Multi-Nyström Method Based on Multiple Kernel Learning for Large Scale Imbalanced Classification.

Area of Science:

• Machine Learning
• Computational Statistics

Background:

• Kernel methods are effective for nonlinear problems but struggle with class imbalance due to high computational costs.
• The Nyström method scales kernel methods but requires many landmark points for accuracy, impacting efficiency.

Purpose of the Study:

• To develop a more efficient and accurate Nyström-based method for large-scale class imbalance problems.
• To improve the performance of multiple kernel learning (MKL) algorithms in imbalanced learning scenarios.

Main Methods:

• Proposed a multi-Nyström method using mixtures of Nyström approximations to manage subkernel matrices.
• Embedded mixture weight optimization within MKL algorithms for enhanced low-rank approximation.
• Selected landmark points based on imbalance distribution to mitigate skewness sensitivity.
• Provided kernel stability analysis to bound model solution error.

Main Results:

• Achieved higher classification accuracy on large-scale imbalanced datasets.
• Demonstrated a significant speedup in MKL algorithm execution.
• The proposed method effectively addresses the limitations of standard Nyström methods.

Conclusions:

• The multi-Nyström method offers an efficient and accurate solution for large-scale class imbalance problems.
• The approach enhances the scalability and performance of kernel methods in practical applications.