User-Level Depression Detection Using Long-Context Transformers and Behavioral Data
Küçük Resim Yok
Tarih
2026
Yazarlar
Dergi Başlığı
Dergi ISSN
Cilt Başlığı
Yayıncı
Ieee-Inst Electrical Electronics Engineers Inc
Erişim Hakkı
info:eu-repo/semantics/openAccess
Özet
Depression poses a major global health challenge, while most computational detection approaches remain constrained by the 512-token input limit of standard Transformer models, restricting their ability to capture longitudinal user behavior. Moreover, aggressive text preprocessing may remove expressive cues that are informative for mental health assessment. This study presents a controlled user-level benchmarking framework for depression detection from social media text, examining the effects of preprocessing strategy, behavioral metadata integration, and context length. A two-stage experimental design is adopted: first, Aggressive Cleaning (AC) and Selective Cleaning (SC) are compared across four 512-token Transformer backbones (BERT, RoBERTa, DistilBERT, and MentalBERT), with and without metadata; second, Longformer and BigBird are evaluated at both 512 and 4096 tokens under the SC setting. The results show that SC provides a more favorable context-preserving preprocessing strategy, while long-context modeling yields the strongest overall gains. The best-performing configuration, Longformer with 4096-token input and behavioral metadata fusion, achieves 90.36% accuracy, 87.15% F1, and 0.9667 AUROC. Additional t-SNE and SHAP analyses indicate that extended context improves the organization of user-level representations, while behavioral metadata provides a complementary rather than dominant contribution, with night posting ratio emerging as one of the most influential features. Overall, the findings show that context-preserving preprocessing and long-context sequence modeling are decisive factors for improving large-scale, text-centered social media-based depression detection.
Açıklama
Anahtar Kelimeler
[Keyword Not Available]
Kaynak
Ieee Access
WoS Q Değeri
Q2
Scopus Q Değeri
Q1
Cilt
14
