Support vector machine with quantile hyper-spheres for pattern classification.

Area of Science:

• Machine Learning
• Pattern Recognition
• Data Mining

Background:

• Traditional support vector machines (SVMs) can be sensitive to noise, particularly around decision boundaries.
• Developing robust classification models that maintain high accuracy in the presence of noisy data is crucial.

Purpose of the Study:

• To introduce a novel Quantile Hyper-Sphere Support Vector Machine (QHSVM) for improved pattern classification.
• To enhance model robustness, generalization, and noise insensitivity.
• To propose a density estimation method for a scalable variant, ρ-QHSVM.

Main Methods:

• Formulation of QHSVM using two quantile hyper-spheres with a shared center for distinct sample classes.
• Utilization of pinball loss instead of hinge loss to mitigate noise sensitivity.
• Incorporation of margin maximization, inner-class sample clustering, and quadratic programming optimization.
• Development of a local center-based density estimation for the ρ-QHSVM variant.

Main Results:

• QHSVM demonstrates distinct advantages in classification accuracy.
• The ρ-QHSVM variant offers adjustable execution speed, making it suitable for large-scale datasets.
• Experimental validation on artificial, benchmark, and real-world defect datasets confirms QHSVM's efficacy.

Conclusions:

• QHSVM offers a robust and accurate approach to pattern classification, outperforming traditional methods in noisy conditions.
• The ρ-QHSVM variant provides a scalable solution for large datasets without compromising accuracy.
• The proposed methods contribute to advancing noise-resilient and efficient machine learning classification techniques.