This thesis explores the emerging field of causal discovery in time series. Through the use of graphical models, the thesis addresses complex problems of causal inference, fundamental in many areas of science, medicine and engineering. The work focuses on methodologies to identify causal relationships in multidimensional time series, using advanced approaches such as dynamic Bayesian models, Granger causality, and structural equation models with temporal components. The thesis proposes an integrated approach that combines theory and practical applications. Particular attention is given to structure learning methods, such as PCMCI+ and the VARLiNGAM method, which are essential for identifying causal relationships in temporal data. Furthermore, various practical applications and artificial cases are explored to demonstrate the effectiveness of these methods in analyzing real time series. These ready-to-use examples provide a direct view of the implementation and efficacy of the models in realistic contexts. The thesis also discusses the implications of these techniques in enhancing the accuracy and reliability of forecasts in various sectors. The main contribution of this thesis lies in the deepening of the theory of dynamic causal models and their practical application. The comparability between the methods used, based on data assumptions, offers new perspectives and tools for deepening the study of time series and causal discovery. The work concludes by highlighting the importance of integrating causal theory and time series analysis to address complex challenges in various fields of scientific research.
Questa tesi esplora il campo emergente della scoperta causale nelle serie temporali. Attraverso l’uso di modelli grafici, la tesi affronta complessi problemi di inferenza causale, fondamentali in molti ambiti della scienza, medicina e dell’ingegneria. Il lavoro si concentra sulle metodologie per identificare relazioni causali in serie temporali multidimensionali, utilizzando approcci avanzati come i modelli Bayesiani dinamici, la causalità di Granger e i modelli di equazioni strutturali con componenti temporali. La tesi propone un approccio integrato che combina teoria e applicazioni pratiche. Viene data particolare attenzione ai metodi di apprendimento della struttura, come il PCMCI+ e il metodo VARLiNGAM, essenziali per identificare le relazioni causali nei dati temporali. Inoltre, vengono esplorate diverse applicazioni pratiche e casi artificiali per dimostrare l’efficacia di questi metodi nell’analisi di serie temporali reali. Questi esempi pronti all’uso offrono una visione diretta dell’implementazione e dell’efficacia dei modelli in contesti realistici. Il contributo principale di questa tesi sta nell’approfondimento della teoria dei modelli causali dinamici e nella loro applicazione pratica. La confrontabilità tra i metodi utilizzati, in funzione delle assunzioni dei dati, offre nuove prospettive e strumenti per approfondire lo studio delle serie temporali e della scoperta causale. Il lavoro conclude evidenziando l’importanza dell’integrazione tra teoria causale e analisi di serie temporali per affrontare sfide complesse in diversi campi della ricerca scientifica.
Exploring causal discovery in time series: graphical models and structure learning
SPARACINO, RICCARDO
2022/2023
Abstract
This thesis explores the emerging field of causal discovery in time series. Through the use of graphical models, the thesis addresses complex problems of causal inference, fundamental in many areas of science, medicine and engineering. The work focuses on methodologies to identify causal relationships in multidimensional time series, using advanced approaches such as dynamic Bayesian models, Granger causality, and structural equation models with temporal components. The thesis proposes an integrated approach that combines theory and practical applications. Particular attention is given to structure learning methods, such as PCMCI+ and the VARLiNGAM method, which are essential for identifying causal relationships in temporal data. Furthermore, various practical applications and artificial cases are explored to demonstrate the effectiveness of these methods in analyzing real time series. These ready-to-use examples provide a direct view of the implementation and efficacy of the models in realistic contexts. The thesis also discusses the implications of these techniques in enhancing the accuracy and reliability of forecasts in various sectors. The main contribution of this thesis lies in the deepening of the theory of dynamic causal models and their practical application. The comparability between the methods used, based on data assumptions, offers new perspectives and tools for deepening the study of time series and causal discovery. The work concludes by highlighting the importance of integrating causal theory and time series analysis to address complex challenges in various fields of scientific research.File | Dimensione | Formato | |
---|---|---|---|
Tesi_depositata.pdf
accessibile in internet solo dagli utenti autorizzati
Dimensione
1.47 MB
Formato
Adobe PDF
|
1.47 MB | Adobe PDF | Visualizza/Apri |
Tesi_Riccardo_Sparacino.pdf
accessibile in internet solo dagli utenti autorizzati
Dimensione
1.79 MB
Formato
Adobe PDF
|
1.79 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/10589/214231