A Bi-directional Graph Convolutional Network Rumor Detection Model Based on Textual Data and Propagation Tree Node Enhancement

This paper proposes a bi-directional graph convolutional network based on textual data enhancement and propagation tree node enhancement for performing rumor detection accurately and efficiently in social networks. This method not only mines the deep layer features of text and avoids the problem of feature construction but also generates hidden features that are difficult to discover manually. Meanwhile, it considers the effects of semantic context and labels on shielded position words, thus generating different expressions of the same emotion polarity at the position of shielded words according to context. Therefore, it can represent the information in social networks more flexibly and accurately and improve the generalization ability of trained classifiers. Moreover, a propagation tree node enhancement method based on event propagation structure is proposed. The formed hidden layer feature vectors after each node pass through the graph convolutional layer are merged with the hidden layer feature vectors of the last layer node, and a new feature matrix is obtained. In addition to better using the abundant information contained in the source post as well as various forwarding nodes, it can also obtain more accurate node expression from the relationships among various nodes. Therefore, it can improve the accuracy of rumor detection.