Variable selection in discriminant partial least-squares analysis.

Area of Science:

• Chemometrics
• Multivariate Data Analysis
• Spectroscopy

Background:

• Variable selection is crucial for interpretable multivariate classification.
• Existing methods may not fully capture the most relevant features in high-dimensional data.

Purpose of the Study:

• Introduce a novel variable selection method for discriminant partial least-squares (VS-DPLS) analysis.
• Enhance the understanding and interpretability of classification models using spectral data.

Main Methods:

• Developed a VS-DPLS method extending DPLS by retaining key elements in the weight vector.
• Determined optimal DPLS factors using cross-validation.
• Applied an iterative VS-DPLS procedure to assess variable uniqueness and identify important spectral regions.

Main Results:

• Achieved excellent results on four high-dimensional spectral datasets (FTIR and PyMS).
• The iterative procedure identified important spectral regions rather than just individual variables.
• Demonstrated improved classification model interpretability.

Conclusions:

• The VS-DPLS method offers a powerful approach for variable selection in chemometrics.
• This technique enhances the interpretability of multivariate classification models.
• Identifies key spectral regions critical for accurate classification.