CAPÍTULO 3: IMPLEMENTACIÓN SOBRE UN SISTEMA
3.4 VERIFICACIÓN SOBRE EL SISTEMA HETEROGENEO
3.4.2 Adaptación del kernel
We want to point out multiple opportunities for future research, including applications of the interruptibility model, increasing its accuracy and applicability, and investigating new ways to reduce interruption cost for both the interruptee and interrupter.
Interruptibility Model in Practice. Based on the findings from RQ 1 and RQ
2, an obvious next step is to use the determined interruptibility model from RQ 1 for the FlowLight (RQ 2 ). While the simple algorithm of the FlowLight
was already accurate enough to reduce interruption cost, the model developed in RQ 1b is based on more detailed data from computer interactions and also biometric measurements and thus can be more accurate for a broader range of activities. These activities include situations where a knowledge worker is highly focused but not interacting actively with the computer, such as times spent understanding source code or reading a document, or working with pen and paper. The general interruptibility model trained across multiple participants can be used to achieve a high accuracy without the need of an initial training phase which solves the cold-start-problem. In addition, we could incorporate a feature to collect interruptibility ratings for a few days to train it to the individual and further increase accuracy. Given the variety of sensors used in our study from more to less physically and privacy invasive, we can further take into account the users’ preferences on which sensors and features to use. For instance, users might prefer using a biometric tracker over a computer interaction tracker or vice versa.
While the major focus of the FlowLight is to reduce in-person interruptions, the interruptibility model could also be used to support the handling of computer- based interruptions, such as emails or instant message notifications. Most existing work has focused on finding naturally occurring task boundaries to display
1.6 Opportunities and Future Work 23 interruptions (e.g. [Iqbal and Bailey, 2008]), while our model would provide a more continuous interruptibility measurement and only display interruptions when the interruptibility level is above a certain threshold. Such a continuous measurement would allow deferring computer-based interruptions not only to task boundaries but to other times when interruption cost are particularly low, and addressing computer-based and in-person interruptions at the same time.
External interruptions, either coming from in-person interruptions or from notifications at the computer, only make up half of the interruptions that a knowledge worker experiences in a day. The other half stems from self- interruptions, such as switching to check a news website [Czerwinski et al., 2004]. These self-interruptions can also have a big impact on the developer’s focus and performance. Since our interruptibility measurement is presumably also related to focus, we might be able to use it to reduce the cost of self-interruptions and better support developers in their work. For instance, by automatically detecting when a knowledge worker’s focus is decreasing, we might be able to intervene, e.g. by reducing distracting content on the screen that might cause self-interruptions or by suggesting to take a break. Furthermore, by knowing when a knowledge worker is more or less focused during the day, we might be able to optimize the work day by scheduling highly demanding tasks during times of high focus.
Increasing the Applicability and Accuracy. In our work, we focused on soft- ware developers as one coherent group of knowledge workers. We see great potential in targeting the sensors and features towards the specific kind of work of the users’ job roles, possibly increasing the accuracy of the model. As an example, in our work, we added the time spent in activities related to soft- ware development as a feature for our interruptibility model. Future research could investigate meaningful features for other job areas, e.g. designers or other engineers.
Further, we focused on sensing interruptibility while a knowledge worker is working with the computer or close by. Yet, to cover a broader range of activities, such as having discussions with colleagues or while interacting with other devices, integrating additional data sources can potentially increase our
model’s accuracy. Data sources such as interactions with mobile devices, audio or location logs can be valuable and have already been explored in sensing interruptibility (see [Turner et al., 2015] for a review). Obviously, further data sources also add privacy concerns and future research could focus on evaluating costs, benefits and predictive power of each data source and integrate valuable sources into one holistic model.
Our research serves as a good starting point evaluating a variety of biometric sensors in office workplaces. As new and improved biometric sensing technologies are emerging frequently given the fast advances in their development, it might be possible to collect and use additional biometric data in the field and over a prolonged time period. As an example, using attentive states detected with EEG data can be valuable additional features in the prediction of interruptibility, increasing its accuracy even more. Additionally, the capabilities of regular devices to sense biometric features are growing, e.g. web cams have been used to predict emotions based on facial expressions [Bahreini et al., 2016] and cognitive load based on pupil dilation [Samara et al., 2017].
Further Ways to Reduce Interruption Cost. Our approach helps to protect the interruptee from interruptions at inopportune moments. However, after the interruption, it might still take some time to refocus and remember the context of the suspended task. A potential research direction could therefore be to leverage computer interaction data such as recently used programs, files, or websites to summarize or highlight relevant contextual information of the suspended task when the user returns from an interruption. Such an aid can potentially reduce the resumption lag—the time to resume the primary task—and thus decrease the cost of interruptions even more [Iqbal and Horvitz, 2007, Rule et al., 2015].
Further, researchers have shown that additional to the interruptee’s interrupt- ibility state, other characteristics such as the urgency, importance and context of both the primary and interrupting tasks are important factors to find optimal moments for interruptions [Arroyo and Selker, 2011, Grandhi and Jones, 2010]. In our approach, the interruptee and interrupter are aware of the interruptee’s in- terruptibility state, but still need to assess these additional factors by themselves
1.7 Background and Related Work 25