Research On Prediction Models Of Information Diffusion In Social Networks

Online social networks facilitate people's access to massive information,while it also brings challenges to platform management.When information is disseminated on social net-works,an information cascade with a graph structure is formed through dissemination behav-iors such as retweet and sharing,which benefits the further spread of information.Predicting future development of information diffusion at the early stages has important values for the promotion of high-quality content,the blocking of rumors,and the enhancement of platform viscosity.The dissemination process in offline social networks is similar to the diffusion pro-cess of online social networks,and can also form a similar cascading structure.Therefore,the online information diffusion models can be applied to the research on the prevalence of offline virus transmissions.Early analysis virus diffusion can provide a reference for the prevention and control of the epidemic,which is of great social significance.In the current research on the prediction of information diffusion,there are different pre-diction scenarios and some unresolved problems according to the timing of the prediction,the perspective of the research,and the completeness of the data.Among them,there are challenges in temporal-spatial modeling of information cascades,learning of the bidirectional social influ-ences,the in-depth modeling of the structure and content characteristics that affect the user's retweet behavior,and modeling under incomplete data.This article considers the calculable fac-tors of information diffusion and applies ideas of computational communication and machine learning methods to address the dynamics of the graph structure in the dissemination process,the bidirectional social influence,the complex relevance of influencing factors,and the lack of complete actual offline dissemination data.We established computational information diffusion models based on graph structure of cascades,and carried out the research of information prop-agation prediction.The main work and innovations of this article are summarized as follows:In the scene of early popularity predictions of online information diffusion,the temporal characteristics and structure types of the information cascades reflect the timeliness,dissemina-tion speed,and popularity of information,which are important to the growth of the popularity of information diffusion.Existing models sample and learn the characteristics of the cascade from the static network structure,which leads to the problem of poor prediction results and large-scale model parameters.To solve the problems,this paper proposes a recurrent graph perception neural network leveraging on the cascade structural properties.The node characteristics are in-tegrated into the whole representation of the subgraphs of cascades,highly reducing the model parameters.The model can effectively capture the temporal and spatial dynamics of cascades and greatly reduce the prediction error of popularity prediction.In addition,the social influence of information cascades is bidirectional.This article proposes a bidirectional graph sequence attention networks,which learns the bidirectional characteristics of the cascade structure from the local and the global perspectives.In the early stage of information diffusion,the model improves the ability to identify the social role characteristics of nodes,and solves the problem that the current model can only learn the characteristics of the cascade structure from a single direction.The popularity of Information diffusion is the macroscopic performance of user forwarding behavior.The study of user forwarding behavior is not only of great value for link prediction and information recommendations,but also important for understanding the macroscopic popular-ity growth of cascades.In the scenario of predicting the user's forwarding behavior in advance of information diffusion,the user's forwarding behavior is affected by both social roles and information content topics.Furthermore,the social roles played by users in information dis-semination are also closely related to the information content topics.The same users may play different social roles on different topics.However,most methods of user behavior modeling methods lack joint modeling of social roles and text topics,so that they fail to learn the compli-cated patterns that affect reposting behavior of users.We propose a deep cascading model for micro information diffusion,which applies representation learning to deeply mine social roles and text topic features.The model overcomes the problem of simple considerings of influence factors in traditional learning methods and greatly improves the prediction effect.In the early prediction of the information of offline social networks,the spread of informa-tion or viruses is closely related to the graph structure of social relationships and the intensity of people's activities.However,these data are difficult to obtain directly and there is a lack of complete training data when modeling the problem.At the same time,it is difficult to ob-tain the real spreading popularity directly.Therefore,it is necessary to estimate and forecast the popularity based on observed data,and then predict the future popularity growth based on the current state.We take the transmission predictions of the new coronavirus as an example to study the modeling of offline information diffusion prediction.The assumptions of existing prediction methods that are based on infectious disease models are too simple to achieve a good performance.Most of the models do not consider the influence of offline social relationship network on dissemination,and it is difficult to estimate the true value and predict the trend of future spread precisely.In this paper,we apply the research results of the information diffusion models to the modeling of coronavirus transmission.A novel virus transmission model based on the diffusion of social network is proposed.The model solves parameters such as social net-works,population activities,and infection rates by fitting part of the observed data.It estimates the current popularity of virus and predicts the trend of future epidemics more accurately.The model can also restore the early virus transmission process,overcomes the problem that the cur-rent infectious disease model cannot evaluate the current situation intuitively,and also reveals the ”iceberg effect” and its reasons caused in the early diffusion of the virus.In summary,this article is oriented to three typical scenarios of social network information dissemination.Based on the graph structure modeling of social networks and information cas-cades,the spatial-temporal dynamics,bi-directionality social influence,feature relevance and data missing of complete data are studied respectively.This research applies machine learning methods to improve the prediction performance of information dissemination,and applies the information diffusion models to the new coronavirus transmission modeling to obtain analysis results with social value.This research has significant importance for understanding and man-agement of information diffusion on social network platforms and the prevention and control of new infectious disease.