Peramalan Jumlah Kasus COVID-19 Menggunakan Joint Learning

Abstract

COVID-19 is a dangerous illness because it spreads quickly and easily. Vaccines are already available but the pandemic isn’t likely to end soon. Forecasting is hoped to help handle the pandemic. Deep learning, specially LSTM, has been used to forecast COVID-19 case count in some regions. However, deep learning models generally need a lot of training data while COVID-19 daily data are scarce. However, COVID-19 pandemic happens in many regions. This research aims to use joint learning with data from other regions to improve model performance with fewer data and to use the model to forecast until 9 months since the date of last data taken. Joint learning was done by making models share some parts and training the models together. To overcome the different data scale and pandemic age in the regions, the data was first transformed into discrete SIRD variables and was evaluated using RMSSE. Joint learning failed to improve the model performance in this research. The proposed model performance was signficantly better than ARIMA-SIRD and SIRD model but wasn’t better than normal encoder-decoder LSTM. The models only reached RMSSE below one occasionally. Additionally, it was found that doing joint learning with all regions without selecting them by clustering can make the model performance worse instead. It was also found that RMSSE is too sensitive to a mostly stagnant time-series due to its division by the error of one-step naïve forecast.