A Multiple Kernel Learning Model Based on p-Norm.

Area of Science:

• Machine Learning
• Computational Science
• Pattern Recognition

Background:

• Support Vector Machines (SVMs) effectively address linearly inseparable data using kernel functions.
• Multiple Kernel Learning (MKL) enhances SVM classification accuracy by combining kernels.
• Existing MKL methods often use L1-norm (sparse solutions) or L-norm (non-sparse, noise-sensitive) constraints on kernel weights.

Purpose of the Study:

• To propose a generalized Multiple Kernel Learning (MKL) method based on the p-norm.
• To extend the L1/L2-norm based Generalized Multiple Kernel Learning (GMKL) by incorporating L1- and L-norms.
• To evaluate the proposed L1/L-norm MKL method's classification accuracy compared to existing GMKL.

Main Methods:

• Developed a novel generalized MKL framework utilizing p-norm constraints on kernel combination weights.
• Formulated the method by joining L1- and L-norms, encompassing the L1/L2-norm GMKL as a special case (p=2).
• Conducted experiments to compare the classification performance of the proposed L1/L-norm MKL against the L1/L2-norm GMKL.

Main Results:

• The proposed L1/L-norm MKL method achieved higher classification accuracy than the L1/L2-norm GMKL.
• The method successfully integrated properties from both L1-norm (sparsity) and L-norm (information retention) constraints.
• Experimental results validated the superior performance of the generalized p-norm MKL approach.

Conclusions:

• The generalized L1/L-norm MKL offers a more effective approach for improving SVM classification accuracy.
• This framework provides a flexible way to balance sparsity and information preservation in kernel combinations.
• The proposed method represents a significant advancement in Multiple Kernel Learning techniques.