Abstract
Mobile devices such as smartphones, smartwatches, and other wearable devices are equipped with a rich set of sensors that can collect human behavioral and physiological data continuously and unobtrusively. Data collected by using the embedded sensors (e.g., accelerometer, GPS sensor, and Bluetooth sensor) in mobile devices has been leveraged in a plethora of healthcare-related fields, including but not limited to physical state inference, mental health monitoring, and mobile intervention. Despite the recent achievements and advancements in mobile health (mhealth), wide adoption of mhealth remains a challenge. First of all, handcrafted feature engineering and conventional deep neural networks (e.g., Multi-Layer Perceptron, Convolutional Neural Network) are restricted to generating sufficient representations of raw mobile sensing data, making it difficult to capture complex interdependence within human behaviors. Secondly, complete responses of high-frequency ecological momentary assessments (EMAs) in the wild are impractical due to heavy user burden and low user engagement, resulting in sparsely annotated mobile sensing data at different levels of granularity. Last but not least, current centralized training of machine learning models can expose sensitive information of mobile users to privacy risks due to data breaches and misexploitation, preventing widespread use of mobile sensing. This research demonstrates a set of deep graph learning systems to overcome the above mentioned challenges and presents a state-of-the-art modeling paradigm in mobile sensing from a topological perspective.
