Introduction
Behavioral sensing—the use of mobile and wearable sensors to track human behavior and well-being—is increasingly utilized in mental health and education. However, traditional fairness evaluations often focus on broad demographic categories like age or gender, overlooking more nuanced "context-induced harms." This research identifies two critical gaps: situated identity-based harm (biases against identities relevant to specific environments, such as first-generation students) and situation-based harm (performance drops caused by physical or technological settings, like room temperature or device type).
A Six-Step Framework for Responsible Sensing
To address these gaps, the researchers developed a human-centered, iterative framework designed to identify and mitigate harms throughout the sensing lifecycle. Unlike generic machine learning audits, this framework emphasizes user engagement and longitudinal monitoring.
The framework consists of:
- Comprehensively understanding the context through user engagement.
- Establishing evaluation criteria using specific metrics like False Negative Rate disparities.
- Evaluating data inclusivity to ensure training sets reflect diverse contexts.
- Assessing algorithm sensitivity to contextual changes.
- Conducting harm evaluation by combining quantitative metrics with qualitative user feedback.
- Continuous maintenance to update models as environments evolve.
Case Studies and Methodology
The framework was validated through post-hoc audits on two real-world datasets:
- GLOBEM: A four-year longitudinal dataset of ~500 college students tracking depression via passive sensing (steps, sleep, phone use).
- En-Gage: A dataset of 23 high school students capturing physiological and environmental data (temperature, CO2) to predict classroom engagement.
Key Insights: Uncovering "Unseen" Biases
The study revealed significant disparities that traditional audits miss:
- The "Unseen" Bias: In depression detection, algorithms showed significant performance disparities for first-generation students and immigrants—groups rarely included in standard fairness checks.
- Environmental Inequity: In the engagement study, models were significantly less accurate for students in uncomfortable thermal environments. For example, the Mean Squared Error (MSE) for emotional engagement was $0.822$ for students who felt too cold, compared to $0.631$ for those who were comfortable. This suggests that students with less access to climate-controlled environments may be unfairly penalized by automated monitoring.
- The Fairness "Whack-a-Mole": Efforts to mitigate bias for one group (e.g., first-generation students) often introduced new biases against others (e.g., non-male or non-heterosexual participants). This "bias shift" highlights the complex trade-offs inherent in optimizing for fairness.
Conclusion and Future Work
While the framework successfully identifies context-induced harms, the researchers note limitations regarding data scarcity for small sensitive groups (e.g., students with disabilities) and the potential for harms arising from user interface (UI/UX) design. Future work will focus on validating the framework in "live" deployments and expanding its scope to include privacy and accountability dimensions in behavioral sensing.
Summary updated