Bayesian Semi-Supervised Learning (BSSL) for spectral variable selection.

Area of Science:

• Spectroscopy and Analytical Chemistry
• Machine Learning and Data Science
• Chemometrics

Background:

• Variable selection is critical for spectra analysis but is challenged by noise and distribution shifts.
• Ensuring the reliability of variable selection methods under real-world conditions is essential.
• Existing methods may lack robustness against external disturbances in spectral measurements.

Purpose of the Study:

• To develop a robust Bayesian Semi-Supervised Learning (BSSL) framework for reliable variable selection and model updating in spectra analysis.
• To enhance the adaptability and predictive performance of variable selection models using unlabeled data.
• To improve the interpretability and robustness of variable selection in the presence of spectral uncertainties.

Main Methods:

• A Bayesian variable selection approach utilizing maximum likelihood estimation (MLE) and posterior variance for uncertainty assessment.
• A semi-supervised learning framework incorporating unlabeled target samples for adaptive model updating.
• Implementation of a pseudo-labeling update strategy to adjust neighboring variables and refine model performance.

Main Results:

• The proposed BSSL framework demonstrated effectiveness on publicly available datasets.
• Selected variables accurately identified sensitive spectral regions relevant to target analytes.
• Model updates concentrated on neighboring spectral regions, indicating localized adaptation.
• The method achieved enhanced predictive performance, interpretability, and robustness.

Conclusions:

• The BSSL framework offers a reliable solution for variable selection and model updating in spectra analysis.
• The approach effectively handles uncertainties arising from noise and distribution shifts.
• The method provides interpretable results by identifying key spectral regions and demonstrates robust performance.