Learning from Others: A Data Driven Transfer Learning based Daily New COVID-19 Case Prediction in India using an Ensemble of LSTM-RNNs

Abstract

Accurate prediction of the number of COVID-19 infected cases per day is fast becoming a critical necessity globally to mitigate the burden on various health systems. In a densely populated country like India which has currently the second highest number of infections and limited medical support, it is a need for the authorities to know the statistics beforehand to address these issues more effectively. In this article, a data driven transfer learning based model is proposed that takes into account the conditions of different countries which have witnessed the COVID-19 infection. We have taken four countries to be the source domain for transfer learning scenario namely, the United States of America, Spain, Brazil and Bangladesh. We have pre-trained four different LSTM-RNN models with each of the country’s data and have re-trained (fine tuned) each of the models using only a very small portion of Indian data on COVID-19. Predictions of these four models are averaged to get the actual prediction. It is seen that such an ensemble model outperforms all the compared models and accurately predicts even the daily cases. This may be due to the fact that the four LSTM-RNNs used here could successfully take into account the diversities of conditions. As India is a diverse nation with variety of climates, it makes more sense to incorporate such transfer learning techniques.