Unlocking the nonlinear TOD-metro ridership relationship: A novel machine learning approach embedding spatiotemporal heterogeneity

Machine learning approaches to unlocking the TOD-metro ridership relationship have attracted great attention due to the strong capability of such approaches to handle the underlying nonlinearity and complexity in this relationship. Considering the peculiarities of spatiotemporal heterogeneity in metro ridership, however, one prominent challenge remains unsettled, namely, the issue that traditional machine learning algorithms are designed to be ‘aspatial’ and thus only produce global estimations. In this paper, a geographical and temporal random forest regression algorithm (GTRFR) is developed, which extends the traditional random forest (RF) as a disaggregation of a number of local submodels and computes an individual random forest regression for each location i at time j using neighboring observations across time and space. It further employs this algorithm to unlock the nonlinear TOD-metro ridership relationship in the case of the Hangzhou metropolitan area. The results show that the GTRFR outperforms the traditional RF in explaining the TOD-metro ridership relationship. Particularly, the nonlinear TOD-metro ridership relationship is unlocked from two major aspects: (1) the relative importance of TOD structural factors across time and space and (2) spatially and temporally varying threshold effects in the effects of the TOD structural factors. The findings portray a much broader picture of the mechanisms underlying the TOD-metro ridership relationship. This paper contributes to the argument that accounting for spatiotemporal heterogeneity should be beneficial to applying machine learning algorithms to transport geography.