The paper explores federated learning (FL) for detecting depression using patient-authored multilingual textual data. It utilizes five datasets in English, Arabic, Spanish, Russian, and Korean, each varying in size. Findings reveal that FL in IID scenarios matches centralized and local model performance. Multilingual data favor non-IID configurations, prompting exploration of data partitioning strategies to address volume and label distribution differences among clients. Although extreme quantity imbalances reduce model performance, FL emerges as a promising privacy-centric approach for mental health diagnostics.
All Keywords
【저자키워드】 Depression, mental health, social media, natural language processing, federated learning, Multilingual,
【저자키워드】 Depression, mental health, social media, natural language processing, federated learning, Multilingual,