B1923
Title: A recursive approach to variable selection with functional data
Authors: Jose Luis Torrecilla - Universidad Autonoma de Madrid (Spain) [presenting]
Carlos Ramos Carreno - Universidad Autonoma de Madrid (Spain)
Alberto Suarez - Universidad Autonoma de Madrid (Spain)
Abstract: The increasing volume and complexity of data have made the use of methodologies for dimensionality reduction commonplace. In this context, variable selection techniques have proven to be very useful alternatives, as they provide interpretable reductions with significant predictive power. variable selection for supervised classification is studied when data are functions. In this setting, the continuous structure of the data makes feature selection a particularly appealing choice. One of these techniques is the maxima hunting method (MH), which performs variable selection by identifying the local maxima of a dependence function between the predictive functional variables and the class label. MH presents good performance and some valuable properties, including certain optimality results. However, the relevance of each variable is assessed individually, and the method has some estimation issues as well. A recursive extension of MH is presented which addresses these limitations by subtracting the expectation of the process conditioned on the already selected variables. The new methodology overcomes the limitations of the original MH and introduces some intriguing properties. The empirical performance is illustrated with simulations and real examples.
