Regularized Simple Multiple Kernel k-Means With Kernel Average Alignment.

Area of Science:

• Machine Learning
• Data Mining
• Clustering Algorithms

Background:

• Multiple Kernel Clustering (MKC) seeks an optimal kernel from basic kernels, often assuming linear combinations.
• Existing MKC methods like SimpleMKKM lack the ability to integrate prior knowledge, potentially leading to inaccurate partitions.

Purpose of the Study:

• To propose a novel MKC algorithm, Regularized Simple Multiple Kernel K-Means with Kernel Average Alignment (R-SMKKM-KAA), that incorporates prior knowledge.
• To enhance clustering performance by preventing learned partitions from deviating significantly from expected ones, especially when ground truth is unavailable.

Main Methods:

• Introduced a regularization term based on average partition alignment to guide the kernel learning process.
• Developed an efficient algorithm to optimize the regularized objective function.
• Leveraged prior knowledge from average partitions to improve unified kernel learning.

Main Results:

• The proposed R-SMKKM-KAA algorithm demonstrated significant improvements in clustering performance across nine common datasets.
• Incorporating prior knowledge via average alignment effectively regularized the learning process.
• The method benefits from both basic kernel combinations and prior knowledge integration.

Conclusions:

• R-SMKKM-KAA effectively addresses the limitations of existing MKC methods by integrating prior knowledge.
• The proposed approach offers a robust and effective strategy for improving clustering accuracy in various datasets.
• Average partition alignment serves as a strong baseline for guiding MKC algorithms.