Closed-Form Gaussian Spread Estimation for Small and Large Support Vector Classification.

Area of Science:

• Machine Learning
• Computational Science

Background:

• Support Vector Machines (SVMs) with Gaussian kernels are powerful for classification.
• Tuning the kernel spread parameter (gamma) is crucial but often computationally expensive.
• Existing methods require iterative training, limiting scalability for large datasets.

Purpose of the Study:

• To develop a direct, non-iterative method for calculating the optimal Gaussian kernel spread.
• To significantly accelerate the training and application of SVMs, especially for large-scale problems.
• To improve the efficiency and performance of SVM classification.

Main Methods:

• Formulation of a direct analytic expression to compute the kernel spread.
• Minimization of the difference between Gaussian and ideal kernel matrices.
• Integration with random sampling for handling large datasets.

Main Results:

• The proposed Direct Gamma Tuning (DGT) method achieves performance comparable to state-of-the-art approaches.
• DGT is one to two orders of magnitude faster than existing methods on small datasets.
• On large datasets (up to 31 million patterns), DGT is faster and outperforms linear SVM.

Conclusions:

• DGT provides a highly efficient and effective solution for tuning Gaussian kernel SVMs.
• The method demonstrates significant speedups and performance improvements, particularly for large-scale classification.
• DGT offers a practical alternative to computationally intensive iterative optimization techniques.