Graphical Models via Univariate Exponential Family Distributions

Area of Science:

• Statistical Modeling
• Machine Learning
• Network Analysis

Background:

• Undirected graphical models, including Gaussian and Ising models, are widely used but limited for non-Gaussian/non-categorical data.
• Selecting the appropriate graphical model subclass can be challenging for diverse data types.

Purpose of the Study:

• To introduce a general subclass of graphical models based on exponential families for broader applicability.
• To develop robust M-estimators for fitting these models and analyzing their performance.
• To demonstrate the models' utility in learning network structures from real-world data.

Main Methods:

• Deriving multivariate graphical model distributions from univariate exponential family distributions (e.g., Poisson, exponential).
• Developing and applying a class of M-estimators for model fitting.
• Conducting rigorous statistical analysis to prove exact structure recovery with high probability.

Main Results:

• The proposed M-estimators accurately recover the true graphical model structure.
• The methodology is effective for non-Gaussian and non-categorical data.
• Successful application in learning genomic and proteomic networks.

Conclusions:

• The developed graphical models and M-estimators offer a flexible and powerful approach for network inference.
• This framework expands the applicability of graphical models to a wider range of data distributions.
• The findings have significant implications for network analysis in various scientific domains.