Research On Robust Rumor Detection Models

With the rocketing progress of the Internet,the way that people obtain information changes a lot,the generation and propagation of information are also become decentralized.Everyone has the ability of spreading or misrepresenting message,which leads to the situation that false information’s generation and dissemination becoming more frequent and speediness.In order to reduce the harm to society caused by the propagation of the false information,the rumor detection task is getting more and more attention.In recent years,a great deal of deep learning methods for rumor detection task have been proposed as the deep learning developed and they get excellent performance.However,the existing methods still have some problems as follows:(1)the characteristics of rumor data varies greatly at different times or events,and the robustness of rumor detection model is insufficient to unfamiliar rumor data;(2)The mainstream rumor detection data sets have insufficient data,So the rumor detection methods can’t have the adequate data to fully learn the rumor characteristics,resulting in the lack of robustness of the model.To address the above issues,in this paper,the main research contents and results achieved are as follows:(1)A self-training rumor detection transfer learning method is proposed.Since rumor data varies greatly at different times or events,we regard the classification task of rumor detection on new data as a domain adaptation task.We take the existing labeled data as the source domain data,and the unfamiliar data as the target domain data.The model trained on the source domain data uses the self training algorithm to predict the pseudolabel on the target domain data for iterative learning.At the same time,a method based on training loss for filtering noisy label is proposed,which removes the untrusted false labels from the training data and reduces the introduction of noise.Experiments show that our transfer learning framework can make the rumor detection model get better performance on the target domain data.In addition,this paper is proposed for the first time to focus on the migration performance of the rumor detection model,and conduct the transfer learning task of rumor detection on the Twitter15,16 datasets.(2)A rumor detection model based on the pre-trained method of graph data augmentation is proposed.The existing public rumor detection datasets are insufficient in data volume.At the same time,the difficulty in collecting rumor data and the high cost of manual labeling lead to dilemma on large-scale data collection.In this case,in order to make full use of the existing datasets,we augment the dataset with three graph-based data augmentation methods for the rumor detection task.At the same time,we add a pretraining process based on deep clustering to the model,which reduces the introduction of noise caused by data augmentation.Experiments show that our model and data augmentation method can further improve the performance of the model on existing datasets and have the higher robustness.