Research And Implementation Of Multi-event Prediction Method Based On Graph Neural Network

Social events refer to social activities that are premeditated or spontaneously gathered by different actors at a specific time,place,and carried out in different forms such as appeals,cooperation,and protests.Improper handling may lead to violent conflicts and social instability.Multi-event prediction aims to predict multiple events with different types that may occur in the future based on historical social events.It is the basis for exploring the internal evolution mechanism and development pattern of social events.It can assist the government to carry out crisis event early warning,policy intervention and personnel guidance for adverse social events in advance.Current research focuses more on a single type of event prediction,ignoring the extensive correlations between multiple events in terms of structural and semantic features,or taking insufficient consideration of the long-term temporal dependence of social events.The paper aims to improve the above shortcomings based on the graph neural network and further improve the performance of multi-event prediction.The main research contents and innovations are as follows:(1)Multi-event Prediction based on Knowledge-aware attention and Temporal Graph Convolution.Aiming at the problem that the relation and semantic differences between multi-event are not fully represented in the existing methods,which leads to the problem of unsatisfactory multi-event prediction performance,the paper proposes a multi-event prediction method based on knowledge-aware attention and temporal graph convolution.The method constructs multi-event actors,event types,and time as a temporal event graph to model the potential correlation of multi-event in structure relation and time.Then,a graph convolution method based on knowledge-aware attention is designed to assign different importance values to the neighborhood relations and neighborhood entities of nodes in the temporal event graph,to obtain a more reasonable event graph embedding representation.To capture more discriminative event semantic features,the paper uses the topic model LDA to extract topic keywords of different event types to enhance the semantics of event graphs,and performs temporal encoding and prediction through a recurrent encoder.Compared with the existing models,the model performs better on multi-event prediction tasks,and at the same time,the effectiveness and rationality of the model are further verified through ablation experiments and a case analysis.(2)Enhance Long-term Temporal Dependence based on Dilated Graph Convolution for Multi-event Prediction.The outbreak of social events is usually accompanied by complex temporal patterns,such as epidemic disease transmission events.It requires a large amount of historical information from the past to predict the future,which requires the model to have the ability to capture long-term time-dependent features.Therefore,the paper proposes to employ dilated graph convolutions to enhance long-term temporal dependencies,and then solve the multi-event prediction task.Based on the temporal event graph,the event semantics is modeled as a semantic context graph to highlight the dynamic development of event semantics;then a graph convolutional network is used to capture the relation and semantic embeddings represent sequence;and then the extended causal convolutions is used product,at the temporal level to capture long-term-time dependence features across different time intervals.We conduct extensive Experiments on social events and epidemic disease Covid-19 event datasets and the in-depth analysis of historical window sizes,the results verify the model’s predictive performance advantages and its advantages in long-term temporal feature capture.