Reinforced Angle-based Multicategory Support Vector Machines.

Area of Science:

• Machine Learning
• Computational Statistics

Background:

• Support Vector Machines (SVMs) are effective binary classifiers.
• Existing Multicategory SVMs (MSVMs) often require k functions for k classes and rely on quadratic programming solvers, presenting computational challenges.

Purpose of the Study:

• To propose a novel group of MSVMs, termed Reinforced Angle-based MSVMs (RAMSVMs).
• To develop an MSVM that utilizes k-1 functions and an angle-based prediction rule.

Main Methods:

• Introduced Reinforced Angle-based MSVMs (RAMSVMs) using k-1 functions.
• Proved Fisher consistency for RAMSVMs.
• Implemented RAMSVMs using an efficient coordinate descent algorithm on the dual problem.

Main Results:

• Demonstrated that RAMSVMs can achieve Fisher consistency.
• Showcased the efficiency of the coordinate descent algorithm for RAMSVM implementation.
• Numerical experiments confirmed competitive computational speed and classification performance.

Conclusions:

• RAMSVMs provide a computationally efficient and high-performing alternative for multicategory classification.
• The proposed method addresses limitations of existing MSVM approaches.