Beyond Screen Time: Predicting Social Media’s Academic Impact Through Machine Learning and Student Risk Segmentation
, Wanchana Joobanjong, Nathaphan Meemuk, Theerawut TantiathimongkhonThis study investigates how psychological dimensions of social media use affect university students’ academic performance. The focus extends beyond screen time to examine addiction and interpersonal conflict as predictors of academic decline.
Social media now serves as an informal knowledge-sharing platform in higher education. Most existing research uses descriptive statistics and cannot identify which students are at risk. Few studies distinguish between usage duration and the psychological factors that drive different outcomes across individuals.
Secondary survey data from 705 university students were analyzed. Methods included Random Forest classification, SHAP feature attribution, K-Means clustering, and hierarchical clustering. Five-fold stratified cross-validation was used to assess model robustness.
Psychological friction, specifically social media conflict and addiction, predicts academic decline far more strongly than daily usage hours. This challenges institutional policies that focus mainly on restricting screen time. The study also introduces a three-tier student risk segmentation framework to support targeted intervention design.
The Random Forest model achieved perfect classification accuracy (1.00), confirmed across all five cross-validation folds. SHAP analysis identified interpersonal conflict as the dominant predictor (importance = 0.40), followed by addiction score (0.27) and mental health score (0.19). Daily usage hours contributed only marginally (0.05). Clustering revealed three behaviorally distinct student segments: Low-Risk, At-Risk, and High-Risk.
Universities should shift digital welfare policy away from usage restrictions and towards psychological support. Practical steps include early detection systems based on indicators of conflict and addiction, digital literacy embedded in core curricula, and differentiated counseling for At-Risk and High-Risk students.
Future studies should validate this framework across diverse institutional and cultural contexts. Longitudinal designs would help assess causal direction. Federated learning could enable multi-institutional collaboration while protecting student privacy. Integration with Learning Management System logs would also strengthen the feature set.
Universities can treat social media as a manageable educational variable rather than an uncontrolled risk. This is achievable when support systems are built around psychological indicators rather than usage metrics alone. The approach has broader relevance for digital health policy in higher education.
Replication across diverse student populations remains a priority. Platform-level behavioral differences warrant a dedicated study, particularly one that contrasts short-form video platforms with messaging applications. Incorporating physiological stress indicators and longitudinal behavioral data would substantially improve predictive validity.
Back