Semisupervised Learning Process Based on a Laplacian Regularized One Class Support Vector Machine with Dynamic...

Area of Science:

• Machine Learning
• Spectroscopy
• Data Science

Background:

• Semisupervised learning is crucial for NIR data classification when labeled data is scarce.
• Existing methods struggle with unknown classes and sparse training data in industrial and scientific contexts.
• One-class Support Vector Machines (OSVM) and Laplacian regularized OSVM (LapOSVM) have limitations in handling unlabeled and unknown data.

Purpose of the Study:

• To present a nonconventional semisupervised learning method, LapDRegOSVM, for classifying near-infrared (NIR) data.
• To address challenges of unknown data classes and sparse labeled training data.
• To improve classification accuracy and reliability by leveraging unlabeled data and refining decision rules.

Main Methods:

• Developed LapDRegOSVM, combining spectral segmentation with a Laplacian regularized one-class support vector machine and a dynamic decision rule.
• Utilized parallel LapDRegOSVM procedures for identifying single known classes from mixed data.
• Enhanced traditional OSVM and LapOSVM by incorporating manifold regularization and redefining decision rules using dynamic thresholds or D-constrained K-means clustering.

Main Results:

• LapDRegOSVM demonstrated superior performance compared to standard OSVM and LapOSVM in utilizing unlabeled data.
• The refined decision rule, especially with D-constrained K-means, significantly improved classification accuracy, particularly for 'not available' (NA) data.
• Achieved high accuracy and reliability in identifying expected classes within NIR spectra, even with numerous unknown classes.

Conclusions:

• LapDRegOSVM offers a robust semisupervised classification approach for NIR data, effectively handling sparse labels and unknown classes.
• The method's dynamic decision rule and manifold regularization provide significant advantages over traditional OSVM techniques.
• This approach enables reliable identification of known spectral classes while classifying unknown classes as 'NA', a valuable capability for real-world applications.