Learning how people move through deep learning and epidemic signatures
Introduction
In the past years, human mobility has received a surge of interest from the research community, mainly in the context of COVID-19. Posed as a dynamic temporal graph modelling problem, the main body of literature on human mobility bases itself on link-based approaches. Most of these models are not able to accurately reproduce the typical dynamics of contacts between individuals, as they fail to learn a governing graph generative process, and, separately, treat the resulting graph as an end in itself rather than as the substrate on which other dynamical processes, such as epidemics, actually unfold. Our research proposes to model dynamic human contacts from first principles, to investigate how the underlying generative process can be learned, and to develop a framework for quantifying the class of processes which generate temporal graphs, including through the dynamics they support.
The Random Walker Induced temporal Graph (RWIG) model is conceptually intuitive: a collection of random walkers traverses a Markov graph in discrete time steps, with co-location of any pair of walkers in a Markov state resulting in a link in their contact graph. RWIG is simple, interpretable, and analytically tractable, but it is an open question whether a first-order Markov process is expressive enough to capture real human mobility, or whether more expressive, data-driven sequence models, such as temporal GNNs or Transformers, learn genuinely better generators, rather than simply overfitting to noise in the training window. At the same time, a temporal contact graph is not just structure: it is the substrate an epidemic spreads on. Two generative processes that look similar in isolation can drive very different SIS (susceptible–infected–susceptible) outbreak dynamics, which suggests that a process's epidemic signature, e.g. how prevalence rises and falls as connectivity changes, could itself become a compact, learnable way to characterize and distinguish between classes of temporal-graph-generating processes.
Project
The student will pursue one (or, time permitting, both) of two closely related directions:
- Learning the generator. Compare RWIG's interpretable, hand-built Markov walker against more expressive, data-driven sequence models (temporal GNNs, Transformers) as generators of temporal contact graphs, and investigate whether their added expressiveness yields genuinely better-fitting mobility patterns or mainly memorises the training window.
- Epidemics as a probe. Build and study an SIS epidemic simulator running on temporal contact graphs from different generative classes (RWIG and its extensions), characterize how network topology and its time-variation shape epidemic dynamics, and explore whether the resulting epidemic curve can serve as a signature of the underlying graph-generating process.
The student is expected to:
- Develop a good understanding of the RWIG model and, depending on the chosen direction, either a more expressive sequence-model alternative or an SIS epidemic model on temporal networks
- Perform large-scale simulations of temporal contact graphs and/or of SIS epidemics running on top of them, across a range of input parameters and generative classes
- Conduct both theoretical and empirical analysis appropriate to the chosen direction: of the probability distribution of RWIG-generated graphs, and/or of how epidemic dynamics relate to the connectivity and generative class of the underlying temporal graph
Requirements
You are in the final stage of your degree in computer science, mathematics, electrical engineering or a similar degree. You are expected to have strong programming skills and experience in Markov theory and network science; familiarity with deep sequence models (GNNs/Transformers) is a plus for direction 1, and with epidemic or dynamical-systems modelling a plus for direction 2. You are pragmatic and focused on making things work. Next to technical expertise, we value a high level of independence and ability to work remotely, strong communication skills, and a results-driven attitude.
To apply
Contact David Almasan,