Performance of Algorithms for Chemometric Calibration

• Activity Number: 436
• Type: Contributed
• Date/Time: Wednesday, August 5, 2009 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Stanislav Zakharkin*+ and David Ryan and Nicole Mueth and Kyoungmi Kim
• Companies: Solae and Solae and Solae and University of California, Davis
• Keywords: partial least squares ; support vector machines ; random forest ; artificial neural nets ; near-infrared spectroscopy ; chemoinformatics

Abstract

In this study, we examined six algorithms implemented in R (PLS, SVM, ANN, elastic net, random forest, and boosted trees) and two implemented in ISI software (modified PLS and LOCAL). In selecting a robust and efficient calibration method, we compared algorithms using 5 real data sets that represent a wide diversity of composition and spectral properties, instrument configurations (transmission and reflection), and reference methods ranging from basic compositional characteristics to more empirical functional attributes. The utility of each algorithm was evaluated by comparing standard errors of cross-validation and standard errors of prediction on the independent validation sets. No algorithm performed uniformly better than others. Overall, PLS, LOCAL, SVM, and ANN consistently showed comparable performance, while performance of tree-based methods and elastic net was less stable.