Robust One-Class Kernel Spectral Regression.

Area of Science:

• Machine Learning
• Pattern Recognition
• Data Mining

Background:

• The kernel null-space technique is an effective one-class classification (OCC) method.
• Its applicability is limited by susceptibility to training data corruption and inability to rank observations.
• Existing OCC methods lack robustness and sample conformity ranking.

Purpose of the Study:

• To address the limitations of the kernel null-space technique in one-class classification.
• To enhance robustness against training data contamination.
• To introduce functionality for ranking training samples based on model conformity.

Main Methods:

• Regularizing the null-space kernel Fisher methodology using a regression-based formulation.
• Analyzing the effect of Tikhonov regularization in Hilbert space, framing the problem as sensitivity analysis.
• Investigating the effect of solution sparsity and proposing iterative algorithms with recursive label confidence updates.

Main Results:

• The proposed methodology significantly enhances robustness against training data contamination compared to baseline and other OCC approaches.
• The method effectively ranks training samples according to their conformity with the model.
• Iterative algorithms were developed for both regularization schemes.

Conclusions:

• The regularized null-space kernel Fisher methodology offers a more robust and informative approach to one-class classification.
• This technique overcomes key limitations of the traditional kernel null-space method.
• The findings have implications for applications requiring reliable anomaly detection and data quality assessment.