Design of a multiple kernel learning algorithm for LS-SVM by convex programming

Area of Science:

• Machine Learning
• Computational Statistics
• Optimization Theory

Background:

• Least squares support vector machine (LS-SVM) performance is sensitive to kernel choice and regularization parameter selection.
• Traditional cross-validation methods are computationally expensive and lack flexibility for multiple kernel learning scenarios.
• Efficient model selection for LS-SVM with multiple kernels remains a significant challenge.

Purpose of the Study:

• To develop an efficient method for multiple kernel learning in LS-SVM.
• To integrate regularization parameter optimization within a unified framework for automatic model selection.
• To address the computational cost and inflexibility issues of cross-validation for LS-SVM.

Main Methods:

• Formulation of the multiple kernel learning problem for LS-SVM using semidefinite programming (SDP).
• Development of a unified framework to optimize both the kernel selection and the regularization parameter simultaneously.
• Implementation of an automatic model selection process derived from the SDP formulation.

Main Results:

• The proposed SDP-based approach effectively handles multiple kernel learning for LS-SVM.
• The unified framework enables automatic optimization of the regularization parameter alongside kernel selection.
• Experimental validations demonstrate the efficacy and efficiency of the developed method.

Conclusions:

• Semidefinite programming provides a robust framework for multiple kernel learning in LS-SVM.
• The integrated optimization approach leads to an automated and more flexible model selection process.
• This work advances LS-SVM methodology by offering an efficient solution for complex kernel and parameter tuning.