Spatial transferability of machine learning based models for ride-hailing demand prediction

Abstract

Accurate prediction of ride-hailing demand is crucial to provide quality service to consumers, to effectively schedule vehicles, and to maintain a well-functioning transportation system. As information of ride-hailing demand in most of the cities is not available, assessing the spatial transferability of ride-hailing demand models is an important research problem. To address this problem, this study aims to develop a ride-hailing demand prediction model using trip information available from ride-hailing service providers and to test the spatial transferability of the model. Using aggregated trip data, we have developed ride-hailing generation and attraction prediction models using several well-known machine learning algorithms such as random forest, extreme gradient boost, support vector machine, and artificial neural network for two study areas including the New York City and Chicago with similar built environment and land use characteristics. The random forest and extreme gradient boost models have superior performance for predicting ride-hailing demand with both the training and testing data in the intra-city level. The developed models for the New York City are later used to predict the ride-hailing demand of Chicago using two different transfer learning approaches. A knowledge transfer approach shows better transferability potential of ride-hailing demand models with reduced error rates. An analysis of prediction errors suggests that the models achieve better accuracy to predict demand on areas near central business districts or during peak periods.