Inference of Pattern Variation of Taxi Ridership Using Deep Learning Methods: A Case Study of New York City

摘要

Taxis constitute an important component of the public transportation infrastructure in large metropolitan areas. However, when seen within a supply and demand framework the operation of taxi transportation system is far away from its optimal equilibrium, yielding a missed cost of opportunity for customers, drivers, and city planners. The key for optimizing its market lies in forecasting taxi demand with high geospatial-temporal precision. In this paper taxi pickup pattern is predicted by utilizing a deep learning approach that leverages long short-term memory (LSTM) neural networks. This study is based on publicly available taxi data for the New York City. Pickup data is binned based on geospatial and temporal informational tags, which are then clustered using principal component analysis. The spatiotemporal distribution of the taxi pickup demand is studied within short-term periods (next one hour) as well as long-term periods (next 48 hours) within each cluster. The performance and robustness of the proposed technique is evaluated through a comparison with adaptive boosting and decision tree regression models fitted to the same dataset. Numerical results show the dominance of the LSTM model on the short-term horizon and relatively smaller errors for the long-term prediction.