Gaussian Graphical Model Exploration and Selection in High Dimension Low Sample Size Setting.

Area of Science:

• Statistics
• Machine Learning
• Computational Biology

Background:

• Gaussian graphical models (GGMs) are crucial for analyzing conditional correlations in random vectors.
• Existing inference methods, nodewise and penalized likelihood, face challenges with small sample sizes, leading to inaccurate graph structures.

Purpose of the Study:

• To develop and evaluate a novel composite procedure for GGM inference that enhances accuracy, particularly in low-sample scenarios.
• To compare the proposed method against established nodewise and penalized likelihood approaches.

Main Methods:

• A composite procedure combining a nodewise exploration scheme with an overall likelihood criterion for graph selection.
• Validation using synthetic data to assess graph structure accuracy and KL divergence.
• Application to real-world datasets in brain imaging and biological nephrology.

Main Results:

• The proposed composite method yields graphs closer to the true structure compared to existing methods when sample sizes are small.
• The new approach results in distributions with improved Kullback-Leibler (KL) divergence from the true distribution.
• Demonstrated effectiveness on brain imaging and nephrology data, aligning results with domain-specific knowledge.

Conclusions:

• The composite GGM inference procedure offers superior performance over traditional methods, especially under data scarcity.
• This method provides more accurate graphical representations of conditional dependencies in complex systems.
• The algorithm's applicability is validated across diverse scientific domains, including neuroimaging and bioinformatics.