<![CDATA[Adolescent depression is an increasingly alarming global mental health issue, particularly amid the rapid adoption of social media platforms. This study compares the performance of four machine learning algorithms Random Forest, XGBoost, Support Vector Machine (SVM), and Logistic Regression, in classifying adolescent depression risk using 12 digital behavioral and lifestyle features. The dataset comprises 1,200 records with severe class imbalance (97.4% vs 2.6%), addressed through Synthetic Minority Over-sampling Technique (SMOTE) with k=5. Eight experiments (four algorithms × two conditions) were conducted and evaluated using Precision, Recall, F1-Macro, AUC-ROC, and 5-fold cross-validation. Results demonstrate that SMOTE consistently enhanced all model performances, with an average F1-Macro improvement of 25.7%. XGBoost achieved perfect F1-Macro and AUC-ROC scores (1.0000) under both conditions, while SMOTE yielded the greatest gains for SVM (+67.9%) and Logistic Regression (+63.7%). Feature importance analysis identified sleep hours, stress level, anxiety level, and daily social media usage as dominant predictors, while academic performance contributed negligibly. These findings underscore the critical importance of imbalanced data handling and provide a methodological framework for developing digital behavior-based early depression detection systems.]]>
Copyrights © 2025
