Abstract Title
Harnessing Model Diversity and Prediction Similarity for Selecting Multivariate Calibration Tuning Parameters
Additional Funding Sources
The project described was sponsored by the National Science Foundation under Grant No. CHE-1506417 and is gratefully acknowledged by the authors.
Abstract
Spectral multivariate calibration offers a cost-effective mechanism to obtain sample analyte values of a substance (e.g. protein level). However, calibration requires varying one or more tuning parameters in order to identify the most accurate model. Model selection is particularly difficult for model updating where spectral and reference information in both the original (primary) conditions and new (secondary) conditions are combined in order to better predict new spectra. Secondary situations can be new instruments, temperatures, or other condition affecting the shape and magnitude of the spectra relative to the primary conditions and analyte values. This poster uses model diversity while maintaining similar analyte prediction values to choose a set of acceptable models. The model selection technique is tested across the calibration method partial least squares and four model updating methods: two require a small set of secondary samples with analyte values and two do not require the secondary analyte values (unlabeled data). Results are presented across a variety of datasets and conditions showing that the cosine of the angle between models in combination with model vector 2-norms and prediction differences are key to selecting models.
Harnessing Model Diversity and Prediction Similarity for Selecting Multivariate Calibration Tuning Parameters
Spectral multivariate calibration offers a cost-effective mechanism to obtain sample analyte values of a substance (e.g. protein level). However, calibration requires varying one or more tuning parameters in order to identify the most accurate model. Model selection is particularly difficult for model updating where spectral and reference information in both the original (primary) conditions and new (secondary) conditions are combined in order to better predict new spectra. Secondary situations can be new instruments, temperatures, or other condition affecting the shape and magnitude of the spectra relative to the primary conditions and analyte values. This poster uses model diversity while maintaining similar analyte prediction values to choose a set of acceptable models. The model selection technique is tested across the calibration method partial least squares and four model updating methods: two require a small set of secondary samples with analyte values and two do not require the secondary analyte values (unlabeled data). Results are presented across a variety of datasets and conditions showing that the cosine of the angle between models in combination with model vector 2-norms and prediction differences are key to selecting models.
Comments
W61