Graph Estimation Based on Neighborhood Selection for Matrix-variate Data

TL;DR

This paper introduces a new regression-based approach for estimating matrix graphical models, effectively recovering true network structures in high-dimensional matrix-variate data, with proven theoretical guarantees and practical applications to EEG data.

Contribution

The paper proposes a novel regression-based method for matrix graphical model estimation, with theoretical guarantees for exact edge recovery and improved support detection over existing methods.

Findings

Method achieves high probability of exact edge recovery.

Simulation studies show superior support recovery performance.

Application to EEG data reveals meaningful brain network structures.

Abstract

Undirected graphical models are powerful tools for uncovering complex relationships among high-dimensional variables. This paper aims to fully recover the structure of an undirected graphical model when the data naturally take matrix form, such as temporal multivariate data. As conventional vector-variate analyses have clear limitations in handling such matrix-structured data, several approaches have been proposed, mostly relying on the likelihood of the Gaussian distribution with a separable covariance structure. Although some of these methods provide theoretical guarantees against false inclusions (i.e. all identified edges exist in the true graph), they may suffer from crucial limitations: (1) failure to detect important true edges, or (2) dependency on conditions for the estimators that have not been verified. We propose a novel regression-based method for estimating matrix graphical models, based on the relationship between partial correlations and regression coefficients. Adopting the primal-dual witness technique from the regression framework, we derive a non-asymptotic inequality for exact recovery of an edge set. Under suitable regularity conditions, our method consistently identifies the true edge set with high probability. Through simulation studies, we compare the support recovery performance of the proposed method against existing alternatives. We also apply our method to an electroencephalography (EEG) dataset to estimate both the spatial brain network among 64 electrodes and the temporal network across 256 time points.