A comparison of methods for multiclass support vector machines.

Area of Science:

• Machine Learning
• Computer Science

Background:

• Support vector machines (SVMs) are primarily designed for binary classification.
• Extending SVMs to multiclass classification remains an active research area with various proposed strategies.
• Existing multiclass SVM methods often involve combining binary classifiers or solving a single, large optimization problem.

Purpose of the Study:

• To implement decomposition methods for two "all-together" multiclass SVM approaches.
• To compare the performance of these "all-together" methods against established binary classification-based methods.
• To evaluate the suitability of different multiclass SVM strategies for large-scale problems.

Main Methods:

• Decomposition implementations for two "all-together" multiclass SVM methods.
• Comparative analysis using "one-against-all," "one-against-one," and directed acyclic graph SVM (DAGSVM) as baseline binary classification-based methods.
• Empirical evaluation on large-scale datasets to assess computational efficiency and performance.

Main Results:

• The "one-against-one" and DAGSVM methods demonstrated greater suitability for practical applications.
• Methods that consider all classes simultaneously generally require fewer support vectors, especially on large datasets.
• Computational expense limits large-scale comparisons of "all-together" methods.

Conclusions:

• "One-against-one" and DAGSVM are recommended for practical multiclass classification tasks.
• "All-together" multiclass SVM approaches offer advantages in support vector reduction for large problems.
• Further research is needed to address the computational challenges of large-scale "all-together" multiclass SVM methods.