RAPID: Dynamic Graph Neural Networks for Modeling and Monitoring COVID-19 Pandemic

新型冠状病毒COVID-19已成为人类历史上最大的流行病之一，并对公共卫生、社会和经济产生了持久影响。美国的病例数已超过100万例，总死亡人数超过5万人。迫切需要进行研究和开发，以便对病毒的传播进行预测性了解，从而使缓解方法能够减轻COVID-19的负面影响。传统的流行病学模型在构建预测模型时通常只考虑少数特征，可能无法捕捉这种新流行病的潜在危险因素和各种干预机制的效果。在这个项目中，研究人员开发了新的机器学习方法，可以同时建模和预测COVID-19的传播，检测和监测风险因素，并评估干预措施在时间和空间上的有效性。新模型吸收并整合了不同来源的异构和快速积累的数据，如出版物，新闻，人口普查，社交媒体和疫情观察跟踪器。它采用了一种新的上下文语言模型，从大量的文本数据中准确识别命名实体和关系，并构建知识图来提取潜在的风险因素。构造了一个动态图。每个位置节点可以具有一组静态和时间相关属性。事件、个人行为、社会活动、干预被映射到活动节点，其中边连接到当时对应的位置节点。一种新的动态图神经网络被训练来随着时间的推移对所有位置进行联合预测。具有显著关注权重的活动节点代表主要的风险因素或有效的干预机制。该项目将导致预测模型和所有源代码的公开传播，立即有利于对抗COVID-19。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

The novel coronavirus, COVID-19, has become one of the biggest pandemics in human history and has generated lasting impacts on public health, society, and economy. The number of cases in the United States has passed 1 million with a total number of deaths over 50 thousand. There is an urgent need for research and development that can bring a predictive understanding of the spread of the virus, thereby enabling mitigation methods to alleviate the negative effects of COVID-19. Traditional epidemiological models usually take into consideration only a small number of features in building a prediction model, which may not be able to capture potential risk factors and effects of various intervention mechanisms of this new pandemic. In this project the investigators develop novel machine learning methods that can simultaneously model and predict the COVID-19 spread, detect and monitor risk factors, and evaluate effectiveness of interventions over time and space. The new model ingests and integrates heterogeneous and rapidly accumulating data across diverse sources, such as publications, news, census, social media, and outbreak observation trackers. It employs a new contextualized language model to accurately recognize named entities and relations from vast text data and build knowledge graphs to extract potential risk factors. A dynamic graph is constructed. Each location node may have a set of static and time-dependent attributes. Events, individual behaviors, social activities, interventions are mapped to activity nodes with edges connecting to the corresponding location nodes at the time. A novel dynamic graph neural network is trained to perform joint predictions of all locations over time. Activity nodes of significant attention weights represent major risk factors or effective intervention mechanisms. The project will result in public dissemination of the prediction model and all source codes, immediately benefiting the combat against COVID-19.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.