Evaluation of the effect of data pre-treatment procedures on classical pattern recognition and principal components analysis: a case study for the geographical classification of tea.

Abstract

A simple transformation that uses the half-range and central value has been used as a data pre-treatment procedure for principal component analysis (PCA) and pattern recognition techniques. The results obtained have been compared with the results from classical normalisation of data (mean normalisation, maximum normalisation and range normalisation), autoscaling and the minimum-maximum transformation. Three data sets were used in the study. The first was formed by determining 17 elements in 53 tea samples (901 pieces of data). The second and third data sets arose from two long-term drift studies performed to examine instrumental stability at standard and robust conditions. The instruments used were an inductively coupled plasma atomic emission spectrometer and an inductively coupled plasma mass spectrometer. Each drift diagnosis experiment consisted of replicate determinations of a test solution containing 15 analytes at 10 mg l-1 over 8 h without recalibration. Twenty-nine emission lines were determined 99 times, thus, each data set was formed by 2881 pieces of data. Data pre-treatment was applied to the three data sets prior to the use of principal component analysis, cluster analysis, linear discrimination analysis and soft independent modelling of class analogy. The study revealed that the half-range and central value transformation resulted in a better classification of the tea samples than that achieved using the classical normalisation. The loadings in the PCA for the long-term stability study, under both standard and robust conditions, were found to be similar to the drift trends only when the minimum-maximum transformation and the mean or maximum normalizations were used as data pre-treatments.