Deep Attention Model Based On Ternary Features For Rumor Identification

The development of social networks has also brought about the proliferation of rumor information.Because of the defects of manual method,the automatic rumor identification method is extremely important for public safety and public opinion monitoring.In the existing rumor identification method,the traditional feature-based machine learning model has high requirements for expert knowledge and does not work well on other data sets.Moreover,the existing end-to-end method model does not fully exploit the features of events in propagation process.In view of the above problems,this paper proposes a time-synthesis representation of information in the social network with the ternary feature characteristics,based on the text content of the information in the event period,the user network of the event in propagation and the feedback signal of the user.In this paper,the supervised feature extraction is performed on the text topic of the information,the user network topology of the event and the event feedback signal.The heterogeneous feature is used to represent the propagation mode of the event.The main research contents have the following four points.1)Distributed representation learning of user network nodes based on space-time similarity kernel methodThe flow of information in different types of social media greatly influence the propagation pattern of the event,which is the existing methods ignored.For the dissimilarity of adjacent nodes and adjacency sparseness in open social networks,an improved skewed random walk algorithm based on the similarity of space-time kernel is proposed for distributed representation of nodes,which provides the basis for distributed representation of user network and rumor identification model.The proposed method is based on the similarity of local spatial structure in the open social network,considering the time interval of information interaction and the similarity of nodes under the relationship of time and space.Moreover,for the requirements of the rumor identification task,the attributes of the user nodes and the global features of the users group in the network are considered.2)Vectorization method of network topology based on distributed memory modelThe existing research methods are often too simple to use the characteristics of information dissemination.These methods cannot get the propagation characteristics in the whole information period comprehensively and have difficulty in using complex structural information.This paper verifies the distinction of user distribution patterns in the propagation network between rumor information and non-rumor information.Based on the distributed representation of user network nodes,the paragraph2 vec method is used to extract the vectors of user distribution in different periods of information propagation.The unsupervised distributed learning model is used to learn the network topology in propagation of event.3)Research on user view feedback signal based on seed wordsThe current rumor identification model algorithm often focuses on the characteristics of information content in a single piece of information or topic.However,with the spread of information,the event tends to form sub-events with different public opinion on social networks.The change of viewpoint in the process of information dissemination has a significant correlation with the propagation pattern of event.The information has obvious time effect with the user's feedback signal in propagation.This paper proposes an automated acquisition method for user's feedback signal of event in the propagation to extraction of the user's attitude expression in the information propagation.4)Deep self-attention rumor identification model based on ternary elements featuresIn this paper,the influence of the ternary features on the propagation mode in each time slice is analyzed,and the influence of the propagation mode at different time periods on rumor identification in the propagation is explored.Moreover,in the endto-end rumor identification model,the bi-gru layer is used for heterogeneous feature aggregation and the self-attention layer focuses on the potential information of the event propagation.This paper focuses on heterogeneous feature aggregation with the text content of the information,the user network of the event and the feedback signal of the users.The distributed representation learning method of the node and the vectorization method of the network topology are proposed.Moreover,the method proposes an automated acquisition method for user's feedback signals.Based on these studies,a model based on deep self-attention mechanism for rumor identification is designed.