Flexible working patterns are developing fast since the Covid-19 pandemic. However, preceding the pandemic, a growing diversity of commuting behaviours existed among the population. Some of these behaviours may have served as trendsetters of commuting patterns that align with the emerging flexible working patterns, albeit not completely identical to those observed today. The challenge for travel behaviour modellers is that in the existing travel surveys, such heterogeneity is hard to estimate due to their low presence in the data samples.

This dissertation develops an Early Stopping Bayesian Data Assimilation (ESBDA) estimator which enables an established behaviour model to be well adapted to a new context characterised by significantly lower data samples. We carry out the development using the Mixed-Logit as the main behaviour model, which is a remarkably versatile tool in revealing insights into inter and intra-individual behavioural heterogeneity. ESBDA makes it possible for existing models to learn from small samples of travellers who show distinct travel choices, establishing predictive models while existing methods require more data.

The prevalent Mixed-Logit estimators—Maximum Simulated Likelihood (MSL), Hierarchical Bayes (HB), and Variational Bayes (VB)—demonstrate susceptibility to overfitting and underfitting in small datasets. ESBDA integrates Bayesian inference and Machine Learning techniques to address these issues. Our method is evaluated against MSL, HB, VB, and Bayesian Data Assimilation (BDA) without early-stopping in 15 experiments involving two Mixed-Logit models and varying sample sizes. MSL, HB and VB are found problematic in six experiments each. BDA malfunctions in five. ESBDA stumbles in just one, showcasing its efficacy with smaller samples.

Overall, ESBDA consistently outperforms existing methods in estimate quality, convergence speed, out-of-sample predictability, and behavioural insights. Particularly, its advantages are magnified with diminishing sample sizes. ESBDA is thus proven to be a more practical, economical and relatively time-saving tool for analysing travel behaviour where there are low samples of distinct, heterogeneous travel behaviour.

ESBDA is then applied to commuter samples from the London Travel Demand Survey to examine the effects of five commuter attributes that are cogent to the evolution of flexible working patterns: (1) workplace type (fixed/variable), (2) National Statistics Socio-Economic Classification (NS-SEC), (3) timing of the commute, (4) Origin-Destination (OD) and (5) age of the commuters. These are some of the key dimensions that the data can yield in identifying the extent of flexible working. A systematic, tree-like model transfer framework is established to help establish behavioural models for 6 levels of increasingly narrow segmentation of commuter sub-groups.

The findings from the ESBDA model have uncovered insights that existing travel choice estimations have not been able to reveal. Even a rather coarse segmentation of the commuters has identified starkly different travel preferences, such as distinctly different levels of willingness to pay for time savings between those who commute to fixed or variable workplaces, given the same NS-SeC, timing of the trip, OD of the commute and age of commuters. This suggests that significant impacts may yet be in store for the provision of transport infrastructure and services if / when flexible working scales up. ESBDA has also opened up a new way to study small samples of distinctly heterogeneous travel behaviour in large metropolitan scale surveys – and as the emerging behaviours grow, this approach can be used to investigate new sub-variants which are yet to emerge.