Title: Nonparametric variable selection and screening with a large number of functional predictors
Authors: Jan Gertheiss - Clausthal University of Technology (Germany) [presenting]
Abstract: The situation where a very large number, like thousands, or tens of thousands, of functional predictors are available in a regression or classification problem is considered. A typical example are time course gene expression data used for classifying patients into different groups. Standard approaches for variable (pre)selection in this context are typically based on univariate, ANOVA-type testing, followed by ranking covariates according to the associated p-values. We propose an alternative, nonparametric method: the functional nearest neighbor ensemble. Nearest neighbor based estimates are used to build an ensemble, with each ensemble member representing a specific functional covariate. A lasso-type penalty is used to select ensemble members, and hence functional predictors with the highest predictive power. The proposed ensemble method is illustrated in simulation studies and on real world data, and compared to standard approaches described above.