Graph deep learning for irregular spatiotemporal data

Marisca, Ivan

Back

Doctoral thesis

Graph deep learning for irregular spatiotemporal data

Marisca, Ivan
Alippi, Cesare (Degree supervisor)

2025

PhD: Università della Svizzera italiana

Spatiotemporal graph neural networks

English Irregular spatiotemporal data – with observations acquired unevenly across space and time – are common in real-world applications such as climate modeling, social media analytics, and transportation systems. In these settings, capturing underlying structural dependencies is critical for effective pattern recognition and learning. While the existing literature has predominantly focused on regularly sampled data, irregular sampling introduces significant challenges that require novel methodological approaches. My doctoral research aims to advance predictive modeling for irregular spatiotemporal data through a graph-based representation, focusing specifically on imputation, filtering, and prediction across both temporal and spatial dimensions. Graphs naturally lend themselves to modeling irregularities, with edges encoding relationships between data-generating processes, making graph deep learning a suitable and powerful processing framework. Imputation and filtering methods are proposed to reconstruct missing observations and improve data quality by enforcing spatiotemporal consistency. Prediction approaches are explored to forecast future observations by leveraging underlying dependencies. A theoretical analysis of information propagation in spatiotemporal architectures completes the thesis, providing insights and guidelines for the design of future models.

Collections