Research On Prediction Of Content Information Cascade Influence Based On Propagation Structure

Thanks to the rapid development of Internet technology,the channels and ways for users to obtain information have been greatly expanded,which greatly facilitates the dissemination of content.At the same time,users can publish,generate or forward content based on the Internet platform,so that every user can act as a creator,consumer and disseminator of information.Therefore,the prediction of content influence has become a major research hotspot.However,due to the diversity of content,the complexity of communication structure and many other influences,the task of content influence prediction is challenged.How to establish an accurate prediction model has become a major difficulty,and it is also an urgent problem to be solved in various researches.This paper carefully sorts out the existing research methods and paradigms through a full investigation of the existing relevant literature,and on this basis,in-depth research and analysis of many factors that affect influence prediction,and based on two classic content dissemination scenarios:Social content dissemination and dissemination of academic papers,using deep learning technology to model the key factor in information dissemination—the dissemination structure,that is,to use the dissemination structure of the content to predict the influence.The main research contents and contributions of this paper are as follows:·1)Summarize the existing content information influence prediction methods.By sorting out a large number of research works,this paper firstly summarizes the research background and significance of content information influence prediction;secondly,the content information influence prediction methods mainly include three categories:prediction methods based on feature extraction,prediction methods based on generative models and prediction methods based on deep learning.And the research status,advantages and disadvantages of various methods are described in detail.·2)In this paper,we propose a Dynamic Mapping(DMM)mechanism to map the sub-cascade propagation structures of different periods into Degree Interval Sequences(DIS).Through this mechanism,the entire propagation structure of content information is divided into dynamic sub-cascade propagation network sets according to the time order,and further transformed into dynamic sub-cascade propagation network sets into dynamic degree distribution sequence sets,which can reduce the complex high-dimensional network-type data structure into onedimensional vectors,significantly reduce the space occupied by the model,improve the computational efficiency,and at the same time can significantly preserve the The information of the original network propagation structure can be substantially preserved.Then,in order to further extract the temporal information of degree distribution sequences,we introduce Temporal Convolutional Neural Network(TCN)for the first time in the field of content information influence prediction,specifically,in this paper,we feed the set of degree distribution sequences obtained by dynamic mapping mechanism to TCN to learn the structural information of the whole propagation network.Thanks to the causal convolution and inflation convolution mechanisms of the temporal convolutional neural network,the temporal convolutional neural network can capture the dependencies between degree distribution sequences and can better learn the temporal structure information.Finally,the output of the temporal convolutional neural network is fed into a multilayer perceptron to obtain the final impact prediction results.The effectiveness of our method CasTCN is verified by detailed comparison experiments,ablation experiments and parameter sensitivity experiments,and the efficiency of our proposed method CasTCN is demonstrated by giving the number of model parameters and computational effort.·3)For academic citation networks,this paper focuses on predicting the influence of academic content by using deep learning methods to characterize the static propagation network structure information formed by early citations of academic content.Specifically,this paper defines the concept of network layers for academic citation networks,and characterizes the degree sequences generated by different layers using a multilayer perceptron,and inputs the characterization results into a Bidirectional Recurrent Neural Network(BRNN)to extract dependencies,and finally sends the output results to the output layer to obtain prediction results.The effectiveness of our proposed method is verified by adequate comparison experiments on two real-world datasets and parameter sensitivity experiments,and the propagation structure of the network can be captured by our proposed degree sequence is preliminarily verified by analysis.Thanks to the great success of the deep learning method in the field of content information dissemination,this paper developed a cascading communication influence prediction framework for social content CasTCN and a cascading communication citation prediction framework for academic content DeepCCP based on the deep learning method.In order to overcome the problem of large parameters and low operating efficiency of the traditional deep learning method,this paper uses the deep learning method to directly learn the overall structure information of the cascaded propagation network,directly obtain the whole graph representation,and skip the single node in the learning network Indicates the step of further aggregation,which greatly reduces the amount of model parameters and improves the efficiency of model operation.