Enhancing user sequence representation with cross-view collaborative learning for depression detection on Sina Weibo

Detecting depression through social media is a crucial task in the era of mobile media. Several methods have been proposed for modeling user behavior sequences among social media users with depression. However, existing methods often overlook the dynamic dependencies between user behaviors. Moreover, previous research has overlooked the importance of incorporating multi-perspective information in personalized user modeling. Therefore, we propose a cross-view collaborative learning method for user behavior sequence modeling. Specifically, we built a post-level behavior graph for each individual and employed a graph neural network (GNN) to model the contextual dependencies between user behaviors. We propose using Long Short-Term Memory (LSTM) networks to capture the sequential evolution of user behavior. Additionally, we introduce a node-level attention mechanism to aggregate node representations and obtain graph-level semantic representations. Next, we leverage a BERTopic-based model to extract personalized topics and interesting knowledge from each individual's posts. These knowledge elements are then aggregated to form user-level topic representations. We propose a cross-view collaborative learning method to integrate temporal behavior and topic semantic representations dynamically. This method effectively addresses the semantic alignment and fusion issues across views and enhance the ability of our model to detect depression. Finally, to evaluate the proposed model, we constructed a well-annotated Chinese dataset based on Sina Weibo for user-level depression behavioral modeling. Extensive experimental results and analyses on both datasets demonstrated the effectiveness and advancement of our model for user-level depression detection task.