Rumor Detection Model with Deep Cross Fusion of Text and Propagation Graph Structures

The generation and spread of rumors are phenomenon that cannot be eradicated in human society. However, the wanton spread of rumors will harm the development of society, especially in the current era of information. The existing rumor detection models or frameworks are basically based on deep neural networks, which can extract text and propagation features and other information aiding rumor assessment, and then combine these features to filter rumors. Although these methods are effective, it is difficult to distinguish rumors in complex contexts or instances where the signs of forgery are not so obvious. Therefore, based on the interactive characteristics of text content and Propagation graph structure, this research proposes a rumor detection model with deep cross fusion. First, the text content and the information of propagation graph structures are encoded while ensuring that they are complete and independent, thereby extracting important information; Then, the separately encoded text content is deeply cross fused with the propagation graph structures, so that the information can penetrate and complement each other, and at the same time, a self-attention operation is performed after the cross fusion of each layer. Finally, the encoding from both components is fully integrated, then a classification feature vector is introduced into the output result of each part, and the classification result is output after multiple layers of fully connected layers. The experiments on three classical rumor detection datasets show that the model has a significant improvement in both accuracy and specified delay time, and the maximum accuracy is improved by 1.21%.