Interpretable Perturbator for Variable Selection in near-Infrared Spectral Analysis.

Area of Science:

• Analytical Chemistry
• Chemometrics
• Machine Learning

Background:

• Near-infrared (NIR) spectral analysis is crucial for quantitative modeling.
• Variable selection is essential for optimizing NIR models and enhancing interpretability.
• Deep learning offers advanced methods for complex data analysis.

Purpose of the Study:

• To develop a novel perturbator method for variable selection in NIR spectral analysis.
• To leverage deep learning strategies for improved interpretation of spectral data.
• To create a criterion for evaluating variable importance in quantitative models.

Main Methods:

• A deep learning predictor was trained to establish target predictions from spectra.
• A perturbator was trained using the predictor's output to derive perturbation-positive (P+) and perturbation-negative (P-) features.
• Variable importance was assessed using the weights (σ) of the perturbator layer.
• Cross-validation was employed to determine optimal variable subsets for quantitative models.

Main Results:

• The perturbator method achieved comparable or superior performance (root mean squared error) to existing methods on three NIR datasets.
• Selected spectral variables demonstrated interpretability, identifying key features related to prediction targets.
• The method effectively optimized quantitative models and provided insights into spectral explanations.

Conclusions:

• The developed perturbator method is an effective tool for optimizing quantitative NIR models.
• This approach offers an efficient way to explain the spectra of multicomponent samples.
• The integration of deep learning and perturbation strategies enhances both model performance and interpretability in spectral analysis.