KMAN Publication Inc. (KMANPUB) KMAN Counseling & Psychology Nexus 3041-9026 4 Serial Number 6 2026 08 20 Explainable Semi-Supervised Learning for Depression Subtype Detection in Social Media Explainable Semi-Supervised Learning for Depression Subtype Detection in Social Media 1 18 EN Journal Article 2026 05 02 <p>Depression detection from social-media text has been extensively studied; however, the overwhelming majority of prior research focuses on binary classification under fully supervised settings, overlooking clinically meaningful subtype distinctions and the scarcity of reliable labeled data. This study proposes a Hybrid FixMatch–Self-Training Semi-Supervised Framework for fine-grained depression subtype detection, targeting five clinically relevant categories: Major Depressive Disorder, Bipolar Depression, Psychotic Depression, Atypical Depression, and Postpartum Depression. The proposed architecture is designed to integrate supervised learning on a limited set of subtype-labeled tweets with two complementary semi-supervised mechanisms: (i) FixMatch-based consistency regularization and (ii) iterative self-training to leverage large volumes of unlabeled mental-health discourse from Twitter and Reddit. In addition, token-level explainability is planned to ensure alignment between model predictions and clinically recognized symptom patterns. In the present submission, this full architecture was evaluated as a reduced-scope, fully-reproducible proof of concept: a classical TF-IDF + linear-SVM supervised backbone combined with iterative self-training, run on CPU hardware without GPU access or connectivity to pretrained-weight repositories. Under this configuration, self-training did not outperform the purely supervised baseline (Section 3.7–4.9). Validating the full RoBERTa-encoder, FixMatch consistency branch, cross-platform robustness, and clinician-rated SHAP explainability described in Section 2 remains future work and is not claimed as completed here. By reframing subtype detection as a semi-supervised representation learning problem, this work addresses a methodological gap in computational mental-health research. The framework establishes a scalable and interpretable foundation for fine-grained digital mental-health modeling, with implications for risk screening, large-scale monitoring, and clinically informed AI systems.</p> https://www.journals.kmanpub.com/index.php/psychnexus/article/download/5925/11375