Explainable Machine Learning for Predicting Social Media Addiction: Implications for Digital Well-Being Among University Students

Abstract

Excessive and poorly regulated social media use among university students has been linked to reduced concentration, academic underperformance, and declining digital well-being. However, most predictive approaches to identifying at-risk students rely on opaque, black-box machine learning models that offer little insight into the factors driving their predictions, limiting their trustworthiness and practical adoption in educational settings. This study addresses this gap by developing an explainable machine learning approach to predict digital distraction among university students and examining its implications for digital well-being. Using survey data from 400 randomly sampled university and senior-high-school students in South Sulawesi, Indonesia, a composite Digital Distraction Score was constructed from four core behavioral indicators and dichotomized into High (n = 228) and Low (n = 172) distraction classes via median split. A Random Forest classifier was trained on twelve behavioral, psychological, and demographic predictors, excluding the items used to construct the target variable. Model interpretability was achieved using SHapley Additive exPlanations (SHAP), providing both global and individual-level explanations of model predictions. The model achieved 72.0% accuracy and an AUC of 0.805 on the held-out test set, with stable performance confirmed through 5-fold cross-validation (69.75% ± 2.9%). Feature importance and SHAP analyses consistently identified perceived mental health impact, stress/anxiety when disconnected from social media, and guilt after prolonged use as the strongest predictors of digital distraction—substantially outweighing usage-volume and demographic variables. These findings suggest that digital distraction is primarily an affective-regulatory phenomenon rather than a simple function of screen time, and support the design of digital well-being interventions centered on emotional self-regulation. The study demonstrates that explainable machine learning can serve as a transparent, human-supervised decision-support tool rather than an autonomous diagnostic system.