Faculty Seminar Series
Partial least squares regression, which has been around for about four decades, is
a dimension-reduction algorithm for fitting linear regression models without
requiring that the sample size be larger than the number of predictors. It was
developed primarily by the Chemometrics community where it is now ingrained
as a core method, and it is apparently used throughout the applied sciences.
And yet it seems fair to conclude that PLS regression has not been embraced by
the statistics community, even as a serviceable method that might be useful
occasionally. Nor does there seem to be a common understanding within
statistics as to why this rather enigmatic method should not be used, although
bumptious discussions of PLS failings can be found in some applied areas.
Perhaps this is as it should be — perhaps not.
This talk is intended as a relatively informal overview on PLS regression, including
historical context, personal encounters, methodology, relationship to envelopes
and, near the end, a few recent asymptotic results for high-dimensional
Seminar: September 6 th , 2018
140 Nolte Center