A hybrid depression detection model and correlation analysis for social media based on attention mechanism

As a serious mental illness, depression can be extremely harmful to the physical and mental health of individuals. However, detecting depression can be challenging due to the reluctance of the depressed to actively express themselves. Fortunately, in modern society, online social platforms provide an opportunity for genuine self-expression in our daily lives. Leveraging the power of social data, we can identify potentially depressed users efficiently and accurately. This lays a strong foundation for subsequent interventions. In this paper, we propose a hybrid model that comprehensively considers the features and post texts of users by utilizing a simplified multi-head attention mechanism for detecting user depression. Compared to traditional models, such as Decision Tree and Random Forest, the simplified multi-head attention mechanism achieves higher classification accuracy while offering enhanced interpretability at the individual level. To verify the validity of our model, we apply it to the Weibo User Depression Detection Dataset (WU3D) containing approximately 1,150,000 posts from around 21,000 users. The dataset has been classified by human experts as either depressed or not. The results show that our model both explains the association of each feature of a single user well and achieves better performance than traditional methods. Notably, the final F1-score of our new model on the test set is 0.9473. Furthermore, by visualizing attention scores, we conduct a correlation analysis between different features and results for individual users, which can facilitate further analysis of the behavior and psychological characteristics of specific depressed users by professionals.