Research On Fine-Grained Fake News Detection Based On Multi-Source Feature Dynamic Fusion

The rapid development of Internet technology encourages more users to browse news and express opinions on various network platforms.Netnews has exploded,among which various fake news has emerged.Deep learning-based models have achieved great success in fake news detection by extracting semantic features of news,but some things could still be improved.On the one hand,news contained a wealth of information,but these different types of information were difficult to model uniformly.On the other hand,some existing models focused on improving the post-detection performance of news,ignoring the early detection.In addition,the research on noise reduction information was also insufficient.To solve these problems,this thesis designed heterogeneous graph models from the fusion of multi-source information and constructed a dual-channel attention network,respectively,and applied them to the fine-grained classification task of fake news detection.In order to solve the problems such as insufficient use of multi-source information and loss of high-order structural features of news,this thesis proposed a fake news detection model integrating multi-source information local-global relationships.The model combined news content and users information for modeling,which constructed new-word and new-user subgraphs.First,we designed the subgraph convolutional network and the projection layer to extract news content features and propagation features,respectively.Then,we incorporated local attention into the model to distinguish the weights of different nodes.Finally,the global attention network was designed to measure the weight of the two subgraphs and finally realized the detection of fine-grained vacation news.In order to solve the problems that the existing models ignored the interaction between various components of news,introduced a large amount of noise in feature fusion,and failed to take into account the early detection performance of fake news,this thesis designed a dual-channel heterogeneous graph attention fake news detection model based on dynamic feature fusion.First,we constructed a heterogeneous graph.On this basis,we modeled the semantic information between news and words.Furthermore,we modeled the communication structure information between news and users.Second,based on the constructed news-word heterogeneous graph,we designed a semantic information extractor to extract the high-order features of news.Moreover,we integrated multi-head attention mechanism to distinguish the weight of semantic features of different nodes.In addition,based on the constructed news-user heterogeneous graph,we designed a propagation-diffusion extractor with time-stamped and users’ attribute information to obtain the news’ s vertical and horizontal diffusion feature.Then we integrated the attention mechanism to learn the weights of different nodes and propagation and diffusion paths.Finally,the feature dynamic fusion component combined news semantic features and propagation-diffusion features to obtain the global relationship information of nodes.In addition,the noise information is reduced effectively.Experiments on three datasets of Weibo,Twitter15 and Twitter16 show that the models proposed in this thesis have better performance in detecting fake news.In addition,our models were applicable to both short and long texts.Furthermore,the models realized the early detection of fake news.