Research On Misinformation Detection Methods Based On Graph Neural Network

With the development and popularity of the mobile Internet,a large amount of misinformation is spreading uncontrollably on social media,which may not only cause losses for individual users,but also cause social panic.It has become necessary to identify misinformation and stop the uncontrolled spread of misinformation on social media.Malicious bots spreading misinformation have infiltrated social media platforms,creating an enormous amount of misinformation that is widely circulated among users.Stopping the spread of misinformation on social media platforms is an important task that needs to be addressed to create a purer online environment and maintain cyberspace security.To address this problem,this thesis divides the issue into two tasks: detecting misinformation bots and detecting misinformation content.The following work has been carried out in this thesis:(1)For the task of misinformation bot detection,this thesis proposes a model based on Bayesian graph local extrema convolution algorithm.With the advancement of information technology,misinformation bots are becoming more and more anthropomorphic and realistic,making them difficult to be distinguished.When aggregating the features of neighboring nodes,this method retains only the differences between these nodes and their neighbors,which mitigates the iterative propagation of noise in the model.On the other hand,this method focuses more on the difference between two users and enhances the ability to distinguish between bots and real users.The Bayesian graph local extrema convolution algorithm is particularly suited for this task because it learns better from uncertain data through the posterior distribution when the parameters are updated.This makes it more effective at preventing bot spoofing.Experiments show that the Bayesian graph local extrema convolution algorithm achieves outstanding bot detection results and has better robustness than other graph convolution algorithms.(2)For the misinformation content detection task,a model based on contrastive learning and long-tail strategy is proposed in this thesis.A large number of misinformation bots repost each other on social platforms,and these reposts act as edges of the social network making the degree of bot nodes higher.Existing graph neural network algorithms treat all nodes equally,which can result in high-degree nodes dominating neighborhood aggregation,while long-tail nodes with low degree are often real users with only a small number of followers and retweets.To address this issue,this thesis proposes a long-tail strategy for this situation,further enhancing the learning and importance of long-tail nodes through a self-attention mechanism and recurrent neural networks.The effectiveness and generality of this algorithm are also demonstrated through experiments of porting the longtail strategy to past algorithms.Experiment results at different propagation durations also demonstrate the need for early misinformation detection.To tackle the more complex task of misinformation content detection,this thesis proposes to use a contrastive learning framework for misinformation content detection on social media platforms.The framework enriches the dataset with three data augmentation operations,allowing the model to explore a broader range of hidden differences between information contents.Extensive experiments on two publicly available misinformation datasets demonstrate the model’s effectiveness in this thesis.