Research On Epidemic Forecast Based On Graph Neural Network

Epidemics refer to infectious diseases that spread beyond a certain threshold.So far,people have suffered from disasters caused by various epidemics.From the early smallpox virus to the recent COVID-19 virus,each virus has seriously threatened the safety of people’s lives and property.Therefore,it is increasingly important to predict the epidemic accurately.Epidemic forecasting aims to predict the number of cases,morbidity,or mortality of an epidemic in different regions,times,and populations.The methods of epidemic forecasting have become increasingly mature with the development of science and technology,but there are still a series of problems.For example,the real-time mobility factor is seldom considered,the research domain is single,the long-term forecasting is poor,and it cannot be used for unknown prediction tasks.Because of the problems existing in the current epidemic forecasting methods,under the premise of using the graph neural network,this thesis has done the following three works:(1)Due to the lack of consideration of real-time mobility data in existing research,this thesis proposes a new computing framework based on graph neural networks-Information Aggregation Network(IAN).This thesis considers both the features of regional case data and the features of personnel mobility data between regions.At the same time,in order to optimize the early prediction model of each country,the transfer learning method(TL)is added on the basis of this framework.Experimental results on datasets from four European countries show that IAN and IAN-TL are significantly better than traditional methods and can effectively reduce prediction errors.(2)To address the problems of a single research domain and poor long-term prediction,this thesis uses a method that combines human-mobility factors and timeseries factors-the Spatio-temporal convolutional network(G-TCN).Specifically: firstly,this thesis uses data on the mobility of people between regions to generate graphs of regional relationships.Secondly,to process the spatial information at each moment,this thesis applies to multiple graph convolutional neural network modules to aggregate multilayer neighborhood information.And this thesis inputs the information obtained by graph convolutional neural network modules at different moments into temporal convolutional network modules,which are used to process the time-series information.This thesis tests the proposed G-TCN method using datasets from four countries.The experimental results show that G-TCN has lower prediction errors than other compared methods and can better fit the trend of COVID-19 development.(3)Most current epidemic forecasting methods require predefined graph structure information and cannot be used for unknown prediction tasks.Therefore,this thesis proposes a Multivariate Spatio-temporal Graph Neural Network(MSGNN)to solve the problem.First,this thesis utilizes a hidden correlation layer to extract potential relationships between variables,thereby generating graph sequences at different times.Then,this thesis inputs the obtained graph series into the graph processing module to obtain more valuable information at each level of the multiple neighborhood levels.Besides,a long and short-term memory network module is used to capture the time-series dependencies within each variable.This thesis evaluates the proposed method on COVID-19 and other datasets.The experimental results show that the method proposed in this thesis obtains more accurate forecasting results and outperforms other comparative methods.