Graph Deep Learning for Irregular Spatiotemporal Data

Irregular spatiotemporal data consist of observations collected asynchronously across different times and spatial locations. Unlike regular data, their inherent irregularity makes traditional methods designed for discrete-time sequences and Euclidean spaces partly ineffective. 

In this talk, we explore how graph deep learning leverages the existence of relational dependencies to tackle both interpolation (imputation) and extrapolation (prediction) challenges. We will cover techniques for reconstructing missing data from a limited set of observations by enforcing spatiotemporal consistency, as well as introduce forecasting methods for predicting future values from sparse inputs. Finally, we will see how virtual sensors readings can be inferred from the graph structure provided some exogenous information is available.