Interpretable Temporal Attention Network for COVID-19 forecasting

The worldwide outbreak of coronavirus disease 2019 (COVID-19) has triggered an unprecedented global health and economic crisis. Early and accurate forecasts of COVID-19 and evaluation of government interventions are crucial for governments to take appropriate interventions to contain the spread of COVID-19. In this work, we propose the Interpretable Temporal Attention Network (ITANet) for COVID-19 forecasting and inferring the importance of government interventions. The proposed model is with an encoder-decoder architecture and employs long short-term memory (LSTM) for temporal feature extraction and multi-head attention for long-term dependency caption. The model simultaneously takes historical information, a priori known future information, and pseudo future information into consideration, where the pseudo future information is learned with the covariate forecasting network (CFN) and multi-task learning (MTL). In addition, we also propose the degraded teacher forcing (DTF) method to train the model efficiently. Compared with other models, the ITANet is more effective in the forecasting of COVID-19 new confirmed cases. The importance of government interventions against COVID-19 is further inferred by the Temporal Covariate Interpreter (TCI) of the model.