A Penalized Matrix Normal Mixture Model for Clustering Matrix Data.

Area of Science:

• Statistics
• Machine Learning
• Data Science

Background:

• Matrix data, common in fields like medical imaging, presents clustering challenges due to high dimensionality and complex row/column correlations.
• Conventional methods often transform matrix data into vectors, exacerbating the high-dimensionality problem and losing interpretability of underlying structures.

Purpose of the Study:

• To extend existing regularized models for matrix data clustering by incorporating covariance regularization.
• To address the curse of dimensionality in large image datasets and improve the reflection of conditional correlation structures.

Main Methods:

• A penalized matrix normal mixture model is proposed, utilizing lasso-type penalties on both mean and covariance matrices.
• An expectation-maximization algorithm is developed for parameter estimation.
• The method is evaluated using simulated and real-world datasets, measuring performance with clustering accuracy (ACC) and adjusted rand index (ARI).

Main Results:

• The proposed method demonstrates competence in parsimonious modeling and capturing conditional correlations.
• Theoretical properties, including estimator consistency and limiting distributions, are derived.
• Experimental results show superior performance compared to conventional clustering methods, indicated by higher ACC and ARI values.

Conclusions:

• The penalized matrix normal mixture model offers an effective approach for clustering high-dimensional matrix data, particularly for image analysis.
• The method successfully balances model parsimony with the accurate representation of complex data structures.
• The enhanced performance metrics (ACC, ARI) validate its advantage over existing techniques.