Research On Prediction Model Of Message Dissemination Range Based On Deep Learning

With the rapid development of online social network platforms and mobile network devices,the way people share and obtain messages has gradually changed.Diversified social network platforms have greatly facilitated the generation and dissemination of messages,and have intensified the competition between massive amounts of information for user attention.The competition also highlights the importance of predicting the popularity of news.Understanding how news is spread in social networks and what factors drive the successful spread of news,and accurately predicting the size of the population that the news may affect,is a challenging but extensive in real life Application work.At present,the research on message propagation prediction,also called message cascade prediction,is mostly based on feature methods,generation methods and deep learning methods.Traditional feature-based methods need to manually extract features and then use machine learning methods for further learning,which is cumbersome and features subjective.The method based on deep learning generally uses end-to-end deep learning to autonomously learn the hidden features in the message dissemination process,avoiding the complicated feature engineering in machine learning,and improving the accuracy of prediction.The Hawkes process in the generation method is a widely used classic method of modeling the message propagation process.It mainly uses the impact of previous events on future events to model the probability of occurrence of new events.It is based on the user's influence,self-incentive mechanism,The three factors of time decay effect can well explain the forwarding phenomenon in the process of message propagation.Based on these,this paper proposes two models for predicting the spread of messages.First,this paper proposes a message propagation prediction model based on the spatiotemporal attention mechanism.According to a given message cascade graph,multiple propagation sequences are obtained by random walk.The random walk process samples nodes according to the degree of nodes.The obtained propagation sequence can reflect the flow of message propagation and contain the information of the message forwarder and For local structure information,put the obtained propagation sequence into a two-way GRU to obtain a vector representation of the sequence.Secondly,the corresponding forwarding time interval of each sequence is obtained according to the forwarding process of the message in the observation time window in the early stage.This article believes that for the sequence of the same length obtained by random walk,the smaller the forwarding time interval,the greater the contribution to the message propagation.The force mechanism combines the timing information to select the more important propagation path for the cascade.In addition,since the message is not isolated in the process of dissemination,this article also combines the influence of other messages on the target message to predict the range of message dissemination.Secondly,this paper proposes a message propagation prediction model based on the deep Hawkes process.Use deep learning methods to learn the three important factors used in the Hawkes process to model message dissemination,namely,the user's influence in social networks,the influence of past events on future events(self-incentive mechanism),and the influence of this influence over time The phenomenon of weakening(time decay effect).This paper proposes to use a graph convolutional network based on the spatial domain to directly process the cascaded graph,obtain node representations to represent the user's influence,and then obtain the true forwarding path of the message according to the cascaded graph,and input it into the recurrent neural network to obtain Considering the propagation path representation of the self-excitation mechanism,the time attenuation factor is added when the propagation path representation is aggregated,which means that the influence of the previous forwarding is weakened with the passage of time.In this way,the high accuracy of the deep learning method in prediction and the high understanding of the Hawkes process in explaining the information dissemination mechanism are combined to narrow the gap between prediction and understanding in the process of message dissemination,and further improve the accuracy of prediction.