Unsupervised origin-destination flow estimation for analyzing COVID-19 impact on public transport mobility

The outbreak of COVID-19 caused unprecedented disruptions to public transport services. As such, this paper proposes a methodology for analyzing COVID-19 impact on public transport mobility. The proposed methodology includes: (1) a new unsupervised machine learning (UML) method, which utilizes a decoder-encoder architecture and a flow property-based learning objective function, to estimate the origin-destination (OD) flows of public transport systems from boarding-alighting data; and (2) a temporal-spatial analysis method to analyze OD flow change before and during COVID-19 to unveil its impact on mobility across time and space. The validation of the UML method showed that it achieved a coefficient of determination of 0.836 when estimating OD flows using boarding-alighting data. Upon the successful validation, the proposed methodology was implemented to analyze the impact of COVID-19 on the mobility of the New York City subway system. The implementation results indicate that (1) the rise in the number of weekly new COVID-19 cases intensified the impact on the public transport mobility, but not as strongly as public health interventions; and (2) the inflows to and outflows from the center of the city were more sensitive to the impact of COVID-19.