Reduced multiple empirical kernel learning machine.

Area of Science:

• Machine Learning
• Computational Statistics

Background:

• Multiple Kernel Learning (MKL) offers flexibility for heterogeneous data by using multiple kernels.
• However, MKL suffers from high time and space complexity, limiting real-world applications.
• Empirical Kernel Mapping (EKM) is one MKL approach, but it is less explored.

Purpose of the Study:

• To propose a novel Reduced Multiple Empirical Kernel Learning Machine (RMEKLM).
• To significantly reduce the time and space complexity of EKM-based MKL.
• To maintain the effectiveness and efficiency of MKL for classification tasks.

Main Methods:

• The proposed RMEKLM utilizes Gauss Elimination to extract feature vectors.
• These vectors are used to span a reduced, orthonormal subspace of the original feature space.
• The method ensures the spanned subspace is isomorphic to the original feature space.

Main Results:

• RMEKLM successfully reduces both time and space complexity for EKM-based MKL.
• The method demonstrates a simpler computation and reduced storage requirements, particularly for testing.
• Experimental results confirm RMEKLM's efficient and effective performance in complexity and classification.

Conclusions:

• RMEKLM is the first method to reduce time and space complexity for EKM-based MKL.
• Gauss Elimination provides a stable and efficient way to generate a basis for the feature space.
• RMEKLM offers a computationally simpler and more storage-efficient alternative for MKL applications.