Multi-scale Graph Convolutional Networks For Fake News Detection

With the rapid rise of the internet social platform,the efficiency of information dissemination has been greatly improved.However,the convenience of communication has also contributed to the generation and dissemination of fake news.Fake news will not only weaken the credibility of the media,but also cause confusion in the social order and affect people's daily life.Automatically detecting fake news from the complicated network content is an actual natural language processing problem that needs to be solved urgently,and has positive significance to reduce or even eliminate the negative effects of fake news.Fake news refers to speech or reports that do not reflect the real situation and have false components.They are often released for political or economic benefits and difficult to accurately identify.Because the news content of social platforms such as Twitter and Weibo is relatively short,only considering the news content and its language features cannot reach the satisfactory detection results.An efficient and accurate model is required to detect and identify this type of news.Using limited information to automatically detect and identify fake news is a major challenge in the field of fake news detection.In order to use limited information to improve detection performance,most of the existing research is based on the assumption of independence between samples.Text and other non-text features are simultaneously modeled,ignoring the similarity between news,which is a vital factor that may improve the performance of the classification.This paper mainly studies the problem of effectively using abundant non-text information to assist the detection of fake news under the condition of short news content.As the text of the online social platform is shorter,it has the characteristics of less information content,diverse expression forms,and incomplete structural components.In order to effectively utilize non-text auxiliary information of news to improve detection performance,this paper proposes a semi-supervised learning framework for detecting fake news with the help of using the similarity between news in different context feature.Starting from the relationship of news samples,the framework views news samples as independent nodes in the graph network,uses abundant non-text auxiliary information to determine the weights of edge,and builds a multi-dimensional graph network structure.In view of the shortcomings of graph convolutional networks in combining neighborinformation and learning representations,an improved multi-scale graph convolutional networks is proposed to combine the characteristics of neighbors at different distances to capture information of different granularities in a single relationship graph.The multi-scale information approach improves the diversity of each node's representation.Finally,the information fusion is carried out through the attention mechanism to improve the expression consistency of the information at different granularities,and the fusion generates a node representation with stronger generalization ability.Experiments performed on LIAR,the largest public benchmark dataset in the field of fake news detection,shows that the multi-scale graph convolutional networks for fake news detection framework can use the different granularity similarity of news nodes to distinguish the true and fake degree of news.The proposed M-GCN model has achieved good classification performance,which verifies the effectiveness of the detection framework.