Abstract
Cyclic human behaviors, such as circadian rhythms, have significant implications for health outcomes. These rhythms have primarily been studied in laboratory and controlled settings with small and sparse amount of data collected in short periods. With the advancements in mobile and wearable devices, it is possible to collect longitudinal, continuous, and fine-grained biobehavioral data from individuals in the wild. While this data enables rigorous modeling of cycling human behavior, existing time series modeling techniques are insufficient for this task as they assume that the input time series is strictly stationary, are unable to independently learn non-dominant cycles, require the number and types of cyclic patterns, and cannot process massive sensor data. This dissertation presents a novel computational framework for modeling cyclic human behaviors from multi-modal mobile sensing data. The developed framework is designed to be adaptable in processing data with diverse time granularity and is able to automatically detect and model cyclic human behaviors from multi-modal mobile sensing data. Moreover, it can model similarities and differences in rhythms across individuals and time, predict various outcomes related to well-being and health, identify activity sub-patterns for a single person or across a population, and facilitate an interpretable description of human behavior. Importantly, this is the first framework that addresses the variability of cyclic human behaviors using non-stationary time series. To evaluate the effectiveness of the proposed framework, both open-source and self-collected mobile sensing datasets are used. Through this evaluation, the framework's ability to process multi-modal mobile sensing data and accurately model cyclic human behaviors is demonstrated. Overall, this work has the potential to significantly advance the understanding of cyclic human behaviors and their relationship to health outcomes in real-world settings.
