International Choice Modelling Conference, International Choice Modelling Conference 2017

Font Size: 
Exploring travelers’ route choice behavior from GPS trajectories: a path size logit model with sampling of alternatives
Weiliang Zeng, Tomio Miwa, Mutsumi Tashiro, Takayuki Morikawa

Last modified: 28 March 2017


Travelers’ route choice behavior has become one of the crucial issues in urban transportation system, given the fact that the network-wide traffic state may be influenced by the route choice behavior. There are often multiple routes between an origin-destination (OD) pair. Travelers may choose different routes based on the individual preferences. Many factors may influence travelers’ route choice, such as distance, average travel time, travel time reliability, comfortableness, safety, fuel consumption, etc. We usually assume that a decision-making of route choice is a reflection of potential preferences for each available route and the traveler chooses the route with the highest utility.

Most of the current studies analyzed the route choice behavior based on Stated Preference (SP) questionnaire survey data. The individual characteristics and route choice preferences in hypothetical situations can be collected by SP. However, respondents have to assume choice set instead of experiencing the route choice practically in SP survey. Actual route choice behavior on real-world transportation network can not be adequately investigated. Respondents may simply answer questions that they would not realistically pursue. Also, inherent limitations of questionnaire survey related to honest, accurate and bias-free reporting are difficult to avoid. In recent years, advancements in traffic information collection technologies, including trip records from in-vehicle GPS device, can facilitate the investigation of influences dominates route choice decisions. In this study, rather than questionnaire dataset, we aim to explore travelers’ route choice behavior from large-scale GPS trip records.

There are still two main challenges introduced by using general GPS trip records. First, there are usually a large number of alternatives between an OD pair, resulting in an intractable problem of estimation of a discrete choice model with a full choice set. How to sample alternatives from the full choice set and the size of the sampled choice set needs to be tackled. Second, different from data collected from hypothesis-oriented questionnaires, how to indicate the impact of factors on route choice behavior in different situations is crucial. To overcome these difficulties, we propose to apply the multinomial logit (MNL) model with sampling of alternatives to explore travelers’ route choice preference from real-world GPS trajectories.

First, we generate the route choice set by using a stochastic path generation algorithm in an experienced transportation network. We assume that local commuters are good at choosing a path due to their experiences accumulated over years. When facing multiple route choices, most of the people who are unfamiliar with the local traffic condition may simply select the shortest path while the experienced commuters may utilize their driving experience to choose the best one based on their preferences. Naturally, the link usage frequency reflects the driving experience which can be expressed by the degree of familiarity. We propose to penalize the link travel times by the link usage frequency. Then, the experienced transportation network can be constructed by using the penalized link travel time. A biased random walk algorithm is applied to generate a set of shortest paths with selection probabilities on the experienced transportation network with penalized link travel times.

Second, we estimate the MNL model with sampling of alternatives using classical conditional maximum likelihood estimation (MLE). The conditional probability that an individual n will choose alternative i conditional on the sampling subset  for the individual can be derived using the Bayes theorem.Since the traditional MNL model is restricted by the independence from irrelevant alternative (IIA) property, which does not hold the correlation problem of overlapping routes, we use the modified MNL model, i.e., the path size logit (PSL) model in which an additional term is introduced to capture the correlation of routes, to overcome the overlapping problem.

We take the GPS trajectories recorded in Toyota city, Japan as the experiment data. The data is collected from 153 private cars in 1 month, from March 1st to March 31st, 2011. After map-matching and some basic data cleaning work, 7245 trip records with 5667 OD pairs are extracted. The route characteristics include path length, average path travel time, path travel time variability, fuel consumption, and average speed. A correlation matrix is computed to detect potential collinearity between all pairs of these explanatory variables included in model estimation. We only choose path length, average speed, and path travel time variability as the explanatory variables because path length, average path travel time, and fuel consumption are highly correlated. To investigate the impact of route characteristics on utility function, we normalize the path length, average speed, and path travel time variability so that a larger estimated coefficient indicates the corresponding variable has a greater impact on travelers’ utility. To investigate the impact of individual heterogeneity and journey attributes on route choice preference, we group the OD pairs by gender, age, departure time, and OD Euclidean distance, and then estimate the route choice models respectively. We compare the choice probabilities of four representative paths, i.e., shortest distance, maximum average speed, and least path travel time variability. Finally, the effect of individual heterogeneity and journey attributes on route choice preference can be demonstrated by the difference of choice probability.



Conference registration is required in order to view papers.