Local Modeling by Adapting Source Calibration Models to Analyte Shifted Target Domain Samples Without Reference...

Area of Science:

• Analytical Chemistry
• Chemometrics
• Machine Learning

Background:

• Spectral multivariate calibration models predict analyte amounts but fail when target samples shift from the calibration domain.
• Recalibration with new reference values can be costly and challenging.
• Existing methods like regression domain adaptation struggle with significant analyte distribution shifts.

Purpose of the Study:

• To develop a more robust regression transfer learning method for updating spectral calibration models to shifted target domains.
• To address the limitations of existing model updating techniques in analytical chemistry.

Main Methods:

• Introduced Null Augmentation Regression Constant Analyte (NARCA), a regression transfer learning method.
• NARCA leverages unlabeled repeat spectra of a single target sample to adapt existing calibration models.
• NARCA operates as a local modeling method, assuming near-constant analyte amounts for repeat spectra.

Main Results:

• NARCA was evaluated as a regression transfer learning method across five near-infrared (NIR) datasets.
• The method demonstrates effectiveness in updating models to shifted target domains without requiring new analyte reference values.

Conclusions:

• NARCA offers a simpler and potentially more cost-effective approach to model updating compared to full recalibration.
• This method shows promise for improving the applicability of spectral calibration models in diverse analytical scenarios.