Fusion and DGCN for Optimizing Depression Detection in Social Media Texts

Background: Social media texts offer rich data for early depression detection but pose challenges due to fragmented expressions and complex social dynamics. Existing single-modality models struggle to capture semantic subtleties and social influence critical for accurate identification. Methods: We propose a fusion model integrating: A Multi-head Self-Attention (MHA) mechanism to extract emotional semantics and contextual dependencies from text. A Dynamic Graph Convolutional Network (DGCN) modeling evolving social influence through user interaction graphs. An attention-based fusion module dynamically weighting textual (MHA) and social (DGCN) features for joint optimization. Results: Evaluated on CLPsych 2015 and Reddit datasets. chieved 0.915 F1-score, outperforming LSTM (0.847) and BERT (0.881). Detected mild depression with 82% recall (9% gain over BERT). Social modeling boosted severe depression recall to 91%. Multimodal fusion improved mild case detection by 7% versus text-only models. Conclusion: The fusion of MHA and DGCN significantly enhances depression detection by jointly modeling linguistic cues and social dynamics. This approach demonstrates particular strength in identifying subtle emotional expressions prevalent in social media.