ST-COVID: a Deep Multi-View Spatio-temporal Model for COVID-19 Forecasting

Abstract

The outbreak of COVID-19 has caused a dramatic loss of human life worldwide. Reliable prediction results are crucial on pandemic prevention and control in the early stage. However, it is a very challenging task due to insufficient data and dynamic virus spread pattern. Unlike most existing works only considering local data for a given region, we propose a spatio-temporal prediction model (ST-COVID) for COVID-19 forecasting to borrow experience from historical observations of other regions. Specifically, our proposed model consists of two views: spatial view (modeling global spatial connectivity with neighbor regions in geography and semantic space via GCNs), temporal view (extracting local and global latent temporal trend via CNNs and GRU). Extensive experiments on two real-world datasets at state and county level in US indicate that the proposed model outperforms over nine baselines in both short-term and long-term prediction.