Distribution-Based Cluster Structure Selection.

Area of Science:

• Data Science
• Machine Learning
• Cluster Analysis

Background:

• Ensemble methods aim to unify multiple cluster structures from diverse datasets.
• Identifying and selecting relevant cluster structures is crucial for accurate ensemble outcomes.
• Existing methods may not effectively filter suboptimal cluster structures for ensemble generation.

Purpose of the Study:

• To develop a method for selecting optimal cluster structures within an ensemble.
• To propose a novel framework for distribution-based cluster structure ensemble (DCSE).
• To introduce a technique for selecting a subset of cluster structures (DCSSS) to enhance ensemble representativeness.

Main Methods:

• Cluster structures are represented using Gaussian mixture models, with parameters estimated via the expectation-maximization algorithm.
• Distribution-based distance functions are developed to quantify similarity between cluster structures.
• A DCSE framework and a DCSSS technique are proposed for unified cluster structure generation and selection.
• A distribution-based normalized hypergraph cut algorithm is employed for final result generation.

Main Results:

• The proposed DCSE framework demonstrates effective performance on real-world datasets.
• The DCSE framework integrated with the DCSSS technique shows improved performance compared to baseline methods.
• Nonparametric tests confirm the efficacy of DCSE against competing algorithms.

Conclusions:

• The DCSE framework provides a robust approach to generating unified cluster structures.
• The DCSSS technique significantly enhances the performance and representativeness of the ensemble clustering.
• The proposed methods offer a valuable contribution to the field of ensemble clustering and data analysis.