Picture of scraped petri dish

Scrape below the surface of Strathprints...

The Strathprints institutional repository is a digital archive of University of Strathclyde research outputs. Explore world class Open Access research by researchers at Strathclyde, a leading technological university.

Explore

The successive projections algorithm for spectral variable selection in classification problems

Pontes, M J C and Galvao, R K H and Araujo, M C U and Moreira, T and Neto, O D P and Jose, G E and Saldanha, T C B (2005) The successive projections algorithm for spectral variable selection in classification problems. Chemometrics and intelligent laboratory systems, 78 (1-2). pp. 11-18. ISSN 0169-7439

Full text not available in this repository. (Request a copy from the Strathclyde author)

Abstract

The Successive Projections Algorithm (SPA) has been shown to be a useful tool for variable selection in the framework of multivariate calibration. In this paper, the collinearity minimization role of SPA is exploited in the context of classification methods for which collinearity is a known cause of generalization problems. For this purpose, a cost function associated to the average risk of misclassification by Linear Discriminant Analysis (LDA) is used to guide SPA selection. The proposed approach is illustrated in two classification problems. The first problem involves four types of vegetable oils (corn, soya, canola, sunflower). In this case, UV-VIS spectrometry is adopted to emphasize the ability of SPA-LDA to deal with low-resolution spectra with strong overlapping, which are associated to the wide absorption bands in this region. In the second problem, NIR spectrometry is employed to discriminate diesel samples with respect to the concentration level of sulphur. This application illustrates the use of SPA-LDA in a large-scale variable selection scenario. In these two examples, SPA-LDA is compared with the commonly used SIMCA classification method, as well as with a genetic algorithm (GA). The results show that SPA-LDA is superior to SIMCA and comparable to GA-LDA with respect to classification accuracy in an independent prediction set. Moreover, SPALDA is found to be less sensitive to instrumental noise than GA-LDA.