Spatial Transferability of Neural Network Models in Travel Demand Modeling

Abstract

Neural network (NN) models have been widely used in travel demand modeling in recently years. However, there are few studies about the spatial transferability of NN models. In this paper, the spatial transferability of NN models in travel demand modeling, especially in mode choice models, is analyzed. This paper first discusses the performance of naïve transfer when no data are available in an application context. Then, a NN model adaptation method is proposed using the classification adjustment weight vector when limited local data are available. Using the 27/28 Transportation Planning Board—Baltimore Metropolitan Council Household Travel Survey data, five NN models are built using trips within five areas in the Washington, DC, and Baltimore regions. Each of the five NN models is applied to the other four areas to study spatial transferability using both individual-level and aggregate-level performance measures. The result shows that the naïve transfer of NN models can perform very well between areas that share many similarities. It also indicates the transferability of NN models is not symmetric. The performance of the proposed adaptation method is evaluated for different sample sizes of local training data. For transfer between areas that have significant differences, the proposed NN model adaptation method can improve performance significantly, even with a small sample size, compared to naïve transfer.