Supervised discretization for decluttering classification models

Presented here is the first demonstration of supervised discretization to 'declutter' multivariate classification data in chemical sensor applications. The performance of multivariate classification models is often limited by the non-informative chemical variance within each target class; decluttering methods seek to reduce within-class variance while retaining between-class variance. Supervised discretization is shown to declutter classes in a manner that is superior to the state-of-the-art External Parameter Orthogonalization (EPO) by constructing a more parsimonious model with fewer parameters to optimize and is, consequently, less susceptible to overfitting and information loss. The comparison of supervised discretization and EPO is performed on three classification applications: X-ray fluorescence spectra of pine ash where the pine was grown in three distinct soil types, laser induced breakdown spectroscopy of colored artisanal glasses, and laser induced breakdown spectroscopy of exotic hardwood species. Supervised discretization 'declutters' chemical sensor data to sharpen the observed differences between classes and improve multivariate classification models.