Premium
Non‐linear partial least squares. Estimation of the weight vector
Author(s) -
Hassel Per A.,
Martin Elaine B.,
Morris Julian
Publication year - 2002
Publication title -
journal of chemometrics
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.47
H-Index - 92
eISSN - 1099-128X
pISSN - 0886-9383
DOI - 10.1002/cem.747
Subject(s) - covariance , weighting , partial least squares regression , mathematics , weight , linear model , set (abstract data type) , data set , linear regression , statistics , algorithm , computer science , mathematical optimization , medicine , lie algebra , pure mathematics , radiology , programming language
A key issue in non‐linear partial least squares (PLS) is the calculation of the weight vectors. The weight vector can be derived in a number of ways; for example, as in linear PLS, through an optimization routine as in error‐based non‐linear PLS, or through the covariance criterion as in Spline‐PLS. An alternative approach to the estimation of the weight vector in non‐linear PLS is proposed based on the theory of the weighted average, the reciprocal variance criterion. Unlike the covariance‐based criterion that gives equal consideration to both the response and the process variables in the calculation of the weight vector, the proposed approach gives emphasis to the response variables. The new weighting criterion performs well where the variability in the response variables can be captured by a few underlying phenomena. The methodology is illustrated through its application to simulated spectral data and an industrial NIR data set. It is shown that for the case studies the proposed approach outperforms the covariance criteria of PLS in terms of performance on a validation data set. Copyright © 2002 John Wiley & Sons, Ltd.