The maximum vector-angular margin classifier and its fast training on large datasets using a core vector machine.

Area of Science:

• Machine Learning
• Pattern Classification
• Computational Complexity

Background:

• Kernelized classification methods like SVM and SVDD face computational challenges with large datasets due to O(n^2) or O(n^3) complexity for kernel matrix computation.
• Efficient training on large datasets remains a significant challenge in pattern classification.

Purpose of the Study:

• To introduce a novel classification method, the Maximum Vector-Angular Margin Classifier (MAMC), designed for effective training on large datasets.
• To demonstrate the equivalence of kernelized MAMC to the kernelized Minimum Enclosing Ball (MEB) problem.

Main Methods:

• Proposed the Maximum Vector-Angular Margin Classifier (MAMC) based on vector-angular margin to find an optimal vector c.
• Established the equivalence between kernelized MAMC and kernelized Minimum Enclosing Ball (MEB).
• Extended MAMC to Maximum Vector-Angular Margin Core Vector Machine (MAMCVM) for fast large-dataset training.

Main Results:

• The kernelized MAMC is proven to be equivalent to the kernelized MEB problem.
• MAMC offers flexibility in controlling the sum of support vectors, similar to v-SVC.
• The MAMCVM extension facilitates efficient training on large datasets.

Conclusions:

• MAMC provides a scalable and flexible approach to pattern classification for large datasets.
• The proposed methods, including MAMCVM, demonstrate effectiveness on artificial and real-world datasets.
• The connection to MEB offers a new perspective for kernelized classification algorithms.