Penalized model-based clustering with unconstrained covariance matrices.

Area of Science:

• Machine Learning
• Bioinformatics
• Statistical Modeling

Background:

• High-dimensional data analysis, particularly for microarrays, benefits from clustering.
• Noise variables can obscure true clustering patterns, necessitating variable selection.
• Existing regularization methods for clustering often overlook variable dependencies.

Purpose of the Study:

• To propose a regularized Gaussian mixture model that handles general covariance matrices.
• To improve clustering accuracy and variable selection by accounting for variable dependencies.
• To address challenges in estimating large covariance matrices.

Main Methods:

• Developed a regularized Gaussian mixture model incorporating general covariance matrices.
• Employed regularization for simultaneous parameter estimation and variable selection.
• Utilized a derived Expectation-Maximization (E-M) algorithm with graphical lasso for parameter estimation.

Main Results:

• The proposed model effectively accounts for dependencies among variables within clusters.
• Simultaneous shrinkage of means and covariance matrices enhances clustering and variable selection.
• Numerical examples, including microarray gene expression data, validate the method's utility.

Conclusions:

• The novel regularized Gaussian mixture model offers improved clustering and variable selection in high-dimensional data.
• Accounting for variable dependencies leads to more accurate cluster shapes and orientations.
• The E-M algorithm with graphical lasso provides an effective solution for parameter estimation.