CMStatistics 2015: Start Registration
View Submission - CMStatistics
B1684
Title: New developments of sparse PLS regressions Authors:  Jeremy Magnanensi - University of Strasbourg (France) [presenting]
Myriam Maumy - IRMA/Universite of Technology of Troyes (France)
Nicolas Meyer - Universite de Strasbourg (France)
Frederic Bertrand - Universite de technologie de Troyes (France)
Abstract: Processes based on the so-called Partial Least Squares (PLS) regression, which recently gained much attention in the analysis of high-dimensional genomic datasets, were recently developed to perform variables selection. Most of these processes rely on some tuning parameters that are usually determined by Cross-Validation (CV), which raises important stability issues. We developed a new dynamic bootstrap based PLS process for significant predictors selection, suitable for both PLS regression and its extension to Generalized Linear (GPLS) regression frameworks. It depends on a single tuning parameter, the number of components, which is determined by bootstrap, avoiding the use of CV. We also developed an adapted version of the Sparse PLS (SPLS) and SGPLS regression processes, using bootstrap for the determination of the numbers of components. We benchmarked their variable selection accuracy, as well as their stability concerning the tuning parameters and their predictive ability. This benchmarking is performed on simulations for PLS framework and on real microarray gene expression datasets for PLS-logistic classification.