Calibration with Empirically Weighted Mean Subset

Abstract

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^−ω_γ, 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.