Fake News Detection And Classification Model Based On Fusion Algorithm

With the development of the internet and multimedia technology,people’s access to information has changed dramatically.More and more people are using social media to find out what is happening in the world,but social media is also full of inaccurate and falsified news.Some news is even used to mislead public opinion in the first place and to deliberately undermine the credibility of governments.It is therefore essential to use automatic fake news detectors to prevent users from receiving fake news and to prevent the serious negative effects this can cause.In turn,technological developments have led to a trend of diversifying the forms of news,whereas current work has mostly focused on only one form of news.Moreover,the fake news detection task is not yet the same as a normal sentence classification task? in most cases,the former needs to identify news from different domains that the model has never encountered before,while existing deep learning models generally reason directly through the content of the news,with little consideration of domain knowledge.To address the above challenges,this thesis proposes corresponding fusion models for fake news detection according to different news forms and tasks.For example,models based on Bi LSTM and fully-connected layers are used for tasks with single text-image pairs,and models incorporating adversarial training,sample balancing,and data enhancement algorithms are used for tasks with multi-text and multi text-image pairs,etc.Both are compared with mainstream models based on relevant realistic datasets and relevant ablation experiments are conducted.In addition,this thesis proposes a fusion model for social context-based fake news detection incorporating multi-task,data augmentation,graph neural networks,fusion features,and pre-trained models for news forms that contain social context.The model uses the posts characteristics of users involved in news dissemination as information for graph neural network points and provides three graph neural networks to aggregate this information.The model then extracts news text features using a fine-tuned pre-trained model and fuses them with the global graph attribute features of the propagation network to obtain news features.Finally the features of the resulting graph neural network were fused with the news features and augmented by dropout layers.Notably,the model also proposes a multitask module with a consistent direction of convergence and a new loss function is designed to optimise the model.The model is tested on two realistic fake news datasets and sota results are obtained,and ablation experiments demonstrate the effectiveness of various parts of the model.The thesis also subjects the model to a fake news early detection task,and the experimental results also demonstrate the stability and accuracy.