Social Media User Behavior Modeling and Content Distribution Optimization Using GAT and Temporal Graph Networks

To address the challenges of effectively capturing the dynamic nature of user behavior and complex interactions in social media content distribution, this paper proposes a joint modeling approach that integrates a graph attention network (GAT) and a temporal graph network (TGN). Current static graph-based approaches suffer from two major flaws: ignoring the temporal evolution of user interests and failing to handle heterogeneous node relationships, which limits recommendation accuracy and user dwell time. This study first constructs a heterogeneous user-content graph using the GAT and employs a multi-head attention mechanism (8 heads) to quantify the interaction weights between users, posts, and topic nodes. Secondly, a temporal graph network module is developed, employing a temporal encoder and a gated recurrent unit (GRU) to model the dynamic features of user behavior sequences over a 28-day time window. Furthermore, a dynamic negative sampling strategy is proposed to enhance the model's ability to recognize implicit feedback by adaptively adjusting the negative sample ratio (1:3 to 1:5). Finally, a multi-task learning framework is constructed to jointly optimize click-through rate (CTR) and dwell time metrics. A/B testing on Twitter and Weibo datasets demonstrates that the proposed model maintains a stable daily CTR of 10.35%-11%, with a peak average dwell time of 31.8 seconds. Experiments have shown that this method can effectively capture the spatiotemporal evolution of user behavior, providing a new technical approach for personalized recommendations and content resource allocation on social media platforms.