Transfer learning in high-dimensional semiparametric graphical models with application to brain connectivity...

Area of Science:

• Statistics
• Machine Learning
• Neuroimaging Analysis

Background:

• Transfer learning enhances statistical efficiency by utilizing related datasets.
• Estimating undirected semiparametric graphical models is crucial in various fields, including neuroimaging.
• Existing methods often assume Gaussian distributions, limiting their applicability to real-world data like fMRI.

Purpose of the Study:

• To develop a transfer learning algorithm for estimating undirected semiparametric graphical models.
• To address the limitations of Gaussian distribution assumptions in graphical model estimation.
• To infer functional brain connectivity patterns in Attention Deficit Hyperactivity Disorder (ADHD) patients using fMRI data.

Main Methods:

• Proposed Trans-Copula-CLIME algorithm for transfer learning in graphical model estimation.
• Characterized graph similarity using the sparsity of a divergence matrix.
• Employed nonparametric rank-based correlation coefficient estimators for robustness against non-normality.
• Established theoretical convergence rates for the proposed estimator.

Main Results:

• Trans-Copula-CLIME demonstrates significant advantages over non-transfer learning methods when auxiliary data is similar and abundant.
• The method shows improved performance for non-Gaussian data distributions.
• Successfully applied to infer functional brain connectivity in ADHD patients using multi-site fMRI data.

Conclusions:

• Trans-Copula-CLIME offers a robust and efficient approach for semiparametric graphical model estimation via transfer learning.
• The method's ability to handle non-Gaussian data makes it suitable for complex neuroimaging datasets.
• This work advances the analysis of functional brain connectivity in ADHD by integrating multi-site fMRI data.