Capturing spatiotemporal long-range dependencies in dynamic graphs

In this thesis, models to address the problem of long-range spatiotemporal dependencies in dynamic graphs are presented and tested on both real and synthetic datasets. These models are obtained through a combination of Graph Neural Networks with the NeuralODE paradigm, exploiting theoretical properties of the dynamical system that arise from the ODE describing the model to guarantee stability during training and more efficient propagation of information both in the spatial dimension (between different nodes of the graph) and the temporal dimension (different timestep of the dynamical graph).