Forecasting urban travel demand with geo-AI: a combination of GIS and machine learning techniques utilizing uber data in New York City

Abstract

Presently, online ride-hailing firms and municipal transportation systems substantially influence urban traffic management. Considering the increasing use of these services, it is essential to assess and predict travel demand to optimize service efficiency. This study analyzes Uber data from New York City, employing automated machine learning algorithms to assess and forecast ride demand inside the city. In conventional transportation planning, forecasting travelers’ destination selections is a crucial stage. Traditional methodologies, such as physical models like gravity models, are constrained in their capacity to encompass the comprehensive array of elements affecting travel behavior, frequently addressing just two or three variables. This study employs machine learning methodologies for predicting passenger destination choices, illustrating that these algorithms can include a wider array of variables, hence providing enhanced forecast accuracy relative to conventional approaches. The utilization of ArcGIS Pro and Python modules for automation enhanced the efficiency of spatial analysis. A range of machine learning methodologies, such as decision trees, Light-GBM, XG-Boost, Cat-Boost, and hybrid models, were utilized to predict demand. The selected source demand forecasting model is an ensemble model, with a R² of 0.94 and a Mean Absolute Error of 91.63. The optimal model for predicting destination demand was identified as a composite model comprising eight distinct algorithms. The model’s R² is 0.95, while the Mean Absolute Error is 67.12. Moreover, the examination of environmental factors indicated that proximity to recreational activities, median age, and population density had the most substantial influence on predicting travel demand.