A Study on Social Media Depression Detection Methods Based on Multi-Level User Representation Learning

Abstract

Social-media posts record not only what users say but also when and how they participate. These two sources of evidence can support computational screening for depression-related patterns, although the task is complicated by indirect language, overlapping class boundaries, and incomplete behavioral records. This study examines whether combining information at several levels produces more informative user representations than relying on isolated posts. The proposed workflow begins with BERT-based encoding of post content. Random downsampling is used in the pre-trained-model experiments to reduce the imbalance between the two labels. Posting frequency, activity-time distribution, and historical-post volume are then mapped into the same feature space as the text representation. Post-level vectors are aggregated for each user, and a graph neural network is applied to a similarity graph so that the final representation also reflects relationships among users. A supervised contrastive objective is included to encourage compact within-class representations and clearer separation between classes. Across the reported test results, user-level models were markedly stronger than tweet-level baselines. The user-level BERT + GNN configuration obtained an accuracy of 0.9404, an F1-score of 0.9399, and an AUC of 0.9703. These results suggest that historical context and relational structure are useful for this classification setting; they should not, however, be interpreted as a substitute for clinical assessment.