Discrete and Parameter-Free Multiple Kernel k-Means.

Area of Science:

• Machine Learning
• Data Mining
• Clustering Algorithms

Background:

• Multiple Kernel k-means (MKKM) variants improve upon Kernel k-means (KKM) by integrating diverse data sources.
• Existing MKKM methods often involve multi-stage optimization, leading to information loss and suboptimal clustering.
• Many MKKM approaches neglect kernel correlations, resulting in redundant kernel fusion and reduced data diversity.

Purpose of the Study:

• To develop a novel Discrete and Parameter-free Multiple Kernel k-means (DPMKKM) model.
• To address the limitations of existing MKKM methods, including information loss and redundant kernel fusion.
• To enhance clustering performance by improving kernel fusion and data diversity.

Main Methods:

• The proposed DPMKKM model employs an alternative optimization method for direct cluster assignment, bypassing discretization.
• Kernel correlation is implicitly measured via a regularization term to reduce redundancy and enhance diversity.
• A coordinate descent technique is utilized to optimize the algorithm's time complexity, improving efficiency.

Main Results:

• The DPMKKM model directly obtains discrete cluster assignments without a post-discretization step.
• Implicit regularization effectively reduces kernel redundancy and boosts the diversity of information sources.
• Experimental results on real-world datasets demonstrate the superior effectiveness of the DPMKKM model compared to existing methods.

Conclusions:

• The DPMKKM model offers a more efficient and effective approach to multiple kernel k-means clustering.
• Its parameter-free nature and reduced complexity make it highly suitable for practical applications.
• The method successfully addresses key limitations of previous MKKM algorithms, leading to improved clustering outcomes.