mdsajjadullah/explainDepression-social-media-xai

Cleaned, balanced, and clinically annotated social media dataset for explainable depression detection research.

Combined from real Twitter and Reddit posts, engineered with DSM-5-aligned clinical lexicon features, and used to train a DistilBERT model achieving 96.17% accuracy.

━━━━━━━━━━━━━━━━━━━━━━━━━━━

DATASET STATS Total Rows → 40,770 Class Balance → 50% Depressed / 50% Not Depressed Feature Columns → 12 Sources → Twitter + Reddit

━━━━━━━━━━━━━━━━━━━━━━━━━━━

SOURCE DATA Twitter → stevenhans/depression-and-anxiety-in-twitter-id (6,982 rows) Reddit → reihanenamdari/mental-health-corpus (27,977 rows)

━━━━━━━━━━━━━━━━━━━━━━━━━━━

COLUMNS text → Cleaned post/tweet text source → twitter or reddit textlength → Word count binarylabel → 1=depressed | 0=not depressed lexhopelessness → DSM-5 hopelessness match count lexanhedonia → DSM-5 anhedonia match count lexsocialisolation → DSM-5 social isolation match count lexfatigue → DSM-5 fatigue match count lexsuicidalideation → DSM-5 suicidal ideation match count lexcognitivedistortion → DSM-5 cognitive distortion match count desscore → Depression Explanation Score clinicaltokencount → Total clinical term matches

━━━━━━━━━━━━━━━━━━━━━━━━━━━

REAL DES SCORES (from trained model SHAP) Suicidal Ideation → 0.2402 strongest signal Social Isolation → 0.0726 Hopelessness → 0.0615 Cognitive Distortion → 0.0543 Fatigue → 0.0438 Anhedonia → 0.0172

━━━━━━━━━━━━━━━━━━━━━━━━━━━

PREPROCESSING APPLIED ✅ URL removal ✅ @mention stripping ✅ Emoji and non-ASCII removal ✅ HTML entity cleaning ✅ Lowercase normalisation ✅ Random Over-Sampling for class balance

━━━━━━━━━━━━━━━━━━━━━━━━━━━

TRAINED MODEL Available at: huggingface.co/your-username/explainDepression-distilbert Accuracy: 96.17% | F1: 0.9617 | AUC: 0.9937

━━━━━━━━━━━━━━━━━━━━━━━━━━━

INTENDED FOR RESEARCH ONLY. Not for clinical use without professional oversight.