Second-order advantage when applying different artificial neural networks after unfolded principal component analysis/residual bilinearization

ALEJANDRO CESAR OLIVIERI; HÉCTOR GOICOECHEA,; MARIA CULZONI,; ALEJANDRO GARCÍA-REIRIZ,; PATRICIA DAMIANI,

Montpellier, Francia

Congreso; 11th conference on Chemometrics in Analytical Chemistry; 2008

When processed by adequate algorithms, second-order instrumental data allow analysts to obtain the useful second-order advantage, a property with immense implications in the analysis of real samples of complex composition [4]. Recently, attention has been focused on the possibility of extracting the second-order advantage from non-linear second-order information. This requires the coupling of two separate methods, which are able to accomplish the following successive tasks: 1) model the calibration and test data so that the contribution of unexpected components, not present in the calibration set, is removed from the test sample, and 2) model the non-linear relationship between calibration data and analyte concentration, and interpolating the pre-processed test data for prediction purposes [1]. The first of these two tasks can be achieved using a combination of unfolded principal component analysis (U-PCA) and residual bilinearization (RBL) [1]. On the other hand, in order to accomplish the second task, a variety of non-parametric regression techniques is available. Among them, artificial neural networks (ANN) can be conveniently exploited [8]. The latter constitute a popular methodology to take into account non-linearities in the relationship between instrumental data and concentration, and can be further sub-divided into: 1) multilayer perceptron networks (MLP) with back-propagation learning algorithm, 2) radial basis functions (RBF) networks, and 3) support vector machines (SVM), among other [3]. MLP have been widely applied in calibration problems [3,5]. On the other hand, RBF and SVM have been recently introduced as promising neural network methodologies, able to perform nonlinear multivariate function estimation and nonlinear regression tasks [2,9]. It was of interest, therefore, to check whether the newly developed non-linear models RBF and SVM may approximate better the non-linearities present in several analytical systems. In the present report, we process three experimental systems using U-PCA/RBL to filter the test sample from unexpected component contributions. We then model the score-concentration relationship with the three artificial neural network models mentioned above: MLP, RBF and SVM. A comparison of figures of merit for these three methods allows one to get insight into the best non-linear data processing technique in each of the studied cases.