Square Root Graphical Models: Multivariate Generalizations of Univariate Exponential Families that Permit Positive...

Area of Science:

• Statistics
• Machine Learning
• Data Science

Background:

• Existing multivariate graphical models struggle to represent positive dependencies in data.
• Real-world phenomena often exhibit positive correlations, such as linked airport delays.

Purpose of the Study:

• To introduce Square Root Graphical Models (SQR), a novel class of parametric graphical models.
• To enable the modeling of both positive and negative dependencies in multivariate data, overcoming limitations of prior methods.

Main Methods:

• Developed SQR models as multivariate generalizations of univariate exponential family distributions.
• Derived an exponential generalization allowing arbitrary dependencies with a mild parameter condition.
• Created a Poisson generalization with unconstrained dependencies.
• Employed node-wise regressions with L1 regularization for parameter estimation.
• Utilized sampling for likelihood approximation.

Main Results:

• The exponential SQR generalization successfully models positive and negative dependencies, akin to Gaussian covariance matrix properties.
• The Poisson SQR generalization handles both dependency types without parameter constraints.
• Demonstrated the model's efficacy on synthetic and real-world airport delay data.

Conclusions:

• SQR models offer a flexible and powerful framework for multivariate data analysis.
• The ability to capture positive dependencies enhances the applicability of graphical models to real-world datasets.