Feature selection using distributions of orthogonal PLS regression vectors in spectral data.

Area of Science:

• Chemometrics
• Data Analysis
• Machine Learning

Background:

• Feature selection is crucial for developing parsimonious and predictive chemometric models.
• Partial Least Squares (PLS) regression is a standard method for multivariate data analysis.
• Orthogonal Projections to Latent Structures (OPLS) enhances PLS interpretability by removing irrelevant variation.

Purpose of the Study:

• To present a novel feature selection method for multivariate data using orthogonal PLS regression (OPLSR).
• To assess the significance of input features' effects on the response variable Y.
• To improve the interpretability and predictive power of chemometric models.

Main Methods:

• Combining orthogonal signal correction with PLS regression (OPLSR).
• Generating empirical distributions of feature effects via permutation tests.
• Comparing the proposed method with the false discovery rate method using simulation studies.

Main Results:

• The OPLSR-based feature selection method effectively identifies significant variables.
• Demonstrated performance in a simulation study with a complex network structure.
• Successful application to real-world Near-Infrared (NIR) spectra and mass spectrometry data.

Conclusions:

• The proposed OPLSR feature selection method is effective for chemometric data analysis.
• It provides a robust approach for identifying influential features.
• The method enhances the interpretability and predictive accuracy of multivariate models.