Rumor Detection Models Integrating Contextual Semantics Based Propagation Path And User Information

With the widespread popularity of mobile devices,the rapidly developing social media platforms provide users with more convenient and fast interaction,but they have also inevitably become a hotbed for the spread of various types of rumor information.The rampant spread of rumors on social media poses a huge hidden danger to public and social security.The need of reality gradually highlights the urgency and importance of the task of rumor detection.Previous approaches using deep learning techniques to model text features on time series ignore important structural information in the process of rumor propagation.Works in recent years have proposed propagation structure-based approaches that integrate structural information and text semantics into the learning of the model,effectively improving the effectiveness of rumor detection.However,existing propagation structure-based approaches focus on modeling semantic dependencies among posts with direct reply relationships,resulting in insufficient ability to learn all possible interactions among posts on a propagation path,and simply treats different posts and different paths as equally important,resulting in model that fail to extract key features from them.In addition,the approach relying on text semantics ignores the role played by users in the process of rumor propagation,resulting in insufficient features used in the model and the detection effect needs to be further improved.In this paper,we address the above issues,and the main research contents and results achieved are as follows.(1)A node-and-path dual-attention rumor detection model DAN-Tree based on the propagation tree structure is proposed.The model first uses Transformer coding blocks as feature extractors to learn the possible interactions among all posts on the propagation path.After that,it captures key posts and key paths in the propagation tree by post-level attention and path-level attention,and achieves the organic integration of deep structure and semantic information on the propagation tree.In addition,the DAN-Tree model uses path oversampling techniques and structural embedding to enhance the learning of deep structural information on the propagation tree.Experimental results on four rumor datasets show that the DAN-Tree model achieves better detection than other baseline models,verifying the importance of focusing on all possible interactions between posts and the effectiveness of dual-attention mechanisms in the DAN-Tree model.(2)A rumor detection model DAN-Tree++ that incorporates user features and propagation structure is proposed.The model introduces user features on the basis of the DAN-Tree model.On the one hand,user features are fused on text features,and user credibility is introduced into the learning of the model.On the other hand,global user features are fused on the communication structure features to introduce the overall characteristics of user features in the rumor propagation process.Experimental results on two datasets with user information show that the DAN-Tree++ model achieves superior performance over other baseline models and achieves the best results on the early detection task,verifying the effectiveness of the strategy of fusing user features for the rumor detection task.