Title: Recursive maxima hunting: Variable selection in FDA
Authors: Jose Luis Torrecilla - Universidad Autónoma de Madrid (Spain) [presenting]
Abstract: In a world of big and complex data, the use of methodologies for dimensionality reduction is a commonplace. In this context, variable selection techniques have been proved to be very useful alternatives, since they provide interpretable reductions with important predictive power. We study variable selection for supervised classification, and we are interested in the case of having data that are functions. In this setting, one of the alternatives is the maxima hunting method (MH) which performs variable selection by identifying the maxima of a dependence function between the predictive functional variable and the class label. MH presents a good performance and some valuable properties, however, the relevance of a variable is assessed individually and it has some estimation issues. We present a recursive extension of MH which solves these limitations by subtracting the expectation of the process conditioned to the already selected variables. The new methodology entails some interesting properties and the improvement is illustrated and assessed with simulations and real examples.