Computational Support for Longitudinal Well-Being

Abstract

As artificial-intelligent-powered devices have become more embedded in our lives, they offer an unprecedented ability to passively sense our daily behavior at a high resolution. These everyday devices are already equipped with machine learning techniques to monitor our basic health behaviors, such as physical activity and heart rate, and provide suggestions accordingly. However, they are still far from understanding our high-level, longitudinal behaviors, such as mental well-being. Early research about longitudinal behavior modeling and intervention is still facing a set of deployability challenges before being ready for real-world deployment. For behavior modeling, these challenges include interpretability (revealing human-readable insights about behavior), personalization (adapting models to every individual), and generalizability (ensuring models work robustly on new users and contexts). Furthermore, the results and insights of behavior models need to be connected with intervention techniques to influence users’ behavior and improve their well-being. With mental well-being as the main application, my research is targeted at these deployability challenges by (1) collecting and releasing the first multi-year passive sensing datasets, (2) developing new behavior modeling techniques that are interpretable, personalized, and generalizable, and (3) designing and deploying a novel intervention technique based on behavior models’ insights to improve user well-being. Combining these efforts, I propose the vision of “computational longitudinal well-being”, where interactive systems based on everyday devices can precisely and robustly understand, model, and influence long-term human behavior for better health and well-being.