A Two-Track Study of Interpretable Machine Learning and Efficient Transformers for Student and Social Media Depression Analysis

Abstract

Depression among college-age populations is frequently studied through two disconnected lenses: structured survey instruments that quantify academic and lifestyle stressors, and unstructured social-media narratives that surface the same distress in the participants' own words. This paper connects the two lenses in a single, reproducible pipeline. In the first track, we revisit the Kaggle Student Depression dataset (27,901 rows, 20 attributes) and benchmark a Logistic Regression baseline against XGBoost, a Multi-Layer Perceptron, a one-dimensional Convolutional Neural Network, and a Long Short-Term Memory network, with class balance reinforced through CTGAN-based synthetic augmentation. In the second track, we move from tabular indicators to raw narrative text and fine-tune and compare four parameter-efficient transformer encoders — DistilBERT, ALBERT, ELECTRA-small, and MiniLM — on the Reddit Mental Health Dataset (RMHD) labelled corpus (800 posts, four balanced classes) for assigning Reddit posts to one of four psychosocial root causes: Drug and Alcohol, Early Life, Personality, and Trauma and Stress. Track one shows that suicidal ideation, academic pressure, and financial stress are the dominant risk drivers (Logistic Regression pseudo-R² = 0.479; best ROC-AUC = 0.93 for CNN and XGBoost), while study satisfaction and age act as protective factors. Track two finds that DistilBERT, despite being the largest of the four candidates, generalises best on this small corpus (66.7% accuracy, 0.641 macro-F1, 0.890 macro ROC-AUC on a held-out test set of 120 posts), outperforming the more parameter-efficient ALBERT (60.0% accuracy), ELECTRA-small (40.8%), and MiniLM (29.2%) under an identical fine-tuning budget of three epochs on a single Colab T4 GPU, indicating that aggressive distillation and small hidden sizes trade away representational capacity this task still needs at low sample counts.