Identifying Depressive Content in Decentralized Social Networks: A Case Study on Bluesky

TLDR

Evaluating a hybrid pipeline combining XGBoost and fine-tuned transformers (BERT-base, RoBERTa-base, and MentalBERT), the best model achieves an F1-score of 80.5% on Bluesky and 90.8% on Twitter, highlighting that core linguistic cues remain informative across decentralized platforms.

Abstract

Depression is a global mental health concern, yet scalable screening remains challenging. While prior work focuses on centralized platforms, little is known about depressive signals in newer networks like Bluesky. We present a manually annotated dataset of 4,898 public English Bluesky posts (14.5% labeled as depressive). Evaluating a hybrid pipeline combining XGBoost and fine-tuned transformers (BERT-base, RoBERTa-base, and MentalBERT), our best model achieves an F1-score of 80.5% on Bluesky and 90.8% on Twitter. This highlights that core linguistic cues remain informative across decentralized platforms. The proposed system is intended for decision support rather than diagnosis, requiring clinical validation and ethical oversight.