Abstract
Partial least squares (PLS) is a popular approach for soft modeling in industrial applications. PLS is a
method for constructing predictive models consisting of a set of score vectors (latent variables). The
score vectors are constructed to model relations between multivariate descriptor and uni- or multivariate
response blocks of data where a criterion of maximal covariance is used. This talk will give an overview
of PLS and its different forms. The nonlinear, kernel-based, extension of PLS will be considered throughout
the talk in parallel to the linear PLS model. Existing relations of PLS to canonical correlation analysis,
Fisher discriminant analysis and principal component analysis will be highlighted. The talk will focus on
statistical perspective of the PLS regression model and its shrinkage properties with respect to ordinary
least squares regression. Several aspects of PLS associated with multiple multivariate response regression,
geometric interpretation, variables selection and the use of the PLS model in discrimination tasks will also
be mentioned in the talk. Finally, several successful applications of the use of the PLS regression and
classification models on electroencephalogram data will be described.
Go back