Use of Wishart Prior and Simple Extensions for Sparse Precision Matrix Estimation

Area of Science:

• Statistics
• Machine Learning
• Computational Biology

Background:

• Estimating precision matrices is crucial for understanding Gaussian distributions and their graphical models.
• Existing methods for sparse precision matrix estimation can be computationally intensive or lack robustness.
• Bayesian approaches offer a principled framework for incorporating prior knowledge and quantifying uncertainty.

Purpose of the Study:

• To develop a simple, rapid, and accurate Bayesian procedure for estimating Gaussian precision matrix elements.
• To introduce a decision rule for enhancing the estimation of sparse precision matrices and associated graphical structures.
• To compare the proposed method with existing Wishart-based and graphical lasso approaches.

Main Methods:

• Utilized a conjugate Wishart prior for efficient computation of the analytic posterior (mode and uncertainty).
• Developed a decision-rule step to identify and shrink near-zero precision matrix elements to zero, promoting sparsity.
• Employed simulated datasets for comparative analysis against alternative methods, including graphical lasso.
• Applied an empirical Bayes procedure for selecting prior hyperparameters, particularly in high-dimensional settings with sparsity.

Main Results:

• The proposed conjugate Wishart prior method provides a fast and analytic computation of posterior precision matrix estimates.
• The decision-rule significantly improves the performance of sparse precision matrix estimation and graph recovery.
• The method demonstrates competitive or superior performance compared to established Wishart-based approaches and graphical lasso in simulations.
• Empirical Bayes hyperparameter selection proves effective for high-dimensional sparse scenarios.

Conclusions:

• The conjugate Wishart prior with a decision rule offers an efficient and effective Bayesian approach for precision matrix estimation.
• This method is particularly advantageous for sparse precision matrices and high-dimensional data.
• The approach facilitates improved graph structure recovery and uncertainty quantification in Gaussian models.