Multiple Kernel Clustering With Neighbor-Kernel Subspace Segmentation.

Area of Science:

• Machine Learning
• Data Mining
• Clustering Algorithms

Background:

• Multiple Kernel Clustering (MKC) is widely used but often overlooks the neighborhood structure of base kernels.
• This oversight can negatively impact clustering performance, especially with noisy data.

Purpose of the Study:

• To propose a novel neighbor-kernel-based MKC algorithm that leverages intrinsic kernel relationships.
• To enhance clustering robustness and accuracy by preserving neighborhood structures and mitigating noise.

Main Methods:

• Defining a 'neighbor kernel' to maintain block diagonal structure and resist noise/outliers.
• Linearly combining neighbor kernels to form a consensus affinity matrix.
• Employing exact-rank-constrained subspace segmentation for refining shared structures.

Main Results:

• The proposed algorithm effectively preserves block diagonal structures, enhancing subspace segmentation.
• Experimental results on benchmark datasets demonstrate superior clustering performance compared to existing methods.
• The approach shows improved robustness against noise and outliers.

Conclusions:

• The neighbor-kernel-based MKC approach offers a significant improvement over traditional methods.
• Coupling neighbor kernel definition with subspace segmentation leads to synergistic performance gains.
• The algorithm provides an effective solution for robust and accurate data clustering.