基于泊松回归的传染病感染病例数预测

Predicting the number of infected cases for epidemics plays an important role in the epidemic early warning system. Of all techniques for the prediction task on the number of infected cases, the Poisson regression model is an important and widely used one. Existing studies based on Poisson regression models do not consider the special characteristic of epidemic data. For example, those studies do not take the external factors, which affect the disease diffusion, into consideration. In the modeling process, those studies do not make full use of the time series information and cannot handle big data. Moreover, the existing Bayesian Poisson regression models cannot be applied to epidemic data due to its special properties. Our project will focus on four aspects in the study of Poisson regression models for this prediction task: (1) A study on Poisson regression models based on the internal and external factors which affect disease diffusion; (2) A study on Poisson regression models by utilizing the time-series information; (3) A study on large-scale training algorithms for Poisson regression models to deal with the big data problem; (4) A study on Bayesian Poisson regression models applicable to epidemic data. Based on the four aspects, our project will propose a diffusion-network-based Poisson regression model, a multi-task extension of the Poisson regression model based on time-series information, a distributed optimization algorithm for Poisson regression models based on cloud computing, and a new Bayesian Poisson regression model for epidemic data. Our project will produce about 3-5 high-quality papers on international or domestic journals and conferences. This project will also have a long-term benefit in terms of education by training about 1-2 postgraduate students.

传染病感染病例数预测在传染病预警中起着重要作用。泊松回归是常用的研究传染病感染病例数的一种机器学习技术。已有的基于泊松回归的研究并不能很好地考虑传染病数据的特殊性，主要体现为忽略影响传染病传播的外部因素，没有充分利用传染病数据的时序性，没有考虑到传染病数据的大规模性，已有的贝叶斯泊松回归模型并不适用于传染病数据。本项目重点在以下四个方面对应用于传染病感染病例数预测的泊松回归模型进行研究：（1）基于影响传染病传播的内外因素的泊松回归模型研究；（2）基于传染病数据时序性的泊松回归模型研究；（3）针对大规模传染病数据的泊松回归模型训练算法研究；（4）设计适用于传染病数据的贝叶斯泊松回归模型。项目组将针对以上四个方面分别提出一种基于传播网络的泊松回归模型、一种基于多任务学习的泊松回归模型、一种基于云计算的分布式优化算法以及一种贝叶斯泊松回归模型。本项目可望发表高质量论文3-5篇，培养研究生1-2名。

传染病感染病例数预测在传染病预警中起着重要作用。泊松回归是常用的研究传染病感染病例数的一种机器学习技术。已有的基于泊松回归的研究并不能很好地考虑传染病数据的特殊性，主要体现为忽略影响传染病传播的外部因素，没有充分利用传染病数据的时序性，没有考虑到传染病数据的大规模性，已有的贝叶斯泊松回归模型并不适用于传染病数据。本项目对传染病病例数以及其涉及的多任务学习进行研究。本项目共发表CCF A类会议文章12篇，CCF B类会议文章1篇，CCF A类刊物1篇，CCF C类刊物1篇。