Clustering ensemble method integrating Gaussian mixture model and three-way decision (GMM-3WD-CE).

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Clustering ensembles improve quality by integrating multiple results.
• Existing methods struggle with boundary uncertainty and lack a unified framework.

Purpose of the Study:

• To propose GMM-3WD-CE, a novel clustering ensemble integrating Gaussian Mixture Model (GMM) and three-way decision (3WD) theory.
• To develop a multi-level uncertainty modeling framework for enhanced clustering.
• To address limitations in handling boundary uncertainty in existing methods.

Main Methods:

• Generates diverse base clusterings using a multi-algorithm strategy.
• Constructs a weighted co-association matrix using silhouette, Caliński-Harabasz, and Davies-Bouldin indices.
• Employs ICL criterion for GMM model selection and Otsu algorithm for adaptive thresholding to define core, boundary, and trivial domains.
• Applies differentiated label-assignment strategies for consensus clustering.

Main Results:

• GMM-3WD-CE achieves statistically significant average improvements in NMI and ARI over PCPA and MCLA.
• Demonstrates competitive performance against the SDGCA baseline with a notable NMI advantage.
• Ablation studies confirm the contribution of each component, with statistical significance validated by Wilcoxon and Friedman tests.

Conclusions:

• GMM-3WD-CE offers a robust framework for clustering ensembles by effectively modeling uncertainty.
• The method provides superior performance and statistical significance compared to existing approaches.
• Runtime and scalability analyses characterize computational trade-offs for practical application.