基于双尺度图分层池化的微博谣言检测模型OA

How to accurately identify rumors on social networks has become a hot research issue.In the event-level rumor detection task based on graph neural network,the expressiveness of the graph convolution operator and graph pooling significantly affects the classification results.By comprehensive-ly considering the textual and user information within posts as well as the propagation structure between posts,and modeling various features using an event graph as the carrier,we propose a novel rumor de-tection model named GATv2-DSAPool.The model uses hierarchical pooling as the infrastructure.A dy-namic attention mechanism is introduced in graph attention network to capture the spatial structure of post propagation.A dual-scale attention is designed in graph pooling to compute the node scores.The set of nodes are selected to generate a coarsened graph based on the score ranking,which effectively en-codes the global features of events.Experimental results on two real microblog datasets show that the model outperforms the state-of-the-art benchmark models in terms of F1-score and accuracy.