Near-infrared spectral expansion method based on active semi-supervised regression.

Area of Science:

• Analytical Chemistry
• Chemometrics
• Machine Learning

Background:

• Near-infrared (NIR) spectroscopy is vital in agriculture and pharmaceuticals.
• Limited training data due to laborious sample collection hinders NIR model optimization.
• Existing models struggle with small, labeled datasets.

Purpose of the Study:

• To address the challenge of limited training data in NIR spectral analysis.
• To develop a novel model that augments sample datasets effectively.
• To improve the accuracy and practical application of NIR spectrum analysis models.

Main Methods:

• Integration of active learning (AL) and semi-supervised learning (SSL) techniques.
• Development of the Safer Active Semi-Supervised Sample Augmentation Learning Model (Safer-AS³A).
• Utilizing AL, SSL, and co-training for high-quality pseudo-label generation.

Main Results:

• Safer-AS³A outperforms comparable models in accuracy and robustness with limited labeled samples.
• NIR spectral datasets were effectively expanded, improving model performance.
• Significant R² improvements observed in PLSR, BRR, SVR, and RR models on Diesel and Shoot datasets.

Conclusions:

• The Safer-AS³A model effectively expands NIR spectral datasets.
• This method considerably improves the performance of NIR spectral analysis.
• Presents new avenues for NIR analysis efficiency, precision, and sample diversification.