This dissertation presents novel models and methods for the statistical analysis of spatio-temporal point patterns observed on non-standard spatial domains. Central to the proposed approach is a penalized maximum likelihood framework for non-homogeneous Poisson point processes, which enables the estimation of density and intensity functions on irregular planar regions, curved surfaces, and linear networks. The methodology is grounded in rigorous theoretical results, including the existence and uniqueness of the proposed estimator. The design of a regularization based on roughness penalties in space and time ensures smoothness while preserving consistency as the number of observations increases. Extensive simulation studies show that the proposed method accurately captures complex spatio-temporal variations without relying on restrictive assumptions such as first-order separability. The approach combines computational efficiency and flexibility through advanced numerical techniques and iterative optimization algorithms. The extension to linear networks further highlights the versatility of the method, with its theoretical properties supported by metric and quantum graph theory and validated by numerical experiments. Applications in epidemiology, seismology, and urban mobility demonstrate its practical relevance. A hands-on tutorial illustrates the seamless integration of the C++ implementation in the fdaPDE library with an R interface, facilitating application to real-world datasets. Overall, the methodology provides a flexible, theoretically rigorous, and computationally feasible framework for spatio-temporal point pattern analysis, offering a powerful alternative to current state-of-the-art techniques.
Questa tesi presenta modelli e metodi innovativi per lo studio di dati puntuali spazio-temporali osservati su domini spaziali non standard. Al centro dell'approccio proposto vi è un modello di massima verosimiglianza penalizzata per processi puntuali di Poisson non omogenei, che consente la stima delle funzioni di densità e intensità su regioni planari irregolari, superfici curve e grafi. La metodologia si fonda su risultati teorici rigorosi, inclusa l'esistenza e l'unicità dello stimatore proposto. Il termine di penalizzazione garantisce la regolarità delle stime nello spazio e nel tempo, preservando la consistenza al crescere del numero di osservazioni. Studi di simulazione dimostrano che il metodo proposto cattura accuratamente variazioni spazio-temporali complesse, inclusi comportamenti multimodali e fortemente localizzati, senza assumere ipotesi semplificative come la separabilità di primo ordine in spazio e tempo. L'approccio coniuga efficienza computazionale e flessibilità grazie all'uso di tecniche numeriche avanzate e algoritmi iterativi di ottimizzazione. L'estensione ai grafi evidenzia ulteriormente la versatilità del metodo, con proprietà teoriche supportate dalla teoria dei grafi metrici e validate tramite esperimenti numerici. Applicazioni in epidemiologia, sismologia e mobilità urbana dimostrano la rilevanza pratica del metodo. Un tutorial pratico illustra l'integrazione dell'implementazione in C++ nella libreria fdaPDE con un'interfaccia in R, facilitando applicazioni a dataset reali. La metodologia fornisce un quadro flessibile, teoricamente rigoroso e computazionalmente praticabile per l'analisi di dati spazio-temporali, rappresentando una valida alternativa alle tecniche presenti in letteratura.
Statistical learning for spatio-temporal point patterns on non-standard spatial domains
PANZERI, SIMONE
2025/2026
Abstract
This dissertation presents novel models and methods for the statistical analysis of spatio-temporal point patterns observed on non-standard spatial domains. Central to the proposed approach is a penalized maximum likelihood framework for non-homogeneous Poisson point processes, which enables the estimation of density and intensity functions on irregular planar regions, curved surfaces, and linear networks. The methodology is grounded in rigorous theoretical results, including the existence and uniqueness of the proposed estimator. The design of a regularization based on roughness penalties in space and time ensures smoothness while preserving consistency as the number of observations increases. Extensive simulation studies show that the proposed method accurately captures complex spatio-temporal variations without relying on restrictive assumptions such as first-order separability. The approach combines computational efficiency and flexibility through advanced numerical techniques and iterative optimization algorithms. The extension to linear networks further highlights the versatility of the method, with its theoretical properties supported by metric and quantum graph theory and validated by numerical experiments. Applications in epidemiology, seismology, and urban mobility demonstrate its practical relevance. A hands-on tutorial illustrates the seamless integration of the C++ implementation in the fdaPDE library with an R interface, facilitating application to real-world datasets. Overall, the methodology provides a flexible, theoretically rigorous, and computationally feasible framework for spatio-temporal point pattern analysis, offering a powerful alternative to current state-of-the-art techniques.| File | Dimensione | Formato | |
|---|---|---|---|
|
phdthesis_Panzeri.pdf
solo utenti autorizzati a partire dal 12/02/2029
Descrizione: PhD_Thesis
Dimensione
81.01 MB
Formato
Adobe PDF
|
81.01 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/249918