| The development of Web2.0 technology has led to an exponential increase in the amount of information on social networks,among which the amount of rumor information has also surged,and rumors have severely affected the network order and social stability.Carrying out rumor detection can reasonably guide internet public opinion and purify the internet environment,which is beneficial to social stability.Most of the existing studies focus on studying the statistical characteristics of rumor information,such as the number of words and characters in the text,but they do not consider the dynamic changes of the information during the transmission process,that is,regularity of characteristics over time.In addition,rumor detection research has evolved from traditional research based on single-modal data to more complex research based on coexistence of multi-modal data.However,most existing studies only focus on the text content,ignoring the information contained in the pictures and the relationship between the pictures and the text.Therefore,studying the dynamic information changing process and mining associations between multi-modal data is of great significance to improve the performance of rumor detection.This topic researches rumor detection methods in social networks.This research is based on the research project of the Beijing Education Commission's scientific research and postgraduate training-social-sensed cross-media data analysis and mining research.This thesis studies rumor detection in two scenarios:dynamic information changes and coexistence of multi-modal data.The main research contents and innovations include:1.Aiming at the fact that the rumor information released by users in rumor detection changes with time and leads to the problem of one-sidedness based on statistical features only,this paper proposes a rumor detection framework based on dynamic information embedding.Recurrent Neural Networks(RNN)were used to model the rumor time series and learn the dependence of rumor information at different time stages,thereby discovering the distinctive characteristics of rumors that are different from non-rumors over time.First,the attention mechanism is used to distinguish the importance of rumor vector expression at different time steps,and a rumor detection algorithm based on dynamic time series is proposed.Then,considering that local information interaction can provide information when determining the importance of information at different time steps,a dynamic time series rumor detection algorithm based on local features is further proposed.Finally,experiments on the Sina Weibo rumor dataset show that the proposed algorithm improves the precision of rumor detection by nearly 4%.2.Aiming at the problem of sparseness of single-modal content in the context of multi-modal data coexistence in rumor detection,this paper proposes a rumor detection algorithm based on the fusion of text and image information.In order to solve the problem of sparseness of single-modal content and fully explore the relationship between different modalities of rumor information,this paper builds a network model based on joint learning of text and image information.The model includes three parts:text content understanding based on FastText network,image visual feature extraction based on deep CNN,and text and image fusion representation learning.First,the text features and image features are extracted through the text content understanding based on the FastText network and the visual feature extraction based on the deep CNN,and then text features,image features and the correlation between them are obtained by fusion representation learning.The classification results are obtained finally.On the Sina Weibo rumor dataset based on text and image modals,the precision improvement of nearly 3%was achieved,which verified the importance of combining different modal information to improve the performance of rumor detection in social networks. |
PDF Full Text Request