Rumor Detection Method Based on Temporal-structural Features and User Relationship Enhancement

To address the challenges in existing methods such as difficulty in capturing spatiotemporal features, loss of node information, and insufficient utilization of user data, we propose a novel rumor detection method called TSU-RD, which integrates temporal-structural features and user relationship enhancement. Firstly, a Bi-directional Gated Recurrent Unit (BiGRU) enhanced with spatial information is used to extract local spatiotemporal representation vectors from the propagation sequence. Then, a new node-edge encoding scheme is designed, and an improved Graph Attention Network (GAT) is employed to capture the global structural representation of the event. Finally, user feature representations are constructed, and a self-attention module guided by semantic and interaction information is applied to obtain user representation vectors. The final event representation is generated by fusing the above three types of feature vectors. Experimental results on Weibo2016 and CED datasets demonstrate that the proposed method consistently outperforms existing representative approaches in both accuracy and F₁ score.