Answers are grounded in this paper and cite their sources.
Text-based depression risk detection is an important area of computational mental health because written language may contain signals associated with psychological distress and can support scalable risk-screening research. This study evaluates three term frequency-inverse document frequency (TF-IDF)-based classical machine learning classifiers - Logistic Regression, Multinomial Naive Bayes, and Linear Support Vector Machine (SVM) - against a small instruction-tuned zero-shot Fine-tuned Language Net-Text-to-Text Transfer Transformer (FLAN-T5)-base baseline on a public Reddit depression dataset. Performance was assessed using accuracy, precision, recall, F1-score, and receiver operating characteristic-area under the curve (ROC-AUC). Linear SVM achieved the highest accuracy (0.9651), F1-score (0.9644), and ROC-AUC (0.9937), while Logistic Regression obtained a very similar precision (0.9743 versus 0.9747). Naive Bayes achieved higher recall than Linear SVM (0.9687 versus 0.9544), and the FLAN-T5 baseline produced recall of 1.0000 because it predicted every test sample as depression; its accuracy was 0.4958, and ROC-AUC was 0.5000. Therefore, the findings support the effectiveness of TF-IDF-based linear classifiers for this dataset and experimental configuration, but they should not be interpreted as a general comparison between classical machine learning and large language models (LLMs). The FLAN-T5 result represents only the tested zero-shot prompt, truncation, decoding, and score-parsing setup. Additional tuning, repeated validation, stronger transformer/LLM baselines, and cross-dataset testing are required before broader conclusions can be drawn.
No takes yet. Share an insight, caveat, or question.
Kumari et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: