This work addresses the problem of the sensitivity of Principal Component Analysis (PCA) to the presence of outliers, illustrating and comparing some of the main robust methods developed in the literature to overcome this limitation. After introducing the theoretical foundations of classical PCA and key concepts of robust statistics, including the influence function and the breakdown point, three approaches for constructing a robust PCA are presented: the plug-in method, based on replacing classical estimators with robust estimators such as MCD and OGK, the projection pursuit method, and ROBPCA. The methods are subsequently applied to and compared using two well-known real-world datasets from the literature, the Bank and Octane datasets. Furthermore, two Monte Carlo simulation studies are carried out to evaluate their ability to detect outliers. The results show that robust methods outperform classical PCA when the data are contaminated, successfully identifying outliers.
Relazione che affronta il problema della sensibilità dell'analisi delle componenti principali (PCA) alla presenza di valori anomali, illustrando e confrontando alcuni tra i principali metodi robusti sviluppati in letteratura per ovviare a tale criticità. Dopo aver introdotto i fondamenti teorici della PCA classica e i concetti chiave della statistica robusta, tra cui la funzione d'influenza e il punto di rottura, vengono presentati tre approcci per la costruzione di una PCA robusta: il metodo plug-in, basato sulla sostituzione degli stimatori classici con stimatori robusti quali MCD e OGK, il metodo del projection pursuit, e la ROBPCA. I metodi vengono quindi applicati e confrontati su dataset reali noti in letteratura, il dataset Bank e il dataset Octane, e vengono effettuate due simulazioni tramite Monte Carlo per valutare le capacità di identificazione degli outlier. I risultati mostrano che i metodi robusti risultano generalmente superiori alla PCA classica in caso di dati contaminati, riuscendo a individuare i valori anomali.
Metodi robusti per l'analisi delle componenti principali
TESCARO, MORGANA
2025/2026
Abstract
This work addresses the problem of the sensitivity of Principal Component Analysis (PCA) to the presence of outliers, illustrating and comparing some of the main robust methods developed in the literature to overcome this limitation. After introducing the theoretical foundations of classical PCA and key concepts of robust statistics, including the influence function and the breakdown point, three approaches for constructing a robust PCA are presented: the plug-in method, based on replacing classical estimators with robust estimators such as MCD and OGK, the projection pursuit method, and ROBPCA. The methods are subsequently applied to and compared using two well-known real-world datasets from the literature, the Bank and Octane datasets. Furthermore, two Monte Carlo simulation studies are carried out to evaluate their ability to detect outliers. The results show that robust methods outperform classical PCA when the data are contaminated, successfully identifying outliers.| File | Dimensione | Formato | |
|---|---|---|---|
|
Tescaro_Morgana.pdf
accesso aperto
Dimensione
920.46 kB
Formato
Adobe PDF
|
920.46 kB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/112214