MULTIPLE MULTIVARIATE PLS REGRESSION

• prediction when a high degree of correlation among the variables in both the predictor and response spaces exist

• PLS2 is inherently designed to deal with several response variables, however, almost none theoretical understanding of the properties of such model exist

• the curds & whey procedure (C&W) (Breiman & Friedman’97): the use of CCA between predictors and responses to decorrelate response variables ⇒ univariate (shrinkage) regression on decorrelated responses

• experimental evidence exists that C&W in the PLS2 framework may improve prediction accuracies
(Xu & Massart’03)

Sumber : citeseerx