[Study on combinatorial optimization of spectral principal components using successive projections algorithm].

Area of Science:

• Spectroscopy and Chemometrics
• Analytical Chemistry
• Data Science

Context:

• Near-infrared (NIR) spectroscopy is a valuable tool for analyzing food products like milk powder.
• Principal Component Analysis (PCA) is commonly used for dimensionality reduction in spectral data.
• Selecting the most informative principal components (PCs) is crucial for accurate predictive modeling.

Purpose:

• To evaluate the Successive Projections Algorithm (SPA) for selecting optimal PCs from NIR spectra of milk powder.
• To compare the performance of predictive models using SPA-selected PCs versus a broader range of PCs.
• To develop accurate models for predicting fat and protein content in milk powder.

Summary:

• Principal Component Analysis (PCA) was applied to short-wave near-infrared spectra of milk powder.
• The Successive Projections Algorithm (SPA) was used to determine the optimal combination of principal components (PCs) for predicting fat and protein content.
• Least-squares support vector machine (LS-SVM) models utilizing SPA-selected PCs demonstrated superior prediction accuracy for both fat and protein content compared to models using the first 4 to 8 PCs.
• Specific optimal PC combinations were identified for fat (PC1, PC2, PC4, PC5, PC6, PC7) and protein (PC1, PC2, PC3, PC4, PC5, PC8) prediction.
• The SPA method proved to be fast, effective, and simple, requiring minimal parameter debugging.

Impact:

• The study validates SPA as a robust technique for feature selection in chemometric analysis.
• Achieved high prediction accuracy (e.g., R² of 0.989 for fat, 0.9876 for protein) using the optimized PC combinations.
• Demonstrates the potential for simplified and efficient quality control in milk powder analysis through optimized spectral data processing.