Towards transferable activity-based models – an activity scheduling case study evaluating transfer learning of human behaviours

Abstract

Activity-based transport models are expensive. New models or scenarios require expensive data collection, re-specification, re-calibration and estimation. Adopting models in new locations or adapting them for novel scenarios is challenging. We collect and consolidate 3.6 million worldwide household travel surveys. We train a deep generative scheduling model on these schedules. We show that this model can transfer learning about human scheduling behaviours temporally and, to a very limited extent, spatially. Potentially enabling predictions in novel settings without new data collection, re-specification, re-calibration, or re-training. We share experiments demonstrating (i) transfer learning temporally and spatially, allowing models to perform better given additional data from other years and, to a limited extent, other places, and (ii) the ability to make out-of-context predictions in novel scenarios, such as for missing data or a new location. Our work demonstrates a promising route towards foundational behavioural models that can make high-quality predictions in novel scenarios. For certain situations, such as where standard models are infeasible due to cost, time or data constraints, such models would be valuable.