Predictive-property-ranked variable reduction with final complexity adapted models in partial least squares modeling...

Area of Science:

• Chemometrics
• Spectroscopy
• Data Analysis

Background:

• Partial least-squares regression (PLS) is widely used for chemometric modeling.
• Variable reduction methods enhance PLS1 model performance by removing uninformative variables.
• Few methods exist for variable reduction in multiple-response PLS2 models.

Purpose of the Study:

• To investigate variable reduction for PLS2 models using an adapted FCAM method (FCAM-PLS2).
• To evaluate four new predictor-variable properties derived from PLS2 regression coefficients.
• To assess the effectiveness of these properties across diverse spectral and simulated datasets.

Main Methods:

• Adaptation of the Predictive-Property-Ranked Variable Reduction with Final Complexity Adapted Models (PPRVR-FCAM) method for PLS2.
• Development and application of four predictor-variable properties: mean of absolute coefficients, norm of coefficients, and their significances.
• Testing on six datasets including UV-vis, NIR, NMR, and simulated data with correlated and uncorrelated responses.

Main Results:

• The four developed properties are applicable for variable reduction in FCAM-PLS2.
• Models derived from the four properties showed similar predictive abilities.
• The norm of PLS2 regression coefficients exhibited the best selectivity, retaining fewer, more informative variables.
• The significance of the mean of PLS2 regression coefficients was the least selective property.

Conclusions:

• FCAM-PLS2 is an effective method for variable reduction in PLS2 modeling.
• The norm of PLS2 regression coefficients is a highly effective property for selecting informative variables.
• This approach enhances the interpretability and efficiency of PLS2 models.