Variable selection in near-infrared spectroscopy: benchmarking of feature selection methods on biodiesel data

Area of Science:

• Analytical Chemistry
• Spectroscopy
• Chemometrics

Background:

• Near-infrared (NIR) spectroscopy is a widely adopted analytical tool across diverse sectors.
• Effective data analysis in vibrational spectroscopy relies heavily on variable (wavelength) selection.
• Feature selection is crucial for extracting meaningful information from complex spectral data.

Purpose of the Study:

• To evaluate and compare the performance of 16 distinct feature selection methods.
• To assess the efficacy of these methods for predicting key biodiesel fuel properties.
• To determine the impact of feature selection on calibration model accuracy.

Main Methods:

• Tested 16 feature selection algorithms including stepwise multiple linear regression (MLR-step), interval partial least squares (iPLS), successive projections algorithm (SPA), and genetic algorithms (GAs).
• Employed linear calibration models: multiple linear regression (MLR) and partial least squares regression (PLS/PLSR).
• Included a non-linear calibration model (artificial neural networks - ANN-MLP) for comparative analysis.

Main Results:

• The study systematically compared the performance of various feature selection techniques for biodiesel property prediction.
• Results demonstrated that appropriate feature selection significantly enhances the accuracy of predictive models.
• The findings are applicable to optimizing data analysis in various spectroscopic techniques beyond NIR.

Conclusions:

• Feature selection is a critical step for improving the predictive power of spectroscopic methods.
• The choice of feature selection algorithm can substantially impact the accuracy of biodiesel property prediction.
• Optimized feature selection can enhance results from techniques like Raman, UV-Vis, and NMR spectroscopies.