An Interpretable Machine Learning Framework to Understand Bikeshare Demand before and during the COVID-19 Pandemic in New York City

Aug 4, 2026

Bikesharing demand is modeled using an interpretable machine learning framework built for a large-scale system in New York City across two periods: pre-pandemic (March 2019 to February 2020) and during-pandemic (March 2020 to February 2021). Two Extreme Gradient Boosting models estimate hourly demand from 39.9 million Citi Bike transactions, enriched with weather data and holiday adjustments. SHAP-based interpretation quantifies each variable’s positive or negative contribution, highlighting female user share and time of day, while month effects shift under pandemic conditions.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages25

Reading Time1 h 3 min

AccessFree to download

Introduction

Modeling framework overview

Literature review

Themes considered in prior work

An Interpretable Machine Learning Framework to Understand Bikeshare Demand before and during the COVID-19 Pandemic in New York City Majbah Uddin, PhD, Corresponding author National Transportation Research Center Oak Ridge National Laboratory 1 Bethel Valley Road Oak Ridge, TN 37830 ORCiD: 0000-0001-9925-3881 [Email: ](Email: uddinm@ornl.gov)[uddinm@ornl.gov](Email: uddinm@ornl.gov) [Ho-Ling Hwang](Ho-Ling Hwang), PhD National Transportation Research Center Oak Ridge National Laboratory 1 Bethel Valley Road Oak Ridge, TN 37830 ORCiD: 0000-0001-7651-6263 Email: [hwanghlc@gmail.com](hwanghlc@gmail.com) Md Sami Hasnine, PhD Assistant Professor Department of Civil and Environmental Engineering Howard University 2300 Sixth Street, NW \#1026 Washington, DC, USA, 20059 ORCiD: 0000-0003-4110-5047 [Email: ](Email: mdsami.hasnine@howard.edu)[mdsami.hasnine@howard.edu](Email: mdsami.hasnine@howard.edu) An Interpretable Machine Learning Framework to Understand Bikeshare Demand before and during the COVID-19 Pandemic in New York City In recent years, bikesharing systems have become increasingly popular as affordable and sustainable micromobility solutions. Advanced mathematical models such as machine learning are required to generate good forecasts for bikeshare demand. To this end, this study proposes a machine learning modeling framework to estimate hourly demand in a large-scale bikesharing system. Two Extreme Gradient Boosting models were developed: one using data from before the COVID-19 pandemic (March 2019 to February 2020) and the other using data from during the pandemic (March 2020 to February 2021) . Furthermore, a model interpretation framework based on SHapley Additive exPlanations was implemented. Based on the relative importance of the explanatory variables considered in this study, share of female users and hour of day were the two most important explanatory variables in both models. However, the month variable had a higher importance in the pandemic model than in the pre-pandemic model. Keywords: bikeshare; machine learning; SHAP; New York City; Citi Bike Introduction Bikesharing systems have become increasingly popular as affordable and sustainable micromobility solutions in recent years. In the United States, large cities such as New York City, San Francisco (SF), Washington DC, Chicago, and Boston experienced their highest bikeshare demands in 2020 and 2021. For example, according to the US Department of Transportation website accessed on July 5, 2022, at least 16.7 million trips were made at selected docked bikeshare systems in New York City from January to August 2021, 43% more than in January to August 2020. Although bikesharing ridership was already increasing exponentially before the COVID-19 pandemic began, the pandemic is largely responsible for the skyrocketing bikesharing demand. Considering only pandemic period, bikesharing demand decreased early in the pandemic (2020), but it increased later, particularly in 2021 (Wang and Noland 2021a) . Many previous studies have explored the factors that affect bikesharing demand, particularly sociodemographic factors, weather, and the built environment (Babagoli et al. 2019; El-Assi et al. 2017; Heaney et al. 2019; Reilly et al. 2020a; Wang and Noland 2021a,b; Xu and Chow 2020). A few recent studies have examined bikeshare trip characteristics during the pandemic (Basu and Ferreira 2021; Jobe and Griffin 2021; Padmanabhan et al. 2021) . Another study examined equity concerns related to bikeshare demand (Ursaki and Aultman-Hall 2015) . Because the majority of the described studies relied on either descriptive statistics or simple aggregate modeling techniques (e.g., linear regression), they are difficult for policymakers and planners to use for forecasting and policymaking. Understanding the factors that affect bikeshare demand is extremely important for urban planning and policymaking. Advanced mathematical models such as machine lea

FAQs

What data and time periods are used to estimate bikeshare demand?

The study uses 39.9 million Citi Bike transactions aggregated by hour in New York City, covering March 2019 to February 2021, split into pre-pandemic (March 2019–February 2020) and during-pandemic (March 2020–February 2021).

How does the framework make demand forecasts?

Two Extreme Gradient Boosting (XGBoost) models are trained separately for pre-pandemic and during-pandemic periods. The aggregated demand data are augmented with weather variables and adjusted for holidays.

How is the “black box” interpreted?

A SHAP (SHapley Additive exPlanations) interpretation framework is applied to measure how much each explanatory variable contributes to predicted bikeshare demand, including whether contributions are positive or negative.

Aug 4, 2026

Upload by Emma Wilson