Robust Multicategory Support Vector Machines using Difference Convex Algorithm.

Area of Science:

• Machine Learning
• Computational Statistics

Background:

• Support Vector Machines (SVMs) are widely used for binary classification.
• Generalizing SVMs to Multicategory SVM (MSVM) presents challenges, with existing methods often being suboptimal or lacking robustness.
• Many MSVMs do not account for outliers or achieve Fisher consistency.

Purpose of the Study:

• To propose novel, robust MSVM methods within an angle-based framework.
• To address limitations of existing MSVMs, including suboptimality and sensitivity to outliers.
• To develop efficient and stable MSVM classifiers.

Main Methods:

• Developed two robust MSVM methods utilizing truncated hinge loss functions.
• Employed an angle-based classification framework, avoiding explicit sum-to-zero constraints for improved efficiency.
• Utilized the difference convex algorithm (DCA) for efficient computation of the proposed classifiers.

Main Results:

• The proposed MSVM classifiers demonstrate Fisher consistency.
• These methods effectively alleviate the impact of outliers, leading to more stable classification performance.
• Theoretical and numerical results show competitive performance against existing methods, particularly for datasets with outliers.

Conclusions:

• The novel angle-based MSVMs offer a robust and efficient alternative for multicategory classification.
• These methods provide enhanced stability and performance in the presence of outliers.
• The proposed approach advances the field of robust machine learning classification.