[Paper Review] Sparse Median Graphs Estimation in a High Dimensional Semiparametric Model

Fang Han, Han Liu+1 more|arXiv (Cornell University)|Oct 11, 2013

Statistical Methods and Inference27 references4 citations

TL;DR

This paper proposes a novel semiparametric framework for estimating sparse median graphs in high-dimensional, non-i.i.d. settings where data arise from multiple heterogeneous distributions. By leveraging rank-based estimators like Kendall’s tau within a nonparanormal model and defining the median graph as the sparsest graph minimizing total Hamming distance to all sample graphs, the method achieves consistent graph recovery with a provable upper bound on convergence rate, validated on synthetic and real fMRI data including ABIDE and ADHD-200 datasets.

ABSTRACT

In this manuscript a unified framework for conducting inference on complex aggregated data in high dimensional settings is proposed. The data are assumed to be a collection of multiple non-Gaussian realizations with underlying undirected graphical structures. Utilizing the concept of median graphs in summarizing the commonality across these graphical structures, a novel semiparametric approach to modeling such complex aggregated data is provided along with robust estimation of the median graph, which is assumed to be sparse. The estimator is proved to be consistent in graph recovery and an upper bound on the rate of convergence is given. Experiments on both synthetic and real datasets are conducted to illustrate the empirical usefulness of the proposed models and methods.

Motivation & Objective

• To address the lack of theoretically justified population graph estimation in settings with non-identically distributed, high-dimensional network data.
• To develop a unified framework for inference on aggregated network data when individual graphs are heterogeneous and non-Gaussian.
• To define and estimate a sparse median graph that captures common structural patterns across a population of undirected graphical models.
• To establish theoretical consistency and convergence rate bounds for the proposed estimator under high-dimensional asymptotics.
• To empirically validate the method on real neuroimaging datasets (ABIDE, ADHD-200) to detect group differences in brain connectivity.

Proposed method

• Introduces a semiparametric model based on the nonparanormal family, allowing for non-Gaussian, non-i.i.d. data through unspecified monotone transformations.
• Defines the sparse median graph as the sparsest graph minimizing the sum of Hamming distances to all observed graphs in the sample.
• Employs rank-based estimators (specifically Kendall’s tau) to estimate the correlation structure robustly under non-Gaussianity.
• Applies a penalized likelihood approach with ℓ1-constraints to enforce sparsity in the estimated concentration matrix, ensuring a sparse median graph.
• Derives a consistent estimator via optimization over the space of sparse graphs, with theoretical guarantees on convergence rate.
• Uses a two-step procedure: first estimate pairwise correlations via rank statistics, then recover the median graph via sparse precision matrix estimation.

Experimental results

Research questions

• RQ1 Can a robust, consistent estimator be developed for the median graph in high-dimensional, non-i.i.d. settings where data are not identically distributed?
• RQ2 How can the median graph be defined and estimated when the underlying distributions vary across subjects, especially in non-Gaussian settings?
• RQ3 What is the theoretical convergence rate of the sparse median graph estimator in terms of Hamming distance?
• RQ4 Does the proposed method outperform classical correlation-based approaches (e.g., Pearson) in detecting group differences in brain connectivity?
• RQ5 Can the method reliably identify structural differences in brain networks between clinical groups (e.g., controls vs. ASD patients) in real fMRI data?

Key findings

• The proposed sparse median graph estimator is consistent in graph recovery under high-dimensional asymptotics.
• An upper bound on the convergence rate of the estimator with respect to Hamming distance is established, providing theoretical justification for its performance.
• On the ABIDE dataset, Kendall’s tau-based estimation produced a sparse median graph with 948 edges in controls and 945 in cases, showing minimal difference but clearer group differentiation than Pearson.
• The method revealed that in the ASD group, remote brain regions were more likely to be linked, suggesting altered long-range connectivity patterns.
• Visualization of the difference graph showed more red edges (present only in cases) in long-range connections, indicating potential hyperconnectivity in the case group.
• Empirical results on both synthetic and real data confirm the method’s robustness and utility in detecting subtle network differences across clinical populations.