Machine Learning Analysis of Social Media Usage Patterns and Mental Health Indicators

Abstract

The rapid growth of social media has created new opportunities for connection but has also raised concerns about its impact on mental health. This study investigates how demographic factors, digital behaviour, and self-reported psychological indicators jointly relate to users’ mental states. Using an open dataset of 5000 social media users, we analyse numerical and categorical variables including age, gender, daily screen time, social media time, counts of positive and negative interactions, sleep duration, physical activity, anxiety, stress, mood, and a three-level mental-state label (Healthy, At_Risk, Stressed). Descriptive statistics and correlation analysis show that longer daily screen and social media time are strongly associated with higher stress and anxiety and lower mood, while sleep and physical activity display the opposite pattern. K-means clustering applied to combined behavioural and psychological features reveals three coherent user profiles that align with the Healthy, At_Risk, and Stressed categories, highlighting a clear gradient from balanced to high-risk digital lifestyles. A decision-tree classifier trained only on behavioural features (excluding anxiety, stress, and mood to avoid target leakage) achieves an overall accuracy of about 97% on a held-out test set and provides interpretable if then rules linking specific usage patterns to mental states. The results emphasise that intensive, unbalanced social media use especially when coupled with reduced sleep and low physical activity is strongly linked to adverse mental-health outcomes, and they illustrate how simple machine-learning models can support early risk detection based on non-intrusive behavioural data.