Research On News Event Prediction System Based On Graph Neural Network

Internet news data has the characteristics of strong timeliness,high freedom and strong fluidity,which can reflect the social attitude towards the development of an event in time.Therefore,it is of great significance and value to study the event prediction in the news field.Previous studies mainly focus on event prediction based on temporal event sequence,ignoring the influence of structural relationship between words on event prediction.This makes event prediction suffer from insufficient feature extraction and low prediction accuracy.In this thesis,in order to better extract event features as well as to improve the accuracy of event prediction,we first design an event prediction algorithm based on dynamic graph attention network,which encodes both temporal information and the relationship between words within the event to fully extract the features of the event;then we design an event prediction algorithm based on feature fusion graph neural network,which incorporates the prior knowledge of the event sequence on the basis of the first step,so that the event prediction direction without bias or error;finally,an event prediction system is designed and implemented.The main research work of this thesis is as follows:(1)A dynamic graph attention network-based news event prediction model(DGAT)is proposed.Firstly,news events are constructed as news event graphs,and then the graph attention network is used to encode the news event graphs and combine the temporal feature encoding module to encode both the event information of the current event sequence and the structural information between words to obtain the feature vector of the current event sequence.Finally,the vector is input into the decoder to get the final prediction result.In this thesis,the proposed model is validated using the Chinese Sina news topic dataset.The experimental results demonstrate that the proposed model can effectively utilize the information of the current event sequence,and the error rate is reduced by 1.8% and 1.26% on the two datasets,respectively,compared with previous algorithms.(2)A feature fusion dynamic graph neural network-based news event prediction model(FFDGAT)is proposed.News event sequences contain less information,and only information of the current event sequence is extracted,and the DGAT model may forget the past events.To address the problems of DGAT model,FFDGAT model is designed on the basis of DGAT model.The model adds the semantic background of the current event sequence as prior knowledge,encodes the prior knowledge by gated graph neural network,and fuses the obtained feature vectors with those obtained by DGAT model to achieve news event prediction.The experimental results demonstrate that the prior knowledge can provide more guidance information for event prediction,and the error rate is reduced by 2.21% and 1.58% on the two datasets,respectively,compared to the DGAT model.(3)Design and implement a news event prediction system based on graph neural networks.The above two points are to improve the event prediction effect from the algorithm level.To facilitate user use of the proposed news event prediction model,based on the above research content,this thesis constructs a news event prediction system that can process user-input news event data and perform news event prediction.The system mainly includes user information management,news event prediction,visualization interaction,and other functions.