Vector Space Representation of Bluetooth Encounters for Mental Health Inference

Social interactions have multifaceted effects on individuals' mental health statuses, including mood and stress. As a proxy for the social environment, Bluetooth encounters detected by personal mobile devices have been used to improve mental health prediction and have shown preliminary success. In this paper, we propose a vector space model representation of Bluetooth encounters in which we convert encounters into spatiotemporal tokens within a multidimensional feature space. We discuss multiple token designs and feature value schemes and evaluate the predictive power of the resulting features for stress recognition tasks using the StudentLife and Friends & Family datasets. Our findings motivate further discussion and research on bag-of-words approaches for representing raw mobile sensing signals for health outcome inference.