Provider: ingentaconnect
Database: ingentaconnect
Content: application/x-research-info-systems
TY - ABST
AU - Ã–jelund, Henrik
AU - Madsen, Henrik
AU - Thyregod, Poul
TI - Calibration with Empirically Weighted Mean Subset
JO - Applied Spectroscopy
PY - 2002-07-01T00:00:00///
VL - 56
IS - 7
SP - 887
EP - 896
KW - BEST SUBSET
KW - SHRINKAGE
KW - CALIBRATION
KW - REGULARIZATION
KW - MEAN SUBSET
KW - NIR
KW - MODEL AVERAGING
N2 - In this article a new calibration method called empirically weighted mean subset (EMS) is presented. The method is illustrated using spectral data. Using several near-infrared (NIR) benchmark data sets,
EMS is compared to partial least-squares regression (PLS) and interval partial least-squares regression (iPLS). It is found that EMS improves on the prediction performance over PLS in terms of the mean
squared errors and is more robust than iPLS. Furthermore, by investigating the estimated coefficient vector of EMS, knowledge about the important spectral regions can be gained. The EMS solution is obtained
by calculating the weighted mean of all coefficient vectors for subsets of the same size. The weighting is proportional to SS_{
γ
}
^{-ω
}, where SS_{
γ
}
is the residual sum of squares from a linear regression with subset *γ* and *ω* is a weighting parameter estimated using cross-validation. This construction of the weighting implies
that even if some coefficients will become numerically small, none will become exactly zero. An efficient algorithm has been implemented in MATLAB to calculate the EMS solution and the source code has been
made available on the Internet.
UR - http://www.ingentaconnect.com/content/sas/sas/2002/00000056/00000007/art00012
M3 - doi:10.1366/000370202760171563
UR - http://dx.doi.org/10.1366/000370202760171563
ER -