Interpretable data-driven demand modelling for on-demand transit services

Abstract

In recent years, with the advancements in information and communication technology, different emerging on-demand shared mobility services have been introduced as innovative solutions in the low-density areas, including on-demand transit (ODT), mobility on-demand (MOD) transit, and crowdsourced mobility services. However, due to their infancy, there is a strong need to understand and model the demand for these services. In this study, we developed trip production and distribution models for ODT services at Dissemination areas (DA) level using four machine learning algorithms: Random Forest (RF), Bagging, Artificial Neural Network (ANN) and Deep Neural Network (DNN). The data used in the modelling process were acquired from Belleville’s ODT operational data and 2016 census data. Bayesian optimalization approach was used to find the optimal architecture of the adopted algorithms. Moreover, post-hoc model was employed to interpret the predictions and examine the importance of the explanatory variables. The results showed that the land-use type was the most important variable in the trip production model. On the other hand, the demographic characteristics of the trip destination were the most important variables in the trip distribution model. Moreover, the results revealed that higher trip distribution levels are expected between dissemination areas with commercial/industrial land-use type and dissemination areas with high-density residential land-use. Our findings suggest that the performance of ODT services can be further enhanced by (a) locating idle vehicles in the neighbourhoods with commercial/industrial land-use and (b) using the spatio-temporal demand models obtained in this work to continuously update the operating fleet size.