A Novel Variable Selection Method Based on Binning-Normalized Mutual Information for Multivariate Calibration.

Area of Science:

• Analytical Chemistry
• Chemometrics
• Spectroscopy

Background:

• Variable selection is crucial for accurate multivariate analysis of near-infrared (NIR) spectra.
• Improving model interpretability and performance relies on effective wavelength selection.
• Existing methods may not adequately address noise or feature enhancement in complex spectral data.

Purpose of the Study:

• To introduce a novel variable selection method, binning-normalized mutual information (B-NMI), for NIR spectral analysis.
• To enhance the selection of informative wavelengths by reducing measurement error effects.
• To evaluate the performance and robustness of B-NMI against established methods.

Main Methods:

• Development of the binning-normalized mutual information (B-NMI) method based on information entropy theory.
• Application of data binning to NIR spectra to mitigate minor errors and enhance features.
• Utilizing normalized mutual information to quantify wavelength-reference value correlations.
• Validation using diverse datasets: ternary solvent mixtures, fluidized bed granulation, gasoline octane, and corn protein.

Main Results:

• B-NMI effectively identified the most significant wavelengths in complex real-world NIR spectra.
• The proposed method demonstrated superior performance compared to traditional techniques like BIPLS, VIP, CC, UVE, and CARS.
• B-NMI significantly improved the stability and robustness of variable selection outcomes.
• Data binning successfully reduced the impact of minor measurement errors on spectral features.

Conclusions:

• Binning-normalized mutual information (B-NMI) offers a powerful and reliable approach for variable selection in NIR spectroscopy.
• The B-NMI method enhances the interpretability and predictive accuracy of multivariate models.
• This technique provides a robust solution for analyzing complex spectral data across various applications.