An improved boosting partial least squares method for near-infrared spectroscopic quantitative analysis

Area of Science:

• Chemometrics
• Spectroscopy
• Data Analysis

Background:

• Partial Least Squares (PLS) regression is widely used for predictive modeling.
• Standard PLS methods can be sensitive to outliers in calibration datasets, compromising predictive accuracy.
• Existing boosting PLS techniques still face challenges with data containing outliers.

Purpose of the Study:

• To develop a robust and enhanced boosting Partial Least Squares (PLS) method.
• To improve the prediction ability of PLS models, particularly in the presence of outliers.
• To apply the improved method for quantitative analysis of near-infrared (NIR) spectral data.

Main Methods:

• Incorporation of a robust step to reduce the influence of outliers.
• Utilization of a loss function based on relative errors for updating sampling weights.
• Employing ensemble prediction via weighted mean of boosting series models.

Main Results:

• The improved boosting PLS method demonstrated enhanced robustness and prediction ability.
• The method showed marked superiority when applied to NIR spectral datasets with outliers.
• Weighted mean ensemble prediction proved more effective than weighted median.

Conclusions:

• The proposed robust boosting PLS method effectively addresses outlier issues in calibration datasets.
• This enhanced method significantly improves quantitative analysis using NIR spectral data.
• The approach offers a more reliable and accurate predictive modeling strategy for industrial applications.