Theory-driven or data-driven? Modelling ride-sourcing mode choices using integrated choice and latent variable model and multi-task learning deep neural networks

Abstract

Ride-sourcing services have had a disruptive impact on urban mobility. However, the perceived risk of contracting the COVID-19 virus while using these services has negatively affected people's willingness to travel by this mode. Therefore, it is essential to understand the factors influencing ride-sourcing usage during and after the pandemic. This study utilized data collected through stated preference experiments to model mode choice decisions during and after the pandemic. The study applied both theory-driven integrated choice and latent variable (ICLV) models and data-driven multi-task learning (MTL) deep neural network framework. The study found that the MTL models achieved the highest prediction accuracies. Additionally, econometric information was derived from both ICLV and MTL models. The marginal effects of level-of-service (LOS) variables were largely agreed between the ICLV and MTL models. However, only the latent variables from the ICLV models presented meaningful behavioural interpretations. The study found that individuals who believed there was greater risk associated with ride-sourcing during the pandemic were less likely to use these services. The ICLV model interpretations also indicate that the perceived safety of using ride-sourcing services is higher during the post-pandemic period compared to during the pandemic period. This finding provides reassurance regarding the recovery and growth of ride-sourcing usage in the post-pandemic era.