The graphical lasso: New insights and alternatives

Area of Science:

• Statistics
• Machine Learning
• Computational Biology

Background:

• The graphical lasso (GLASSO) is widely used for structure learning in Gaussian graphical models.
• GLASSO utilizes L1 regularization for precision matrix sparsity.
• GLASSO exhibits convergence issues, where the estimated precision matrix may not invert to the covariance matrix.

Purpose of the Study:

• To investigate and explain the convergence behavior of the GLASSO algorithm.
• To propose novel algorithms that overcome GLASSO's limitations.
• To compare the performance of new and existing algorithms for Gaussian graphical model structure learning.

Main Methods:

• Analysis of GLASSO's dual formulation and block coordinate ascent.
• Development of primal algorithms (P-GLASSO and DP-GLASSO) using block coordinate descent.
• Comparative study of algorithms focusing on coordinate sub-problem solutions.

Main Results:

• GLASSO solves the dual penalized likelihood, targeting the covariance matrix.
• Proposed primal algorithms (P-GLASSO, DP-GLASSO) optimize the precision matrix directly.
• DP-GLASSO demonstrates superior performance across various metrics.

Conclusions:

• The convergence issues in GLASSO stem from its dual optimization approach.
• Primal algorithms offer a more direct and stable approach to graphical lasso.
• DP-GLASSO is recommended as a superior alternative for Gaussian graphical model structure learning.