Does the Night Give Power? Social Media Depression and Its Symptom Detection Considering Temporal Patterns

Given the increasing importance of depression and social media, it's imperative to develop an artifact that enables the detection of mental health disorders in social media. Existing research focuses on exploring the application of semantic features and writing style in depression detection, while ignoring the dependencies between temporal patterns and tweets. Furthermore, few studies investigate the heterogeneity and homogeneity among depression and various symptoms. To address these challenges, we propose a deep learning model Temporal Transformer Multi-task Contrastive Network (TTMCNet) for depression detection modeling. Inspired by temporal landmarks theory (TLT), we adapt traditional Transformer by introducing a temporal relative positional encoding (TRPE) and a decaying term. After that, we construct a multi-task contrastive learning part to explore the heterogeneity and homogeneity of depression. Empirical evaluation shows that our proposed method outperforms other benchmark models. This study contributes to design science literature, particularly in depression detection.