Locally Weighted Ensemble Clustering.

Area of Science:

• Data Science
• Machine Learning
• Clustering Algorithms

Background:

• Ensemble clustering combines multiple base clusterings for improved robustness.
• Existing methods often fail to account for base clustering reliability and local cluster diversity.
• Addressing these limitations is crucial for enhancing consensus performance without data features.

Purpose of the Study:

• To propose a novel ensemble clustering approach utilizing ensemble-driven cluster uncertainty estimation and local weighting.
• To develop methods for evaluating cluster reliability and exploiting local diversity within ensembles.
• To enhance ensemble clustering performance, particularly when data features are unavailable.

Main Methods:

• Estimating cluster uncertainty using an entropic criterion across the ensemble.
• Introducing an ensemble-driven cluster validity measure.
• Presenting a locally weighted co-association matrix to summarize diverse clusters.
• Proposing two novel consensus functions that leverage local diversity.

Main Results:

• The proposed approach effectively estimates cluster uncertainty and incorporates local diversity.
• The locally weighted co-association matrix provides a robust summary of ensemble clusters.
• Experimental results demonstrate the superiority of the novel consensus functions.
• The method shows strong performance across various real-world datasets.

Conclusions:

• The novel ensemble clustering approach effectively addresses limitations of existing methods by considering cluster reliability and local diversity.
• The proposed uncertainty estimation and local weighting strategies enhance the robustness and accuracy of consensus clustering.
• This method offers a significant advancement in ensemble clustering, especially in feature-absent scenarios.