Scalable and Energy Efficient Software Architecture for Human Behavioral Measurements

Abstract

Understanding human behavior is central to many professions including engineering, health and the social sciences, and has typically been measured through surveys, direct observation and interviews. However, these methods are known to have drawbacks, including bias, problems with recall accuracy, and low temporal fidelity. Modern mobile phones have a variety of sensors that can be used to find activity patterns and infer the underlying human behaviors, placing a heavy load on the phone's battery. Social science researchers hoping to leverage this new technology must carefully balance the fidelity of the data with the cost in phone performance. Crucially, many of the data collected are of limited utility because they are redundant or unnecessary for a particular study question. Previous researchers have attempted to address this problem by modifying the measurement schedule based on sensed context, but a complete solution remains elusive. In the approach described here, measurement is made contingent on sensed context and measurement objectives through extensions to a configuration language, allowing significant improvement to flexibility and reliability. Empirical studies indicate a significant improvement in energy efficiency with acceptable losses in data fidelity.