clusterBMA: Bayesian model averaging for clustering.

Area of Science:

• Computational statistics and data mining.
• Machine learning and pattern recognition.
• Bioinformatics and computational biology.

Background:

• Traditional ensemble clustering methods often ignore model selection uncertainty.
• Bayesian model averaging (BMA) offers probabilistic interpretation and uncertainty quantification.
• Existing ensemble methods can be sensitive to model and parameter choices.

Purpose of the Study:

• To introduce clusterBMA, a novel weighted model averaging method for unsupervised clustering.
• To enable probabilistic cluster allocation and quantify model-based uncertainty.
• To combine results from diverse clustering algorithms, including hard and soft clustering.

Main Methods:

• Utilizes clustering internal validation criteria to approximate posterior model probabilities for weighting.
• Employs symmetric simplex matrix factorization on a combined posterior similarity matrix for final allocations.
• Implements the clusterBMA method in an accompanying R package.

Main Results:

• clusterBMA outperforms existing ensemble clustering methods on simulated data, particularly for high-dimensional data with low cluster separation (1.16-7.12x improvement in Adjusted Rand Index).
• Achieves comparable or superior performance to benchmarked methods across various dimensionality and cluster separation conditions.
• Demonstrates utility in a case study identifying probabilistic clusters from electroencephalography (EEG) data.

Conclusions:

• clusterBMA provides a robust approach to ensemble clustering, enhancing accuracy and providing valuable uncertainty measures.
• Probabilistic allocation and uncertainty quantification are crucial for clinical relevance and statistical communication in applied settings.
• The method effectively integrates results from multiple clustering algorithms, offering a more reliable clustering solution.