Journal of Chemometrics最新文献

Due to the significant cost and time involved in identifying barley lines with superior malting quality, the malting industry is searching for accurate and rapid methods to expedite the selection of superior barley lines that meet breeder's goals. The aim of this study is to compare partial least squares regression (PLSR) with advanced statistical models (Bayesian and machine learning) and reliably assess their performance in predicting malt quality traits from near infra-red (NIR) spectral data using barley grains. Using spectral data as predictors and the malt quality traits as references, PLSR outperformed Bayesian and PCA-ANN models for diastatic power (DP), alpha amylase (AA), malt extract (ME), wort protein (WP), soluble to total protein (S/T) ratio, and free amino nitrogen (FAN). WP had the best prediction performance for all models, with the best-performing model, PLSR, having $�{�R�}^{�2�}$ (RPD) values of 0.55 (1.5). The influential wavelength regions identified based on the variable importance in projection (VIP) scores and coefficient estimates for PLSR and Bayesian models, respectively, were comparatively similar for all malt quality traits. Based on these findings, PLSR analysis and wavelength selection techniques would enhance the future design and optimization of NIR prediction models in malt quality improvement programs.

High-dimensional compositional data are commonplace in the modern omics sciences, among others. Analysis of compositional data requires the proper choice of a log-ratio coordinate representation, since their relative nature is not compatible with the direct use of standard statistical methods. Principal balances, a particular class of orthonormal log-ratio coordinates, are well suited to this context as they are constructed so that the first few coordinates capture most of the compositional variability of data set. Focusing on regression and classification problems in high dimensions, we propose a novel partial least squares (PLS) procedure to construct principal balances that maximize the explained variability of the response variable and notably ease interpretability when compared to the ordinary PLS formulation. The proposed PLS principal balance approach can be understood as a generalized version of common log-contrast models since, instead of just one, multiple orthonormal log-contrasts are estimated simultaneously. We demonstrate the performance of the proposed method using both simulated and empirical data sets.

Yellow peaches are soft, and they bruise easily; the bruised areas of them are prone to breed bacteria and molds, so the consumption and the safety of related products of yellow peaches are affected by the difference in the storage time of light bruises in them. In order to accurately distinguish of the storage time of light bruises in yellow peaches, the spectra of the sample bruised region were combined with texture features extracted based on gray-level co-occurrence matrix (GLCM), and the deep learning algorithm was used for modeling. A total of 80 samples were prepared in the experiment, and the hyperspectral images of them were acquired at four time periods (2, 8, 24, and 48 h), and the reflection spectral data as well as the texture features of the bruised samples were extracted from the hyperspectral images. First, the random forest (RF) and extreme gradient boosting (XGBoost) models were built based on spectral, texture, and spectral features combined with texture features (Feature Fusion 1), respectively, and the best model discrimination was the RF model under Feature Fusion 1, with an overall accuracy of 98.33%. In order to remove the redundant information of spectrum, the UVE and CARS algorithms were used to screen the normalized spectral feature data, and then, the texture features were combined again (Feature Fusion 2), and the RF and XGBoost models were built. The results show that the optimal model for distinguishing the storage time of yellow peaches after bruising is the RF model under Feature Fusion 2 (CARS), with an overall accuracy of 98.33%. In summary, this study shows that spectral features combined with texture features can be used to effectively improve the model's discrimination of storage time after bruising of yellow peaches, and it also provides a certain theoretical basis for hyperspectral imaging technology to discriminate storage time after bruising of fruits.

The article from this special issue was previously published in Journal of Chemometrics, Volume 35, Issue 10, 2021. For completeness we are including the title page of the article below. The full text of the article can be read in Issue 35:10 on Wiley Online Library: https://doi.org/10.1002/cem.3370

Many real-world data mining applications involve using imbalanced datasets to obtain predictive models. Imbalanced data can hinder the model performance of learning algorithms in rare cases. Although there are many well-researched classification task solutions, most of them cannot be directly applied to regression task. One of the challenges in imbalanced regression is to find a suitable evaluation and optimization standard that can improve the predictive ability of the model without severe model bias. Based on the importance of rare cases, this study proposes a new evaluation metric called adapted squared error relevance (ASER) by defining new relevance function and weighting functions. This metric weights data points by defining the importance of rare cases and assigns different weights to losses of the same size at different rare cases, thus enabling the model selected by this evaluation metric to better predict rare cases. ASER is compared with SER on 32 real datasets and 9 simulated datasets to verify the predictive performance of the selected model at rare cases. The experimental results show that the new evaluation metric ASER can obtain a high prediction performance at rare cases, while also not losing too much prediction accuracy in common cases.

The complementary nature of analysis of variance (ANOVA) Simultaneous Component Analysis (ASCA+) and Tucker3 tensor decompositions is demonstrated on designed datasets. We show how ASCA+ can be used to (a) identify statistically sufficient Tucker3 models; (b) identify statistically important triads making their interpretation easier; and (c) eliminate non-significant triads making visualization and interpretation simpler. For multivariate datasets with an experimental design of at least two factors, the data matrix can be folded into a multi-way tensor. ASCA+ can be used on the unfolded matrix, and Tucker3 modeling can be used on the folded matrix (tensor). Two novel strategies are reported to determine the statistical significance of Tucker3 models using a previously published dataset. A statistically sufficient model was created by adding factors to the Tucker3 model in a stepwise manner until no ASCA+ detectable structure was observed in the residuals. Bootstrap analysis of the Tucker3 model residuals was used to determine confidence intervals for the loadings and the individual elements of the core matrix and showed that 21 out of 63 core values of the 3 × 7 × 3 model were not significant at the 95% confidence level. Exploiting the mutual orthogonality of the 63 triads of the Tucker3 model, these 21 factors (triads) were removed from the model. An ASCA+ backward elimination strategy is reported to further simplify the Tucker3 3 × 7 × 3 model to 36 core values and associated triads. ASCA+ was also used to identify individual factors (triads) with selective responses on experimental factors A, B, or interactions, A × B, for improved model visualization and interpretation.

The increasing scale of modern chemical plants keeps popularizing investigation as well as application of distributed process monitoring approaches. With a goal of directly quantifying the normal relations between different blocks divided from the whole process, a novel multi-block modeling strategy called block-wise residual generator is proposed, which trains a residual generator for each block through using the partial least squares algorithm with single one output, so that the relation between the corresponding block and the others is quantified as a regression model in a block-wise manner. The deviations caused by the abnormal samples to the normal relations quantified for different blocks could thus be efficiently captured by the residuals generated from the block regression models, which then provide sensitive information for fault detection and contribution-based fault diagnosis. Moreover, the proposed method is applicable for both disjoint and overlapped block divisions, and the direct consideration of individually quantifying relations between different blocks can always guarantee its salient monitoring performance, as validated through comparisons with classical distributed process monitoring methods.

Scatter corrections are commonly applied to refine near-infrared (NIR) spectra. The aim of this study is to assess the impact of measurement errors when using ordinary least squares (OLS) for multiplicative scatter correction (MSC). Any measurement errors attached to the set-mean spectrum may attenuate the OLS slope and that in turn will affect the estimate of the intercept and the adjustment of the spectra when using MSC methods to mitigate scattering. A corrected least squares slope may be used instead to prevent this problem, although the impact of this approach on the final outcome will depend on the relative size of the measurement errors in the individual spectra and the set-mean spectrum. The errors-in-variables or type II regression model (also known as Deming regression) and its special cases, major axis (MA) and reduced major axis (RMA), are discussed and illustrated. The extent of OLS slope bias or attenuation is demonstrated as is the resulting MSC spectral distortion. Further modification to the MSC transformation method is also suggested. The influence of scattering correction (by MSC, standard normal variate (SNV) and detrending) and of using the maximum likelihood estimate of the slope for MSC on the prediction of chemical composition of Lucerne herbage from NIR spectra was assessed. The predictive performance was slightly improved by the use of scattering corrections with fairly minor differences among methods. Nonetheless, it seems well worth considering the use of type II regression models for assessing MSC application aiming at improving the goodness of prediction from NIR spectra.

In various situations requiring empirical model building from highly multivariate measurements, modelling based on partial least squares regression (PLSR) may often provide efficient low-dimensional model solutions. In unsupervised situations, the same may be true for principal component analysis (PCA). In both cases, however, it is also of interest to identify subsets of the measured variables useful for obtaining sparser but still comparable models without significant loss of information and performance. In the present paper, we propose a voting approach for sparse overall maximisation of variance analogous to PCA and a similar alternative for deriving sparse regression models influenced closely related to the PLSR method. Both cases yield pivoting strategies for a modified Gram–Schmidt process and its corresponding (partial) QR-factorisation of the underlying data matrix to manage the variable selection process. The proposed methods include score and loading plot possibilities that are acknowledged for providing efficient interpretations of the related PCA and PLS models in chemometric applications.