Online PCA in High Dimension: A Comparative Study

• Activity Number: 43
• Type: Contributed
• Date/Time: Sunday, August 9, 2015 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): David Degras* and Hervé Cardot
• Companies: DePaul University and Université de Bourgogne
• Keywords: PCA ; Online algorithm ; High Dimension ; R package

Abstract

With the ever increasing amount of available data in many domains, online techniques that do not necessitate to store all data in memory become warranted to compute summary statistics such as principal components. Recursive algorithms for principal components, that allow automatic update at each new observation, have been widely studied in various fields of research: statistics and probability, numerical analysis, signal processing with applications in various fields such as astronomy, process monitoring, etc.. In a unified framework, we review the principal procedures developed in the literature and compare them. We also indicate how to adapt these new recursive techniques to functional data. All procedures are made available thanks to an R package posted on CRAN.