Using a Layered Ensemble of Physics-Guided Graph Attention Networks to Predict COVID-19 Trends
Using a Layered Ensemble of Physics-Guided Graph Attention Networks to Predict COVID-19 Trends
复制标题
DOI：
10.1080/08839514.2022.2055989
复制
发表时间：
2022-04
影响因子：
2.8
通讯作者：
Connie Sun;Vijayalakshmi K. Kumarasamy;Yu Liang;Dalei Wu;Yingfeng Wang
中科院分区：
文献类型：
--
作者：
Connie Sun;Vijayalakshmi K. Kumarasamy;Yu Liang;Dalei Wu;Yingfeng Wang
摘要2019冠状病毒病疫情迅速蔓延，并对世界上大多数国家造成重大影响。在多个尺度上提供COVID-19的准确预测将有助于为公共卫生决策提供信息，但最近的预测模型通常用于州或国家一级。此外，传统的数学模型受到简化假设的限制，而机器学习算法很难推广到看不见的趋势。这激发了对混合机器学习模型的需求，这些模型集成了领域知识以进行准确的长期预测。我们提出了一个三层的地理信息集成，一个广泛的同行学习框架，用于预测国家，大陆和全球层面的COVID-19趋势。作为基础层，我们开发了一个国家级的预测器，使用混合图形注意力网络，它结合了修改后的SIR模型，自适应损失函数和边缘权重的移动性数据。我们汇总了163个国家的GAT，以训练合奏的大陆和世界MLP层。我们的研究结果表明，将定量准确的方程和真实世界的数据来模拟社区间的相互作用，提高时空机器学习算法的性能。此外，我们证明，整合地理信息（大陆组成）提高了我们的分层架构中的世界预测的性能。
ABSTRACT The COVID-19 pandemic has spread rapidly and significantly impacted most countries in the world. Providing an accurate forecast of COVID-19 at multiple scales would help inform public health decisions, but recent forecasting models are typically used at the state or country level. Furthermore, traditional mathematical models are limited by simplifying assumptions, while machine learning algorithms struggle to generalize to unseen trends. This motivates the need for hybrid machine learning models that integrate domain knowledge for accurate long-term prediction. We propose a three-layer, geographically informed ensemble, an extensive peer-learning framework, for predicting COVID-19 trends at the country, continent, and global levels. As the base layer, we develop a country-level predictor using a hybrid Graph Attention Network that incorporates a modified SIR model, adaptive loss function, and edge weights informed by mobility data. We aggregated 163 country GATs to train the continent and world MLP layers of the ensemble. Our results indicate that incorporating quantitatively accurate equations and real-world data to model inter-community interactions improves the performance of spatio-temporal machine learning algorithms. Additionally, we demonstrate that integrating geographic information (continent composition) improves the performance of the world predictor in our layered architecture.