Two criteria for model selection in multiclass support vector machines.

Area of Science:

• Machine Learning
• Computer Science

Background:

• Efficient model selection is crucial for multiclass support vector machine (SVM) classification in practical applications.
• Existing methods may incur significant computational overhead, especially with numerous model parameters.

Purpose of the Study:

• To develop novel, efficient model selection criteria for multiclass SVM classification.
• To address the computational demands of current model selection techniques.

Main Methods:

• Two new model selection criteria were developed by combining or redefining the radius-margin bound from binary SVMs.
• The approach links multiclass SVM test error rates to those of binary SVMs and considers class separability.

Main Results:

• Extensive experiments show the proposed criteria achieve effective model selection across most datasets.
• The new criteria demonstrate performance comparable to k-fold cross-validation.
• Significant reduction in computational overhead was observed compared to benchmark methods.

Conclusions:

• The developed criteria offer an efficient alternative for multiclass SVM model selection.
• These methods provide a practical solution for optimizing SVM models with reduced computational burden.