Research On Information Popularity Prediction Based On Graph Neural Networks

With the rapid development of Internet technology,social networks,as an important information carrier,have greatly changed the way of information dissemination,enabling everyone to play the role of creators,disseminators,and consumers of information.Efficient prediction of information popularity has become one of the research hotspots.However,due to the variety of content,the complexity of network structure,and the diversity of influencing factors,the prediction task faces many challenges.How to establish an efficient prediction model is the key problem to be solved.Based on the full investigation of relevant literature,this paper deeply studies and analyzes the factors that influence the information popularity,and uses deep learning technology to model the key factors in the process of information dissemination to accurately capture the growth pattern of popularity.The main research contents and innovations are as follows:(1)The existing methods of information popularity prediction are summarized.Through combing a lot of research,this paper firstly generalizes the research background and significance of information popularity prediction in social networks,and briefly introduces the information dissemination mechanism.Secondly,information popularity prediction methods mainly include three types: feature extraction based,generation model based,and deep learning based prediction approaches,the research achievements,advantages,and disadvantages of each method are described in detail.(2)A collaborative prediction model based on graph neural networks is proposed.To make full use of the structural features,temporal information,and node attributes of information dissemination to predict information popularity,an endto-end model CCas GNN based on graph neural networks is proposed.Firstly,to capture the temporal information in the propagation process,we use the position encoding function to calculate the position code of each node according to the order of nodes in the propagation sequences.Secondly,we jointly learn the node representation vectors using Graph Convolutional Network(GCN)and Graph Attention Network(GAT)to introduce network structural features and node attributes.In addition,the position code is added to the neural network operation layer to fuse the time series information,then the cascade representation vector is obtained by pooling operation,and finally,it is sent to the multi-layer perceptron to acquire the prediction result.The model effectively combines spectral-based and spatial-based graph neural networks and improves the prediction accuracy while ensuring time efficiency with the help of position encoding.(3)A subgraph-based deep learning prediction model is proposed.To better capture the structural features and temporal information during information dissemination,we propose a deep learning based end-to-end model Cas Seq GCN.By defining the active state of a node,we divide one cascade into multiple subgraphs,each containing the topology of the cascade and the status of nodes.The representation of nodes in one subgraph can be learned by the classical Graph Convolutional Network(GCN).Then,through the aggregation algorithm based on dynamic routing proposed in this paper,all nodes are aggregated into a subgraph representation vector,which contains the structural features of the subgraph and the status information of nodes.A sequence of the subgraph representation is sent to the long short-term memory network(LSTM)to extract the temporal order of spreading.Finally,the output of LSTM is fed into the multi-layer perceptron to obtain the prediction result.This model avoids large-scale feature engineering and effectively combines structural features and temporal information to improve the prediction performance.In order to verify the effectiveness and feasibility of the above models,a series of contrast experiments and ablation studies are designed in this paper.The relevant results demonstrate their excellent prediction performance and the necessity of each module.