A Systematic Evaluation of Deep Learning Methods for Transit Passenger Demand Prediction

Sunday, June 16, 2024

3:00 PM - 4:15 PM ET

Room: Galleria

Primary Author(s)

Zhiwei Chen, Ph.D.

Assistant Professor
Drexel University, United States

Presenter(s)

Zhiwei Chen, Ph.D.

Assistant Professor
Drexel University, United States

Co-Author(s)

TL

Tianzhuo Li, n/a

Graduate Research Assistant
South China University of Technology
Guangzhou, Guangdong, China (People's Republic)
yH

yucong Hu, n/a

Associate Professor
South China University of Technology
Guangzhou, Guangdong, China (People's Republic)

Submitter(s)

ZC

Zhiwei Chen

Drexel University
Philadelphia, Pennsylvania, United States

Abstract: Recent breakthroughs in Artificial Intelligence (AI), particularly in Deep Learning (DL), have encouraged the development of DL models for various transportation tasks. However, the direct impact of DL model’s predictive performance on transportation decision-making outcomes remains largely unexplored. This study addresses this gap by conducting a large-scale experiment to assess how DL model’s performance for transit passenger demand prediction affects operational design outcomes in transit systems. The experiment involves 22 predictors, including DL, traditional ML, and statistical models, and utilizes three datasets from the Guangzhou Subway system. Analysis of 1540 data points generated reveals a crucial relationship between prediction accuracy and decision-making effectiveness. Specifically, at lower prediction accuracy values, there is a considerable fluctuation in decision-making effectiveness. However, as the prediction accuracy surpasses a certain threshold, the uncertainty diminishes, and decision-making effectiveness stabilizes, irrespective of further increases in prediction accuracy. These findings challenge the prevailing assumption that developing larger DL models to improve passenger prediction accuracy would always benefit decision-making, as the marginal gain appears to be small. This study highlights the importance of considering decision-making effectiveness as a criterion when designing and evaluating large DL models. This study also provides a generalizable empirical benchmark for comparing different prediction approaches, assisting researchers and practitioners in selecting suitable methods for transit demand modeling.

Learning Objectives:

Attendees can expect to learn the following from this session:

understand the impact of deep learning models on transit passenger demand prediction
understand how this prediction affects decision making
provide a benchmark for comparing different deep learning models for transit systems