Tree-based Node Aggregation in Sparse Graphical Models.

Area of Science:

• Network analysis
• Statistical modeling
• Machine learning

Background:

• High-dimensional graphical models often use regularization to reduce network complexity.
• Existing methods primarily focus on reducing the number of edges (edge sparsity).
• There is a need for methods that simplify networks by aggregating nodes.

Purpose of the Study:

• To develop a novel method for estimating graphical models that are both edge-sparse and node-aggregated.
• To introduce the tree-aggregated graphical lasso (tag-lasso) method.
• To leverage side information encoded in a tree structure for data-driven node aggregation.

Main Methods:

• Developed a new convex regularized method: tag-lasso.
• Utilized a tree structure to encode node similarity and guide aggregation.
• Implemented tag-lasso efficiently using the locally adaptive alternating direction method of multipliers.

Main Results:

• The tag-lasso method successfully estimates graphical models that are both edge-sparse and node-aggregated.
• Node aggregation, guided by the tree structure, enhances model interpretability.
• The efficient implementation allows for practical application of the method.

Conclusions:

• Tag-lasso offers a powerful approach to simplify high-dimensional graphical models.
• The method provides enhanced interpretability through data-driven node aggregation.
• Demonstrated practical advantages in simulations and real-world applications in finance and biology.