Semi-Supervised Kernel Mean Shift Clustering.

Area of Science:

• Machine Learning
• Data Mining
• Computational Statistics

Background:

• Mean shift clustering is a powerful nonparametric method, but its unsupervised nature limits performance when distance metrics fail to capture cluster structure.
• Existing semi-supervised clustering methods have had limited success in adapting the mean shift algorithm.
• There is a need for methods that can leverage limited supervision to improve mean shift clustering.

Purpose of the Study:

• To propose a novel semi-supervised framework for kernel mean shift clustering (SKMS).
• To incorporate pairwise constraints into the mean shift algorithm to guide the clustering process.
• To enhance the performance of mean shift clustering in scenarios where the original distance metric is insufficient.

Main Methods:

• Developed a semi-supervised framework for kernel mean shift clustering (SKMS).
• Utilized pairwise constraints to guide the clustering procedure.
• Mapped data points to a high-dimensional kernel space and imposed constraints via linear transformation.
• Modified the kernel matrix by minimizing a log det divergence-based objective function.

Main Results:

• Demonstrated the effectiveness of SKMS on various synthetic and real-world datasets.
• Showcased the advantages of SKMS compared to state-of-the-art semi-supervised clustering algorithms.
• Validated the ability of SKMS to improve clustering accuracy by incorporating pairwise constraints.

Conclusions:

• SKMS offers a robust approach to semi-supervised clustering by effectively integrating pairwise constraints.
• The proposed method addresses limitations of purely unsupervised mean shift clustering.
• SKMS provides a valuable tool for enhancing data analysis when limited supervision is available.