clustvarsel: A Package Implementing Variable Selection for Gaussian Model-Based Clustering in R.

Area of Science:

• Statistics
• Machine Learning
• Data Mining

Background:

• Finite mixture modeling is a statistical framework for cluster analysis using Gaussian mixture models.
• Variable or feature selection is crucial for parsimonious models, leading to efficient estimates and improved clustering.
• Identifying relevant variables enhances model interpretability and clustering performance.

Purpose of the Study:

• To introduce the R package clustvarsel for subset selection in model-based clustering.
• To implement an improved methodology for identifying optimal subsets of variables with clustering information.
• To provide efficient algorithms and discuss optimizations for variable selection in clustering.

Main Methods:

• Utilizes an improved version of the Raftery and Dean (2006) methodology.
• Employs step-wise greedy search or a headlong algorithm for solution space exploration.
• Discusses adjustments for algorithm speed-up and a parallel implementation of the stepwise search.

Main Results:

• The R package clustvarsel facilitates effective subset selection for Gaussian mixture model-based clustering.
• The implemented methodology identifies optimal subsets of variables contributing to cluster structure.
• The package offers efficient and potentially parallelized search algorithms for variable selection.

Conclusions:

• The clustvarsel R package provides a valuable tool for parsimonious variable selection in model-based clustering.
• Efficiently identifying informative variables leads to improved clustering accuracy and model interpretability.
• The package supports advanced search strategies and optimizations for practical data analysis.