Abstract

We propose a new class of parameterizations for spatio-temporal pointprocesses which leverage Neural ODEs as a computational method and enableflexible, high-fidelity models of discrete events that are localized incontinuous time and space. Central to our approach is a combination ofrecurrent continuous-time neural networks with two novel neural architectures,i.e., Jump and Attentive Continuous-time Normalizing Flows. This approachallows us to learn complex distributions for both the spatial and temporaldomain and to condition non-trivially on the observed event history. Wevalidate our models on data sets from a wide variety of contexts such asseismology, epidemiology, urban mobility, and neuroscience.

 

Quick Read (beta)