Localized Incomplete Multiple Kernel k-Means With Matrix-Induced Regularization.

Area of Science:

• Machine Learning
• Data Mining
• Clustering Algorithms

Background:

• Existing Localized Incomplete Multiple Kernel k-means (LI-MKKM) methods do not fully leverage the diversity and complementarity of base kernels.
• This limitation can reduce the effectiveness of incomplete kernel imputation and subsequent clustering performance.

Purpose of the Study:

• To propose an improved LI-MKKM algorithm, named LI-MKKM with matrix-induced regularization (LI-MKKM-MR).
• To address the limitations of previous methods by incorporating a regularization term that manages correlations among base kernels.

Main Methods:

• Introduction of a matrix-induced regularization term to handle correlations among base kernels, promoting selection of diverse yet moderately different kernels.
• Development of a three-step iterative algorithm for optimization and convergence analysis.
• Theoretical justification of effectiveness through generalization error bound derivation.

Main Results:

• Experimental results on public datasets demonstrate that LI-MKKM-MR consistently outperforms state-of-the-art clustering algorithms.
• The proposed regularization term effectively decreases the selection of similar kernels and increases the selection of moderately different kernels.

Conclusions:

• LI-MKKM-MR offers superior clustering performance compared to existing methods.
• The matrix-induced regularization effectively enhances kernel selection for improved clustering accuracy.