Interpretable Multi-class Trees for Travel Choice Mode Analysis

Analysis of travel mode choice is fundamental to forecast travel demand when planning intervention on the supply system. Commonly, this is conducted via Random Utilities Models (RUMs) which relies on the random utility theory. Recently, the large availability of travel demand data, mainly from smartphones, has increasingly led to the use of machine learning models that find their ideal context of use in big data. Although such models are generally capable of good performance, their intrinsic black-box nature is a critical aspect. Hence, to fruitfully apply Machine Learning approaches to travel mode choice, some enhancements must be considered. Recently, interpretable machine learning approach has been proposed. The main idea is to improve the output results of random forests with some additional aids. This paper provides a new framework to approach interpretable machine learning using tree-based methods in combination with classical models. The basic rationale with the main theoretical aspects can be found in old papers. A fresh approach integrating logistic regression and latent budget models will be demonstrated for a multi-class response problem typical in transportation studies.