Q Guo

Feature selection in principal component analysis of analytical data

A feature selection method is proposed to select a subset of variables in principal component analysis (PCA) that preserves as much information present in the complete data as possible. The information is measured by means of the percentage of consensus in generalised Procrustes analysis. The best subset of variables is obtained by applying a genetic algorithm (GA) to optimise the consensus between the subset and the complete data set in order to avoid exhaustive searching. The method was evaluated on a standard data set known as the Alate data, and on a high-dimensional industrial gas chromatography (GC) data set. The results showed that the proposed method successfully identified structure-bearing variables in both data sets and that it leads to a better subset of variables than other studied feature selection methods.

Exploratory chemometric analysis of the classification of pharmaceutical substances based on chromatographic data.

A chemometric study has been conducted on a published data set consisting of the retention times of 83 substances, from five pharmacological families, on eight HPLC systems. Principal component analysis, clustering and sequential projection pursuit were applied. In this way it was investigated to what extent the combination of chromatography and chemometrics allows one to make conclusions about pharmacological activities of (candidate) drugs and what the contribution is of the different HPLC systems considered.

The robust normal variate transform for pattern recognition with near-infrared data

The standard normal variate transform (SNV) is applied to pretreat NIR data for pattern recognition. Eleven NIR data sets are analysed. The results show that SNV improves classification results in most of the cases by reducing the within-class variance. Because of the closure problem, SNV leads to artefacts and is difficult to interpret in simple methods of wavelength distance and univariate direct discrimination (DD). A proposed robust normal variate transform (RNV) gives more reasonable results than SNV. Because of the artefacts, SNV sometimes gives worse results for regularised discriminant analysis (RDA) than using the original data. In this case, RNV leads to improved results, and in general, it performs better than SNV, even when SNV gives better results than using the original data. However, the drawback of RNV is that the applied percentile needs to be optimised. A proposal for quick selection of the percentile is given.

Feature selection in sequential projection pursuit

A feature selection method is proposed to select a subset of variables in sequential projection pursuit (SPP) analysis in order to preserve as much sample clustering information as possible. The inhomogeneity of the complete data is explored by SPP, and the retained inhomogeneity information of a candidate subset is measured by means of the percentage of consensus in generalised procrustes analysis. The best subset is obtained by applying a genetic algorithm (GA) which optimises the consensus between the subset and the complete data set. An improved algorithm is proposed which enables analysis of high-dimensional data. The method was studied on three high-dimensional industrial data sets. The results show that the proposed method successfully identified inhomogeneity-bearing variables and leads to better subsets of variables than the other studied feature selection methods in preserving interesting clustering information.

Data structures and data transformations for clustering chemical data

The quality of a clustering of chemical data is determined by a proper choice of distance measures and data transformations. The latter aspect is often neglected and its importance is shown here. It is also shown that the V-shaped data structure that is often obtained in a principal component analysis of chemical data may indicate that the clustering of the raw data can lead to classifications that are not relevant from a chemical point of view and that the log double centering transform should be considered as a possible alternative.

The Neural-Gas network for classifying analytical data

This article introduces the Neural-Gas network, a fast neural net-based method for clustering, and shows how it is applied to gas chromatographic patterns of Maillard reaction products. The advantages of Neural-Gas are compared to the K-means clustering method and one of the best-known neural methods for clustering, the Kohonen self-organising maps. Some novel combinations with visualization techniques are also presented.

Structure preserving feature selection in PARAFAC using a genetic algorithm and Procrustes analysis

Abstract In this paper, a method is proposed to select subsets of variables in parallel factor analysis (PARAFAC), such that information in the complete multi-way data set is preserved as much as possible. The information retained is measured by means of the percentage of consensus in Procrustes analysis. The best N -way subset is obtained by applying a genetic algorithm (GA) to optimize the consensus between the subset and the complete N -way data set in order to prevent exhaustive searching. The method was applied to two industrial data sets: a three-way sensory data set and a four-way gas chromatography (GC) data set. The results showed that the proposed method successfully identified structure-bearing variables in both data sets and that it led to better subsets of variables than feature selection based on loadings.

Assessing molecular similarity/diversity of chemical structures by FT-IR spectroscopy.

FT-IR spectra have been investigated for their ability to distinguish compounds which are chemically diverse and to produce clusters of compounds which makes sense chemically. Principal component analysis (PCA) was applied to the analysis of a small database of FT-IR spectra. The effect of the data pretreatment step of log transformation on spectral data pattern was also visualized by using PCA plots. The method of sequential projection pursuit (SPP) was applied to detect inhomogeneities in the data. Finally, cluster analysis of these spectra, depending on unweighted pair-group average linkage, was carried out.

The star plot: an alternative display method for multivariate data in the analysis of food and drugs

The star plot (SP) is a method of displaying multivariate data. It can be used to display data with more than two variables. Combined with principal component analysis (PCA), more than two PCs can be displayed in one plane. Different variants of this method are applied to an atomic absorption spectrometry (AAS) data set and to three near infra-red (NIR) spectral data sets. The results show that SP offers an easy way of visualising the multivariate data for food and drugs in a plane, and it is able to help the analyst to identify and to detect different qualities of food and drug composition. Moreover, when an object is added or removed, the PCs must be computed all over again, which is not the case for the SP-plot. The application of SP to the examples presented in the text suggests that the SP approach can be applied as an alternative method for displaying multivariate chemometric data in place of PCA or, to improve visualisation of the results already obtained with PCA.