CMStatistics 2015: Start Registration
View Submission - CMStatistics
B0663
Title: Robust outlier detection methods for functional data Authors:   - ()
Nicholas Tarabelloni - Politecnico di Milano (Italy)
Abstract: Functional data is a very attractive model for modern applications, yet in view of practical analyses of functional datasets a major issue is the identification of their extreme observations. Due to their infinite-dimensional nature, even after a dimensional reduction, functional data are always high-dimensional compared to their typical sample size. For this reason, even a little contamination of the dataset may lead to unreliable inferential conclusions. From an operational point of view, two main kinds of outlying behaviours are generally considered: the first is related to the amplitude of data (amplitude outliers); the second is related to the phase of data (shape outliers), which can be due to different running times of the units. In principle, outliers can be addressed either by building robust estimators for the quantities involved in the inferential procedures, or by suitably robustifying the dataset, through the use of ad-hoc techniques; an example is the adjusted functional boxplot, based on the notion of statistical depth and adapted to the dataset at hand, which targets amplitude outliers. We explore the possibility to combine the strengths of both approaches in order to build a robust version of the adjusted functional boxplot.