A machine learning approach to analyzing spatiotemporal impacts of mobility restriction policies on infection rates

This study analyzed the impact of a range of policies that restrict travel accessibility and mobility on infection rates for the original strain of the virus during the first year of the COVID-19 crisis. We constructed a multidimensional dataset and developed an effective data-driven predictive model to investigate causality between a policy, mobility, and an infection, drawing upon spatiotemporal perspectives. The multidimensional dataset included daily infections, daily restriction policies, and daily and hourly multimodal travel patterns. We quantified and normalized the dataset in relation to pre-COVID-19 policies and travel activities. A machine learning framework that integrated principal component analysis (PCA) and a Gaussian process regression (GPR) was formulated to evaluate the effectiveness of mobility restriction policies and their optimal implementation time during the infancy stage of the pandemic. In a case study, we selected Seoul in South Korea and Sydney in Australia for model calibrations and validations. Both countries deployed comprehensive urban restriction policies during the worldwide pandemic. The proposed model produced better performance than diverse non-parametric and parametric models to estimate the daily number of infections in the two areas. Furthermore, we discovered effective restriction policies and the best times to implement them to minimize the number of acquired COVID-19 cases by analyzing coefficients in PCA and GPR kernel functions. Our finding has far-reaching policy implications. First, the proposed methods can be used for formulating restriction policies for other regions with diverse population densities as the chosen cities in this case study. Second, our finding contributes to evidence-based policymaking.