Comparison of locally weighted PLS strategies for regression and discrimination on agronomic NIR data

Abstract

In multivariate calibrations, locally weighted partial least squared regression (LWPLSR) is an efficient prediction method when heterogeneity of data generates nonlinear relations (curvatures and clustering) between the response and the explicative variables. This is frequent in agronomic data sets that gather materials of different natures or origins. LWPLSR is a particular case of weighted PLSR (WPLSR; ie, a statistical weight different from the standard 1/n is given to each of the n calibration observations for calculating the PLS scores/loadings and the predictions). In LWPLSR, the weights depend from the dissimilarity (which has to be defined and calculated) to the new observation to predict. This article compares two strategies of LWPLSR: (a) "LW": the usual strategy where, for each new observation to predict, a WPLSR is applied to the n calibration observations (ie, entire calibration set) vs (b) "KNN-LW": a number of k nearest neighbors to the observation to predict are preliminary selected in the training set and WPLSR is applied only to this selected KNN set. On three illustrating agronomic data sets (quantitative and discrimination predictions), both strategies overpassed the standard PLSR. LW and KNN-LW had close prediction performances, but KNN-LW was much faster in computation time. KNN-LW strategy is therefore recommended for large data sets. The article also presents a new algorithm for WPLSR, on the basis of the "improved kernel #1" algorithm, which is competitor and in general faster to the already published weighted PLS nonlinear iterative partial least squares (NIPALS). K E Y W O R D Sdiscrimination, locally weighted calibration, near-infrared spectroscopy, partial least squares, regression

Selecting Milk Spectra to Develop Equations to Predict Milk Technological Traits

“…In this study, we applied an objective, data-driven approach to select the optimal spectra for prediction according to the characteristics of the predictand. This differs to previous studies which have applied the local approach, whereby the number of neighbours used during the PLSR analyses is constant for all the observations in a dataset, regardless of the differences between observations [ 4 , 29 ]. In fact, some predictands can have more similar observations in the data set used for prediction (thus more neighbours) than others.…”

“…Local approaches, whereby the points used to predict a target spectrum are specially selected based on their similarity to the predictand, have the potential to improve the prediction performance over global PLSR prediction in heterogeneous datasets and when non-linear associations are present between the spectra and the trait of interest [ 4 , 11 ]. While local approaches have improved prediction accuracy in many studies [ 13 , 14 , 15 ], the local approaches did not consistently benefit prediction accuracy in the homogeneous milk dataset analysed here.…”

“…Partial least squares regression (PLSR) analysis [ 2 , 3 ] is the principal statistical method used to relate MIRS spectral data to a trait, as PLSR can handle collinear and high-dimensional datasets. Partial least squares regression is efficient in phenotype prediction when spectral data originate from the same source material, for example milk, and when the relationship between the spectral information and the reference data is linear [ 4 ]. Recently, the use of PLSR to relate milk spectral data to various milk and animal traits has been challenged by alternative machine-learning methods [ 5 , 6 , 7 ].…”

Performance of near infrared spectroscopy of a solid cattle and poultry manure database depends on the sample preparation and regression method used

“…A k-nearest-neighbour locally weighted PLS (kNN-LWPLSR) 28 belongs to the family of local regressions, such as LOCAL regression, 29 locally-weighted regression 30 or comparison analysis using restructured near-infrared and constituent data (CARNAC). 31 The local regression methods are based on constructing a specific calibration for each unknown sample from a selection of k calibration samples spectrally close to each unknown validation sample within a large spectral library.…”

ChemHouse: A research and development centre for chemometrics

“…Three seminars were dedicated to local PLS: weighted PLS, ParSketch-PLS and RoBoost-PLS. [1][2][3] Some seminars were held on other non-linear methods, such as artificial neural networks. • Another important topic in ChemHouse is the preprocessing of spectra.…”