Explorations of a Neural Network Model for Binary Choices: An Application to Commuters’ Auto Mode Choices

Abstract

Discrete choice models (DCMs), such as the logit model, have long been a cornerstone in travel behavior studies. However, the linear combination of constant coefficients and explanatory variables of traditional DCMs’ utility limits their ability to capture nonlinear relationships. The mixed logit model and mixed probit model introduce random coefficients to address limitations by capturing a certain degree of nonlinearity. In contrast, neural networks (NNs) autonomously learn nonlinear relationships without prior assumptions, often outperforming DCMs in fitting travel survey data and making predictions. This paper compares the random coefficients model and the NN model in describing choice behavior and explores the variable significance test of NNs in discrete choice contexts. Because of the complexity of DCMs, this paper begins with a binary choice context. Simulations and Monte Carlo experiments on synthetic datasets compare the fitting results of NNs and the random coefficients model, and verify the effectiveness of the variable significance test of NNs in DCMs. The significance test is then applied to London Passenger Mode Choice dataset. Findings suggest that the binary choice model based on NNs is capable of approximating the mixed binary logit (MBNL) model, capturing the average impact of a variable on choice probability. However, it does not offer the same level of detail in depicting the distribution of a variable’s effects as the MBNL model. In addition, the results of the mixed binary probit model proves the generalizability of the NN’s approximation ability. Significance test of NNs in discrete choice demonstrates their effectiveness, while SHapley Additive exPlanations analysis is also conducted for comparison.