Feature selection for functional data

Abstract

We herein introduce a general procedure to capture the relevant information from a functional data set in relation to a statistical method used to analyze the data, such as, classification, regression or principal components. The aim is to identify a small subset of functions that can "better explain" the model, highlighting its most important features. We obtain consistency results for our proposals. The computational aspects are analyzed, a heuristic stochastic algorithm is introduced and real data sets are studied.