Incomplete Multiple Kernel Alignment Maximization for Clustering.

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Multiple Kernel Alignment (MKA) maximization is crucial for Multiple Kernel Clustering (MKC).
• Existing MKC methods struggle with incomplete kernel matrices.
• Handling missing data in kernel matrices is a significant challenge in MKL.

Purpose of the Study:

• To develop a novel framework that integrates kernel matrix imputation with MKA maximization for robust MKC.
• To address the limitations of current MKC algorithms in handling incomplete data.
• To enhance clustering performance in the presence of missing kernel information.

Main Methods:

• A unified learning framework combining kernel imputation and MKA maximization is proposed.
• Clustering guides the imputation of incomplete kernel elements iteratively.
• Completed kernel matrices are combined for subsequent MKC until convergence.

Main Results:

• The proposed algorithm demonstrates superior clustering performance on MKL benchmark datasets.
• Empirical evaluations show significant improvements over existing state-of-the-art methods.
• Theoretical analysis of the clustering generalization error bound supports the findings.

Conclusions:

• The integrated imputation and MKA maximization framework effectively handles incomplete kernel matrices in MKC.
• The proposed method offers a robust and superior alternative to existing MKC algorithms.
• Publicly available code and data facilitate reproducibility and further research.