Geometric algorithms to large margin classifier based on affine hulls.

Area of Science:

• Computational Geometry
• Machine Learning
• Optimization Algorithms

Background:

• Binary data classification is fundamental in machine learning.
• Geometric optimization offers intuitive approaches to classification.
• Existing methods face challenges with separable and inseparable datasets.

Purpose of the Study:

• To introduce a geometric framework for binary classification problems.
• To develop theoretical results for candidate extreme points of reduced affine hulls (RAHs).
• To enable efficient application of nearest point algorithms within the RAH framework.

Main Methods:

• Theoretical analysis of candidate extreme points for reduced affine hulls (RAHs).
• Direct application of nearest point algorithms to RAHs.
• Implementation of Gilbert-Schlesinger-Kozinec and Mitchell-Dem'yanov-Malozemov algorithms within the RAH framework.

Main Results:

• Established theoretical results on candidate extreme points of RAHs.
• Demonstrated successful and efficient application of nearest point algorithms for both separable and inseparable classification problems.
• Achieved significant performance improvements through the proposed methods.

Conclusions:

• The proposed geometric framework and theoretical advancements provide an effective approach to binary classification.
• The RAH framework allows for direct and efficient use of established nearest point algorithms.
• The study highlights the practical utility and performance gains of the novel methods in solving real-world classification tasks.