Constructing and Evaluating Sensor-Based Statistical Models of Human Interruptibility

While people can typically make a rapid assessment of another person’s interruptibility, current systems generally have no way to consider whether an interruption is appropriate. Current systems therefore tend to interrupt at inappropriate times or unduly demand attention. I will present my work examining the feasibility and robustness of sensor-based statistical models of human interruptibility. I have conducted a series of studies to explore the utility of a number of potential sensors, examine human ability to estimate interruptibility, deploy and evaluate actual sensors in office work environments, and develop models of task engagement based on low-level event streams. My studies show that a typical laptop computer, with no additional sensors, can support models that identify «Highly Non-Interruptible» situations significantly better than human observers. Informed and motivated by this result, I have developed Subtle, a tool to enable non-expert development of applications that sense and model human situations.

Outside of my work on human interruptibility, I have a variety of interests in human-centered approaches to machine learning in support of ubiquitous computing applications. I am currently investigating the feasibility of home activity recognition based on low-cost microphone-based sensors placed on a home’s existing water pipes. By providing an unobtrusive and low-cost alternative to installing sensors directly in the living space, this research could enable a variety of practical elder care applications. I am also beginning to investigate anonymous and privacy-sensitive approaches to collecting sensed data in location-based applications, enabling such applications as live traffic monitoring, bus tracking, and conference room availability estimates.