Research On Rumors Detection Based On Transfer Learning

With the rapid development of the Internet,online media represented by social networking sites,Weibo,blogs,and forums has become the largest channel for information generation and exchange in the Internet world,and has also created a hotbed for the generation and dissemination of rumors.The rumors in the online media often aim at attracting traffic to create panic,and thus often have a large spurt and maliciousness,and become a big cancer that destroys the order and environment of the cyberspace.Therefore,the network rumors are effectively identified,which is relatively large.Economic,social and practical significance.Based on the traditional rumor detection work,this paper will focus on the field differences of rumors and the updating of rumor detection models.With the idea of transfer learning,the deep transfer network is used to realize cross-domain detection of network rumors and the cross-time domain detection of network rumors is realized by using ACGAN architecture.The article innovations are as follows:The first is to propose a cross-domain rumor detection model based on deep transfer network.Current rumor detection uses a unified model to detect network rumors,ignoring domain differences in rumors.In this paper,we try to realize the effective detection of network rumors in different fields,and use the deep transfer network for cross-domain modeling for the problem of insufficient annotation data in some fields.In the deep transfer network,it is assumed that the source domain has annotation data,the target domain is an unlabeled dataset,and the text information,user information,and propagation information of the network information are selected as elements for identifying rumors.In the deep migration network,it is assumed that the source domain has the annotation data,and the target domain is the unlabeled dataset.Through the deep migration network,the labeled data in the source domain can be effectively migrated,and the unlabeled target domain is constructed to construct the categorical detection classifier.The second is to propose a cross-time domain rumor detection model based on ACGAN architecture.The characteristics of the network rumors change with time,and the new features of the rumors can be more effectively updated when the rumor detection model is updated.In this paper,when the model is updated,the data set is divided into historical rumor data and current rumor data.The historical data is re-encoded by the generator G in the ACGAN architecture to make the data distribution tend to be rumored in the current period.This cross-time domain data migration guarantees The model can reflect more of the new features of network rumors when it is updated.Experiments show that in the cross-domain rumor detection,compared with the rumor detection method and sub-domains in the non-sub-domain,but not using the rumor detection method of migration learning,the method improves the F1 index by 10.3% and 8.5%,respectively.Compared with the existing methods,it can be seen that the cross-domain rumor detection model proposed in the paper is superior to the unsupervised method in F1 value and stability;in the cross-time domain rumor detection,compared with retraining and pre-processing The training update method,the F1 index of this model is 5.5% and 3.7% higher respectively,which solves the problem of reducing the accuracy of the rumor detection model caused by the change of rumor feature distribution to some extent,and improves the stability of rumor detection.