Social Event Detection Based On The Information Dissemination Characteristics

Online social networks,as a basic service of Internet technology,have not only enriched the topics of people’s discussion and communication in the past 20 years,but also narrowed the distance between people in different regions.With the widespread use of social platforms,the huge user group brings massive and uneven information data,and rumors,defamation and other illegal infringements are also rampant.Therefore,extracting features from social information data to detect events has gradually become an important research topic.Based on this,this thesis will focus on the accurate detection of social events and rumors,analyze the shortcomings of existing research methods,and propose targeted solutions for this.The existing methods still have some shortcomings in utilizing the characteristics of information data in social networks,mainly reflected in:(1)they fail to deal with massive social information well,there is data redundancy and noise,which affects the accuracy of event detection;(2)lack of capturing the spatiotemporal characteristics of social data,crossday data association may become important information for preventing disease transmission at any time;(3)ignore the influence of users with a certain fan base on rumor transmission in social networks;(4)cross-platform information data There are information barriers,lack of a framework that considers multiple factors to eliminate data islands,and combines user influence to detect rumors.This thesis focuses on the following issues:1)Graph neural network based on filtering strategy and information completion:This method uses a filtering strategy for data cleaning,and adopts an information supplement window to rebuild the correlation between cross-day social messages,effectively solving the problems of data noise and cross-day information gap mentioned in(1)and(2)above.2)Bidirectional graph convolutional neural network based on user influence factor:Inspired by information entropy,this thesis proposes the concept of influence factor to describe the influence of users in social networks,and models two types of rumor transmission methods at the same time,Effectively solved the problem that high-fan users have an impact on rumor transmission mentioned in(3)above.3)Graph attention network based on federated learning and influence factor:This method adopts a distributed parallel way to train the model.The parameter features are obtained by combining the influence factor on the client side and input to the server side for aggregation.After that,the global parameter features are fed back to each client.This way breaks through the information barrier formed by(4)above and realizes cross-platform information sharing.In summary,this thesis conducts related research work around social event and rumor detection,takes accurate event and rumor detection as the research goal,and solves problems caused by social network information dissemination characteristics in practical scenarios.The experimental results on real datasets show that the model proposed in this thesis has a significant advantage in detecting social events,and also shows more stable performance in early rumor detection.Therefore,this thesis’s research has high value,accurate detection of social events and rumors also plays an important role in disaster early warning,disease outbreak prevention and rumor monitoring and control.