Semisupervised Learning Process Based on a Laplacian Regularized One Class Support Vector Machine with Dynamic Decision Rule for Near-Infrared Data Classification

This paper presents a nonconventional semisupervised learning method for classifying near-infrared (NIR) data, designed for situations where not all data classes are known and labeled training data are sparse. Such requirements are commonly encountered in both industrial applications and scientific research contexts. The proposed method to tackle this challenge here is a designed process with LapDRegOSVM, which combines spectral segmentation with a Laplacian regularized one-class support vector machine and a dynamic decision rule. The learning process uses parallel LapDRegOSVM procedures, with each procedure identifying a single known class from mixed data. LapDRegOSVM improves upon traditional one-class SVMs (OSVM) and classical Laplacian regularized OSVM (LapOSVM) by leveraging information from unlabeled data through kernel reformation with manifold regularization and decision rule redefinition. A significant advancement of LapDRegOSVM lies in its refined decision rule, implemented via either a dynamic threshold or D-constrained K-means clustering. Results show that LapDRegOSVM outperforms standard OSVM and LapOSVM in utilizing unlabeled data and reregulated decision rule for achieving more accurate classification, particularly in handling "not available" (NA) data. The D-constrained K-means approach to the decision rule also proves superior to static thresholds. This semisupervised classification process achieves high accuracy and reliability in identifying expected classes within NIR spectra even with a substantial number of unknown classes, and all unknown classes remain under "NA" status postclassification, a capability rarely demonstrated by other learning methods.