Weak teachers: Assisted specification of discrete choice models using ensemble learning

Abstract

Mode choice modelling has almost exclusively been tackled using Discrete Choice Models (DCMs). This is in part due to their highly interpretable linear structure, which allows the model to be checked for consistency against established behavioural expectations. However, a key drawback of DCMs is that the utility functions must be specified manually in advance of fitting the model, a process that does not scale well with increasing data complexity. Machine Learning (ML) is increasingly being investigated as an alternative to DCM for modelling mode choice. Whilst ML automates the decision-making process, requiring no utility functions to be specified, it has a crucial limitation in that the resulting models are difficult to interpret and to check for behavioural consistency. In order to address the limitations of both ML and discrete choice models, we propose an assisted specification procedure, in which the aggregate structure of a fitted Ensemble Learning (EL) model is used to inform the utility functions in a DCM. The resulting models are found to have greatly improved performance over manually specified DCMs, outperforming all but the highest performing ML classifier.