| With the development of the Internet,the degree of informatization in today’s society has become more and more remarkable.From the perspective of information volume,the amount of data available on online platforms shows a rapid growth trend.From the perspective of communication efficiency,the popularity of online platforms and the surge in the number of people accessing the Internet have led to a qualitative increase in the speed of dissemination of events on online platforms.On the one hand,this situation makes it difficult for online media workers to quickly and accurately filter out the real hot events from the many events.On the other hand,it leads to users of social platforms not being able to receive the real hot events first.Therefore,the prediction study of event heat has become one of the current research priorities.Most of the current prediction studies on event heat rely excessively on the trend of event heat values in the time dimension,but due to the growth of information dissemination speed,the temporal characteristics have a serious lag in the prediction of event heat.Therefore,in order to achieve the prediction of event heat value based on the content of the event itself without considering the time dimension.This study presents three aspects of constructing a hot event dataset,modeling and inference of event heat based on factor models,and accelerated processing of event heat prediction algorithms.(1)Hot event dataset construction.Firstly,we obtain specific data in Baidu platform according to predefined rules and construct the hot events dataset.The data obtained include five aspects: title of the hot event,brief content of the event,source of the event posting,accompanying image of the event and real-time heat value of the event.It provides a data base for exploring the generation of hot events.(2)Factor model-based event heat modeling and inference.First,on the one hand,the base heat value of the event is modeled based on the topic distribution of the event and the extracted semantic keywords;on the other hand,the interrelationship between the semantic keywords of the event is modeled based on the factor model,so as to realize the simulation of the heat value of the event excitation.And the modeling of event heat values is achieved by multimodal feature fusion.The event heat prediction algorithm is constructed based on the modeled event heat values,and the objective function is designed assuming that the event heat values obey Poisson distribution,and the prediction of the event heat values is achieved by optimizing the objective function.After that,experiments are designed to verify the rationality and effectiveness of the algorithm on the simulated dataset,the Chinese hot events dataset and the English paper citation dataset HEP-PH,and ablation experiments are designed to explore the effects of different terms in the multimodal fusion module on the overall performance of the algorithm.(3)Accelerated processing of event heat prediction algorithms.In order to achieve optimization of the event heat prediction algorithm,this study uses the Laplace transform of the original objective function to achieve a simplification of the objective function without changing the physical meaning of the parameters.The assumption that it obeys Poisson distribution is broken,and a new objective function is obtained.After that,the prediction of the event heat value is achieved by optimizing the new objective function,and the optimization effect of the accelerated processing on the operation efficiency of the algorithm is explored through comparative experiments. |
