Computational Analysis of Student Stress on Social Media using Support Vector Machine and Latent Dirichlet Allocation

Abstract

This study develops a two-stage machine-learning framework to identify academic stressors among Indonesian university students using Twitter data. A Support Vector Machine (SVM) classifier was trained on manually annotated tweets and benchmarked against Naïve Bayes, logistic regression, and random forest, achieving an accuracy of 0.91 and a macro F1-score of 0.914, outperforming all baselines. Tweets classified as stress-related with ≥75% confidence were subsequently analyzed using Latent Dirichlet Allocation (LDA), which generated six coherent stressor categories. The framework reveals both structural academic pressures and culturally specific patterns, including references to “dosen killer” and emerging mental-health vocabulary. Contributions include the first Indonesia-focused stressor map derived from large-scale social media discourse and the integration of confidence filtering to enhance topic quality. While results demonstrate the feasibility of social-media–based stress detection, limitations remain regarding temporal drift, annotation bias, and demographic representativeness. Future research should incorporate real-time streaming pipelines, multimodal annotation, and longitudinal evaluation to enhance robustness and early-warning potential.