Simultaneous Global and Local Graph Structure Preserving for Multiple Kernel Clustering.

Area of Science:

• Machine Learning
• Data Mining
• Computer Vision

Background:

• Multiple Kernel Learning (MKL) surpasses Single Kernel Learning (SKL) for nonlinear clustering by avoiding manual kernel selection.
• Graph-based MKL subspace clustering integrates self-expression learning but often neglects crucial kernel-space graph structures.
• Existing MKL methods fail to adequately preserve the global and local data structures within the kernel space.

Purpose of the Study:

• To introduce Structure-Preserving Multiple Kernel Clustering (SPMKC), a novel MKL approach.
• To address the limitation of ignored graph structures in kernel space within prior MKL clustering methods.
• To enhance clustering performance by preserving both global and local data structures in the kernel space.

Main Methods:

• SPMKC employs a new kernel affine weight strategy for optimal consensus kernel learning from a kernel pool.
• It incorporates a kernel group self-expressiveness term to maintain global data structure.
• A kernel adaptive local structure learning term preserves local data structure within the kernel space.

Main Results:

• SPMKC effectively preserves both global and local data structures in the kernel space.
• The proposed method demonstrates superior clustering performance compared to state-of-the-art MKL clustering techniques.
• SPMKC achieves improved results with a reduced computational cost.

Conclusions:

• SPMKC offers a significant advancement in MKL for nonlinear clustering tasks.
• Preserving data structure within the kernel space is crucial for effective clustering.
• The method shows strong potential for applications in image and text clustering.