Variable selection for clustering with Gaussian mixture models.

Area of Science:

• Statistics
• Machine Learning
• Bioinformatics

Background:

• Variable selection is crucial for effective cluster analysis, particularly in model-based approaches.
• Existing methods often require assumptions about variable relationships, limiting their applicability.
• Identifying the role of each variable enhances the interpretability and performance of clustering models.

Purpose of the Study:

• To propose a generalized model for variable selection in model-based cluster analysis.
• To develop a method that does not assume linear relationships between selected and discarded variables.
• To provide a statistically sound procedure for determining variable roles in clustering.

Main Methods:

• A generalized model is proposed, extending previous work by Raftery and Dean.
• Bayesian Information Criterion (BIC) is employed for model comparison and selection.
• A novel algorithm combines backward stepwise selection for clustering and linear regression to ascertain variable roles.
• Model identifiability and criterion consistency are theoretically established.

Main Results:

• The proposed method effectively identifies relevant variables for cluster analysis without prior assumptions on variable relationships.
• Numerical experiments on simulated datasets demonstrate the procedure's efficacy.
• Application to genomic data highlights the practical utility and performance of the variable selection technique.

Conclusions:

• The developed variable selection procedure offers a flexible and robust approach for model-based cluster analysis.
• The method enhances the interpretability of clustering results by specifying variable roles.
• This technique shows promise for applications in diverse fields, including bioinformatics and data mining.