Regularized matrix data clustering and its application to image analysis.

Area of Science:

• Statistics
• Machine Learning
• Data Mining

Background:

• Clustering matrix-valued data presents unique challenges due to high dimensionality and complex structures.
• Existing methods may not effectively capture the underlying sparsity and covariance patterns inherent in such data.

Purpose of the Study:

• To develop a novel regularized mixture model for effective clustering of matrix-valued data.
• To incorporate sparsity structures in the mean signal and separable covariance structures within clusters.
• To provide a computationally efficient algorithm for parameter estimation.

Main Methods:

• Formulation as a finite mixture model using matrix-normal distributions.
• Inclusion of regularization terms to induce sparsity (e.g., low rankness, spatial sparsity) on cluster means.
• Development of an expectation-maximization (EM) type algorithm for efficient computation.
• Theoretical analysis demonstrating strong consistency of the proposed estimators.

Main Results:

• The proposed regularized mixture model demonstrates excellent performance in simulations.
• The method achieves superior prediction accuracy compared to existing approaches.
• Applications to brain signal studies highlight the scientific interpretability of the clustering solutions.

Conclusions:

• The novel regularized mixture model offers a powerful and interpretable approach for clustering matrix-valued data.
• The method's ability to handle sparsity and complex covariance structures makes it suitable for various applications, including neuroimaging.
• The developed algorithm ensures efficient and reliable estimation for practical use.