Limitations of the State of the Art and Open Challenges
Interruptions are an inevitable part of our daily life. As discussed in this book, the effects of disruption caused by interruptions occurring at inopportune moments have been studied thoroughly in the past. Numerous studies have been conducted to investigate the effect of interruptions on users’ ongoing tasks. More specifically, researchers have found that completion time [21, 22, 81], error rate [60], and even emotions toward the ongoing tasks [2, 7] are adversely affected by interrupting users at inappropriate moments. However, studies have also provided evidence that in order to not miss any newly available important information, people tolerate some disruption [52].
Since the era of desktops, managing interruptions has been a key theme in Human-Computer Interaction research [3, 43, 44, 49]. With the advent of mobile and wearable technologies, the problem of managing interruptibility has become even more pressing as users can now receive notifications anywhere and at anytime. Indeed, numerous research efforts have been carried out with the goal of designing interruptibility management system for mobile environments [40, 52, 70, 75, 83, 85]. Most of the work on mobile interruptibility emphasizes the exploitation of features that can easily be captured through mobile sensors, such as task phases [29, 40, 52], users’ context including location and activity [26, 85, 91], and notification content [70, 75] to infer the right time to interrupt. More specifically, studies have shown that notifications are considered more positively and received a faster response when delivered while a user switches from one activity to another [29, 40]. On the other hand, studies have also demonstrated that machine learning algorithms can learn about users’ interruptibility by exploiting passively sensed contextual information and notification content [75, 85]. Also, studies have demonstrated that the contextual information can be exploited to infer and filter out the irrelevant information from being delivered [70]. Furthermore, a handful of recent studies have investigated the feasibility of predicting the right medium to deliver notification for managing interruptibility [73, 104]. The studies have shown that users’ behavior in terms of handling notifications in multi-device environment can be modeled by exploiting the contextual and phone interaction data [73].
Consequently, existing studies have focused on various challenges concerning the understanding and learning users’ behavior in terms of interactions with notifications. However, the characterization of attentiveness and receptivity of users for mobile notifications is still an open problem. We believe that there is still a considerable scope for improvement, for example by exploiting other physical, social and cognitive factors for modeling users’ notification interaction behavior. For instance, more knowledge about users’ cognitive context could help the system to reduce the amount of notifications delivered to users when they are stressed [74]. We now summarize some key open questions in the area that must be investigated to build intelligent mechanisms that could effectively trigger the right information in a given context.
Let us start from a key open question in this area: should we defer a notification if it is not delivered at an opportune moment and for how long. Until now, interruptibility management studies have focused on inferring if the current moment is opportune to deliver notifications or not. We believe that in order to be more effective, these systems should not just predict users’ current interruptibility, but if the current time is not an opportune one, it should also anticipate the best moment in the nearest future [79]. This would enable the overlying application to decide whether the notification should be deferred until the predicted opportune moment or not.
A recent study has found that users tend to defer notifications related to people and events [103]. They also found that user decision for deferring notifications are influenced by their daily routines. We believe that such findings and prediction techniques similar to the ones employed for other interruptibility management systems could be exploited in order to address this challenge.
7.2 MONITORING COGNITIVE CONTEXT
Previous approaches for interruptibility management focus on exploiting users’ physical context. However, as discussed in this lecture, users’ interruptibility might also be associated with their cognitive context. For instance, previous studies have already have already reported that users’ interruptibility is influenced by cognitive factors such as their mood [78], their engagement with the current task [84], and complexity of the interrupting task [77]. However, none of the studies have exploited this information to build an interruptibility prediction model.
We believe that this limitation exists due to the fact that current mobile and wearable technologies are yet not capable of passively sensing such information. So, there is indeed a need for developing and evaluating mechanisms for automatically capturing the level of users’ engagement with the current task, complexity and difficult of execution of the interrupting task, and similar cognitive factors that might influence interruptibility. One of the potential approaches might be to explore the use of affective computing [87] in order to monitor users’ emotional states.
7.3 LEARNING “GOOD” BEHAVIOR: INTERRUPTIONS FOR POSITIVE BEHAVIOR INTERVENTION
Until now, all interruptibility studies have focused on learning the observed user behavior associated with the sensed contextual information. However, interruptibility management system can also be considered as a key component for behavior change intervention tools that could help prevent and modify harmful behavior of users [59]. In other words, a potential future direction of these systems could also be to gather the knowledge about good behavior and exploit it to improve the behavior of users. Such knowledge about good behavior for interacting with notifications could potentially be obtained by carrying out a large-scale ESM-based study that can query users about the ideal notification-interaction behavior in their current situation. However, there is an inherent problem related to learning “good” vs. “bad” behavior. The problem is inherent in the fact that a machine learning might not distinguish between a behavior that should be promoted and one that should not. As an example, let us consider a learning component that is able to learn the right moment to interrupt by past experience. If upon receiving a notification a user reads emails on their mobile phone while driving, the notification mechanism should not learn this behavior and deliver emails accordingly. Instead, the mechanism should infer that it is a harmful behavior to read notifications while driving and try to avoid sending unnecessary emails. Indeed, if the information is critical it should be delivered immediately regardless of the current situation. However, designing such an ideal notification delivery mechanism is extremely difficult and have never been considered in the scope of any interruptibility study.
7.4 MODELING FOR MULTIPLE DEVICES
A previous study has shown that users’ prefer to receive notifications on specific devices based on their situation [104]. Inline to this work, another study has proposed a solution for modeling users’ behavior in terms of handling notifications in multi-device environment by exploiting the contextual and phone interaction data in order to deliver notifications on the right device. However, they conducted a pilot study by considering only two devices: mobile phones and alternatives (i.e., any device other than phones).
Consequently, the aspect of delivering information through the right medium is not well studied. Almost all of the previous studies have not focused on predicting users’ behavior on receiving cross-platform notifications, which are delivered on multiple devices at the same time. In other words, there is a lack of understanding of the features that determine users’ receptivity to such notifications on a specific device in a given context. Given the fact that users are surrounded by an increasing number of devices that are able to receive a notifications (such as apps in laptops, mobile phones, wearables, smart television sets, and appliances), the design of such mechanism is an open and interesting research area.
7.5 NEED FOR LARGE-SCALE STUDIES
Almost all studies in the area of interruptibility management are conducted with small samples of the population and for short time periods. At the same time, these studies are often publicized through the network of the researchers performing the studies. This could introduce a bias deriving from the self-selected sample of users and thus the behavior of a certain group of user (within a network) might be different from others. Moreover, the validation of the interruptibility prediction mechanisms in these studies is usually performed in an offline fashion (i.e., the evaluation is performed on the collected data a posteriori). For this reason, the results presented in these studies do not have ecological validity as the collected datasets might be biased toward a certain group of population.
Therefore, we believe that there is a need for large scale studies as well as in-the-wild deployments [19, 95] to guarantee the ecological validity and robustness of the proposed interruptibility prediction mechanisms. We also believe that reproducing these studies in different social context and users’ demographics is also essential.
We believe that the open challenges presented in this chapter have useful implications for designing the future interruptibility management systems. In the next chapter, we will summarize the content of this lecture and hope they are useful for realizing an ideal solution of the interruptibility management system.