TFMKC: Tuning-Free Multiple Kernel Clustering Coupled With Diverse Partition Fusion.

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Multiple Kernel Clustering (MKC) is key in unsupervised learning for identifying data groupings.
• Late fusion MKC models offer promising performance but suffer from limited representation capacity due to inflexible fusion mechanisms.
• Existing methods often rely on Eigen-decomposition (EVD) and fine-tuning, which introduce hyperparameters and neglect information across diverse partitions.

Purpose of the Study:

• To address the limitations of inflexible fusion mechanisms and parameter-tuning costs in MKC.
• To propose a novel flexible fusion mechanism for enhanced representation capacity in MKC.
• To develop a method that integrates diverse and complementary information for optimal consensus partitioning.

Main Methods:

• Introduced a tuning-free multiple kernel clustering (TFMKC) method.
• Designed a flexible fusion mechanism that reweights diverse partitions through optimization.
• Transformed the problem from direct optimal partition determination to diverse partition fusion (parameter ensemble).

Main Results:

• TFMKC achieves competitive effectiveness and efficiency compared to existing baselines.
• The proposed method overcomes limitations associated with inflexible fusion and parameter tuning.
• Demonstrated the integration of diverse and complementary information for improved clustering outcomes.

Conclusions:

• TFMKC offers a novel and effective approach to multiple kernel clustering.
• The tuning-free, diverse partition fusion strategy enhances representation capacity and efficiency.
• The method provides a valuable alternative to traditional fine-tuning approaches in MKC.