Robust one-class support vector machine.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• One-Class Support Vector Machine (OCSVM) is effective for one-class classification but sensitive to noise and outliers.
• Existing solutions use bounded loss functions with limitations like discontinuity and non-differentiability.

Purpose of the Study:

• Introduce a novel, continuous, smooth, and differentiable loss function: Quadratic Type Squared Error Loss Function (QTSELF).
• Propose a more robust OCSVM (Q-OCSVM) that handles outliers effectively and improves model optimization.

Main Methods:

• Developed the Quadratic Type Squared Error Loss Function (QTSELF).
• Implemented a robust OCSVM (Q-OCSVM) utilizing QTSELF.
• Applied Rademacher complexity theory for generalization error bound analysis.
• Employed the momentum method for Q-OCSVM optimization.

Main Results:

• Q-OCSVM differentiates samples based on position, applying distinct treatments.
• The model demonstrates enhanced robustness by imposing minimal penalties on noise and outliers.
• Extensive experiments show Q-OCSVM outperforms benchmark techniques.

Conclusions:

• The novel QTSELF and Q-OCSVM offer a more robust and mathematically elegant approach to one-class classification.
• Q-OCSVM provides superior performance compared to existing methods, particularly in the presence of noisy data.