Partial Least Squares Methods: from regularized regression to structural equation models

vendredi 1 mars 2013, 9h30 - 10h30

Salle de réunion, espace Turing

The acronym PLS (Partial Least Squares) (PLS) refers to a family of “soft modelling” methods implemented by various extensions of the Nonlinear estimation by Iterative PArtial Least Squares (NIPALS) algorithm. The basic principles of NIPALS were first developed in order to model the relationships between several blocks of observed variables, each one supposed to be the expression of an underlying latent variable (PLS approach to Structural Equation Models, or PLS Path Modeling – PLS-PM). Then, the NIPALS iteration was exploited to implement a component-based regularized regression technique, known as PLS regression (PLS-R).
Nowadays both these techniques, i.e. the PLS Regression and the PLS Path Modeling, are largely used in applied statistics.
This talk will focus on the computational and methodological aspects of these two methods. Moreover, some recent developments for both the PLS Regression and the PLS Path Modeling will be presented. Namely, a new approach to perform variable selection in PLS Regression, and the use of multidimensional latent variables in PLS-PM will be discussed.