Detecting Perceived Unfair Treatment Among US College Students Using Mobile Sensing: Pilot Machine Learning Study

Experiences of unfair treatment on college campuses are linked to adverse mental and physical health outcomes, highlighting the need for interventions. However, detecting such experiences relies mainly on self-reports. No prior research has examined the feasibility of using mobile sensing via smartphones and wearables for the passive detection of these experiences.

This pilot study explores the potential of using passive sensing to detect daily experiences of perceived unfair treatment (PUT) after they occur. It aims to develop and evaluate machine learning models against naive baselines and establish a benchmark for future research.

We analyzed data from 201 undergraduate students collected over two 10-week academic terms in 2018. PUT was self-reported at the daily level via ecological momentary assessment (EMA) surveys, with 413 of 9629 (4.3%) total responses indicating unfair treatment. We implemented two modeling approaches with distinct training schemes: (1) supervised classification models trained in a user-independent manner using data from different individuals, and (2) anomaly detection models trained in a user-dependent manner using historical data from the same individuals. Classification performance was assessed using stratified group 5-fold cross-validation for user-independent models and a chronological train-test split for user-dependent models.

Of the 201 study participants, 110 reported experiencing unfair treatment at least once. On average, participants reported unfair treatment in 4.66% of their EMA responses (95% CI 3.13% to 6.19%). User-independent classification models showed mixed performance (AUC-ROC [area under the receiver operating characteristic curve]: 0.546-0.640, AUC-PR [area under the precision-recall curve]: 0.047-0.093, F₁-score: 0.070-0.121). Tree-based models, particularly light gradient boosting machine (LightGBM) and Random Forest, outperformed all 3 baselines in AUC-ROC and AUC-PR; LightGBM also improved the F₁-score. In comparison, user-dependent anomaly detection models performed better, with the multiday long short-term memory-AE model (50 features, 7-day window) achieving the highest recall (0.830, +73.3%, P<.001) and F₁-score (0.391, +24.9%, P<.001) without reducing precision (0.256), and improving AUC-PR by 45.9% and AUC-ROC by 21.6% relative to naive baselines (P<.001). Feature importance analysis identified key behavioral patterns for population-level detection, including increased time spent off campus, elevated evening and nighttime activity, reduced indoor mobility on campus, prolonged screen use, delayed sleep onset, and shorter sleep duration.

Mobile sensing shows promise for detecting daily experiences of PUT in college students and identifying associated behavioral patterns. Our findings highlight opportunities for timely interventions through mobile technology to mitigate the impact of these experiences on students' mental health and well-being.