Flexible Regularized Estimation in High-Dimensional Mixed Membership Models

Area of Science:

• Statistics
• Biostatistics
• Machine Learning

Background:

• Traditional cluster analysis assumes data points belong to a single group, which can be overly simplistic for complex datasets.
• Mixed membership models extend finite mixture models, enabling observations to partially belong to multiple components.
• High-dimensional continuous data, common in biomedical research, presents challenges for existing modeling techniques.

Purpose of the Study:

• To propose a novel probabilistic framework for mixed membership models tailored for high-dimensional continuous data.
• To enhance scalability and interpretability of mixed membership analyses.
• To provide a flexible modeling approach that overcomes the limitations of single-group assignment in cluster analysis.

Main Methods:

• A probabilistic representation based on convex combinations of dependent multivariate Gaussian random vectors.
• Approximations of a tensor covariance structure using multivariate eigen-approximations.
• Adaptive regularization via shrinkage priors and establishment of conditional weak posterior consistency for efficient sampling.

Main Results:

• The proposed model demonstrates scalability for high-dimensional data.
• The framework allows for a more nuanced understanding of data where observations can belong to multiple clusters.
• Conditional weak posterior consistency ensures desirable theoretical properties for the model.

Conclusions:

• The developed mixed membership model offers a powerful and flexible tool for analyzing complex, high-dimensional biomedical data.
• This approach provides more natural and informative interpretations compared to traditional clustering methods.
• Applications in autism spectrum disorder brain imaging and breast cancer gene expression highlight the model's utility.