A Data-Driven Approach to Combating Social Media Addiction: Insights from Explainable AI

Abstract

Social media has transformed the way people communicate and connect, but it has also raised worries about psychological dependence and excessive use of technology. This paper proposes an interpretable machine learning framework for predicting social media addiction (SMA) using combined behavioral and psychological data collected from 10,780 respondents. To address class imbalance, ten classifiers were assessed using stratified cross-validation, LASSO-based feature selection, and SMOTE. Experimental results demonstrate that Logistic Regression (83.33%), Gradient Boosting (82.70%), SVM (82.60%), and AdaBoost (82.60%) produced the best dependable outcomes in terms of accuracy and balance. PCA visualization showed clear behavioral patterns, and SHAP and LIME were used for both global and individual-level explanations to increase transparency. Additionally, the causal influence of daily social media use on addiction risk was evaluated using Conditional Average Treatment Effects (CATEs), DoWhy, backdoor and frontdoor modifications, and Directed Acyclic Graphs (DAGs). In contrast to earlier XAI-based research, the suggested approach combines causal, local and global interpretability to convert black-box predictions into useful behavioral insights.

Ähnliche Arbeiten