some important contextual factors that have been used to infer interruptibility include transitions [HI05], engagement with a mobile device [FGB11], and, more generally, time of day, location and activity [PM14b].
However, until now, the focus has been on the context in which a notification has been received and not on the actual content of the notification. After all, not all the notifications are disruptive [ML02]. It is the relevance of the interruption content in the recipient’s current context that partly defines the disruptiveness of an interruption. For
example, a chat notification from a friend can be extremely disruptive if delivered during a meeting. But, an email notification from a project collaborator might be acceptable to the user, and in some cases, considered very useful in the same context.
In this work we discuss how content and context can be used together in order to de- sign intelligent non-disruptive notification mechanisms. More specifically, we investigate how users behave when they receive specific types of content through mobile notifica- tions arriving at different times, and in different contexts. Unlike previous studies such as [PM14b], we do not restrict ourselves to context information provided by mobile sensors only. Instead, to the best of our knowledge, for the first time we also take into account the type of information delivered and the social relationship between the information sender and receiver. It is worth noting that our work is predicated on the hypothesis that the acceptance of a mobile notification depends on what, i.e., the type and the origin of the information contained in a notification and where, i.e., the user’s context in which the notification is delivered. We will consider a notification as accepted if it is handled (i.e., clicked to launch the respective application) by the user within a certain time from its arrival.
4.1.1
Key Contributions
The findings of our analysis can be summarised as follows:
• user’s acceptance of a notification depends on the content and the originator of the notification and, not surprisingly, the time needed to respond to a mobile notification varies according to the user’s physical activity level.
• notification content, together with the sensed context, can serve as a basis for the design of machine learning classifiers that infer a user’s response to a notification. • automated inference of opportune moments to interrupt, as with the above clas-
sifiers, outperforms subjective rules with which our subjects described their inter- ruptibility.
ing fashion, achieving a stable state (i.e., more than 60% precision) after nine days of training.
We conclude that inferring the right moment for interruption should be done in a holistic manner, jointly taking into account the context of the recipient and the notification content. We show that a prediction model trained on a user’s personal data performs far better than a generic prediction model trained on multiple users’ data. Finally, we point out the importance of an automated approach, as humans remain inefficient at formalising and communicating their preferences for mobile notifications in terms of predefined rules.