Supervised discretization for decluttering classification models

Area of Science:

• Analytical Chemistry
• Chemometrics
• Machine Learning

Background:

• Multivariate classification models in chemical sensing are often hindered by non-informative variance within classes.
• Existing decluttering methods aim to reduce within-class variance while preserving between-class variance for better model performance.
• External Parameter Orthogonalization (EPO) is a current state-of-the-art technique for data decluttering.

Purpose of the Study:

• To introduce and demonstrate supervised discretization as a novel method for decluttering multivariate classification data.
• To compare the effectiveness of supervised discretization against the established EPO method.
• To evaluate the performance of supervised discretization in real-world chemical sensor applications.

Main Methods:

• Supervised discretization was developed and applied to multivariate chemical sensor data.
• The method was compared to External Parameter Orthogonalization (EPO) using key performance metrics.
• The approach was validated across three distinct classification tasks: X-ray fluorescence (XRF) of pine ash, laser-induced breakdown spectroscopy (LIBS) of artisanal glasses, and LIBS of hardwood species.

Main Results:

• Supervised discretization demonstrated superior decluttering performance compared to EPO.
• The method resulted in a more parsimonious model with fewer parameters, reducing the risk of overfitting.
• Information loss was minimized, leading to enhanced classification accuracy in the tested applications.

Conclusions:

• Supervised discretization offers a more effective and robust approach to decluttering chemical sensor data than current methods.
• This technique holds significant promise for improving the performance and reliability of multivariate classification in chemical sensing.
• The parsimonious nature of supervised discretization makes it a valuable tool for developing advanced chemometric models.