[A new algorithm for NIR modeling based on manifold learning].

Area of Science:

• Machine learning
• Chemometrics
• Spectroscopy

Context:

• High-dimensional Near-Infrared (NIR) spectra present challenges in data analysis.
• Manifold learning assumes complex data resides on lower-dimensional manifolds.
• Locally linear embedding (LLE) is a key manifold learning technique.

Purpose:

• To develop a novel manifold learning algorithm, least square locally weighted regression (LS-LWR).
• To apply LS-LWR for quantitative prediction of glucose concentrations from NIR spectra.
• To evaluate LS-LWR's performance against established methods like PCR and PLSR.

Summary:

• LS-LWR is proposed, a novel algorithm adapting locally weighted regression (LWR) with least squares-determined weights.
• The algorithm leverages manifold learning assumptions for NIR spectral data.
• LS-LWR was validated using NIR spectra of glucose solutions with varying concentrations.

Impact:

• LS-LWR demonstrates superior predictability for glucose concentration compared to PCR and PLSR, indicated by a lower standard error of prediction (SEP).
• The developed model exhibits good stability and efficiency.
• This research offers an effective approach for quantitative analysis of complex spectral data.