Trading energy for accuracy in mobile interruptiblity inference

Untimely interruptions from our mobile devices may have a significant impact on our work performance, stress and well-being, and in critical situations, such as when driving, can even have fatal consequences. State of the art approaches to inferring interruptiblity of mobile users harness an array of sensors available on our devices. Yet, the energy consumption of these sensors clashes with the need to preserve the most precious of the device's resources - its battery charge. In this work we revisit the sensor-based approach to interruptiblity inference and examine the trade-off between a sensor's energy use and its contribution to interruptiblity modelling. Our findings, based on a two week long field study with 14 users demonstrate that turning on additional sensors indeed improves interruptiblity inference, but at a cost of increased energy consumption. We then propose an interruptiblity management systems that uses the classifier confidence as a knob allowing fine-grain tuning along the trade-off front, thus enabling user- and application- specific energy-optimal interruptiblity management.