Identifying the Effective Restriction and Vaccination Policies During the COVID-19 Crisis in Sydney: A Machine Learning Approach

This study identified effective COVID-19 restriction policies and the best times to deploy them to minimise locally acquired COVID-19 cases in Sydney. We normalised stringency levels of individual COVID-19 policies, usage levels of urban mobility, and vaccination rates to establish unbiased multivariate time-series features. We introduced the time-lag from 1 day to 15 d before when the governments have officially announced the number of locally acquired COVID-19 cases to the multivariate features. This time-lag dimension allows us to decide critical timings for announcing various COVID-19 related policies and vaccinations to control rapidly increasing infections. We used principal component analysis (PCA) to reduce the dimensions of the multivariate features. A Gaussian process regression (GPR) estimated the daily number of locally acquired COVID-19 cases based on the reduced dimensional features. The model outperformed diverse parametric and non-parametric models in estimating the daily number of infections. We successfully identified effective restriction policies and the best times to implement them to minimise the rate of confirmed COVID-19 cases by analysing PCA coefficients and kernel functions in GPR.