Domain-Invariant Partial-Least-Squares Regression.

Area of Science:

• Chemometrics
• Machine Learning
• Spectroscopy

Background:

• Multivariate calibration models struggle with extrapolation due to domain shifts (instrumental, environmental, sample matrix).
• Existing calibration transfer methods are limited to similar analytical devices, lacking generic model adaptation solutions.
• Adaptation to new domains is crucial for maintaining model performance in real-world applications.

Purpose of the Study:

• Introduce domain-invariant partial-least-squares (di-PLS) regression for generic calibration model adaptation.
• Develop a method to align source and target distributions in latent-variable space.
• Enable model adaptation using labeled, partially labeled, or unlabeled target data.

Main Methods:

• Extended ordinary partial-least-squares (PLS) by incorporating a domain regularizer.
• Derived a closed-form solution for a domain-invariant weight vector.
• Applied di-PLS for unsupervised, semisupervised, and supervised model adaptation.

Main Results:

• Demonstrated successful adaptation of calibration models to target domains with varying data distributions.
• Showcased the ability to desensitize models to interfering agents in unsupervised adaptation.
• Validated di-PLS effectiveness on simulated and real-world near-infrared (NIR) spectroscopic data.

Conclusions:

• di-PLS regression provides a versatile framework for robust calibration model adaptation.
• The method effectively handles domain shifts, improving model generalizability.
• di-PLS facilitates model transfer across diverse analytical scenarios, enhancing practical applicability.