New Multivariate Calibration and Classification Processes

Including Multimerits and Multituning Parameters

Dr. John Kalivas
Department of Chemistry
Idaho State University

Multivariate calibration of spectroscopic data is key to many areas of data modeling including nuclear physics, chemistry, astrophysics, planetary exploration, contamination assessment, detection of hidden radioactive material, food analysis including adulteration detection and authentication of product origin, medical diagnosis such as disease detection, and the list goes on. A primary issue with multivariate calibration is calibration maintenance. That is, developing a calibration in one set of environmental, instrumental, physical, and chemical conditions (the primary conditions) and then updating the calibration to now work in new secondary conditions. Presented are some new Tikhonov regularization (TR) variants to solve this issue including calibration without reference samples. Classifying samples into known categories is also a common problem in many fields. Presented is a projection approach that improves on several benchmark processes. Regardless of the purpose, multivariate calibration and classification processes typically involve selecting a value for a tuning parameter (meta-parameter). Selection of the value is typically dependent on one merit (quality indicator). Presented is the new method sum of ranking differences (SRD) to ensemble a collection of model evaluation merits varying across tuning parameter values. No weighting scheme of the merits is needed and it is shown that the SRD consensus ranking of model tuning parameter values allows automatic selection of the final model(s). Lastly, the SRD process can be used with TR calibration methods (and others) based on more than one tuning parameter (penalty term).