Characterizing the Importance of Criminal Factors Affecting Bus Ridership using Random Forest Ensemble Algorithm

Abstract

Public transit systems provide mass movement with substantial traffic operational and environmental benefits. Despite these benefits, they still represent a small market share in the United States. A comprehensive understanding of the determinants of transit ridership is essential for investment allocation to improve safety, mobility, and air quality in an urban area. Except socio-economic factors, crime has been identified as a determinant for the ridership. Most studies found that ridership and crime are linearly correlated to each other, whereas other studies believed the level of crime can result in a nonlinear effect on the ridership. The relationship between ridership and crime remains inconclusive. Besides, the simultaneous relationship between ridership and crime is scarcely addressed and most ridership studies only include one or a few external factors that affect crime opportunity. This paper proposes a random-forest-based feature selection method to characterize the importance of multiple variables, to bus ridership and total crime, respectively, at different levels. A case study in Houston, Texas, USA, for the year 2017 is provided to illustrate the feature selection and modeling process. A total of 110,885 crimes, ridership on 9004 bus stops, and related socio-economic information were collected. Results indicated that a medium or lower level of ridership is positively correlated to crime; the linear relationship can be broken down at a high level; and reducing the total crime per capita can promote bus ridership. Random-forest-based models were developed with the selected determinants, performing with high accuracy in ridership per capita estimates.