Exploring the Predictive Power of Correlation and Mutual Information in Attention Temporal Graph Convolutional Network for COVID-19 Forecasting

Subas Rana, Nasid Habib Barna, John A. Miller

Abstract: Accurately forecasting the COVID-19 spread across all states is crucial for implementing effective measures to control its transmission and minimize its impact. Since the virus’ spread in one state can significantly affect other states over time through connections between them, a graph structure with temporal data is needed to capture the interdependence of COVID-19 spread among the states in the United States. In forecasting tasks that involve complex spatial and temporal dependencies, it is crucial to ensure that the model captures these dependencies accurately. In this study, we implemented an Attention Temporal Graph Convolutional Network based model for COVID-19 mortality long-term prediction which can effectively capture these dependencies. This model incorporates attention that enables us to weigh the significance of different time points and focus on the most informative data, including both adjacent and distant time points that capture the temporal dynamics accurately. For capturing spatial dependencies, we assessed the impact of using Pearson’s correlation and Mutual Information to establish connections between highly dependent states. Our experiments showed that our model, particularly when utilizing mutual information, outperformed the existing baselines and the models that only consider neighboring states resulting in lower sMAPE and MAE values. This emphasizes the importance of selecting the appropriate technique for accurate COVID-19 forecasting in each state. Furthermore, our model achieved the second-highest performance among the forecasting models submitted to the Centers for Disease Control and Prevention.