Clustering ensembles: models of consensus and weak partitions.

Area of Science:

• Machine Learning
• Data Mining
• Computational Statistics

Background:

• Clustering ensembles improve robustness and stability in unsupervised classification.
• Consensus clustering is challenging, with existing approaches from graph-based, combinatorial, or statistical perspectives.

Purpose of the Study:

• To develop a unified representation for multiple clusterings and formulate a categorical clustering problem.
• To propose a probabilistic consensus model using finite mixture of multinomial distributions.
• To define a novel consensus function based on generalized mutual information.

Main Methods:

• Formulating a categorical clustering problem with a unified representation.
• Developing a probabilistic consensus model using finite mixture of multinomial distributions.
• Employing the EM algorithm for maximum-likelihood estimation and defining a new consensus function.

Main Results:

• Demonstrated efficacy of combining weak clustering algorithms (data projections, random splits).
• Analyzed combination accuracy based on component partition parameters and number of partitions.
• Evaluated performance with incomplete information and missing cluster labels.

Conclusions:

• The proposed probabilistic model and consensus function effectively improve clustering ensemble performance.
• The methods are robust to weak clustering components and handle incomplete data scenarios.
• Experimental results validate the effectiveness on real-world datasets.