Inferring Multitasking Breakpoints from Single-Task Data

Inferring Multitasking Breakpoints from Single-Task Data Peter Bogunovich (pjb38@drexel.edu) Drexel University, Department of Computer Science Philadelphia, PA, USA Dario D. Salvucci (salvucci@cs.drexel.edu) Drexel University, Department of Computer Science Philadelphia, PA, USA Abstract Recent research has shown that computer users placed in a de- ferrable multitasking situation generally postpone secondary- task interruptions until points of low mental workload in the primary task. Studies examining this phenomenon have re- lied on empirical data that explicitly show user switch points in the course of multitask performance. This paper addresses a related question: Can these same switch points, found em- pirically in a multitasking context, be inferred solely from single-task data? We investigate this question and propose an approach that analyzes a particular behavioral signature in single-task data—outliers in the distributions of time between task actions—to infer multitasking breakpoints. We evaluate this approach using behavioral data from a user-interface task, showing how the proposed method’s inferences from single- task data match well to the real switch points observed during multitask performance. Keywords: Multitasking; task analysis; data analysis. Introduction Multitasking is a concept that is familar to most computer users. It is not uncommon for a user to switch computing tasks every few minutes. In many cases switching is ini- tiated by an interruption of the current task. For example, a notification of a newly received email may appear on the screen prompting a user to stop what he is doing and look at his email before continuing his previous task. Research has shown that interruptions can increase the overall time spent on a single task. One important source of this increase is the resumption lag, or time required to switch back to the task and resume after the interruption has been addressed (Trafton, Altmann, Brock, & Mintz, 2003; Monk, Boehm-Davis, Ma- son, & Trafton, 2004). Recently it has been shown that it is more beneficial to interrupt at certain points than at others (Adamczyk & Bailey, 2004; Bailey & Konstan, 2006; Cutrell, Czerwinski, & Horvitz, 2000). One particularly strong result states that the performance loss associated with interruption is reduced when interruptions occur at points of low mental workload (Iqbal & Bailey, 2005). This result has obvious im- portance when considering forced interruptions in which the user is required to address the interruption immediately be- fore moving on with the primary task. The relationship between mental workload and interrupt- ibility has been strengthened in further studies of deferrable interruptions (Salvucci & Taatgen, 2010) in which a user is notified of a secondary task but the user can defer processing of this task until a later (presumably more comfortable) time. For example, it has been shown (Salvucci & Bogunovich, 2010) that in this situation users tend to defer switching tasks until a point where there is a drop in mental workload. As ex- emplified by these studies, a detailed analysis of when users switch tasks is critical to a deeper understanding of human multitasking behavior. A particular goal in this line of re- search involves the prediction of breakpoints, the points in a task sequence where the user can most conveniently switch tasks. One approach to breakpoint prediction combines expert coding, feature detection and model prediction (Iqbal & Bai- ley, 2007). This approach begins by observing users in some natural multitasking environment. An expert manually exam- ines user actions and identifies specific features which appear to signal breakpoints. A statistical model is then developed based on these features. Promising results have been ob- tained with his method, however it requires the human coders to identify the perceived breakpoints and features, and does not necessarily make use of the relationship between cogni- tive load and interruptibilty. A successful related approach that makes use of mental workload is to examine the typi- cal execution structure of an action in advance and use this structure to estimate opportune breakpoints (Bailey, Adam- czyk, Chang, & Chilson, 2006). This method still requires expert analysis and it may fail when variation in strategy is introduced. There exists a well-known relationship between cognitive load and pupil dilation (Beatty, 1982). Researchers have made use of this link in another approach to breakpoint detec- tion (Bailey & Iqbal, 2008). In this approach, pupil dilation data is recorded as users perform a task, and subtask bound- aries, where there is an assumed drop in cognitive load, are estimated by changes in dilation. The result is a more general and more automatic estimation of good potential breakpoints that relies less on pre-computed models or experts. Despite these findings, it may not be possible to obtain pupil-dilation in practice for many tasks. In this paper we attempt to infer multitasking breakpoints in a automatic, data-driven manner. In this respect our ap- proach is most similar to (Bailey & Iqbal, 2008), but in- stead of relying on typically inaccessible equipment like eye- trackers, our goal is to come up with the good estimates us- ing only data logs of system events generated by users per- forming a single primary task. Our analysis focuses on the distributions of elapsed time between recorded event pairs, using single-task data collected for a customer-support task