A new approach to improve destination choice by ranking personal preferences

摘要

It is vital to have the right choice-sets when dealing with many alternatives in discrete choice models, which play a critical role in transport modelling. Various approaches have been proposed to address the issue when forming individual choice-sets. While these methods have been continuously improved, they seem not effectively explain how individuals form their choice-sets when facing a large number of alternatives. To know individual choice-sets, one possible way is to ask all of them about their preferred alternatives directly. However, this is costly and impractical for a large population. This paper proposes a novel behavioural choice-set generation approach by ranking personal preferences of destinations using a matrix factorisation model with Bayesian personalised ranking. From a large travel survey, we form a user-zone-visited frequency matrix for shopping locations. We then use the model to factorise the user-zone-visited frequency matrix into two lower-rank latent matrices. The matrix factorisation model is optimised by using Bayesian personalised ranking. After estimation, the model's outputs, which are user-factor and zone-factor latent matrices, can produce top preferred destinations for individuals. Our experiment from a large travel survey with thousands of alternatives shows that the proposed choice-set generation framework can significantly improve the predictive capability of discrete choice model evaluation with even small choice-set sizes.