Latent Class Learning Model Using Probe Person Data: Formulation and Application for Departure Time



Latent Class Learning Model Using Probe Person Data: Formulation and Application for Departure Time

Authors

SAITO Itsumi, Nippon Telegraph And Telephone Corporation, HATO Eiji, Department Of Civil Engeneering,the University Of Tokyo

Description

In urban area, car sharing systems and bike sharing services attract
attention in recent years. To evaluate and control appropriately such a system, it is necessary to predict variability of daily travel-activity
pattern in good accuracy among individual levels.

Specifically, in sharing systems, it is important to stochastic control of the deviation of time-space demand in transportation networks, so we
focus on the intra-inter personal variability of daily departure time and route choice . In departure time choice, arrival time preference is a basic element, however these information can’t be dilectly-observed as position
or departure time data collected through probe person systems using cell phones. So we aim to estimate indirectly the arrival time preference as a latent variable and reveal day-to-day learning mechanism by using repeated observations of the same individual departure time
choice data.
To identify intra-inter personal variability of departure time and
influence which a previous experience has on the following action
(day-to-day learning process), we analyze daily commuting trips
collected by repeated travel activity survey based on GPS-equipped cell
phone. The travel-activity data collected over a period about a month to
respondents who live in Yokohama and Matsuyama city.
Estimation results indicate that the choice set structure based on
observed intra-inter personal variability and stochastic distribution
of preference for arrival time and value of time differs for individuals
characteristics and OD patterns. And also, there are some different
types of segments about the reaction to a previous experience. Travel
time variability and their own experiences by transportation mode
influence estimation results.
Our findings suggest benefits of more exact evaluation or of mobility sharing systems using massive
demand dataset and will offer implications of stochastic control methods improving efficiency of mobility sharing systems such as “person-based” dynamic incentive systems.

Abstract

1. Objective
In urban area, car sharing systems and bike sharing services attract
attention in recent years. To evaluate and control appropriately such a
system, it is necessary to predict variability of daily travel-activity
pattern in good accuracy among individual levels.

Specifically, in sharing systems, it is important to stochastic control
of the deviation of time-space demand in transportation networks, so we focus on the intra-inter personal variability of daily departure time and route choice . In departure time choice, arrival time preference is a basic element, however these information can’t be dilectly-observed as position or departure time data collected through probe person systems
using cell phones. So we aim to estimate indirectly the arrival time
preference as a latent variable and reveal day-to-day learning mechanism by using repeated observations of the same individual departure time choice data.

2. Data/Methodology
To identify intra-inter personal variability of departure time and
influence which a previous experience has on the following action
(day-to-day learning process), we analyze daily commuting trips
collected by repeated travel activity survey based on GPS-equipped cell
phone. The travel-activity data collected over a period about a month to
respondents who live in Yokohama and Matsuyama city.

We aim to compare three types of departure time choice models, which are
latent class model, mixed logit model, and dynamic latent class model. To
employ the latent class model and mixed logit model, we can evaluate the
validity of treating time preferences as latent variables and also we
can evaluate the stochastic variation of preferred arrival time. Further,
dynamic model can reveal the day-to-day learning mechanism of departure
time choices. We estimate the model using maximum likelihood
estimation method based on the EM algorithm.

3. Results/Findings
Estimation results indicate that the choice set structure based on
observed intra-inter personal variability and stochastic distribution
of preference for arrival time and value of time differs for individuals
characteristics and OD patterns. And also, there are some different
types of segments about the reaction to a previous experience. Travel
time variability and their own experiences by transportation mode
influence estimation results.

4. Implications for Research/Policy
We analyze the actual day-to-day departure time choice mechanism by
individual levels. Using the repeated observation of same individual
collected through probe person systems, we can know more detailed
day-to-day real leaning mechanisms. Our findings suggest benefits of
more exact evaluation or of mobility sharing systems using massive
demand dataset and will offer implications of stochastic control methods
improving efficiency of mobility sharing systems such as “person-based”
dynamic incentive systems.

Publisher

Association for European Transport