Xie, S., Hillel, T., and Jin, Y. (2020)

An Early Stopping Bayesian Data Assimilation Approach for improved Mixed Multinomial Logit transferability

9th Symposium of the European Association for Research in Transportation, Lyon, France

Mixed Multinomial Logit (MMNL) models can provide valuable insights into inter and intra-individual heterogeneity in transportation choice modelling. However, the high computational and data requirements for MMNL models has limited the application of MMNL models in practice. These requirements are particularly problematic when investigating the behaviour of specific population sub-groups or market segments, where a modeller may want to estimate separate models for a number of similar contexts, each with low data availability. The same challenges arise when adapting one model to a new location or time period. To overcome these barriers, we establish a new Early Stopping Bayesian Data Assimilation (ESBDA) approach which updates a previously estimated MMNL on a new data sample or subsample through iterative Bayesian inference. This approach therefore enables an existing model from one context to be transferred to a new context with lower data availability. The ESBDA approach is benchmarked against two reference estimators: (i) a standard Bayesian estimator (MMNL); and (ii) a Bayesian Data Assimilation (BDA) estimator without early stopping. The results show that the proposed ESBDA approach can effectively overcome over-fitting and non-convergence. ESBDA models outperform the models estimated by the reference estimators in terms of behavioural consistency of parameter estimates and the out-of-sample predictive performance of the model. Even when using few collected data, ESBDA can still produce suitable and stable MMNL model with parameter estimates consistent with established behavioural theory.