Abstract
Effectively capturing individual- and collective-level human mobility is paramount to solving many modern societal challenges. In the last few years, thanks to the availability of big mobility datasets and the constant improvements in computational infrastructures, the number of models to predict and generate human movements has significantly increased. However, the development of these models has consistently been centered on enhancing performances in accordance with particular evaluation criteria, pushing more in-depth analyses of their real behavior in specific scenarios to the background. With this Thesis, we want to shed light on the ability of mobility models to (i) generalize well in new scenarios and (ii) resist distribution shifts at the test phase. Models with such characteristics are more crucial than ever as the number of episodes in which people change their everyday behaviors (e.g., during a pandemic, when a natural disaster occurs) is increasingly frequent. Each Chapter of this manuscript consists of an article carried out during the Ph.D. Chapters 3 and 4 focus on collective-level models, while Chapter 5 deals with individual-level models. In each Chapter, we propose novel evaluation frameworks designed to measure the performances under such circumstances and new models and methodologies more resilient to distribution shifts and better at generalizing over unseen scenarios.