Towards Human-Centered Behavioral Sensing for Student Support

Abstract

Behavioral sensing technologies hold significant potential to enhance human support systems, especially in high-stakes domains such as education and mental well-being. However, existing research often prioritizes model performance, limiting its ability to address real-world needs, adapt to diverse populations, or surface potential risks. Focusing on student support, this dissertation advances a human-centered approach to behavioral sensing by deepening the understanding of students' needs, examining potential harms embedded in current practices, and developing behavioral models that align with Human-Centered Machine Learning (HCML) principles. This work begins with a six-year longitudinal study that combines passive sensing and self-reported data to capture the everyday academic and mental well-being experiences of college students. This is followed by a mixed-method study examining how the transition to online learning during COVID-19—a major life event that disrupted routines, support systems, and access to resources—impacted students with disabilities and mental health concerns—revealing their needs from the onset of the pandemic through the following academic year. These empirical findings motivate a deeper investigation into the ethical risks and fairness concerns surrounding behavioral sensing in real-world deployment. Through both quantitative and qualitative studies, we provide evidence of algorithmic bias embedded in existing behavioral models and uncover broader ethical challenges across the behavioral sensing lifecycle. This work highlights the unique nature of fairness in behavioral sensing, identifies stage-specific vulnerabilities, and surfaces systemic barriers to fair and accountable system design. Drawing on these insights, we propose a reflexive fairness framework and actionable guidelines to support more ethically aligned development and evaluation practices. Informed by these insights, we develop and evaluate three predictive modeling approaches that identify at-risk students as early as the first week of an academic term. These approaches integrate fairness, interpretability, and generalizability as core design objectives. Our findings demonstrate the feasibility of operationalizing HCML in behavioral modeling, while also highlighting key trade-offs and design tensions that emerge in practice. Finally, we discuss open challenges and future directions for building responsible behavioral sensing systems that are not only technically robust but also ethically grounded, inclusive, and attuned to the lived realities of those they aim to support.