The Holy Trinity: Blending Statistics, Machine Learning, and Discrete Choice with Applications to Strategic Bicycle Planning

Lecture: Transportation Engineering: Civil and Environmental Engineering | May 12 | 4-5 p.m. | 290 Hearst Memorial Mining Building

 Timothy Brathwaite, UC Berkeley

 Institute of Transportation Studies

Abstract:
Across all levels of government in the United States (U.S.), transportation and planning agencies have prioritized encouraging bicycle use. However, despite such admirable goals, actually increasing bicycle usage has been a struggle. For instance, the City of San Francisco set a goal in 2010 to increase its 3.5% bicycle mode share to 20% by 2020 (SFMTA, 2012). Unfortunately, given the 2014 bicycle mode share of 4.0% (U.S. Census Bureau, 2015), San Francisco appears unlikely to meet its mode share goals. Similarly, in 1999, Oakland set a goal of increasing it’s 1990 bicycle commute mode share of 1.1% to 4% in 2010 (City of Oakland, 2007). In 2014, Oakland’s bicycle commute share was only 3.1% (U.S. Census Bureau, 2015). While sad, this pattern is common. Many agencies are interested yet unsuccessful in raising their bicycle commute mode shares.

To successfully make planning and investment decisions regarding bicycle infrastructure projects, agencies must accurately judge how much each possible project is expected to increase bicycle ridership. To support this activity, my research aims to improve bicycle demand models. In this talk, I will focus on three flaws of current mode choice models: (1) the exclusion of roadway-level variables (e.g. on-street bicycle infrastructure measures, traffic speeds, etc.), (2) the assumption of “perfectly rational” decision makers, and (3) the issue of class imbalance (i.e. the relatively small numbers of cyclists in household travel surveys). In addressing these issues, I merge traditional discrete choice with recent advances in statistics and machine learning, making use of methods such as parametric link functions, Bayesian decision trees, and Gaussian Process models. In all cases, these methods are modified and theoretically extended for use in a transportation context. Together, the developed techniques increase the policy relevance and accuracy of bicycle demand models in particular, and they advance the field of choice modeling in general.

Bio:
Timothy Brathwaite is a Ph.D. candidate in transportation engineering in the Civil and Environmental Engineering department from the University of California (UC) at Berkeley, working under the supervision of Professor Joan Walker. Motivated by efforts to predict the demand for bicycling under various policy scenarios, Timothy’s research makes methodological improvements to discrete choice models to account for omitted roadway variables, traveler “irrationality,” and the typically low number of cyclists in household travel surveys. He was the UCCONNECT Outstanding Graduate Student of 2017 and a UC Berkeley 2016 Outstanding Graduate Student Instructor. Previously, he received his Master of City Planning and Master of Science in Civil Engineering from UC Berkeley and his Bachelor of Science in Urban Studies and Planning from the University of New Orleans. Professionally, Timothy has worked on the data science team at Lyft, with transportation consulting firms (Fehr and Peers and Cambridge Systematics), with the bicycle facilities program at the City of Oakland, and with the non-profit "Bike Easy" in New Orleans.

 jmarie@berkeley.edu