Sparse multiple kernel learning for signal processing applications.

Area of Science:

• Signal Processing
• Machine Learning
• Computational Statistics

Background:

• Feature grouping enhances interpretability in signal processing models.
• Existing multiple kernel learning methods lack strict sparsity in kernel selection due to convex formulations.
• Efficient kernel methods often rely on solving dual problems, limiting primal solution sparsity.

Purpose of the Study:

• To develop a novel algorithm for inducing sparsity in group parameters within nonlinear models.
• To generalize group-feature selection to kernel selection in multiple kernel learning.
• To improve the interpretability and efficiency of kernel-based signal processing models.

Main Methods:

• Proposed a log-based concave penalty term added to the primal problem to enforce group sparsity.
• Developed a generalized iterative learning algorithm for primal space parameter estimation.
• Extended the method to nonlinear models using the "kernel trick" to create Sparse Multiple Kernel Learning (SMKL).

Main Results:

• SMKL achieves sparser solutions in terms of the number of kernels used compared to existing methods.
• The algorithm effectively performs group-feature selection and kernel selection.
• Demonstrated high accuracy with a significantly reduced number of kernels across various signal processing applications.

Conclusions:

• SMKL offers a powerful approach for sparse kernel selection in nonlinear models.
• The method enhances model interpretability and efficiency in signal processing tasks.
• SMKL successfully applied to diverse datasets including mass spectra, hyperspectral imagery, and NIR spectra.