Efficient Multiple Kernel Learning Algorithms Using Low-Rank Representation.

Area of Science:

• Machine Learning
• Pattern Recognition
• Data Science

Background:

• Multiple Kernel Learning (MKL) offers flexibility in kernel selection compared to Support Vector Machines (SVM).
• MKL demonstrates superior recognition accuracy but faces computational challenges due to time-consuming calculations.
• Existing MKL algorithms present analytical and computational difficulties.

Purpose of the Study:

• To develop a novel kernel approximation approach for MKL.
• To propose an efficient Low-Rank MKL (LR-MKL) algorithm.
• To enhance the computational efficiency and recognition performance of MKL.

Main Methods:

• Developed a novel kernel approximation technique for MKL.
• Proposed an efficient Low-Rank MKL (LR-MKL) algorithm utilizing Low-Rank Representation (LRR).
• Redesigned binary-class MKL into multiclass MKL using a pairwise strategy.

Main Results:

• The proposed LR-MKL algorithm demonstrates significant efficiency in kernel weight allocation.
• LR-MKL effectively reduces dimensionality while preserving data features through global low-rank constraints.
• Experimental validation on Yale, ORL, LSVT, and Digit datasets confirmed improved recognition performance and efficiency.

Conclusions:

• The LR-MKL algorithm is an efficient method for kernel weight allocation in MKL.
• The proposed approach substantially boosts the overall performance of Multiple Kernel Learning.
• LR-MKL addresses the computational challenges associated with MKL, making it more practical.