Identifying Choice Sets for Public Transport Route Choice Models using Smart-Card data: Generated vs. Empirical Sets

Public transport route choice models are fundamental components in many transport applications, as they model and predict individuals’ route choice behavior and therefore help in assessing and improving public transport network design and performance. A route choice model consists of two main components: 1) choice set generation, which tries to enumerate all possible alternatives between origin and destination pairs; and 2) choice modeling of the chosen alternative from the generated choice set. This study focuses on the first component, as choice sets play a crucial role in understanding travel decision-making behavior. Several studies have shown that the composition and size of choice sets have a significant impact on both choice model estimation and demand prediction (Bovy, 2009; Swait & Ben-akiva, 1987). Inaccurate choice sets could result in misspecification of choice models and introduce biases to forecasted demand (Bovy, 2009; Ortuzar & Willumsen, 2001). Studies on public transport route choice modelling have mainly relied on conventional choice set generation approaches used in road network applications, with some modifications to account for the differences between road and public transport networks (Tan, 2016). Namely, conventional approaches/algorithms such as k-shortest path (van der Zijpp & Catalano, 2005), multiobjective path, simulation (Bekhor et al., 2006), branch and bound (Prato & Bekhor, 2006), labeling (Ben-Akiva et al., 1984), link elimination (Azevedo et al., 1993), and/or doubly stochastic (Nielsen, 2000) have been applied in several studies on public transport route choice models to generate exhaustive choice sets (e.g., Abdelghany & Mahmassani, 1999; Anderson et al., 2017; Benjamins et al., 2001; Friedrich et al., 2001; Tan et al., 2015). However, choice set generation is a complex and challenging task in dense urban public transport networks, due to the large number of possible route/path alternatives in such networks. Enumerating all possible alternatives becomes challenging and impractical. In addition, it is unlikely that all enumerated alternatives are in reality considered by passengers (Gentile & Noekel, 2016). A choice set generation approach should ensure high coverage by generating enough paths to cover passengers’ choices but must also ensure high precision by including only paths that are relevant (Marra & Corman, 2020). In addition, the choice set size and quality of the generated alternatives may significantly affect parameter estimates (Frejinger et al., 2009; Zimmermann & Frejinger, 2020). However, defining relevant paths is not an objective task and cannot be easily cross-checked against actual passengers’ behavior, which complicates the evaluation of choice set quality. More recently, researchers have relied on observed Smart Card (SC) data, collected by Automated Fare Collection (AFC) systems, to generate choice sets (e.g., Arriagada et al., 2022; Lee & Sohn, 2015; Zhang et al., 2018). The high implementation rate of AFC systems in many countries enables them to cover nearly the entire population of travelers, resulting in substantial volumes of travel data over long periods of time (Bagchi & White, 2005). By using SC data, the observed routes are assumed to form the choice set of the corresponding origin-destination pairs. It is assumed that considering SC data over long periods of time should cover all relevant paths that are considered by passengers. This study will generate choice sets for a large multimodal public transport network using both conventional approaches and smart card data. It will evaluate and compare the two generated choice sets based on computational performance, coverage tests, and composition tests (size of choice set, diversity of alternatives, variations of path attributes etc.). In addition, route choice models will be developed using the conventional and observed choice sets and compared on the basis of statistical goodness-of-fit measures, interpretation of parameter estimates, and out-of-sample generalization performance.