Semi-supervised route choice modeling with sparse Automatic vehicle identification data

Abstract

Massive and passive Automatic Vehicle Identification (AVI) data provides samples of whereabouts and movements of vehicles, which is a potential source of information for route choice behavior modeling. However, the AVI observations are too sparse to infer the specific chosen route and OD pair, which discourages its application on route choice model estimation. To tackle this issue, this paper develops a semi-supervised learning method that can train the route choice model with sparse AVI observations. First of all, the likelihood function in Maximum Likelihood Estimation procedure was derived by decomposing the AVI trace into observation pairs. Combined with high-resolution GPS observations, the measurement equation and OD inference model were then defined to deal with the sparsity problem of AVI observations. At the same time, the Mixed Logit model was introduced to capture the correlation and heterogeneity across the choice behavior between different observation pairs. Finally, the relationship between route choice model and the likelihood function was established and the unknown parameters in route choice model can be estimated by seeking a maximum to the log-likelihood function. Empirical studies were conducted with field-testing data in this paper. The estimated results show that the proposed semi-supervised method improved the identification accuracy of route choice model significantly without sacrificing interpretability. The evaluation of the computational efficiency presented the potential of the semi-supervised method to learn route choice behavior for a large-size sample set. The sensitivity analysis was also performed to illustrate how robust the proposed method is. This is the first research that attempts to apply AVI data on route choice model and it endows the high-penetration AVI data with great practical value for modeling the route choice behavior of city-wide samples over a long period.